Enabling Direct Interest-Aware Audience Selection

Transcription

1 Enabing Direct Interest-Aware Audience Seection ABSTRACT Arie Fuxman Microsoft Research Mountain View, CA Zhenhui Li University of Iinois Urbana-Champaign, Iinois Advertisers typicay have a fairy accurate idea of the interests of their target audience. However, today s onine advertising systems are unabe to everage this information. The reasons are two-fod. First, there is no agreed upon vocabuary of interests for advertisers and advertising systems to communicate. More importanty, advertising systems ack a mechanism for mapping users to the interest vocabuary. In this paper, we tacke both probems. We present a system for direct interest-aware audience seection. This system takes the query histories of search engine users as input, extracts their interests, and describes them with interpretabe abes. The abes are not drawn from a predefined taxonomy, but rather dynamicay generated from the query histories, and are thus easy for the advertisers to interpret and use. In addition, the system enabes seamess addition of interest abes provided by the advertiser. The proposed system runs at scae on miions of users and hundreds of miions of queries. Our experimenta evauation shows that our approach eads to a significant increase of over 50% in the probabiity that a user wi cick on an ad reated to a given interest. Categories and Subject Descriptors H.3.3[Information Storage and Retrieva]: Information Search and Retrieva Custering; H.2.8 [Database Management]: Database Appications Data mining Genera Terms Agorithms, Experimentation Keywords Onine Targeted Advertising, Query-Log Custering Work done whie the author was an intern at Microsoft Research. Work done whie the author was at Microsoft Research. Permission to make digita or hard copies of a or part of this work for persona or cassroom use is granted without fee provided that copies are not made or distributed for profit or commercia advantage and that copies bear this notice and the fu citation on the first page. To copy otherwise, to repubish, to post on servers or to redistribute to ists, requires prior specific permission and/or a fee. CIKM 12, October 29 November 2, 2012, Maui, HI, USA. Copyright 2012 ACM /12/10...$ Anitha Kannan Microsoft Research Mountain View, CA ankannan@microsoft.com Panayiotis Tsaparas University of Ioannina Ioannina, Greece tsap@cs.uoi.gr 1. INTRODUCTION In onine advertising, advertisers want to target a specific audience that is more ikey to engage with their campaign. Typicay, advertisers are capabe of describing this audience fairy accuratey. However, in today s onine advertising systems, they do not have the option to expicity specify the characteristics of the users that they wish to target, except for broad demographic information. Instead, they bid on query terms, which act as a proxy for the user interests. But queries can be miseading when taken out of context. For exampe, if a user queries for hemets, it is not obvious if she is ooking for bike hemets or motorcyce hemets. Ads for both wi appear, since this is a term reated to both bike and motorcyce hemet companies. If we knew that the user who posed the query has a ong-term interest in biking, then it woud become cear that the query is more ikey to be about bike hemets. The techniques that we present in this paper aow the bike hemet advertiser to directy specify that she prefers users who are interested in biking and the advertising system to identify the users who are interested in biking. Thus, a match between users and advertisers with intersecting interests can be easiy made. We ca this capabiity direct interest-aware audience seection. Enabing advertisers to directy specify user interests is extremey powerfu. For instance, part of the appea of advertising on socia media sites such as Facebook is the abiity for advertisers to directy seect their audience based on their expressed interests, as we as their ikes and friends 1. An expensive restaurant can seect users who have specified interest in Food and Wine, whie a company that ses outdoors equipment can advertise to users who have decared interest in Camping. This option is not avaiabe when advertising on search engines. The aforementioned restaurant woud have to guess the terms with which a user wi express their interest, and bid on these terms. In the case of Food and Wine, this transates to a arge set of terms reated to restaurants, fine dining, wine seection, entertainment arrangements, etc. This paces a huge burden on the advertisers to come up with the right terms, and they sti run the risk of triggering incorrecty, or missing an important term. Unike socia media users, search engine users do not expicity state their interests and preferences. However, they give abundant impicit information about their interests 1 Seehttps://

2 8/1/2010 Espn fantasy footba 2010 camaro superchargers, fastane,fastane camaro, fast ane performance 8/23/2010 nf payer rankings, cbs sportsine fantasy footba 8/28/2010 whippe superchargers 9/6/2010 whippe superchargers Oakand raiders 9/12/2010 9/13/2010 Houston Texans 10/2/2010 Oakand raiders 10/12/ accord cutch change,87 honda accord parts,advance auto parts 10/29/2010 NFL hats 10/31/2010 Sema 2010 awards 11/4/2010 Redine motorsports 11/5/2010 Oakand raiders 11/15/2010 Ask automotive questions 12/14/2010 Nf sreams free raiders vs cots 12/26/2010 Auto trader Apex motorsports, autotrader (a) A portion of the user history corresponding to queries for ony two of the identified custers, footba and cars. The custers spread over a ong period of time. Queries of the custers are semanticay reated but do not necessariy share terms. For instance, the footba custer has queries oakand raiders and houston texans with no overapping terms. Simiary, the cars custer has queries incuding Sema 2010 awards and redine motorsports. For best visuaization, pease see this figure in coor. 1/31/2011 3/9/2011 Cars.com, autotrader. 