A Bidding Strategy of Intermediary in Display Advertisement Auctions

Master Thesis A Bidding Strategy of Intermediary in Display Advertisement Auctions Supervisor Associate Professor Shigeo MATSUBARA Department of Social Informatics Graduate School of Informatics Kyoto University Yosuke SAITO February 7, 2014

Abstract A Bidding Strategy of Intermediary in Display Advertisement Auctions i Yosuke SAITO Display advertisements are a type of web advertisement, which are usually images or movies that are displayed as a part of web page. In a display advertisement market we consider that media sites (advertising media) are sellers, that ad spaces are goods, and that advertisers are buyers. Previously, a fixed price was set for displaying advertisements and ad agencies mediated the transaction. However, auctions have been introduced to this transaction. Second price auctions are used as the auction method, and we consider that this auction is also a sequential auction in terms of trading the same ad space constantly. We cannot find remarkable features in this method of display advertisement auctions. However, advertisers, who are the end buyers, who do not directly bid, but intermediaries called DSP (Demand Side Platform) do. DSPs have a number of advertisers as their clients, and take bids instead of advertisers to satisfy advertiser s constraints such as a budget constraint. Although a DSP competes against other DSPs, it has the space to coordinate bidding among its clients. We consider that this auction allows a part of bidders colludes with each other in a conventional auction theory. Therefore, an efficient bidding strategy is not present. In this study, I aimed to propose an efficient bidding strategy for the intermediaries. To achieve this goal, we focused on following problems: Proposal on an efficient bidding strategy The strategy which is currently used by DSPs is seemingly reasonable. Nevertheless, they have problems such as an overpayment for ad spaces. Therefore, I must reveal why above-mentioned problems occurred, and propose the bidding strategy to solve this problem. Analysis of auctions for the case of applying the proposed strategy I must measure a performance of the proposed strategy to evaluate whether or not the proposed strategy is more efficient than the existing strategy. In addition, I need to comprehend the effect of the proposed strategy by analyzing auction data of the evaluation in order to predict what is going to happen if the

ii proposed strategy is introduced. In order to solve the first problem, I first pointed the problem that the existing strategy treated display advertisement auctions as single unit auctions despite the fact that display advertisement auctions are originally multi unit auctions, and proposed the bidding strategy that a DSP coordinated bidding among its clients and decided bidding prices to ad spaces on day to day basis. Advertiser s demands have an affect on each other in the proposed strategy, therefore the problem that it is difficult to decide when bidding prices originate. To deal with this problem I used a market model, which is a method to solve resource allocation problems using prices. To solve the second problem, I implemented a computer simulator, and evaluated the performance of the proposed strategy on a virtual auction environment. As a result, I found that the proposed strategy could make more successful bid than the existing strategy in the same budget. Moreover, I analyzed the auction data and found that the proposed strategy improved impressions of advertisements which had high values and impressions of affordable ad spaces. Contributions of this study are as follows: Proposal on an efficient bidding strategy I pointed the problem that the existing strategy treated display advertisement auctions as single unit auctions despite the fact that display advertisement auctions are originally multi unit auctions. Moreover, I proposed the bidding strategy that a DSP coordinated bidding among its clients, and decided bidding prices to ad spaces on day to day basis. Advertisers demands have an affect on each other in the proposed strategy, and the problem that it is difficult to decide bidding prices occurs. To deal with this problem I used a market model. Clarification of bidding coordination using a market model I evaluated the performance of the proposed strategy. As a result, I found that the proposed strategy could make more successful bid than the existing strategy in the same budget. Moreover, I analyzed the auction data, and found that the proposed strategy improved impressions of advertisements which had high values and impressions of affordable ad spaces.

iii Web Web DSP(Demand Side Platform) DSP DSP DSP 2 DSP

iv DSP 2 DSP

A Bidding Strategy of Intermediary in Display Advertisement Auctions Contents Chapter 1 Introduction 1 Chapter 2 Preparation 5 2.1 Second Price Auctions............................. 5 2.2 Sequential Auctions............................... 6 Chapter 3 Related Work 7 Chapter 4 Analysis of Existing Strategy 9 4.1 Feature of Display Advertisement Auctions............... 9 4.2 Formulation of Bidding Problem...................... 11 4.3 Details and Problem of Existing Strategy................ 13 Chapter 5 Proposed Strategy 16 5.1 Basic Idea...................................... 16 5.2 Adverse Effect by Bidding Coordination................. 18 5.3 Bidding Coordination among Advertisers Using Market Model. 21 5.4 Estimation of Evaluated Value of Other DSP............. 24 Chapter 6 Evaluation of Proposed Strategy 28 6.1 Method of Evaluation.............................. 28 6.2 Results of Evaluation.............................. 32 Chapter 7 Analysis of Evaluation Results 37 7.1 Analysis from Viewpoint of Advertisements.............. 37 7.2 Analysis from Viewpoint of Ad Spaces.................. 39 7.3 Bidding Coordination Only in Temporal Variation.......... 41 Chapter 8 Discussion 44 Chapter 9 Conclusion 47 Acknowledgments 50 References 51

Chapter 1 Introduction The online advertisement market centering on display advertisements has recently expanded. Display advertisements are a type of web advertisement, which are usually images or movies displayed as a part of a web page. At first, banner advertisements, which were one of display advertisements, were the mainstream in the online advertisement market. However, search advertisements drastically became popular, and the display advertisements market slowed down. However, the display advertisement market has recently gathered attention again by Google, Yahoo! and other companies, which have made a strong effort to ad technologies. In fact, Google, Yahoo! and so on, have increased their display advertisement market share, moreover, this market have expanded along with growth of smartphones. Jordan considers the model is unrealistic, which assumes that each impression of the ad stochastically leads to a click / conversion independent of other impressions of the same ad, and discusses the model where a sequence of impressions has an influence on a click / conversion by considering a purchasing funnel, which is well known in the marketing literature [1]. Now, an impression intends to display an advertisement on a web page, and a conversion means that website viewers click an advertisement and move to the advertiser s web page, and the advertiser achieves a goal such as merchandise purchases, document requests. Moreover, Muthukrishnan presents several research directions and discusses some insights about them [2]. In a display advertisement market we consider that media sites (advertising media) are sellers, that ad spaces are goods, and that advertisers are buyers. Previously, a fixed price was set for displaying advertisements and ad agencies mediated the transaction. However, auctions have been introduced to this transaction, and the scale has steadily expanded. In this display advertisement auction market, there are SSP (Supply Side Platform), which assists media, and DSP (Demand Side Platform), which assists advertisers. Advertisers can select various ad spaces for impressions. On the other hand, a problem to decide a bidding strategy 1

is emerging because a bidding strategy decides whether or not impressions are efficient in cost. Recently, the bidding strategy have actively been researched. Ghosh studies the design of a bidding agent that implements a display advertising campaign by bidding in such a marketplace [3], and Cui proposes a general divide-and-conquer approach to forecast the bid distribution for any advertising campaign in non-guaranteed delivery exchange system [4]. Now, Figure 1 shows a relationship among players in display advertisement auctions. In Figure 1, advertisers are companies which ask DSPs to display Advertiser! Advertiser! Advertiser! Advertiser: Asking to run an ad! DSP! DSP! DSP: Bidding according to a strategy Ad Exchange! Ad Exchange: Sending a request to bid to DSP and a winning ad to SSP! SSP! SSP! SSP: Getting goods(ad spaces) up on Ad Exchange Media! Media! Media: Consigning to goods to SSP Visitor! Visitor: Viewing an ad Figure 1: Relationship among players in display advertisement auctions advertisements, and ad spaces are frames in web pages of media companies. Display advertisement auctions are conducted in an advertisement market called Ad Exchange. The process is as follows until viewers actually see advertisements [5]. 1. A bidding request occurs when a viewer visits a web page at the moment. 2. The Ad Exchange sends an opening notice of an auction to DSPs. 3. Each DSP decides an advertisement to bid. 4. The DSP having the maximum bidding price wins, and actually delivers the advertisement. 2

5. The advertisement is displayed simultaneously when the viewer opens the web page. Second price auctions are used as this auction method, and we consider that this auction is also a sequential auction in terms of trading the same ad space constantly. We cannot find remarkable features in this method of display advertisement auctions. However, advertisers, who are the end buyers, who do not directly bid, but DSPs do [6]. DSPs have a number of advertisers as their clients, and take bids instead of advertisers to satisfy advertiser s constraints such as a budget constraint. Although a DSP competes against other DSPs, it has the space to coordinate bidding among its clients. We consider that this auction allows a part of bidders colludes with each other in the conventional auction theory. Therefore, an efficient bidding strategy is not present. In this study, I aimed to propose an efficient bidding strategy for DSPs, the intermediaries. As mentioned above, the bidding strategy is important for both DSPs and advertisers. Because it decides a cost-benefit performance of an advertisement. The better bidding strategy is desirable, however, the existing strategy does not consider the influence between advertisers, therefore advertisers cannot get the results as they have expected. The decision problem of the bidding strategy in the conventional auction theory is discussed mostly from the position of bidders, however I try to propose the efficient bidding strategy that the DSP intermediates. Specifically, I propose the bidding strategy which can get a high cost-benefit performance, where the intermediary considers the influence of advertisers and ad spaces, and coodinates biddings. Next, I show the effectiveness of the proposed strategy by the experimental evaluation using the real auction data. Moreover, I discuss the influence of the proposed strategy to advertisers and ad spaces by analyzing the evaluation result multidirectionally. To acheive this goal, I focused on following problems: Proposal on an efficient bidding strategy The strategy which is currently used by DSPs is seemingly reasonable. Nevertheless, they have problems such as an overpayment for ad spaces. Therefore, I must reveal why above-mentioned problems occurred, and propose the bidding strategy to solve this problem. Analysis of auctions for the case of applying the proposed strategy 3

