The active, participatory, and creative Web audience is prevailing today

S o c i a l M e d i a A n a l y t i c s a n d I n t e l l i g e n c e Using Social Media to Predict Future Events with Agent- Based Markets Efthimios Bothos, National Technical University of Athens Dimitris Apostolou, University of Piraeus Gregoris Mentzas, National Technical University of Athens Agents can exploit user-generated content available The active, participatory, and creative Web audience is prevailing today with relatively accessible social media. The explosion of user- generated content on social media provides a wealth of opportunities for tapping into user preferences, assessments, and opinions about content, products, in social media by extracting user sentiments and assessments and using them to make informed transactions on the market. services, or even people. 1,2 Such information is captured typically in ratings denoted with Likert scales (such as 1 5 or 1 7) or other equivalent schemes such as stars or linguistic labels. User preferences also appear in textual content provided in forums, blogs, wikis, and the like and may represent detailed reviews or short statements expressing sentiments. In fact, sentiment analysis has shown that textual messages in microblogs correlate highly to public opinion polls. 3 Social media content can refer to possible future events, either explicitly and implicitly, but it remains unexploited in making predictions and aiding decisions about the future. The ability to predict the outcome of future events quickly and accurately is critical in today s business environment. For example, sales predictions and sensing of future consumer behavior can help businesses alter research and production planning. One emerging method for facilitating predictions about future events is prediction markets (PMs), which are speculative markets that serve to aggregate the beliefs of multiple traders in the price of contracts representing different outcomes of a future event. Contract prices provide a reasonable estimate of what the traders in aggregate believe to be the probability of an event, and as such, markets are able to generate forecasts. Individuals influence these prices by buying and selling contract shares based on their belief about the outcome. At the end of the dealing period, individuals are paid off based on the accuracy of their bids. PMs are characterized by their accuracy, easy deployment, and ability to dynamically incorporate new information available to traders by continuously adjusting an event s price and hence its probability conditioned to the new market information. 4 They are 50 1541-1672/10/$26.00 2010 IEEE IEEE INTELLIGENT SYSTEMS Published by the IEEE Computer Society

Related Work in Prediction Markets Social media content contains an untapped collective wisdom that can be exploited, among other things, for predicting future events. For example, Sitaram Asur and Bernardo Huberman leveraged the chatter of Twitter users to forecast box-office revenues for movies with stunning accuracy. 1 Jonas Krauss and his colleagues interpreted opinionated discussions and the level of buzz about the movie business on the Web to identify Academy Awards winners and box-office revenues. 2 Joshua Ritterman, Miles Osborne, and Ewan Klein proposed a regression framework that when applied to simple features extracted from Twitter posts can reduce the error associated with people s beliefs as modeled with PMs. 3 Existing methods apply regression techniques that model social media content and use these models to predict future events. In contrast, our work does not fit any model to social media content. Instead, we use social media content to feed an agent-based PM to predict future events and test it with predicting the Oscar for Best Motion Picture award. PMs can effectively forecast future events provided certain preconditions hold. James Surowiecki provided a qualitative analysis of participant characteristics necessary for a trustworthy market: diversity of opinion, independence of thought, and decentralization of knowledge. 4 Justin Wolfers and Eric Zitchevitz established a theoretical model and provided an account of sufficient conditions under which market prices aggregate private information held amongst participants. 5 They concluded that when participants are typically well-informed have a better than even probability of making the correct judgment market prices will aggregate participants information into useful information that can be used for decision making. Agent-based systems have been widely used to model emergent market dynamics or long-run convergence determined by the aggregation of individual choices. For example, researchers have used populations of computational agents to test analytic equilibrium models and explore the resulting dynamics. 