3/20/2011 Interest Confidence Cars 0.99 Footba 0.88 Jobs 0.87 Music 0.66 (b)top 4inferred interests for the user. Figure 1: Exampe of inference of user interests furniture trave movies basketba games fishing coupons wedding airines dining diet food recipes hotes cars games cruises baby chiropractic music heath yrics baking commodities brewers divorce timeshare cycing arthritis roofing orthopedic fu Tabe 1: A subset of interest abes identified using our approach through their actions, and more specificay their queries. Users query about anything and everything that is on their mind. Compiing the ong-term (e.g., year-ong) query history of a user reveas a variety of interests: ephemera interests that correspond to short-term tasks such as buying a new washing machine; routine interests that correspond to queries that enabe everyday tasks such as reading a newspaper or checking emai; activities that correspond to ongterm interests of the user such as diet, sports, gaming, and heath care. In this paper, we consider the probem of extracting user interests from query histories for enabing direct interestaware audience seection. Given a coection of mutipe user histories, we wi produce a set of interest abes, and train a mode that assigns interest abes to users. The abes are not drawn from a predefined taxonomy, but rather dynamicay generated from the query histories, and thus easy for the advertisers to interpret and use for targeting specific users. A direct interest-aware audience seection capabiity is important for both sponsored search and dispay advertising. In the former, the targeted user interests woud be provided by the advertisers aongside the usua bid terms; in the atter the interests woud be specified together with demographic and other behaviora targeting information. We contrast our approach to other audience seection approaches, where users may be associated to interests, but these interests cannot be directy used by the advertisers. For exampe in the work of Ahmed et a. [1] interests are represented by topics produced by a topic mode, and used for ad cick prediction. Advertisers do not have the option to directy specify the interests they want to target. Furthermore, the produced interests are not easiy interpretabe, and thus they cannot be used for direct targeting. In a nutshe, our approach is as foows: First, we custer the query history of each individua user in order to identify groups of queries that are about the same interest. For the custering, we use a measure of semantic simiarity between terms that we obtain by expoiting tempora reationships between queries. Figure 1(a) shows an iustrative exampe of two custers obtained using our approach from a user history in our data set. Notice, for instance, the queries oakand raiders and houston texans being custered together but having no terms in common. Given the custering of the query history, we extract a short description for each custer consisting of the most popuar query terms present in the custer. Then, we generate the interest abes by finding terms that occur frequenty across mutipe user histories, and seecting a subset of these terms as our interest vocabuary. Tabe 1 ists some of the terms that our approach extracted as part of the interest vocabuary. We can see that the interests are represented using commony used vocabuary found in search queries. In order to map the custers into this vocabuary, we train a cassifier using massive amounts of automaticay created training data constructed from the queries in the abeed custers. The cassifier can then be appied to new users to map them to the set of interest abes. For the same user shown in Figure 1(a), Figure 1(b) shows the top four interests inferred by our approach. Our contributions incude the foowing: We address the probem of direct interest-aware audience seection. Our approach distinguishes itsef from previous work on earning interests [1] by the fact that users are assigned a concise set of interpretabe interest abes, empowering the advertisers to directy target users using these abes.

3 At the core of our approach is a component for custering queries within a user history that are thematicay reated. Our custering agorithm uses a nove simiarity measure, which makes use of the semantic reationships between terms defined by the tempora co-occurrence of queries across mutipe user histories. Thus, our custering approach expoits both the oca (within a singe user history) and goba (across user histories) reationships between queries for deriving query custers. We impement our approach on a distributed data storage and processing system. Our system runs at scae for miions of users and hundreds of miions of queries. We perform a thorough experimenta evauation that shows that our approach eads to a significant increase of over 50% in the probabiity that a user wi cick on an ad reated to a given interest. The evauation was performed at a arge scae, on 150,000 users using 2 months of ad data and user histories consisting of 16 months of query activity. We note that athough in this work we consider the probem of audience seection, our work can aso be appied to other tasks, such as personaization of search user experience. Inthis case, the user interests coud be used toprovide context for a query, and taior the search engine response to the needs of the specific user. The rest of the paper is structured as foows. In Section 2, we present reated work. In Section 3, we provide an overview of our approach. In Section 4, we present the detais of the modeing phase of our approach. In Section 5, we present experimenta resuts. Finay, in Section 6, we make concuding remarks and give directions for future work. 2. RELATED WORK Computationa advertising is an emerging research fied that considers the appication of computationa and agorithmic techniques to onine advertising. We refer the reader to the course notes of Introduction to Computationa Advertising 2 for a thorough review of the fied. Behaviora targeting, the use of prior user history for improving the effectiveness of an onine campaign, is a prominent research topic within this fied and has received considerabe attention [1, 5, 9, 14, 21, 22]. Pandey et at. [14], Chen et a. [5], and Yan et a. [22] mode the user as a bag of events, such as cicks to pages or queries. Jaworska et a. [9] represent users as a vector of categories, by mapping their web page visits to a predefined taxonomy. A machine-earning mode is then trained in order to predict whether a user wi cick on an ad. Tyer et a. [21] mode the probem of audience seection as an information retrieva probem, where there is a repository of users, and some users that are known to respond we to a campaign are used as queries over the repository. Users are modeed again as a bag of events, queries, and web page cicks. Recent work [12, 17, 2] has aso considered the use of socia network information (friendships, emai communication) for improving behaviora targeting. The scaabiity of the behaviora targeting probem has been addressed either with Map-Reduce impementations [5, 14] or samping [1]. The cosest work to our approach is the recent work by Ahmed et a. [1]. They consider the probem of behaviora 2 targeting in dispay advertising, and use a generative topic mode to define interests over histories of mutipe users. Then, they use the interests of users who have cicked on an ad as features in a cassifier that predicts whether a user wi cick on the ad. Their technique assumes the existence of previous ad cick activity for the given ad. Furthermore, their interest topics are not directy used by the advertisers. In contrast, we associate users to a concise set of interpretabe abes and empower the advertisers to directy specify such interest abes together