I must measure a performance of the proposed strategy to evaluate whether or not the proposed strategy is more efficient than the existing strategy. In addition, I need to comprehend the effect of the proposed strategy by analyzing auction data of the evaluation in order to predict what is going to happen if the proposed strategy is introduced. In order to solve the first problem, I first pointed the problem that the existing strategy treated display advertisement auctions as single unit auctions despite the fact that display advertisement auctions are originally multi unit auctions, and proposed the bidding strategy that a DSP coordinated bidding among its clients and decided bidding prices to ad spaces on day to day basis. Advertiser s demands have an affect on each other in the proposed strategy, therefore the problem that it is difficult to decide when bidding prices originate. To deal with this problem I used a market model, which is a method to solve resource allocation problems using prices. To solve the second problem, I implemented a computer simulator, and evaluated the performance of the proposed strategy on a virtual auction environment. Moreover, I analyzed the evaluation result, and discussed the influence of the proposed strategy to advertisers and ad spaces. I show the composition of this paper as below. In Chapter 2, I introduce the elements of the auction theory such as second price auctions, sequential auctions. In Chapter 3, I introduce the related work of this study. In Chapter 4, I explain a feature of display advertisement auctions first. Next, I formulate the decision problem of the bidding strategy, and remark the existing strategy and its problem. In Chapter 5, I deal with a basic idea of the proposed strategy, the adverse effect of the proposed strategy, and the market model. Next, I explain how I apply the market model to the proposed strategy. Moreover, I remark an estimation method of other DSPs evaluation values. In Chapter 6, I show the effectivity of the proposed strategy by an evaluation using real auction data. In Chapter 7, I analyze the experimental result multidirectionally. In Chapter 8, I discuss about the result of analysis in Chapter 7. Finally, I represent a conclusion in Chapter 9. 4

Chapter 2 Preparation In this chapter, I explain the elements of the auction theory, which are related to display advertisement auctions. With these elements in mind, I think of bidding strategies in this study. First, an auction is one of processes to decide allocation of goods and payment of a buyer when a seller does not know the buyer s values for goods. There are many methods in an auction. An auction is divided into open bidding auctions and secret bidding auctions according to whether or not a bidder can know bidding prices of other bidders. Moreover, we can classify an auction based on whether the number of goods sold in an auction is only one or not. Second price auctions is one of secret bidding auctions, and sequential auctions is one of multiple auctions. As mentioned in Chapter 1, we consider that display advertisement auctions is a combination of second price auctions and sequential auctions. These auctions are explained below. 2.1 Second Price Auctions In the secret bidding auction, where bidders cannot know bidding prices of other bidders with each other, first price auctions is the most famous auction. In this auction, the bidder whose bidding price is the highest can get a good by paying his or her price. Second price auctions is invented by Vickrey [7]. In this auction, the bidder whose bidding price is the highest also becomes a winner, however, the price of payment is the second highest bidding price. For example, we assume that an ad space is sold at an auction, and the DSP A offers 2, and the DSP B offers 1. In this case, the DSP A can get the ad space and the price of payment is 1. In second price auctions where the number of goods is only one, it is known that the equilibrium is the strategy to offer the bidder s evaluation value to the good as the bidding price directly. In fact, the bidders offer their evaluation values to the good in this auction. 5

2.2 Sequential Auctions Sequential auctions is auctions where multiple goods are sold by holding single unit auctions sequentially [8]. As discussed in greater detail below, the existing strategy of the DSP treated display advertisement auctions as single unit auctions despite the fact that display advertisement auctions are originally multi unit auctions. Milgrom widely discusses sequential auctions, however does not deal with the situation focused on in this study [9]. In display advertisement auctions, second price auctions take place, however, we consider that this auction is also multiple sequential auctions. Therefore, there is not the bidding strategy to be the equilibrium, and the more efficient bidding strategy is needed. 6

Chapter 3 Related Work In this chapter, I introduce the previous study regarding collusions written by McAfee and McMillan [10]. Whether there are bidder s incentives to break collusions depends on the method of auctions. Bidders do not have incentives to break collusions in English auctions, second price auctions, and Vickrey auctions. For example, the bidder A has an evaluation value 10, and other bidders have evaluation values which are not greater than 9. In English auctions, the collusion that the bidder A offers 3, and other bidders offer 2 is possible. In this case, if other bidders try to break the collusion, the bidder A can find the betrayal soon, and he or she can raise his or her bid. Collusions are also possible in second price auctions. For instance, the collusion that the bidder A offers 10, and other bidders offer 2 is possible. Other bidders cannot make a profit if they break the collusion. On the other hand, bidders have incentives to break collusions in Dutch auctions and first price auctions. For example, when the bidder A offers the bid which is not greater than 9, other bidders can make a profit by breaking the collusion, and the bidder A cannot prevent them from breaking the collusion. Display advertisement auctions use second price auctions as a protocol, therefore bidders do not have incentives to break collusions. In the case of collusions, the winner and the allocation of profit are not obvious. McAfee and McMillan say that conducting pre-auctions before auctions can solve this problem. McAfee and McMillan analyzed 2 types of a collusion, a weak cartel and a strong cartel. A weak cartel is a collusion in which the bidders cannot make side payments. A strong cartel is a collusion in which the cartel members can make transfer payments. In a weak cartel, the optimal strategy is that bidders whose evaluation values are not lower than the reservation price r offer the value directly. In a strong cartel, the optimal strategy is that the cartel members decide the bidder who has the highest evaluation value in the pre-auction, and the bidder offers the reservation price r, and shares benefits with the members by using side payments. 7

In the same way as collusions described in the previous study, intermediating by the DSP can decrease advertisers payments compared to the case that individual advertisers participate in the auction directly. McAfee and McMillan described the optimal strategy of the cartel members in single unit auctions, however, did not deal with multi unit auctions, which were described in this study. I consider that collusions in multi unit auctions are more difficult problems than single unit auctions. For example, a DSP coordinates biddings between an advertiser A and an advertiser B. As above mentioned, this is considered as a collusion between A and B. In multi unit auctions, we have to consider the coordination that the DSP makes an advertiser A get an ad space, and an advertiser B get an another one. Therefore, we can say the range of the bidding coordination in display advertisement auctions is wider than single unit auctions remarked by McAfee and McMillan. This point is different between collusions in the previous study and the bidding coordination by the DSP in display advertisement auctions. 8

Chapter 4 Analysis of Existing Strategy First, I explain a feature of display advertisement auctions in this chapter. Next, I formulate the decision problem of the bidding strategy and explain the existing strategy and its problem. 4.1 Feature of Display Advertisement Auctions In this section, I explain the targeting type which is one of advertisements parameters, which is a feature of display advertisement auctions. Targeting types are one of advertisements parameters, which are used to determine how to narrow down viewers to bid according to viewers attribute information. As I mentioned the flow of display advertisement auctions in Chapter 1, DSPs can bid by a viewer in this auction. Compared with conventional advertisement delivery systems, display advertisement auctions are superior because the cost-benefit performance of advertisements is higher by coordinating the bidding price by a viewer. On a technical note, DSPs can get viewers attribute information from their browser s cache. Obtaining data to show what web pages the viewer have visited from cache, DSPs can comprehend information of the viewer such as sex, age, interest and so on. For example, DSPs can consider to improve advertising effectiveness when they display automobile advertisements only to 20 s or 30 s men. There are 3 kinds of targeting types as shown in Table 1. As we know from Table 1: Kinds of targeting types Targeting type Broadreach Audience targeting Retargeting Content To display the advertisement to all viewers without narrowing down viewers according to their attribute information. To display only to viewers who seem to be interested in the advertisement. To display only to viewers who have already visited the advertiser s web page. 9

a little thinking about targeting types, narrowing down become tighter in the order of targeting types, broadreach, audience targeting, retargeting. Therefore advertising effectiveness also become higher in the same order. Moreover, bidding prices of broadreach are apt to be set low comparatively. On the other hand, bidding prices of audience targeting and retargeting are apt to be set high because of a chance of a high conversion rate. Figure 2 shows the number of bids and the average bidding price hourly of the advertisement showing features of each targeting type. A horizontal axis shows the time, and a vertical axis shows the number of bids and the average bidding price. Broadreach has the./#'+01(#$'23'()*! '(!!!!" '!!!!!" &(!!!!" &!!!!!" %(!!!!" %!!!!!" $(!!!!" $!!!!!" (!!!!" *+,"-./0,1"23"045" 67,189,"04554-9":14;,"!"!" (" $!" $(" %!".)1#!!#!)"!#!("!#!'"!#!&"!#!%"!#!$"!"!"#$%&#'()**)+&',$)-#! Broadreach '(!!!!"!#!)" '(!!!!"!#!)"./#'+01(#$'23'()*! '!!!!!" &(!!!!" &!!!!!" %(!!!!" %!!!!!" $(!!!!" $!!!!!" (!!!!"!#!("!#!'"!#!&"!#!%"!#!$"!"#$%&#'()**)+&',$)-#!./#'+01(#$'23'()*! '!!!!!" &(!!!!" &!!!!!" %(!!!!" %!!!!!" $(!!!!" $!!!!!" (!!!!"!#!("!#!'"!#!&"!#!%"!#!$"!"#$%&#'()**)+&',$)-#!!"!" (" $!" $(" %!".)1#!!"!"!" (" $!" $(" %!".)1#!!" Audience targeting Retargeting Figure 2: The number of bids and the average bidding price of each targeting type largest number of bid in total in all targeting types. On the other hand, we notice that the average bidding price is lower than other targeting types. Audience targeting is fewer than broadreach in terms of the number of bids, and we also know audience targeting narrows down viewers to bids. In terms of the 10