6 Agent-based systems let researchers run market simulations with theoretical models and reach predictions about future market states. Recently, Johan Perols, Kaushal Chari, and Manish Agrawal proposed integrating agents and PMs as a new method for aggregating heterogeneous sources of information and applied it to combine the decisions of multiple individual classifiers. 7 References 1. S. Asur and B.A. Huberman, Predicting the Future with Social Media, HP Labs report, 2010; www.hpl.hp.com/research/scl/ papers/socialmedia/socialmedia.pdf. 2. J. Krauss et al., Predicting Movie Success and Academy Awards through Sentiment and Social Network Analysis, Proc. European Conf. Information Systems, 2008, pp. 2026 2037. 3. J. Ritterman, M. Osborne, and E. Klein, Using Prediction Markets and Twitter to Predict a Swine Flu Pandemic, Proc. 1st Int l Workshop on Mining Social Media, 2009, pp. 9 17. 4. J. Surowiecki, The Wisdom of Crowds, Doubleday, 2004. 5. J. Wolfers and E. Zitzewitz, Interpreting Prediction Market Prices as Probabilities, Graduate School of Business, Stanford Univ., 2005. 6. J. Seiffertt and D. Wunsch, Intelligence in Markets: Asset Pricing, Mechanism Design, and Natural Computation, IEEE Computational Intelligence, Nov. 2008, pp. 27 30. 7. J. Perols, K. Chari, and M. Agrawal, Information Market- Based Decision Fusion, Management Science, vol. 55, no. 5, 2009, pp. 827 842. increasingly popular among corporations and have been deployed at companies like HP, GE, Motorola, and Eli Lilly in applications such as forecasting sales and project-completion times. PMs are also used on a Web scale to predict the outcome of political, sports, and entertainment events. Nevertheless, they have several limitations. First, they require multiple participants, who need to be recruited and incentivized to participate. Second, substantial evidence from psychology and economics suggests that people tend to overvalue small probabilities and undervalue near certainties; hence, markets are likely to perform poorly when predicting small and high probability events. 5 Third, market participants tend to trade according to their desires and interests rather than objective probability assessments. 5 Fourth, like all markets, PMs are prone to manipulation attempts in which participants try to influence the prices of specific securities for own benefit. 6 In this article, we propose a new approach for making predictions using markets that uses use agents instead of human participants. Our agents exploit user-generated content available in social media by extracting user sentiments and assessments and using them to derive actionable information that is, make informed transactions on the market. We developed an agent-based framework for applications that can benefit from the wisdom of crowds, which is created and disseminated in different types of social media (such as blogs, microblogs, online forums, and Web communities). We demonstrate that our approach matches the prediction accuracy of human-based PMs in forecasting the Academy Award for Best Motion Picture for the last seven years. Our Approach We claim that user-generated content available in social media can provide a solid basis for making predictions about future events. Our focus is on event outcomes that are mutually exclusive and discrete that is, not drawn from a continuous variable. (See the Related Work in Prediction Markets for other approaches and previous work.) Social media content aggregates the collective opinion of users through sentiment and preference NovEMber/DecEMber 2010 www.computer.org/intelligent 51

S o c i a l M e d i a A n a l y t i c s a n d I n t e l l i g e n c e Social media content extraction Social information processing Market-based collective intelligence with artificial agents Rate Comment Discuss IMDb Flixster Yahoo Movies Rotten Tomatoes Google Average Infer polarity Trend detection Processing of rating Sentiment analysis Query analysis Update beliefs Agents Trade Prediction market Twitter Figure 1. Users participate in social media and express their opinions about movies. We parse this information and provide it as input to computational agents that trade in a prediction market in which contracts represent possible Best Motion Picture Oscar award winners. expression in blogs, microblogs, rankings, and so on. To further aggregate opinions and sentiments from varying media and predict future event outcomes we develop computational agents and let them participate in agent-based PMs. Our agents extract user sentiments and assessments available on social media and use them to make informed transactions in the market. Each agent represents a social medium (such as Twitter), interprets the user-generated content, and reflects its beliefs by trading in the market. The functions of our agents were inspired by the Belief-Desire-Intentions (BDI) paradigm, where an agent is characterized by its beliefs, goals (desires), and intentions. An agent will intend to do what it believes and will achieve its goals given its beliefs about the world. 