with their ads. Query ogs are instrumenta in the improvement of search engines, and they have been under intense anaysis in the past few years. There is a vouminous iterature on different aspects of query-og anaysis. One important probem is that of breaking upaqueryhistory intosessions [7, 10, 11, 13, 16, 18], deaing with the fact that tempora coherence does not necessariy impy thematic coherence. This is a chaenging task, since different tasks tend to be intereaved or span ong periods of time. Tempora correation between queries over arge number of sessions has been expoited to define semantic correations between queries [3, 8, 20] for tasks ike reformuations or query suggestions. One key differentiation of our work is that we use tempora correations between queries to define simiarity between terms, and then we utiize this simiarity to custer queries. In contrast, previous works define reationships directy between queries. Reated to our approach is the work by Richardson[19] that discovers ong-term reationships between terms in query histories. 3. OVERVIEW OF OUR APPROACH In this paper, we address the probem of identifying user interests from search query histories, and describing them using a concise vocabuary. Given a user u and their query historyq u, wewanttoassignuseruasetofabesl u, drawn from a arger vocabuary L of possibe interests. The choice of the vocabuary L is of paramount importance in enabing the advertisers to seect the appropriate audience for their campaign. We propose a methodoogy for generating the interest vocabuary L, and an agorithm for mapping the user history to this interest vocabuary. Our approach has two phases: the modeing phase, and the inference phase. In the modeing phase we use query histories from mutipe users to generate the vocabuary of interests, and train a machine earning mode that maps coections of queries to abes in our vocabuary. In the inference phase, we appy our abeing agorithm to user query histories to obtain a abeing of the users in our interest vocabuary. We now discuss the detais of the two phases. 3.1 Modeing User Interests The modeing phase takes as input a coection Q = {Q 1,...,Q m}ofqueryhistoriesofmusers, andproduces a vocabuary of interest abes L = { 1,..., K}, and a mode M that assigns interest abes to coections of queries. This phase can be decomposed into three steps: First, we custer the individua query histories in order to extract themes; Then we use the produced custerings to generate the abe vocabuary. When avaiabe, we aso augment the interest vocabuary with advertiser provided interest abes; Finay, we train a machine earning mode that maps themes into the abe vocabuary. The mode itsef is trained using data obtained automaticay from the custers. The pipeine of these three steps is shown in Figure 2.

4 User query history coection Q User query history Q m Query history custering custers Labe generation custers Mode training Learned mode Advertiser provided abes Figure 2: Modeing phase pipeine. Query history custering custers Interest assignment Interest aggregation <custer, score> Figure 3: Inference phase pipeine. Top K interests of the user Query History Custering: Users express their interests in their search queries, but not necessariy in a temporay and syntacticay coherent way. Queries pertaining to widy different interests are intereaved over short intervas of time, whie queries that refer to the same interest recur over the span of weeks, months, or even years, each time sighty mutated, using different terms and touching different aspects. As a first step towards extracting interests from query histories, we organize queries of individua users into themes: semanticay coherent custers that are potentiay reated to the same interest. We extract these themes by custering the user history. For our custering, we use a simiarity measure that captures the semantic correation between queries, as this is observed over the query histories of miions of users. We discuss our simiarity measure, and custering agorithm in detai in Section 4.1. Formay, given the history Q u of an individua user u, the query history custering step produces a custering C u = {c 1,...,c Tu }, where each custer of queries c C u corresponds to a semanticay coherent theme that is candidate for capturing an interest. Given a coection of user query histories Q = {Q 1,...,Q m} the output of the custering step is a coection of custerings C = {C 1,...,C m}, one for each individua user. Labe Generation: The custering step does a good job in bringing together queries that are semanticay reated, and organizing the user query history into themes. Manua inspection reveas some ceary defined interests: an ongterm engagement in onine gaming, a proonged search for a new house, or a ong-standing quest for medica advice. These groups of queries make intuitive sense to a human observer, but they are not actionabe for advertisers who cannot afford to go through miions of query custers to find the ones that are of interest to their campaign. We thus need a concise way to describe the interests we observe. Using the themes we have identified, we wi extract a set of interest abes, which wi define the vocabuary with which advertisers can seect the users they are interested in. The abe extraction process identifies key terms that can be used for abeing query custers of individua users. It then aggregates these terms over the fu history coection to identify terms that pertain to a arge number of users. These terms wi define the interest abe vocabuary. We describe the detais of this process in Section 4.2. In summary, given the coection of custerings C output from the history custering step, the abe generation step wi produce an interest abe vocabuary L = { 1,..., K} that describes the space of possibe interests of a search user. Optionay, we can aso incorporate advertiser-provided interest abes to the abe set, thus aowing the advertisers to dynamicay modify the set of interest abes. In Section 5.3, we show the performance of our system when advertiser-provided interest abes such as webkinz and ego are incuded. Mode training: We observed that the themes extracted from the custering step aign we with intuitivey defined interests, and we used this fact to create our interest vocabuary L. Given a new user u, with query history Q u, which is custered into a set of themes C u, we want to be abe to map these themes into the space of interest abes L. In this step we train a discriminative mode M that performs this task: given a custer of queries c, it produces a probabiity distribution P( c) over the interest abes L. Inorder totrain the mode we need training data: custers that are abeed within our abe vocabuary L. We obtain this data automaticay from the custering coection C we produced in the custering step. In summary, in the mode training step, we take as input the interest