average bidding price, audience targeting sets the higher price than broadreach. Incidentally, we consider that the reason why the number of bids and the average bidding price is 0 in around 15 o clock is because the advertiser used up the budget in 14 o clock. Retargeting is fewer than audience targeting in terms of the number of bids, and we also know retargeting narrows down viewers to bid tightly than audience targeting. The average bidding price of retargeting is very high. Of course, these advertisements are only a sample of each targeting types, and the trend of advertisements also depends on the budget and other properties. However, we can conclude the trend of each targeting types as shown in Table 2. Table 2: The trend of targeting types Targeting type The number of bids Bidding price Broadreach Large Low Audience targeting Medium Medium Retargeting Small High I explained the advertisements targeting type, which is a feature of display advertisement auctions. We do not include targeting types in the formulation of the decision problem of the bidding strategy, but discuss them in Chapter 7. 4.2 Formulation of Bidding Problem A decision-making problem of each advertiser i is to decide the bidding price b ij to an ad space j. I assume that a function f ij (b ij ) expresses the number of impressions in the ad space j in a day when the advertiser i offers the bidding price b ij to the ad space j. I consider that I can get this function from the analysis of the past auction data. I also assume that icvr icvr ij is given. icvr is a conversion rate per impression, and is defined every combination of the advertiser i and the ad space j. I can define the decision-making problem of each advertiser as below. Decision variable: The bidding price b ij to each ad space j 11

Objective function: To maximize the number of conversions f ij (b ij )icvr ij j Constraint conditions: Budget constraint: f ij (b ij )b ij budget per day j CPA constraint: f ij (b ij )b ij j target CP A f ij (b ij )icvr ij j Budget constraints represent upper limits of the advertiser s budget in a day, and CPA (Cost Per Action) constraints represent upper limits to cost a conversion. Advertisers give the DSP these constraints in advance. The DSP is able to bid as long as it satisfies these constraints. On the other hand, the decision-making problem of the DSP is defined as below. Decision variable: The bidding price b ij of each advertiser i to each ad space j Objective function: To maximize the number of impressions f ij (b ij ) i,j Constraint conditions: For all advertiser i Budget constraint: f ij (b ij )b ij budget per day j CPA constraint: f ij (b ij )b ij j target CP A f ij (b ij )icvr ij j DSPs can display advertisements if they win auctions. DSPs charge advertisers according to the number of impressions, therefore the objective function of the DSP is defined as above. The goal of this study is to find the bidding strategy to solve this optimization problem. 12

4.3 Details and Problem of Existing Strategy The simple way as the DSP s bidding strategy is as below when I assume each advertiser behaves individually. 1. Each advertiser decides the bidding price to each ad space with satisfying the budget constraint and the CPA constraint. 2. Auctions take place inside DSP, and the advertiser who has the maximum bidding price is the winner of this auction inside DSP, and the DSP offers the bidding price to Ad Exchange. 3. The DSP repeats the above process until advertisers use up their budget. Monotonicity that the winning rate increases with increasing the bidding price is expected. However, the phenomenon that the winning rate increases and decreases with increasing the bidding price is found in fact. Figure 3 shows the relationship between the bidding price and the winning rate in an ad space. The horizontal axis shows the bidding price, and the vertical axis shows the winning rate. We can see the above phenomenon in Figure 3. $"!#,"!#+"!#*"!"##"#$%&'()!!#)"!#("!#'"!#&"!#%"!#$"!"!"!#!+"!#$)"!#%'"!#&%"!#'"!#'+" *"++"#$%,&"-)! Figure 3: Relationship between the bidding price and the winning rate 13

In the auction theory, the fact that the winning price does not always decrease monotonically with holding each auction in sequential auctions has been discussed. It is difficult to decide the optimal bidding strategy when monotonicity is not satisfied. I consider that the above bidding strategy of the DSP causes the above vibration of the winning rate. The above bidding strategy is seemingly considered as the rational action, however, there is the problem that the above bidding strategy treated display advertisement auctions as single unit auctions, though display advertisement auctions are originally multi unit auctions. However, from a big-picture viewpoint, if I assume that other DSPs also use the above simple bidding strategy, advertisers having the high bidding price use up their budget at early time and drop out of the auction. Therefore, I consider the winning price drops down as time goes by. Figure 4 shows the relationship between the time and the winning price in an ad space, and the horizontal axis shows the time, and the vertical axis shows the average winning price per hour. We know that the winning price a little goes up and down, how-!#(&"!#(%"!#($"!"##"#$%&"'%()"*+!!#("!#!'"!#!&"!#!%"!#!$"!"!" %" '" ($" (&" $!","-+! Figure 4: Relationship between the time and the winning price ever, it drops down with time advance in totally from Figure 4. This behavior is 14

based on the fact that the auction starts at 00:00:00 and ends at 23:59:59, and a budget of each advertiser is reset in current display advertisement auctions. Of course, the DSP cannot observe when other DSPs reset budgets of their clients, therefore there are some case that the above fact does not fit, however, I assume that all DSPs reset their clients budget at 00:00:00 in this study. I think of the bidding strategy to get impressions more effectively than the above existing strategy with considering the above fact in the next chapter. 15

Chapter 5 Proposed Strategy I first explain tha basic idea of my proposed strategy in this chapter. Next, I remark an adverse effect by the proposed strategy, and explain how to apply the market model to this strategy to solve the adverse effect. Moreover, I present the method to estimate the function representing the relationship between the bidding price and the number of impressions, which is needed to apply the proposed strategy to the display advertisement auctions. 5.1 Basic Idea To deal with the problem mentioned in Section 4.3, we do not consider that each auction of sequential auctions is individual, but a whole auction on day to day basis. As mentioned above, the DSP can coordinate biddings among advertisers, who are its clients. For example, we consider an ad space is sold 3 times. DSPs and their advertisers are as below. DSP A Advertiser A1: value 10 Advertiser A2: value 6 DSP B Advertiser B1: value 7 Advertiser B2: value 3 I assume each advertiser can show his or her advertisement only one time, and both DSPs use the above simple bidding strategy. The auction result is as shown in Table 3. Table 3: Auction by the existing strategy Auction No. 1 2 3 Winner A1 B1 A2 Payment 7 6 3 In this case, if the advertiser A1 does not offer 10, but 6, he or she can be a winner 1 out of 3 (Table 4). In Table 4, the advertiser A1 can get the 16

Table 4: Auction by the proposed strategy Auction No. 1 2 3 Winner B1 A1 A2 Payment 6 3 3 ad space more cheaply than in Table 3. In fact, it is possible to get the same number of impressions in lower cost than the existing strategy by the strategy that bidding prices are not decided per each advertiser, but considering the minimum bidding price to win 2 times in the ad space, and the DSP deals out ad spaces to advertisers afterward. Moreover, the DSP appropriates the surplus for new bids, and can get more impressions in the same budget. Based on this strategy, I formulate the decision-making problem of DSPs, and it is described as below. In this formulation, the function f j (b j ) represents the number of impressions that the DSP gains in a day by bidding b j. I can also get this function from the analysis of the past auction data like f ij, and remark the detail way in Section 5.4. Decision variable: The bidding price b j to each ad space j, and x ij, which means the number of impressions of the advertiser i when getting the ad space j Objective function: To maximize the number of impressions x ij i,j Constraint conditions: For all advertiser i Budget constraint: xijb ij budget per day j CPA constraint: x ij b ij j target CP A x ij icvr ij j The total number of impressions is the sum of impressions of each advertiser: f j (b j ) = i x ij 17