7 Our agents build a world view (belief) by extracting and interpreting social media content and have the goal (desire) of maximizing their financial resources in the market through rational trading and by engaging (intention) in transactions based on the acquired information. This follows the processes of human-based PMs, where human traders receive information signals and, based on their personal assessment of available information, buy or sell contracts to maximize their portfolios. To set up an agent-based PM, we follow a stepwise approach (see Figure 1). The first step refers to identifying social media information sources, which include opinions relevant to the event set, and extracting from them the appropriate information. For a PM to provide trustworthy results, plentiful and diverse information should be available; otherwise the market can lean toward the wrong outcome. Examples might include political events (Who will be the next US president?), sports events (Who will be the winner of the European Champions League?), and media events (Which movie will win the best picture award?). Possible social media sources include online discussion forums, blogs, or microblogs such as Twitter. In a second step, we apply a method for processing content (such as sentiment analysis for textual input). In the third and final step, we create computational agents, the market is set-up, and trading commences. Market prices are shaped through trading. Once the market ends, the predicted event outcome is the one with the highest-priced contract. Predicting Oscar Award Winners We used our approach to predict the Best Motion Picture Oscar award winners. We hypothesized that the chances of a movie winning the Oscar can be determined by the social media information of online communities discussing movies. (Evidence supporting this assumption can be found in an earlier work. 7 ) We focused on Oscar predictions for two reasons. First, movies are of considerable interest among social media users, and many users discuss and review them. Second, we can compare our approach both with the actual past Oscar winners and with humanbased PMs, such as the Hollywood Stock Exchange (www.hsx.com). Social Media Content Extraction Prominent social media information sources with content-rich discussions about movies include Yahoomovies.com, 52 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS

Flixster.com, IMDb.com, and Rottentomatoes.com. All four sources host specific forums per movie in which users can express their opinions with reviews. On Rotten Tomatoes, users can also tag movies, and on IMDb and Flixster, users can also provide ratings. To extract user reviews, tags, and ratings from these four sources, we created customized parsers for each source. Moreover, we executed Web queries for occurrences of phrases of the following form: movie name + best picture + Oscar award + to win. Research has shown that these keywords convey a positive assessment of the movie nomination. 8 Furthermore, we executed targeted queries at the micro-blogging application Twitter. To overcome the restrictions that Twitter enforces on its historical data (such data are only available for seven days), we used the search within a specific website option Google offers by including the keywords site:twitter.com in the queries. All queries were constrained to 15 days prior to the Oscar s award ceremony and were executed on a daily basis with the specific date range. Once we retrieved the aforementioned information, we identified the Oscar s awards ceremony dates for all the years under consideration and excluded user-generated content submitted after the specific dates. In total, we collected 37,125 ratings and 67,584 user reviews from IMDb.com, Flixster.com, and YahooMovies.com; 3,020 tags from RottenTomatoes. com; 28,515 results from Web queries; and 148 tweets. Social Information Processing We create a positivity index per movie and per type of user-generated information (reviews, ratings, and tags). For IMDb and Flixster, we processed the user ratings if they were available; otherwise we inferred the sentiment of the user reviews. For ratings, IMDb uses a 1 to 5 scale, and Flixster a scale of 10 to 50. To avoid bubble effects due to an excessive number of ratings, we averaged the daily ratings per movie and created a daily positivity index per movie. If the daily average rating was higher than a cutoff value, we considered it a positive indication of the success of the specific movie; otherwise, we considered it a negative. The cutoff values we used were 2.5 for IMDb ratings and 25 for Flixster ratings. To classify the polarity of user reviews, we performed sentiment analysis using the Jane16 open source text analyzer implementing various sentiment analysis algorithms. 