vocabuary L produced in the abe generation step, and the custering coection C produced in the custering step, and we produce a machine earning mode M that maps a custer of queries c into the abe vocabuary L. We describe the detais of this step in Section Inferring User Interests At the inference phase, given a user u, with query history Q u, we wi assign asetofabes L u Lfromthevocabuary of abes L produced in the modeing phase. This phase can be decomposed into three steps: the history custering step, the interest assignment step, and the interest aggregation step. The pipeine for the inference procedure is shown in Figure 3. Query History Custering: In this step we extract the main themes in the query history of the user, using the same custering techniques that we described in the modeing phase, which we describe in detai in Section 4.1. Given theinputhistoryq u weproduceacusteringc u = {c 1,...,c Tu }, where each custer c C u corresponds toatheme in the user history. Interest Assignment: In this step we appy the machine earning mode M that we trained in the modeing phase to the custering C u. For every custer c C u, we obtain a probabiity distribution over the abe set L. That is, for each abe L, we obtain the probabiity P( c) that the custer c shoud be abeed with abe. Interest Aggregation: Given the custering C u and the probabiity distributions P( c) defined for each custer c, we can aggregate them in a number of ways to obtain the consoidated user interest profie. In our case, we associate the user u with the set of abes L u that have probabiity P( c) above a certain threshod θ p (set to 0.75 in our experiments), for some custer c C u. Whie being simpe, the experimenta resuts indicate that this aggregation scheme

5 is robust and works we in practice. It is possibe to expoit a variety of other signas for weighting the probabiity scores of the custers such as the custer size, the time interva over which the queries were asked, etc., but we eave this as potentia probem for future investigation. 4. MODELING USER INTERESTS We wi now describe in detais the three steps of the modeing phase: query history custering, abe generation, and mode training. 4.1 Query History Custering An interest manifests itsef in the query history over mutipe queries that cover different aspects of the interest. For exampe, a user who has an interest in footba wi pose mutipe queries about different footba teams, NFL, or game schedues. A these queries are thematicay reated around the interest footba. Identifying such groups of thematicay reated queries poses the chaenge of defining a suitabe simiarity measure between queries. Typica syntactic simiarity measures such as Jaccard coefficient or edit distance are not sufficient, since they do not capture the diversity in the way peope query about a topic. Queries ike wedding gown and fora arrangements are semanticay cose under a wedding interest, yet far apart under any measure of textua simiarity. Tempora affinity is a popuar method for capturing such semantic correations: reated queries are ikey to appear cose in the user history. There is considerabe amount of work in partitioning a user history into sessions, temporay coherent custers of queries [4, 7, 10]. However, tempora coherence does not aways guarantee thematic coherence: thematicay diverse interests may be intereaved over a short period of time. Conversey, a thematicay coherent interest may span severa days, weeks or months in the history of a user. Therefore, sessions do not necessariy capture interests fuy and accuratey. Athough tempora affinity is not sufficient to capture interests in a singe user history, when aggregating miions of user histories, it provides a strong signa for semantic simiarity. This idea has been previousy expored to extract correations between queries for tasks such as query suggestions and query reformuations [23, 3]. In our approach we wi use tempora co-occurrence of queries over mutipe user histories in order to define semantic simiarity between terms. We wi then use this measure of simiarity to group queries into custers, which capture themes in the user history, and are candidates to be mapped to user interests. Formay, et Q denote a coection of user histories. A user history Q u is a sequence Q u = (q 1,t 1),...,(q nu,t nu ) of query, time-stamp pairs (q i,t i), where query q i was posed at timet i. Wepartitionthequeryhistoryintosessions using the usua 30-minute timeout rue: a timeout of more than 30 minutes between two queries defines the beginning of a new session. The session contains the set of queries between two timeouts. Formay, a session is a maxima subset S Q u of the query history, such that for any two query-timestamp pairs (q i,t i),(q j,t j) S, t i t j is ess than 30 minutes. Let S denote the set of a sessions definedover the history coection Q. Each session S S can be thought of as a bag of words, S = {w 1,...,w k }, consisting of a the terms of the queries contained in S. Let P(w i,w j) be the number of sessions where words w i and w j occur together. Let N be the tota numberof co-occurrences, thatis, N = Σ iσ jp(w i,w j). For a pair of terms (w i,w j) we define the co-occurrence frequency f(w i,w j) as the fraction of co-occurrences that contain both terms w i and w j. That is, f(w i,w j) = P(w i,w j ). N Simiary, for a term w i, we define the frequency f(w i) of term w i to be the fraction of co-occurrences that contain term w i. That is, f(w i) = Σ jp(w i,w j ). N In order for two terms to be simiar we woud ike them to have high co-occurrence frequency. However, high cooccurrence frequency by itsef is not sufficient to determine simiarity. Terms that have high frequency on their own are ikey to participate in pairs with high co-occurrence frequency. For exampe, queries facebook and googe are prominent in the search ogs, and they exhibit high cooccurrence frequency with each other and with other terms, yet this does not impy semantic simiarity. To normaize for this effect, we divide the co-occurrence frequency with the probabiity that the two terms co-occur in the same session by chance. This ratio, or more precisey the og of this ratio, is the point-wise mutua information (PMI) between the two terms, a commony used simiarity measure in text mining and natura anguage processing [6]. Formay, it is defined as foows: PMI(w i,w j) = og f(wi,wj) f(w i)f(w j) A known drawback of PMI is that it favors rare co-occurrences. Two terms that appear ony once in the the query histories in the same session, have the highest possibe PMI. This is undesirabe, since we woud ike the pair of terms to have some supportin order tobe deemed simiar. Weaddress this issue by using the discounted PMI (dpmi) measure [15]: dpmi(w i,w j) = PMI(w i,w j) f(w i,w j ) min{f(w i ),f(w j )} f(w i,w j )+1/N min{f(w i ),f(w j )}+1/N Given the simiarity measure between terms, we can extend it to queries, or coection of queries. We represent those as bags of terms. Given two bags of terms X = {x 1,...,x kx } and Y = {y 1,...,y ky }, we define their simiarity as foows: sim(x,y) = 1 X Y (x,y) X Y dpmi(x, y) That is, the simiarity of the two bags of terms is the average dpmi simiarity of the pairs of terms in the cross-product between the two bags. Note that a coection of terms may contain the same term mutipe times. According to our simiarity definition this term wi appear mutipe times in the sum, and thus contribute more to the simiarity. This foows the intuition that terms that are frequent shoud have more impact on the simiarity of the coection of queries. It is aso possibe that X and Y share a term w. In this case we need to define a measure of simiarity of a term to itsef. We compute this usingthedefinitionofpmi, wherewedefinef(w,w) = f(w). Therefore, we have: PMI(w,w) = og 1 f(w) This definition captures nicey the intuition that two coections that share a rare term (e.g., aquarium ) are more