The difference from the formulation in Section 4.2 is to decide a bidding price as DSP, and offer it each time there are bidding requests from Ad Exchange. This strategy can increase the number of impressions compared with the simple bidding strategy, however, advertisers constraints have an influence on each other. In other words, when an advertiser raises a bidding price in an ad space to increase the number of impressions, other advertisers have possibilities not to lose their some impressions. I explain this problem in the next section with using examples. 5.2 Adverse Effect by Bidding Coordination As mentioned in the previous section, the approach in this study is to buy ad spaces when their prices are low because the winning price drops down with time in a big picture. To achieve this, the DSP conducts the bidding coordination, and the adverse effect discussed below becomes possible. I explain this using concrete examples. I assume that there is an ad space, and 18 bidding requests occur in a day. There are also the DSP A and B, and they have advertisers as clients as below. DSP A Advertiser A1: value 9 budget 25 Advertiser A2: value 6 budget 16 Advertiser A3: value 4 budget 8 Advertiser A4: value 1 budget 3 DSP B Advertiser B1: value 8 budget 20 Advertiser B2: value 7 budget 20 Advertiser B3: value 5 budget 15 Advertiser B4: value 2 budget 4 As mentioned above, I assume that all advertisers can bid all 18 bidding requests in this section. Example 1 I consider that the both DSP A and B hold the inner auction, and offer the bidding price of the advertiser who has the maximum value, and we 18

obtain the result as below. For example, I assume the advertiser A1 does not bid although he or she leaves 1 of budget in the 4th auction, because he or she overspends his or her budget if he or she offers his or her value 9. Table 5: Auction result of Example 1 Auction No. 1 2 3 4 5 6 7 8 9 Winner A1 A1 A1 B1 B1 B1 B2 B2 B2 Payment 8 8 8 6 6 6 6 6 6 Auction No. 10 11 12 13 14 15 16 17 18 Winner A2 A2 A2 B3 B3 B3 A3 A3 A3 Payment 5 5 5 4 4 4 2 2 2 Example 2 Unlike in the case of Example 1, I consider the DSP A conducts the bidding coordination and the DSP B does not change the bidding strategy. I assume the DSP A continues to offer 5 ϵ. ϵ represents a minimum bidding unit, and I assume it is close to 0. Although the advertisers A1 and A2 leave Table 6: Auction result of Example 2 Auction No. 1 2 3 4 5 6 7 8 9 Winner B1 B1 B1 B2 B2 B2 B3 B3 B3 Payment 5 5 5 5 5 5 5 5 5 Auction No. 10 11 12 13 14 15 16 17 18 Winner A1 A1 A1 A2 A2 A2 A3 A3 A3 Payment 2 2 2 2 2 2 2 2 2 their budget, the DSP A decides the bidding advertiser to let all advertisers bid. As this result shows, as compared to Example 1, the advertiser A1 can reduce payments from 24 to 6, and A2 can also reduce from 15 to 6 without reducing the number of impressions. Example 3 In Example 2, the advertiser A1 and A2 still leave their budgets. When A1 demands more impressions and continues to bid until using up the budget, A3 cannot bid to display the advertisement. To deal with this, the DSP 19

A offers 5 + ϵ to bid and tries to get more impressions. As a result, we obtain the following result. Table 7: Auction result of Example 3 Auction No. 1 2 3 4 5 6 7 8 9 Winner B1 B1 B1 B2 B2 B2 A1 A1 A1 Payment 5 5 5 5 5 5 5 5 5 Auction No. 10 11 12 13 14 15 16 17 18 Winner A1 A1 A2 A2 A2 B3 B3 B3 A3 Payment 5 5 5 5 5 4 4 4 2 I assume that after A1 and A2 used up their budget, A3 offers the value 4. Because if A3 offers 5+ϵ to bid and the payment exceeds the value 4, the utility becomes negative. The number of impressions of the DSP A is still 9, however, the number of impressions of A1 increases from 3 to 5, and the A3 s amount of impressions decreases from 3 to 1. As just described above, the number of impressions of an advertiser changes by demands of other advertisers in the same DSP. As the above examples show, when DSPs conduct the bidding coordination among ad spaces, DSPs also include coordinating how much each advertiser demands each ad space. In the case that there are many advertisers, we need to solve the large problem, and it can be an obstacle to apply the proposed strategy. In fact, according to real auction data of June 29th, 2012 offered by MicroAd, Inc. who actually operates a DSP, hundreds of millions of bidding requests occur in a day in real display advertisement auctions. Therefore we find that the scale of the market is large. Moreover, there are only several DSPs in the country. When we consider these facts, a DSP has many advertisers as clients, therefore it is expected for the amount of calculation to be enormous. To solve this problem, I propose the method using the market model in the next section. I should add that the real auction is a little more complicated. Even if the bidding request from the same ad space occurs, the set of advertisers to request 20

to bid differs depending on the attribute information of viewers. That is, if the overlapping set of bidding targets is large, the DSP decides just one bidding price for the ad space, however, if the overlapping set of bidding targets is small, the DSP needs to decide bidding prices per advertiser. 5.3 Bidding Coordination among Advertisers Using Market Model The market model is the method to solve resource allocation problems using prices based on the general equilibrium theory as Figure 5 illustrates [11]. In this 4. Updating prices! Price 1! Price 2! Price 3! Ad space 1! Ad space 2! Ad space 3! 1. Noticing prices! 5. Noticing prices again! 3. Declaring demands! Advertiser 1! Advertiser 2! Advertiser 3! Advertiser 4! Demand:! (1, 2, 3)! (1, 2, 3)! (1, 2, 3)! (1, 2, 3)! 2. Calculating demands! Figure 5: Schematic view of market model study, we consider that the DSP is a seller, ad spaces are goods, and advertisers are buyers. Although advertisers ask the DSP to bid instead of them, they still have independency of settings to budgets and targets, therefore, we consider the proposed strategy is with a high affinity of the market model. In the market model, the DSP coordinates bidding prices to ad spaces virtually, and actually applies the bidding prices which finally have been calculated to the real auction. The DSP does not bid to Ad Exchange in mid-flow of calculation using the 21

market model. The process of calculating the bidding prices using the market model is as below. 1. Ad spaces broadcast the current prices. The initial value is the price to be needed to get an impression. 2. Each advertiser decides the demand of impressions under given prices to satisfy the budget constraint and the CPA constraint, which maximizes the number of impressions. 3. Each ad space calculates the total demand from all advertisers. If the total demand its supply, the ad space raises its price by an unit. On the other hand, if the total demand < its supply, the ad space reduces its price by an unit. 4. The DSP repeats the above process until the prices of all ad spaces do not change. 5. The DSP offers the price of each ad space to Ad Exchange and decides the allocation of what advertisements are displayed according to the demands, when the prices converge. One of questions is whether or not this demand coordination mechanism converges when applying the market model. The demand decreases monotonically when the price of the ad space increases. On the other hand, the supply decreases monotonically with increasing the price of the ad space, therefore the equilibrium price is ensured. Moreover, the demand of advertisers is alternative because there is no case that the value of the ad space increases for the advertiser when he or she gets other several ad spaces, so we can find the equilibrium price by the above searching process and solve the problem as Example 2 and Example 3 in the previous section. In Chapter 6, I use the algorithm as shown in Algorithm 1 to evaluate the proposed strategy. GetP ricef orunit is the function to return the bidding price to get the number of impressions which is given as a parameter. I explain the method for acquiring this function in the next section. SolveLinearP rogramming is the function to return the demand of each ad space by solving the linear programming problem. This requires the budget constraint and the CPA constraint as constraint conditions. Moreover, I add 22

the quota constraint to prevent the price from vibrating drastically. Algorithm 1 Calculate a set of bid prices and demands Require: quota: quota {Initialization} for i = 1 to num buyer do for j = 1 to num seller do new demand ij 1 end for end for {Main loop} repeat demand ij new demand ij for all i, j for j = 1 to num seller do price j GetP ricef orunit( i demand ij ) {price update} end for for i = 1 to num buyer do new demand ij SolveLinearP rogramming(quota) {demand update} end for until new demand ij = demand ij for all i, j return price j, demand ij for all i, j If an arbitrary change of the demand is accepted, for example, when the ad space 1 s price is lower than the ad space 2 s price, the ad space 1 collects more purchase requests than the ad space 2. In the next step, the demands are shifted from the ad space 1 to the ad space 2. This phenomenon is repeated, and prevents the stabilization of the prices. If I set a quota small, it takes a long time to converge, and if I set a quota large, the vibration of the price occurs easily. Because of this trade-off, we need to find the suitable quota. I consider the DSP decides the bidding price per ad space in this study. This may also be possible when each advertiser participates in the auction 23

individually without the DSP. However, it is difficult to decide the suitable bidding price in this situation because their actions have an affect on each other. In response, when the DSP intermediates, an advertiser can decide the demand with knowing other advertiser s actions. Currently, the number of DSPs is small, so the advantage by the intermediary is expected. 5.4 Estimation of Evaluated Value of Other DSP To apply the method to decide the bidding price using the market model to the real auctions, estimating to the function f j (b j ) representing the relationship between bidding prices and impressions is needed (GetP ricef oru nit in Algorithm 1 is the inverse function of f j (b j )). This function shows the number of winning in a day with continuing to offer the bidding price b j to an ad space. I need the distribution of the other DSPs bidding price to get the above function. For example, there are 4 bidding requests, and the other DSP has values of 10, 8, 5, 3. In this case, the DSP gets 3 impressions with continuing to offer 9, and gets 2 impressions with continuing to offer 7, and gets 1 impression with continuing to offer 4. When the DSP wins an auction, the payment is the bidding price of the other DSP because second price auctions take place in Ad Exchange. When there are several DSPs, the payment is the maximum bidding price in the other DSPs bidding price. On the other hand, when the DSP fails to win, it is difficult to know the accurate bidding price of the other DSPs. Figure 6 shows the relationship between bidding prices and winning prices of some ad spaces when the DSP wins the auction. The horizontal axis is the bidding price, and the vertical axis is the winning price. Table 8 shows coefficients of correlations between bidding prices and winning prices on each ad space. The bidding price corresponds to the various winning prices from Table 8: Correlation between bidding prices and winning prices on ad spaces Ad space [1] [2] [3] [4] Coefficient of correlation 0.713 0.641 0.542 0.434 Figure 6, and I consider there is no correlations between bidding prices and 24

winning prices. However, coefficients of correlations show relatively high values in Table 8, and it shows weak associations. This is because that the method of display advertisement auctions is second price auctions, therefore the observable winning price is certainly less than the bidding price. I must consider these!#)"!#'"!#("!#&$"!"##"#$%&"'%()"*+!!#'"!#&"!#%"!"##"#$%&"'%()"*+!!#&"!#%$"!#%"!#$"!#!$"!"!"!#%"!#'"!#)"!#*" $" $#%","''"#$%()"*+%!"!"!#%"!#&"!#'"!#("!#$"!#)","''"#$%()"*+! Ad space [1] Ad space [2]!"##"#$%&"'%()"*+!!#'"!#&$"!#&"!#%$"!#%"!#!$"!"!"!#&"!#("!#)"!#*" %" %#&","''"#$%()"*+!!"##"#$%&"'%()"*+!!#!+"!#!*"!#!)"!#!("!#!'"!#!&"!#!%"!#!$"!"!"!#$"!#%"!#&"!#'"!#("!#)","''"#$%()"*+! Ad space [3] Ad space [4] Figure 6: Relationship between bidding prices and winning prices correlations, however I assume the DSP and the other DSPs are symmetric, and consider the following simple procedure as the approximating method for the other DSPs values. 1. The DSP calculates the price of each ad space which the other DSPs will not offer in advance such as the maximum bidding price or the maximum winning price and so on, and considers the price as the upper limit of each ad space. 2. The DSP considers the other DSPs value is distributed uniformly between the bidding price and the above upper limit when the DSP cannot win. 25

A symmetric relation represents there is non-biased in the scale of biddings and ad spaces to bid. As discussed above, I summarize the method to estimate the function f j (b j ) of bidding prices and impressions as below. When the DSP succeeds to win, considering the winning price as the other DSP s value. When the DSP fails to win, considering the price which is distributed uniformly between the bidding price and the upper limit as the other DSPs value. The question how optimistic or pessimistic the above estimation is still leaves, therefore, I compare this estimation to the method for estimating the distribution of the other DSPs values using the maximum-likelihood approach from only the observable auction data [12]. I select the maximum bidding price and the maximum winning price as the above upper limit in each ad space, and call these estimations the pessimistic estimation and the optimistic estimation respectively. Figure 7 shows the function of bidding prices and impressions in an ad space by each estimation. The horizontal axis is the bidding price, and +!!!!!!" *!!!!!!" )!!!!!!" (!!!!!!"!"#$%&&'()&* '!!!!!!" &!!!!!!" %!!!!!!" $!!!!!!" -./012345" #!!!!!!" 67829:80" ;.::929:80"!"!"!,#"!,$"!,%"!,&"!,'"!,("!,)"!,*" +',,')-*#$'.%! Figure 7: Function of bidding prices and impressions by each estimation the vertical axis is the estimated number of impressions. Benchmark shows the 26

estimation in [12], Optimistic shows the estimation using the maximum bidding price as the upper limit in this study, and Pessimistic is the estimation using the maximum winning price as the upper limit. Figure 7 shows the DSP can get f(x) impressions if the DSP continues to offer the bidding price x to this ad space. We find that 2 estimations in this study are more pessimistic than the reasonable benchmark, and not optimistic estimation at least. We have prepared for applying the proposed strategy to the real auction. In the evaluation of the next chapter, I also use the pessimistic estimation and the optimistic estimation as the method to decide the upper limit of the price. 27

Chapter 6 Evaluation of Proposed Strategy In this chapter, I remark the evaluation of the proposed strategy. The proposed strategy has possibility for getting impressions more effectively than the existing strategy, however we do not know the result in the real auction. For example, if the rate of advertisements which originally have low bidding prices is high, the bidding price by the coordination also becomes low. As a result, we easily imagine the number of impressions does not increase by the proposed strategy. Therefore we need to confirm whether or not the proposed strategy is effective in the environment using the real auction data. I explain the evaluation method first, then the evaluation result. 6.1 Method of Evaluation I conducted the evaluation experiment using the real data to evaluate the performance of the proposed strategy. The data used in this experiment are the data in June 29th and 30th, 2012, which are offered by MicroAd, Inc. It is difficult to conduct the experiment using all data, therefore I extracted 100 advertisements and 100 ad spaces from the above data, and made the virtual auction environment, and compared the number of impressions gotten by the proposed strategy to the number of impressions gotten by the simple existing strategy. The advertisements are 100 which have most impressions in June 29th, and the ad spaces are 100 which have most biddings in June 29th. The reason why I use the above selection is that I considered I could make the close situation to the scale of the real auction by using above advertisements and ad spaces. The experiment using the data of June 30th also use the same advertisements and ad spaces to compare the result to the result of 29th. Table 9 shows the scale of the each virtual environment. From this section, the number of impressions Table 9: Scale of the virtual environment Day The number of bids The number of wins 6/29 62,596,551 6,330,943 6/30 56,672,129 5,948,718 28

of the existing strategy means the number of wins in Table 9. I assume the number of auctions in a day is given. There are several DSPs except the DSP in fact, however I consider an aggregation of these DSPs, and call this aggregation the other DSP. I also assume the other DSP does not change its strategy. As regards the evaluation, I developed 3 programs as below. Bidding price estimator Given the real bidding log as input, this estimator complements the other DSP s bidding price when the DSP loses by the estimation method in Section 5.4. Bidding price calculator This program calculates the bidding price to each ad space by the method using the market model as shown in Section 5.3. Strategy Evaluator This Evaluator simulates auctions in a day. Inputs are data from the above estimator and calculator, and an output is a bidding log. I describe each program and explain the flow of the evaluation below. Bidding price estimator Table 10 shows the input data simply. The payment Table 10: Example of bidding log Bid ID Ad ID Ad space ID Bidding price Payment Time 1 50 20 0.4 0.3 2012-06-29 00:00:00 2 25 10 0.5 NULL 2012-06-29 00:00:01 3 79 16 0.25 0.2 2012-06-29 00:00:01 when the DSP cannot win the auction, that is, the other DSP s bidding price is NULL. This program aims to estimate this value. As the method is described in Section 5.4, this program sets up the upper limit of the price in each ad space from the bidding log, and consider that the other DSP s value is distributed uniformly between the bidding price and the upper limit. As the estimation method of the upper limit is also described in Section 5.4, I provide 2 patterns as the upper limit of the price, the maximum bidding price and the maximum 29

winning price in each ad space. As we can know easily, The estimation using the winning price as the upper limit is optimistic. This estimator makes it possible to calculate how much the bidding price for each ad space is needed to get the target number of impressions. This estimator is used in the following 2 programs. Bidding price calculator This program is to calculate the bidding price to each ad space based on the proposed strategy as described in Chapter 5. Inputs are the function of bidding prices and impressions calculated by using the above bidding price estimator, constraint conditions of each ad space (budget constraints and CPA constraints), and conversion rates of each advertisement to each ad space. The output is the bidding price to each ad space. Constraint conditions and conversion rates are given from the real auction data. We must pay attention not to use budgets which advertisers first set as budget constraints, but to use the sum of payments to each ad space which are selected in this simulation. By considering the payments which were actually used in the existing strategy as the budget constraint, I can evaluate the number of impressions gotten by the proposed strategy when the proposed strategy uses the same budget of the existing strategy. Strategy Evaluator This program is to evaluate the proposed strategy. Inputs are the complemented bidding log from the bidding price estimator, the bidding price to each ad space from the bidding price calculator, and advertisements budget constraints from the real data. Table 11 shows the output. When the DSP even loses an auction, the other DSP s bidding price is marked in the biddin log of Table 11 unlike Table 10. Bid ID Ad ID Ad space ID Table 11: Example of result log Bidding price Other DSP s bidding price Winner Payment Time 1 50 20 0.4 0.3 SELF 0.3 2012-06-29 00:00:00 2 25 10 0.5 0.55 ENEMY 0.5 2012-06-29 00:00:01 3 79 16 0.25 0.2 SELF 0.2 2012-06-29 00:00:01 The summarized flow of the above 3 programs and data is as below. 1. The bidding price estimator first receives the real bidding log and reads 30

out the bidding log whose other DSP s bidding prices are complemented. 2. Next, the bidding price calculator receives the budget constraint and the CPA constraint of each advertisement to each ad space, and conversion rates from the real auction data, and also receives the complemented bidding log from the bidding price estimator. This program calculates bidding prices to each ad space from these data. 3. Finally, the strategy evaluator simulates an auction receiving the budget constraint and the CPA constraint of each advertisement to each ad space, conversion rates from the real auction data, and the complemented bidding log from the bidding price estimator, and bidding prices from the calculator. As shown above, outputs are bidding logs in Table 11. I obtain the number of impressions by the proposed strategy from the output bidding log. The number of impressions by the existing strategy is the number of impressions which extracted 100 advertisements actually gained from extracted 100 ad spaces. I compare these number of impressions, and evaluate whether the proposed strategy can get impressions more effectively than the existing strategy. I prepare several combinations of evaluation methods, and they are actually combinations of the environment for calculating bidding prices and the environment for evaluating strategies. Each environment depends on the date of data and the estimation method for the other DSP s bidding prices. In this study, I prepare environments in Table 12 for calculating bidding prices. I prepare Table 12: Environments for calculating bidding prices Date Estimation method Environment [1] 29th The maximum bidding price is a limit Environment [2] 30th The maximum winning price is a limit environments in Table 13 for evaluating strategies. I evaluated 8 combinations of Table 12 and Table 13. The proposed strategy assumes that the DSP decides bidding prices one day using auction data of the previous day, therefore, for example, the combination of the calculation environment [1] and the evaluation 31