6 simiar than two coections that share a frequent term (e.g., facebook ). Note that our fina query simiarity measure makes use of both semantic and syntactic simiarity between queries. Semantic simiarity is expicity introduced by using the tempora co-occurrence of terms, whie syntactic simiarity is a side benefit of reducing the simiarity of queries to comparisons between terms. Equipped with a simiarity measure between queries and sets of queries, we can now appy any standard custering agorithm for grouping the queries into interests. We opt for hierarchica aggomerative custering. The agorithm proceeds iterativey, starting with a set of singeton custers each consisting of a singe query, and at each iteration it merges the custers with the highest simiarity. It continues unti the simiarity of the most simiar pair drops beow a predefined threshod. Therefore, givenacoection of muser queryhistories Q = {Q 1,...,Q m}, we have obtained a coection of m custerings C = {C 1,...,C m}. A custer c C u in the custering of user u is a set of queries that are thematicay reated, and are candidates for defining an interest of user u. In the foowing section, we wi show how we generate the interest abes from the custerings C, and then abe the custers in C u with the appropriate interest abe. 4.2 Labe Generation In the abe generation step, we expoit the coection of custerings C that we obtained in the custering step to automaticay generate a rich set of terms that can serve as the vocabuary L of interest abes. Given the coection C, we first perform a pruning step to remove custers with number of distinct queries beow a certain threshod (set to 30 in our experiments). Such custers are too sma to capture a prevaent interest of the user. Let C p denote the new custering coection. For each custer c C p we produce a set T c containing the top-q most frequent terms, where q = 5 in our experiments (we excude stop words, etc. ). These top frequenty-occurring terms represent a synopsis of the custer, capturing the underying theme of its queries. Let T = T c denote the unionofa terms thatare amongthemostfrequenttermst c of at east one custer, for at east one user. We keep as our abe set L the terms in T that appear in the query history of at east θ u users, where θ u = 100 in our experiments. That is, our abe set consists of terms that appear as a theme for at east 100 users. We aso do minima human inspection to remove certain abes that do not correspond to interests (e.g., map and store ), and canonicaize certain synonymous terms (e.g., recipes and recipe ). Tabe 1 shows a subset of the 300 abes that were generated using this approach. We can see that the abes cover a wide spectrum of interests, and at different eves of granuarity. For instance, whie a abe ike games tends to be more encompassing, a abe such as basketba is more specific. There are subte variants of simiar kind of interests, e.g., food, recipes, baking and dining, to name a few. There are aso time bounded interests such as wedding and roofing. Furthermore, these interest abes are of high vaue to the advertisers. To verify this, we performed the foowing check: we obtained the top 1,000 unigram advertising bid terms (in terms of the revenue that they generate in a major sponsored search engine), and we computed the overap with the ist of generated interest abes. It turns out that 12.5% of these 1,000 top advertising terms are actuay incuded in our ist. It is important to note that our approach is not restricted to using these automaticay-generated interest abes. Any ist of interest abes provided by the advertisers can be used, as ong as it contains terms that appear in the query custers. In Section 5.3, we experimentay show the effectiveness of our approach not ony in the scenario of automaticaygenerated abes but aso in the case of additiona abes provided by the advertisers. 4.3 Mode Training Given the abe set L = { 1,..., K}, in this step we train a machine earning mode M which given a custer c, produces the probabiity P( c) that the custer c beongs to the interest described by abe, for every abe L. We use a muticass ogistic regression mode parameterized by W to compute P( c). The parameter matrix W is a coection of weight vectors {w k }, one for each abe k L such that each component w jk measures the reative importance of the j th feature for predicting k th abe. The muticass ogistic regression earns a mapping from the feature vector of c, denoted by z(c) to the abe y, using the foowing softmax ogistic function: P( k c) = P(y = k z(c),w) = exp(b k +z(c) w k ) 1+ K j=1 exp(b j +z(c) w j ) where b j (1 j K) are bias terms. In our setting, the feature vector z(c) corresponds to a summary constructed from the custer, c. In particuar, we used exica features consisting of a unigrams (terms) appearing in the queries in the custer. We then converted them to booean features representing the presence or absence of these unigrams. We did not consider the frequency of occurrence of these terms as that woud require that the custers are normaized for many factors such as number of queries in the custer, number of unigrams in the custer, etc. In fact, our experimenta evauation shows that these simpe binary features perform effectivey. The weight vector in Equation 1 is earned from a abeed data set D = {(x 1,y 1 ),...,(x n,y n )}, where each pair (x j,y j ) corresponds to the feature extracted from a custer and its corresponding interest abe. In particuar, W is earned so as to maximize the conditiona og-ikeihood of the abeed data: W = argmax W (1) n ogp(y j = k x j,w) (2) j=1 Labeed data for training: Manua construction of a abeed training set can be too expensive and time-consuming. Our approach enabes an effective way to obtain arge amounts of automaticay abeed training data. For some abe L, et C denote the set of custers, from any custering C u C, such that for a c C, T c, that is, the abe is one of the top terms in the custer c. We treat the custers in C as positive exampes for the abe, and we use this data to train our mode. In order for our approach to be successfu, we need custers in C to be homogenous, with highy frequent terms being semanticay reated. We manuay evauated the custerings for their homogeneity and we confirmed