Table 13: Environments for evaluating strategies Date Estimation method Environment [1] 29th The maximum bidding price is a limit Environment [2] 29th The maximum winning price is a limit Environment [3] 30th The maximum bidding price is a limit Environment [4] 30th The maximum winning price is a limit environment [1] is somewhat optimistic because the DSP decides bidding prices under knowing what happens in the day. On the other hand, the combination of the calculation environment [1] and the evaluation environment [3] is a more realistic evaluation because the date of the calculation is different from one of the evaluation. 6.2 Results of Evaluation Table 14 shows the number of impressions by each evaluation environment. Improvement rates show how much the proposed strategy increases impressions compared to the existing strategy. As shown in Table 14, the proposed strat- Calculation environment Table 14: Evaluation result Evaluation environment The number of impressions Improvement rate Evaluation [1] [1] [1] 6,702,407 +5.87% Evaluation [2] [1] [2] 11,839,678 +87.01% Evaluation [3] [1] [3] 7,633,190 +28.32% Evaluation [4] [1] [4] 10,308,194 +73.28% Evaluation [5] [2] [1] 6,391,897 +0.96% Evaluation [6] [2] [2] 9,273,242 +46.47% Evaluation [7] [2] [3] 6,917,054 +16.28% Evaluation [8] [2] [4] 9,125,224 +53.40% egy is beneficial in all evaluations. Especially, the proposed strategy improves impressions in realistic evaluation such as Evaluation [3] and Evaluation [5], therefore I conclude the proposed strategy can get impressions more effectively than the existing strategy. 32

I examine the simulation result of Evaluation [1] then. First, I compare bidding prices of the proposed strategy to the existing strategy s bidding prices, and Table 15 shows the result. In the existing strategy, each advertisement has Table 15: Average and variance of bidding prices of each strategy Average Variance Proposed strategy 3.71 10 2 2.77 10 9 Existing strategy 3.49 10 2 5.84 10 4 a different value from each other to an ad space, therefore the bidding price of the existing strategy is not unique unlike the proposed strategy. In Table 15, the average and the variance of the existing strategy are calculated by using the average values to each ad space. As described in Table 15, there is little difference between the proposed strategy and the existing strategy in terms of the average, however the variance of the proposed strategy is greatly smaller than the existing strategy. I consider this reason is the following phenomenon by the bidding coordination among ad spaces. )!!!!!" (!!!!!",-./.012"03-43156" 7890:;5"03-43156" '!!!!!"!"#$%&&'()&! &!!!!!" %!!!!!" $!!!!!" #!!!!!"!" #" #!" #*" $+" %)" &(" ''" (&" )%" +$" *#" #!!" *+,%$-&%"%).&! Figure 8: The number of impressions per advertisement Ad spaces whose price is high: Demands decrease, and the price also 33

decreases. Ad spaces whose price is low: Demands increase, and the price also increases. By the above phenomenon, the difference among bidding prices of ad spaces becomes small, as a result, the variance also becomes small. Next, Figure 8 shows the result of the comparison of the number of impressions per advertisement between the proposed strategy and the existing strategy. The horizontal axis is the number of advertisements, and the vertical axis is numbers of impressions. We know that the number of impressions in a day is improved by the proposed strategy from Table 14, however, the number of impressions of advertisements is not all improved from Figure 8. This is caused by that advertisements which offer high bidding prices in the existing strategy can buy ad spaces at a lower price by the proposed strategy, but advertisements which offer low bidding prices in the existing strategy must buy ad spaces at a higher price. I explain the detail analysis in the next chapter. Next, Figure 9 shows the result of the comparison of the number of impressions per ad space between the proposed strategy and the existing strategy. The horizontal axis is the number of ad spaces, and the vertical axis is numbers of impressions. Along with advertisements, we know the number of impressions of '#!!!!!" '!!!!!!",-./.012"03-43156" 7890:;5"03-43156"!"#$%&&'()&! &!!!!!" %!!!!!" $!!!!!" #!!!!!"!" '" '!" '(" #&" )*" $%" ++" %$" *)" &#" ('" '!!" *+,&#-.%&! Figure 9: The number of impressions per ad space 34

ad spaces is not all improved from Figure 9. We also find that the more impressions the ad space has in the existing strategy, the more active up and down of impressions in the proposed strategy is. I also explain the detail analysis in the next chapter. Incidentally, the reason why the number of impressions of the existing strategy basically increases in descending order in Figure 8 and Figure 9 is just that I assigned numbers to ad spaces in order of descending impressions or biddings. '!!!",-./.012"03-43156" 7890:;5"03-43156" &!!!"!"#$%&'("#'! %!!!" $!!!" #!!!"!" #" #!" #(" $)" %*" &+" ''" +&" *%" )$" (#" #!!" )*$%&+'%,%#'! Figure 10: The number of conversions per advertisement I explained about impressions so far, however, DSP is consistently the intermediary of advertisers. The number of impressions is surely important, however, it is also important to maximize the number of conversions, which is the objective of advertisers. The DSP earns the trust of advertisers and can get more clients by getting conversions more effectively than the other DSPs. The proposed strategy ensures some conversions by including the CPA constraints into constraint conditions of the linear programming problem in the bidding coordination, however, we do not know actual conversions. Table 16 and Figure 10 show the comparison result of conversions between the proposed strategy and the existing strategy. Table 16 shows the total number of conversions in a day of the proposed strategy and the existing strategy, and Figure 10 shows the 35

Table 16: Comparison of the number of conversions The number of conversions Proposed strategy 41,718 Existing strategy 24,695 comparison of the number of conversions per advertisements. The horizontal axis is the number of advertisements, and the vertical axis is numbers of conversions. Conversions are calculated using conversion rates which are obtained in the past bidding log, so consistently estimated values. We consider the proposed strategy improves not only impressions but also conversions from Table 16. I consider this reason is because the more impressions increase, the more conversions also increase simply. Along with impressions, we know advertisements do not entirely improve their conversions as shown in Figure 10. 36

Chapter 7 Analysis of Evaluation Results In this chapter, I analyze the result of Evaluation [1] to discuss the influence of the proposed strategy to advertisers and ad spaces in display advertisement auctions. First, I analyze the data from viewpoints of advertisers and ad spaces. Next, I examine the number of impressions in the case of the simulation where the DSP conducts the coordination only in temporal variation without the coordination among ad spaces. 7.1 Analysis from Viewpoint of Advertisements I analyzed the evaluation result gained in the previous chapter from a viewpoint of advertisers. In the previous chapter, we find advertisements do not entirely improve their impressions by the proposed strategy, however, we do not still know the features of the advertisements which increase or decrease their impressions. In this section, I categorize advertisements into targeting types which is one of parameters of advertisements as described in Section 4.1, and compare the number of impressions of the proposed strategy to the existing strategy. In Figure 8 in the previous chapter, whether or not an advertisement increases the impression by the proposed strategy depends on the advertisement, therefore, I divide advertisements into targeting types, and examine the rate of improved advertisements by the proposed strategy. Table 17 shows the result. We can find 60% advertisements of audience targeting and retargeting improve Table 17: Rate of improved advertisements by the proposed strategy Targeting type The number of advertisements The number of improved advertisements Improvement Rate Broadreach 17 0 0.00% Audience targeting 53 36 67.92% Retargeting 30 20 66.67% their impressions, however, impressions of all broadreach drop down. Next, I described the point diagram of the proposed strategy s impressions and the ex- 37

isting strategy s impressions per targeting type in Figure 11. The horizontal axis is the number of impressions of the existing strategy, the vertical axis is the number of impressions of the proposed strategy, and a single point in the diagram shows an advertisement. Increasing advertisements by the proposed strategy are above the dotted line in the diagram, and decreasing advertisements are below the line. In broadreach, the advertisement which gains many impressions by the existing strategy decreases impressions by the proposed strategy. On the other hand, in audience targeting and retargeting, the advertisement which gains a little impressions by the existing strategy increases impressions by the proposed strategy. I apply χ square test of independence to examine!"#$%&&'()&*(+*,-%*#$(#(&%.*&,$/,%01! (!!!!!" '!!!!!" &!!!!!" %!!!!!" $!!!!!" #!!!!!"!"!" #!!!!!" $!!!!!" %!!!!!" &!!!!!" '!!!!!" (!!!!!"!"#$%&&'()&*(+*,-%*%2'&3)0*&,$/,%01! Broadreach!"#$%&&'()&*(+*,-%*#$(#(&%.*&,$/,%01! &!!!!!" %#!!!!" %!!!!!" $#!!!!" $!!!!!" #!!!!"!"!" #!!!!" $!!!!!" $#!!!!" %!!!!!" %#!!!!" &!!!!!"!"#$%&&'()&*(+*,-%*%2'&3)0*&,$/,%01!!"#$%&&'()&*(+*,-%*#$(#(&%.*&,$/,%01! %#!!!!" %!!!!!" $#!!!!" $!!!!!" #!!!!"!"!" #!!!!" $!!!!!" $#!!!!" %!!!!!" %#!!!!"!"#$%&&'()&*(+*,-%*%2'&3)0*&,$/,%01! Audience targeting Retargeting Figure 11: Relation between the proposed strategy and the existing strategy per targeting type whether or not there is a relationship between targeting types and the number of improved advertisements by the proposed strategy. As a result, χ 2 = 26.080, and χ 2 0.01 = 11.345 when degree of freedom = 3, therefore the independence 38