7 that this is indeed the case. The homogeneity of the custers foows from the way we have constructed our simiarity function to capture the semantic simiarity of terms. Note that our approach generaizes naturay to the case that the set of abes we want to train against is provided from some externa source (e.g., provided by the advertisers). Let L denote this providedset of abes. For each abe L we can use the process described above to obtain the set C of custers that have in their top terms. Then we can train our mode against these externay provided abes. We experiment with this case in Section EXPERIMENTAL EVALUATION We now report the resuts of a arge-scae, end-to-end evauation that we performed on our system. In Section 5.1, we present the experimenta setup. In Section 5.2, we describe our methodoogy and metrics for evauating on advertising data. Finay, in Section 5.3, we present our key findings. 5.1 Experimenta Setup We now provide the detais of the data and parameters used for the different components of our system. The custering agorithm uses the simiarity measure defined in Section 4.1, which reies on a discounted PMI computation for unigram pairs over user sessions. We computed discounted PMI over the sessions of 2.2 miion users over 16 months of queries from the query og of the Bing search engine. We used the standard definition of session, where a session consists of a consecutive queries unti there is a period of 30 minutes inactivity [4]. To create the set of interest abes, we ran the custering agorithm on 580,000 users. To ensure good quaity keywords, we constrained ourseves to custers with at east 30 queries. As a resut, we obtained 1,042,729 custers mapped to their top-5 keywords. We aggregated this mapping by counting for each keyword the number of users who have at east one custer that is mapped to the keyword. We then produced a ist of a keywords that are associated to at east 100 users; after removing stop-words, puras etc, this resuted in 5,500 keywords. The size of this ist was manageabe enough to be processed editoriay in order to detect highy frequent terms reated to interests. After editoria processing, we obtained a ist of 332 interest abes. To verify that these interest abes are of high vaue to the advertisers, we performed the foowing check: we obtained the top 1,000 unigram advertising bid terms in terms of the revenue that they generate in a major sponsored search engine, and we computed the overap with our ist of interest abes. It turns out that 12.5% of these 1,000 top advertising terms are actuay captured by our ist. This indicates that the interest abes are, indeed, monetizabe. We note that the creation of a meaningfu taxonomy, or cass system, is an extremey hard undertaking that merits its own scientific fied. Compete automation is neary impossibe, and probaby not desirabe, since introducing some human intuition can improve the quaity significanty. Our approach simpifies the vocabuary creation process significanty, by offering a manageabe set of abes for the data anayst to process. Processing is aso simpified, consisting mosty of fitering out uninteresting or very specific terms. This is a consideraby easier task compared to deriving such cass abes from scratch. More importanty, the produced abes capture the underying trends in the data, they can be updated dynamicay as query histories get updated, and they come together with training data. The Mode Training component used the custering of 120,000 users to create 116,839 (user,custer,abe) training exampes. The cassifier uses exica features: we used a feature vector consisting of 56,983 binary features (these features correspond to unigrams that appeared in at east 20 queries among the 120,000 users). A ogistic regression mode was earned using these training exampes. At inference time, the Interest Aggregation component mapped users to the abes for which they had at east one custer with cassification score above a threshod. Uness otherwisestated, weusedthreshodθ p = Asweexpain in the next section, the system was tested on 150,000 users using 16 month of query activity. Running the system at this scae was enabed by our impementation on arge-scae Map-Reduce distributed data processing system. 5.2 Evauation on Advertising data The main goa of this evauation is to study the effectiveness of our interest-aware audience seection system. Our hypothesis is that users who match advertiser-specified interests are (on average) more ikey to cick on ads reated to the interest than the users currenty seected via keyword match. User cick probabiity. Let A denote a set of ads, e.g., the set of ads in a specific advertising campaign, or a the ads reated to a specific interest. Let U denote the set of users that are candidates to be shown the ads in A, and et U A U denote the set of users that are actuay impressed with at east one ad from the set A. Aso et C A U A denote the subset of these users that cicked on at east one of the ads that they were impressed. We define the user cick probabiity of the set A with respect to the user set U as foows: P U(A) = CA U A That is, P U(A) is the fraction of users being impressed with ads from A, that actuay cicked on at east one ad from the set A. This metric is reminiscent of the standard notion of cickthrough rate (CTR) 3 which, ike user cick probabiity, is aso a ratio between cicks and impressions. However, CTR is the probabiity that, given an impression of an ad, the ad wi be cicked, whie user cick probabiity is the probabiity that given a user who is impressed an ad, the user wi at some point cick on the ad. Athough reated, the two metrics capture different information. We beieve that user cick probabiity metric fits nicey with the goa of audience seection, which is to seect users to whom to impress advertisements. Now, et U denote a set of users tagged with an interest abe. Aso, eta denoteasetofadsthatarereatedtothe interest (we discuss ater how we obtain this set). The user cick probabiity P U (A ) is the probabiity that a user who is assigned the interest abe wi cick on an ad reated to the interest. Therefore, mathematicay, our hypothesis is that on average P U (A ) > P U(A ), that is, users associated with interest are more ikey to cick on an ad reated to, than users drawn from the genera popuation of a users U who are impressed with the ad. 3

8 To test our hypothesis, we used the sponsored search ogs of the Bing search engine for a 2-month period, which does not overap with the time period used to compute user interests 4. For each abe in the vocabuary L, we appied our agorithm to the set of users U in the 2-month query og, and we generated a subset of users U U that were assigned this abe. Next, for each abe we need to obtain a set of ads A that are reated to interest. It is not immediate how to obtain such a set, since currenty, advertisers do not provide interest abes, and thus there exists no test set of ads abeed with interest abes that we coud use for the evauation. We tacke this probem by making the foowing approximation: we use the readiy avaiabe, existing bid terms from advertisers as a proxy for interest abes. More specificay, et a be an ad, and et B a denote the set of a bid keywords associated with this ad. We say that ad a is abeed with interest abe if B a. We define the set A as the set of ads that are abeed with the abe. Given the set of ads A, we use U A to denote the subset of users from U that are impressed an ad in A. Ideay, we woud ike to have contro over the set U A in our experiments. However, since we do not have such contro, we define U A = U U A, i.e., the subset of users abeed with, that are impressed an ad from A in the existing ogs. The set C A of users tagged with the interest abe that cicked on the ads in A is defined simiary. The user cick probabiity of A with respect to set U is P U(A ) = C A / U A whie the user cick probabiity with respect to U is P U (A ) = C A / U A. To make it more concrete, consider the foowing exampe. We have a set of four users U = {Aice,Bob,Cathy,David}, and a set of three ads A = {a 1,a 2,a 3}. Advertisement a 1 is tagged with bid keywords B 1 ={ casino, hote }, a 2 with keywords B 2 = { vegas, casino }, and a 3 with keywords B 3 = { vegas, hote }. In our ogs, Aice is shown ads {a 1,a 2,a 3}, Bob is shown ads {a 2,a 3}, Cathy is shown ads {a 1,a 3} and David is shown ad {a 3}. Aice cicked on ads a 1 and a 2, Bob cicked on a 3, and Cathy and David did not cick on any of the ads. Suppose now that our interest abe is casino, and that our agorithm tagged users U = {Aice,Bob,David} with the interest abe. We have that A = {a 1,a 2}, and U A = {Aice,Bob,Cathy} is the set of users that were impressed with an ad in A. The set of users tagged with the abe that are aso impressed with an ad in A is = {Aice,Bob}. Ony Aice cicked on an ad reated to the interest, therefore, C A = C A = {Aice}. The U A user cick probabiity with respect to the set of a users U is P U(A ) = C A / U A = 1/3 whie the user cick probabiity with respect to the set of users U that we tagged with the interest is P U(A ) = C A / U A = 1/2. Therefore, in our exampe, among the users impressed with an ad in A there is a 33% probabiity for one of them to cick on an ad, whie this probabiity increases to 50% when a user is drawn from the set of users tagged with the interest abe. User coverage. We aso consider a measure of coverage which is defined as the fraction of users who are assigned an interest profie (set of abes) of a minimum given size. The goa is to show that a arge fraction of users get assigned 4 We further restricted the users to high engagement users, as determined by the rues of the search engine company. User Cick Probabiity - our system eectronics baseba awscience jewery diet shoes User Cick Probabiity - baseine Figure 4: Scatter pot of User Cick Probabiity of baseine vs. our system. interest abes. Let U be the universe of users considered in an experiment (in our case, the users in the 2-month snapshot of the sponsored search ogs). Let V k be the users in U that are assigned a profie consisting of at east k abes by the interest-aware system. Then, the measure user coverage for profie of size k is defined as foows: UserCoverage k = V k U Reca that in our system, we ony keep the abes whose cassification score for some custer is above a cassification threshod. Thus, there is a cear tradeoff between user coverage and user cick probabiity. A high user coverage means that more users wi be assigned interest abes, but this may come at the expense of retrieving poor-quaity users who might not cick on ads. 5.3 Key Findings We now discuss the main resuts from our experiments. Effect on user cick probabiity. We first compare the user cick probabiity of our system against that of the baseine, for every abe for which there were at east 30 impressed ads in the2monthtime period thatwe considered for the advertising data (126 abes in tota). In Figure 4, we present a scatter pot of the user cick probabiity of the two aternatives we consider for every abe. The x-axis corresponds to the user cick probabiity P U(A ) of the baseine system, and the y-axis corresponds to the user cick probabiity P U (A ) of oursystem. We can observe thatthe points for most abes ie above the diagona (more precisey, for 97 out of 120 abes). This means that our system outperforms the baseine for 81% of the abes. To get an understanding of the performance for individua interest abes, we indicate on Figure 4 some abes for which we do particuary we, and some abes for which the baseine outperforms our approach. For exampe, we have arge gains for abes such as baseba and jewery which represent permanent (or ong-term) interests of users, and for ong term tasks such as wedding. Arguaby, considering entire histories to make a determination of the user interests heps for these ong-term interests. Some of the worst performing abes for our system are broad, high-eve interests which do not necessariy ead to higher user cick probabiities. Exampes incude terms such as science and aw.

9 UserCoverage k Number of abes in profie (k) Figure 6: Average User Cick Probabiity for different cassification score threshods. Figure 5: User Coverage for our interest-aware system. To get an aggregated view of the resuts, we computed the average user cick probabiity over a abes. For our system, the average user cick probabiity is 0.131; for the baseine, it is This represents a 50.5% increase over the baseine. We used a one-taied t-test, and verified that this difference is statisticay significant with 95% confidence, thus estabishing our hypothesis that on average it is more ikey for a user who is tagged with an interest abe to cick on an ad reated to that interest than a user drawn from the genera popuation. Our resuts demonstrate that given an interest of the advertiser, our technique produces an audience that has increased probabiity of cicking to an ad reated to that interest. Of course, this audience must be of significant size. That is, U A, the abeed users to whom the ads reated to are impressed, shoud be a sizeabe fraction of U A a the users to whom an ad reated to were impressed. In our experiments U A is on average 10% of U A, indicating that we capture a sizeabe fraction of impressed users. Effect on user coverage. Since, in our system, a profie consists of a abes above