between targeting types and the number of improved advertisements by the proposed strategy is rejected if the 0.01 significance level is used, and I conclude both sides have a relevance. Moreover, I apply a residual analysis, as a result, the number of improved advertisements of broadreach is small if the 0.01 significance level is used, and the number of improved advertisements of audience targeting is large if the 0.05 significance level is used. I conclude whether the advertisement improves its number of impressions by the proposed strategy depends on the targeting type of the advertisement from the above analysis. 7.2 Analysis from Viewpoint of Ad Spaces In this section, I analyzed the evaluation result from a viewpoint of ad spaces. I focus on the number of bidding requests and the average bidding price as the feature quantities. The number of bidding requests represents the scale of the ad space, and the average bidding price shows demands of advertisers. Because we can consider the higher the average bidding price is, the more factors to be likely to lead to conversions are, such as the large size of the ad space, frequent visit of viewers who have high motivation to purchase something. I divide all ad spaces into ad spaces whose impressions increase by the proposed strategy and ad spaces whose impressions decrease, and examine the number of bidding requests and the average bidding price. First, Figure 12 shows the number of bidding requests of ad spaces whose impressions increase by the proposed strategy and ad spaces whose impressions decrease by the proposed strategy. The horizontal axis is ad spaces whose impressions increase by the proposed strategy and ad spaces whose impressions decrease by the proposed strategy, and the vertical axis is the number of bidding requests. A single point shows an ad space in this figure. I apply two-sided t test using the 0.01 significance level to examine whether or not the difference of the bidding requests between increasing ad spaces and decreasing ad spaces is significant. As a result, t(98) = 0.737, p = 0.463, and there is no difference between increasing ad spaces and decreasing ad spaces, therefore, whether or not the number of impressions of each ad space is improved by the proposed strategy does not depend on the scale of the ad space. 39

&!!!!!"!"#$%&'(&)*)! %!!!!!" $!!!!!" #!!!!!"!"!" +,-%&.)&! #'" /&-%&.)&! '!" ('" Figure 12: Comparison of bidding requests between increasing ad spaces and decreasing ad spaces by the proposed strategy Next, Figure 13 shows the average bidding price of the increasing ad spaces by the proposed strategy and the decreasing ad spaces. The horizontal axis is ad spaces whose impressions increase by the proposed strategy and ad spaces whose impressions decrease by the proposed strategy, and the vertical axis is the average bidding price. A single point shows an ad space in this figure. I consider there is the difference between the distribution of the increasing ad spaces and the distribution of the decreasing ad spaces from Figure 13. I apply one-sided t test using the 0.01 significance level to examine whether or not the difference of the average bidding price between increasing ad spaces and decreasing ad spaces is significant. As a result, t(98) = 11.819, p = 1.455 10 20, therefore, I conclude the increasing ad spaces has smaller average bidding price than the decreasing ad spaces. That is, whether or not the number of impressions of each ad space is improved by the proposed strategy depends on the average bidding price. 40

!#(%"!#($"!"#$%&#'()**)+&',$)-#!!#("!#!'"!#!&"!#!%"!#!$"!"!".+-$#%/#! $)" 0#-$#%/#! )!" *)" Figure 13: Comparison of the average bidding price between increasing ad spaces and decreasing ad spaces by the proposed strategy 7.3 Bidding Coordination Only in Temporal Variation We can consider this strategy using the market model is the combination of the following 2 coordinations. Coordination among ad spaces To shift bids from ad spaces which have high prices to ad spaces which have low prices, and increase the total impressions. Coordination in temporal variation To shift bids from time when prices are high to time when prices are low, and increase the total impressions. In this study, there is the hypothesis that prices are generally high in the morning and low in the evening. In the case of the bidding coordination only in temporal variation without the coordination among ad spaces, I examine how many impressions the proposed strategy can gain. In the proposed strategy, the budget is the sum of payments which are actually used, and advertisers decide demands for ad spaces to maximize their impressions. On the other hand, in the bidding coordination only in temporal 41

variation, by adding the condition that each advertisement must use the budget, which is used in an ad space in the existing strategy, in the same ad space, I achieve the bidding coordination only in temporal variation. Table 12 shows averages and variances of bidding prices to each ad space of coordination only in temporal variation, the normal proposed strategy, and the existing strategy. The average and the variance of the existing strategy are calculated based on the average bidding price to each ad space. The variance of coordination only in Table 18: Average and variance of bidding prices on each strategy Average Variance Only in temporal variation 4.57 10 2 9.44 10 4 Normal proposed strategy 3.71 10 2 2.77 10 9 Existing strategy 3.49 10 2 5.84 10 4 temporal variation is closer to the existing strategy than the normal proposed strategy. This is because bidding coordinations are individually conducted in each ad space. Next, Table 19 shows the number of impressions by coordinations only in temporal variation, also shows the normal proposed strategy and the existing strategy for comparison. The bidding coordination only in Table 19: The number of impressions by coordinations only in temporal variation The number of impressions Only in temporal variation 5,326,277 Normal proposed strategy 6,702,407 Existing strategy 6,330,943 temporal variation is inferior to other strategies as shown in Table 19. First, the reason why the bidding coordination only in temporal variation is inferior to the normal strategy is simply caused by narrowing down a search range of the optimal solution without the coordination among ad spaces. However, the reason why the bidding coordination only in temporal variation is inferior to the existing strategy is not obvious. To reveal this reason, I divide all ad spaces 42

into ad spaces whose impressions increase by the coordination only in temporal variation and the other, and show these average bidding price distribution in Figure 14. I apply two-sided t test using the 0.01 significance level to examine!#(%"!#($"!"#$%&#'()**)+&',$)-#!!#("!#!'"!#!&"!#!%"!#!$"!"!".+-$#%/#! $)" 0#-$#%/#! )!" *)" Figure 14: Comparison of the average bidding prices between increasing ad spaces and decreasing ad spaces by coordination only in temporal variation whether or not the difference of the average bidding prices between increasing ad spaces and decreasing ad spaces by the coordination only in temporal variation is significant. As a result, t(98) = 4.848, p = 4.678 10 6, therefore the average bidding price of the increasing ad spaces by the coordination only in temporal variation is higher than the decreasing ad spaces. That is, I conclude ad spaces which have the high average bidding prices can improve the number of impressions by the coordination only in temporal variation. This result is opposite to the influence of the normal proposed strategy to ad spaces. 43

Chapter 8 Discussion In this chapter, I discuss the result of analysis obtained in the previous chapter. First, I discuss the analysis from a viewpoint of advertisers. As mentioned in Section 7.1, whether the advertisement improves its number of impressions by the proposed strategy depends on the targeting type of the advertisement from the above analysis. I conclude the proposed strategy has an effect on shifting the decline of broadreach to the increase of audience targeting and retargeting. The reason why the phenomenon happens is as below. As described above, broadreach generally has low bidding prices, and audience targeting and retargeting have relatively high values. If the proposed strategy is applied, bidding prices of all advertisements to each ad space become roughly equal. These bidding prices are higher than bidding prices of broadreach, on the other hand, lower than bidding prices of audience targeting and retargeting. As a result, broadreach cannot purchase more impressions than the existing strategy, and audience targeting and retargeting can purchase more impressions than the existing strategy. I have discussed targeting types, but I generally conclude that the proposed strategy have the effect of shifting the decline of advertisements having low values as typified by broadreach to the increase of advertisements having high values as typified by audience targeting and retargeting. There is the problem which effects by the proposed strategy is large, the decline of impressions of advertisements having low values and the improvement of impressions of advertisements having high values. I consider this is closely related to the rate of high values and low values of advertisements although I cannot conclude it because various factors such as CPA constraints, conversion rates will be also related to this problem. That is, I consider the proposed strategy can improve the total impressions if the bidding rate of advertisements having low values is high, because it is said that there are many losing biddings in the existing strategy, therefore many space to improve. However, I also consider too more high bidding rate of advertisements having low values may decrease the total impressions. This reason is as below. When the rate of biddings of advertisements having low values is high, the number of losing biddings 44

increase, however, the bidding prices decided by the proposed strategy is set to be low. As a result, the number of impressions increases less than expected or decreases. Next, I discuss the analysis from a viewpoint of ad spaces. As mentioned in Section 7.2, I concluded whether or not the number of impressions of each ad space is improved by the proposed strategy depends on the average bidding price. The reason why the number of impressions of each ad space by the proposed strategy depends on the average bidding price of each ad space in the existing strategy is considered as below. As mentioned in Section 7.2, we consider the average bidding price represents demands to the ad space. That is, the higher the average bidding price is, the more demands of advertisers are, and it is said that the competition with the other DSP is very keen. As shown in Figure 13, the average bidding prices of ad spaces which improve their impressions by the proposed strategy are low, the average bidding prices of ad spaces which decrease their impressions by the proposed strategy are high conversely. That is, I conclude advertisers shift their budget allocated to highly valuable ad spaces in the existing strategy into relatively popular ad spaces, therefore the DSP can purchase more impressions in total. I also consider too more ad spaces whose average bidding price is low may decrease the total impressions. When the portion of ad spaces having low average bidding prices is large, the space to improve is also large, however, the bidding prices are coordinated to be low by the proposed strategy. Therefore I consider the total number of impressions does not increase. I consider that there is the distribution of the average bidding price of each ad space to be the threshold in the same way as advertisements, and the result of the proposed strategy increases or decreases on the basis of the threshold. Finally, I discuss the analysis of the bidding coordination only in temporal variation. As mentioned in Section 7.3, ad spaces having the high average bidding prices improve the number of impressions by the coordination only in temporal variation although ad spaces having the low average bidding prices improve the number of impressions by the normal proposed strategy. The reason of the above reverse phenomenon is as below. As mentioned above, because of 45

no coordination among ad spaces in the coordination only in temporal variation, the bidding prices tend to be close to the average bidding price of each ad space in the existing strategy. Therefore, the bidding price in the originally low ad space is also low in the coordination only in temporal variation, and the bidding price in the originally high ad space is relatively high in the coordination only in temporal variation. Using these bidding prices, in originally low ad spaces, the DSP cannot really gain impressions because the bidding price is low. On the other hand, in originally high ad spaces, the DSP can gain more impressions because the bidding price is high. As a result, the number of impressions in originally low ad spaces does not increase by the coordination only in temporal variation, and the number of impressions in originally high ad spaces increases by the coordination only in temporal variation. If more impressions are individually gained in each ad space by the coordination only in temporal variation than the existing strategy, the total number of impressions should increase. However, the total impressions decrease because many ad spaces decrease their number of impressions by coordination only in temporal variation. However, given that the number of impressions in the proposed strategy is larger than the existing strategy, I conclude the bidding coordination among ad spaces is important to increase the total impressions. That is, the market model contributes to improve the number of impressions. 46

Chapter 9 Conclusion In this study, I aimed to propose the efficient bidding strategy of intermediary, the DSP in display advertisement auctions. The feature of display advertisement auctions is that advertisers, who are the end buyers, who do not directly bid, but intermediaries called DSP do. DSPs have a number of advertisers as their clients, and take bids instead of advertisers to satisfy advertiser s constraints such as a budget constraint. That is, although the DSP competes against other DSPs, it has the space to coordinate bidding among its clients. We consider that this auction allows a part of bidders colludes with each other in a conventional auction theory. Therefore, an efficient bidding strategy is not present. I first analyzed the current display advertisement auctions. As a result, I found the phenomenon that the winning rate increases and decreases with increasing the bidding price. Next, I analyzed the strategy used by the DSP, and pointed the problem. The problem is that the existing strategy treated display advertisement auctions as single unit auctions although display advertisement auctions are originally multi unit auctions. To solve the above problem, I proposed the bidding strategy that the DSP coordinated biddings among advertisers, and the DSP decided bidding prices to ad spaces on day to day basis. That is, the DSP continues to fixed bidding price to each ad space all day, and obtains ad spaces whose prices are lower than the price. However, the DSP needs to coordinate demands among advertisers to apply this strategy to the real auction. Because there are multiple demands of advertisers to a single ad space generally, these demands have an influence on each other. Moreover, Moreover, there are many ad spaces, and the DSP generally have many advertisers. Therefore, it is difficult to formulate a simple linear programming problem and solve this because of the enormous amount of calculation. To deal with this problem, I used the market model, which is the method to solve resource allocation problems using prices. The proposed strategy has possibility for getting impressions more effectively than the existing strategy, however we do not know the result in the real 47

auction. To evaluate the performance of the proposed strategy, I implemented a computer simulator, and compared the number of impressions of the proposed strategy to the existing strategy. As a result of a simulation, I found that the proposed strategy could make more successful bid than the existing strategy in the same budget. Moreover, the proposed strategy also improved the number of conversions although the proposed strategy aimed at maximizing the number of impressions. Finally, I analyzed the evaluation result, and discussed the influence of the proposed strategy to advertisers and ad spaces. I analyzed the data from viewpoints of advertisers and ad spaces. I also examined the number of impressions in the case of the simulation where the DSP conducts the coordination only in temporal variation without the coordination among ad spaces. In the analysis from a viewpoint of advertisers, I found that whether the advertisement improves its number of impressions by the proposed strategy depends on the targeting type of the advertisement from the above analysis. In the analysis from a viewpoint of ad spaces, I found that whether or not the number of impressions of each ad space was improved by the proposed strategy depended on the average bidding price. Finally, as a result of the bidding coordination only in temporal variation, the bidding coordination only in temporal variation was inferior to the existing strategy in terms of the number of impressions. From this result, I found that the bidding coordination among ad spaces using the market model contributed to increase the total impressions. Contributions of this study are as follows: Proposal on an efficient bidding strategy I pointed the problem that the existing strategy treated display advertisement auctions as single unit auctions despite the fact that display advertisement auctions are originally multi unit auctions. Moreover, I proposed the bidding strategy that a DSP coordinated bidding among its clients, and decided bidding prices to ad spaces on day to day basis. Advertisers demands have an affect on each other in the proposed strategy, and the problem that it is difficult to decide bidding prices occurs. To deal with this problem I used a market model. Clarification of bidding coordination using a market model I evaluated 48

the performance of the proposed strategy. As a result, I found that the proposed strategy could make more successful bid than the existing strategy in the same budget. Moreover, I analyzed the auction data, and found that the proposed strategy improved impressions of advertisements which had high values and impressions of affordable ad spaces. 49

Acknowledgments First, the author would like to express sincere gratitude to the supervisor, Associate Professor Shigeo Matsubara at Kyoto University, for his consecutive instruction, pertinent advice and earnest argument. The author would like to express sincere appreciations to the advisers, Professor Kazuyuki Moriya at Kyoto University, Associate Professor Michinori Hatayama at Kyoto University for his valuable advice. The author would like to express deep gratitude to MicroAd, Inc. for giving precious data. The author would like to tender his acknowledgements to Professor Toru Ishida, Assistant Professor Hiromitsu Hattori, the member of Ishida and Matsubara Laboratory at Kyoto University, for his technical and constructive advice. Finally, the author would like to thank all members of Ishida and Matsubara Laboratory for their various support and discussion. Especially, the author wish to thank Yuya Itoh to help to develop a simulator for this research. 50

References [1] Jordan, P., Mahdian, M., Vassilvitskii, S. and Vee, E.: The Multiple Attribution Problem in Pay-Per-Conversion Advertising, Algorithmic Game Theory (Persiano, G.(ed.)), Lecture Notes in Computer Science, Vol. 6982, Springer Berlin Heidelberg, pp. 31 43 (2011). [2] Muthukrishnan, S.: Ad Exchanges: Research Issues, Internet and Network Economics (Leonardi, S.(ed.)), Lecture Notes in Computer Science, Vol. 5929, Springer Berlin Heidelberg, pp. 1 12 (2009). [3] Ghosh, A., Rubinstein, B. I., Vassilvitskii, S. and Zinkevich, M.: Adaptive bidding for display advertising, Proceedings of the 18th international conference on World wide web, WWW 09, New York, NY, USA, ACM, pp. 251 260 (2009). [4] Cui, Y., Zhang, R., Li, W. and Mao, J.: Bid landscape forecasting in online ad exchange marketplace, Proceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining, KDD 11, New York, NY, USA, ACM, pp. 265 273 (2011). [5] Kaizu, K.: Proposing a bidding strategy by analyzing the change of other bidder s bids in advertisement auctions, A Graduation Thesis of Computer Science Course, Undergraduate School of Informatics and Mathematical Science, Faculty of Engineering, Kyoto University (2012). [6] Feldman, J., Mirrokni, V., Muthukrishnan, S. and Pai, M. M.: Auctions with intermediaries: extended abstract, Proceedings of the 11th ACM conference on Electronic commerce, EC 10, New York, NY, USA, ACM, pp. 23 32 (2010). [7] Vickrey, W.: COUNTERSPECULATION, AUCTIONS, AND COMPET- ITIVE SEALED TENDERS, The Journal of Finance, Vol. 16, No. 1, pp. 8 37 (1961). [8] Klemperer, P.: Auction Theory: A Guide to the Literature, Journal of Economic Surveys, Vol. 13, No. 3, pp. 227 286 (1999). [9] Milgrom, P. R. and Weber, R. J.: A Theory of Auctions and Competitive Bidding, Econometrica, Vol. 50, No. 5, pp. 1089 1122 (1982). 51

[10] McAfee, R. P. and McMillan, J.: Bidding Rings, The American Economic Review, Vol. 82, No. 3, pp. 579 599 (1992). [11] Wellman, M. P.: A Market-Oriented Programming Environment and its Application to Distributed Multicommodity Flow Problems, CoRR, Vol. cs.ai/9308102 (1993). [12] Itoh, Y.: Estimating the distribution of bidder s evaluation values in advertisement auctions, A Graduation Thesis of Computer Science Course, Undergraduate School of Informatics and Mathematical Science, Faculty of Engineering, Kyoto University (2014). 52