a cassification threshod, users may get profies of different sizes (in terms of the number of abes), or even no profie at a. We measured user coverage for different minimum profie sizes for our interest-aware system. In Figure 5, we give size k of the profie in terms of number of abes on the x-axis; and the user coverage for profies that have at east k abes, UserCoverage k, on the y-axis. We can observe that the coverage is as high as 0.93 and 0.78 for histories with at east 1 and 2 abes respectivey. It remains at reasonabe eves even for histories of size at east five (0.27). This impies that the number of users retrieved by our system is significant for the different abes. Sensitivity to cassification threshod. We aso performed a sensitivity anaysis of the cassification score threshod θ p. In particuar, we considered different instantiations of our system, where we varied the cassification threshod θ p and measured the corresponding average user cick probabiity. The resuts are shown in Figure 6. We can observe that the average user cick probabiity is not overy sensitive to the cassification threshod: for cassification threshod Interest Baseine Interest-aware Vegas Disney Hawaii Lego Webkinz Tabe 2: Performance on advertiser-provided interests. The first coumn shows the User Cick Probabiity of the Baseine, whie the second coumn the User Cick Probabiity of our method. θ p = 0.95 it is 0.158, decreasing genty to for threshod θ p = 0.6. Enabing advertisers to extend the vocabuary of interests. Sofar, we havepresentedresutsonasetofinterest abes derived from the custers themseves (Labe Generation component). However, our system is by no means restricted to that ist: it enabes advertisers to fexiby provide new interest abes and add them to the interest abe set. To show that this is possibe, we performed an experiment where we added some highy monetizabe interests to the ist of abes, and then performed the same evauation as before but with the extended ist. In particuar, we added three popuar tourist destinations (Vegas, Disney, and Hawaii), and two popuar toys (Lego and Webkinz). We show the resuts on Tabe 2. We can observe that our system outperforms the baseine on the five interests. This means that cick-through rate increases for interests that are fexiby added by the advertisers to the interest vocabuary. 6. CONCLUSIONS In this paper, we presented a system for direct interestaware audience seection. Our system takes the query histories of search engine users as input, extracts their interests, and describes them with interpretabe abes that enabe advertisers to easiy target users. We showed the effectiveness of our approach through a arge-scae evauation using advertising data. The resuts indicate that our system associates users to interest abes that are highy usefu for advertisers to better target reevant users. Our system can ead to an increase in user cick probabiity of over 50% compared to the baseine system. We are panning to extend our work in mutipe directions. Extensions incude empoying additiona signas such as cicked URLs and timestamps, and studying the effect

10 of interests on conversion rates, in addition to cicks. One particuary interesting direction invoves buiding upon the output of the custering agorithm to infer a time signature for different types of interests, based on the distribution over time of the queries in an interest custer. This can enabe us to better understand the nature of users interests. For instance, one may expect that the time signature of a timebounded task such as panning a wedding woud be different from that of a more permanent interest such as gardening. Acknowedgements. We woud ike to thank Patrick Pante, Michae Gamon, and Maria Christoforaki for their insightfu feedback. We woudasoiketothankjoshbonisandgilapidshafririfor providing support and access to mutipe resources needed for this project. 7. REFERENCES [1] A. Ahmed, Y. Low, M. Ay, V. Josifovski, and A. J. Smoa. Scaabe distributed inference of dynamic user interests for behaviora targeting. In KDD, pages , [2] E. Bakshy, D. Eckes, R. Yan, and I. Rosenn. Socia infuence in socia advertising: evidence from fied experiments. In Proceedings of the 13th ACM Conference on Eectronic Commerce, EC 12, pages , New York, NY, USA, ACM. [3] P. Bodi, F. Bonchi, C. Castio, D. Donato, A. Gionis, and S. Vigna. The query-fow graph: mode and appications. In CIKM, pages , [4] L. Catedge and J. Pitkow. Characterizing browsing strategies in the word-wide web. In WWW, pages , [5] Y. Chen, D. Pavov, and J. F. Canny. Large-scae behaviora targeting. In KDD, pages , [6] K. Church and P. Hanks. Word association norms, mutua information, and exicography. In ACL, pages 22 29, [7] D. He, A. Göker, and D. J. Harper. Combining evidence for automatic web session identification. Information Processing and Management, 38: , [8] H. Hwang, H. W. Lauw, L. Getoor, and A. Ntouas. Organizing user search histories. IEEE Transactions on Knowedge and Data Engineering, 24(5): , [9] J. Jaworska and M. Sydow. Behavioura targeting in on-ine advertising: An empirica study. In WISE, pages 62 76, [10] R. Jones and K. L. Kinkner. Beyond the session timeout: automatic hierarchica segmentation of search topics in query ogs. In CIKM, pages , [11] A. Kotov, P. N. Bennett, R. W. White, S. T. Dumais, and J. Teevan. Modeing and anaysis of cross-session search tasks. In SIGIR, pages 5 14, [12] K. Liu and L. Tang. Large-scae behaviora targeting with a socia twist. In CIKM, pages , [13] Q. Mei, K. Kinkner, R. Kumar, and A. Tomkins. An anaysis framework for search sequences. In CIKM, pages , [14] S. Pandey, M. Ay, A. Bagherjeiran, A. Hatch, P. Ciccoo, A. Ratnaparkhi, and M. Zinkevich. Learning to target: what works for behaviora targeting. In CIKM, pages , [15] P. Pante and D. Lin. Discovering word senses from text. In KDD, pages , [16] B. Piwowarski, G. Dupret, and R. Jones. Mining user web search activity with ayered bayesian networks or how to capture a cick in its context. In WSDM, pages , [17] F. J. Provost, B. Daessandro, R. Hook, X. Zhang, and A. Murray. Audience seection for on-ine brand advertising: privacy-friendy socia network targeting. In KDD, pages , [18] F. Radinski and T. Joachims. Query chains: earning to rank from impicit feedback. In KDD, pages , [19] M. Richardson. Learning about the word through ong-term query ogs. ACM Trans. Web, 2:21:1 21:27, [20] E. Sadikov, J. Madhavan, L. Wang, and A. Y. Haevy. Custering query refinements by user intent. In WWW, pages , [21] S. K. Tyer, S. Pandey, E. Gabriovich, and V. Josifovski. Retrieva modes for audience seection in dispay advertising. In CIKM, pages , [22] J. Yan, N. Liu, G. Wang, W. Zhang, Y. Jiang, and Z. Chen. How much can behaviora targeting hep onine advertising? In WWW, pages , [23] Z. Zhang and O. Nasraoui. Mining search engine query ogs for query recommendation. In WWW,