9 The tool processes the set of parsed files, with each file containing one review, and outputs the review polarity (positive or negative) and certainty (on a scale of 0 to 100) about the result. On Rotten Tomatoes, each user review is tagged as a tomato (positive) or splash (negative). In this case, we considered the available tags as votes in a type of approval voting and counted positive and negative votes to create the positivity index. Yahoo Movies classified a movie in one of 13 classes (from A+ to F). To process these classification, we mapped classes to a 1 to 13 value scale, with A+ denoting 13 and F denoting 1. The average daily rating was used in this case too. Movies do not have a common release date. Some movies can appear in the theaters up to 12 or more months before the Oscar s awards ceremony, while others appear shortly before the ceremony. Social media content appears soon after a movie is released, and user contributions increase during the first few days. To treat all the nominees equally, we introduced the concept of market time. Because the Oscar s are held in US, we considered a movie s US release date as time 0 in market time. We processed only those user contributions that were available in a certain time window and used that same window for all the movies. Market-Based Collective Intelligence with Artificial Agents Most research on PMs treat market prices as probabilities and hence model user beliefs after probabilistic distributions. In our case, the distribution of agent beliefs is modeled after a Dirichlet distribution, the multivariate generalization of the beta distribution. The selection of a multivariate distribution is justified as agents valuate more than two events. (For every year agents had to choose between five Oscar nominees, with the exception of 2010 with 10 nominees.) The Dirichlet distribution probability density function returns the belief that the probabilities of K rival events are p j (1 j K) given that each event has been observed o j 1 times. In our case, we considered as rival events each one of the Oscar nominees to win the best picture award. These events depend on the positive social media information (positive signals) generated by the users. If a positive signal is received for movie j on day d (denoted by the movie daily positivity index), we considered that there was an observation of the event that movie j wins the Oscars and we increment o j, the relevant parameter of the Dirichlet distribution, by one. Agents beliefs b j are given by the expected value of movie j winning the Oscar as provided by the Dirichlet distribution. Agent Trading. We assigned one agent to each social media information source, and it processed the content generated only by the information NovEMber/DecEMber 2010 www.computer.org/intelligent 53

S o c i a l M e d i a A n a l y t i c s a n d I n t e l l i g e n c e source assigned to it. Trading is performed as follows. Agents evaluate the content generated per day and update their belief function. Then they submit buy or sell offers for the movies they consider most successful. We use three methods to model agents trading behavior. In the first method, which we denote B (for buy), agents do not have constraints on their wealth or utility function. They buy one contract per day of each movie j they consider a favorite according to their beliefs that is, b j > p j, where b j is the agent belief for movie j and p j the current market price for the same movie. In the second method, which we denote BS (for buy-sell), agents compare the each movie s current market price against their internal belief and buy one share if b j > p j or sell if b j < p j. If b j = p j, no transaction occurs. Agents do not have constraints on their wealth in BS. In the third method, which we denote L (for log utility), we enforce a constraint on agents wealth and employ a log utility function. 10 Agents are price takers and trade so as to maximize their subjectively expected utility. That is, agent i with wealth w i decides how many movie contracts q ij of movie j to buy at price p j by solving the following problem: Max Ui = bij log wi qikpk p j j ( 1 ) k j with the following restrictions: wi wi qij pj = if( maxbet) Bet if( w j max i > maxbet) q ij 0 (1) (2) where max Bet is the wealth constraint. Market Mechanism. In PMs a market mechanism is employed to allow trading on the virtual contracts. The main purpose of such a mechanism is to serve traders buy or sell orders. Although several mechanisms can be used such as call auctions or parimutuel mechanisms, the majority of PMs use either continuous double auction (CDA) or a market-maker mechanism to ensure liquidity. In our work, we use a market-maker mechanism and, more specifically, the logarithmic market scoring rule (LMSR) 11 market maker due to its wide acceptance by prominent PMs. Market makers ensure the market is liquid by always having a published price at which they are willing to buy or sell. They make a profit by maintaining a spread between the prices at which they buy and sell; the quoted price is the price at which traders can buy or sell. Automated market makers set the price according to a rule that tells them how much to raise or lower the price when a buy or sell order is processed, respectively. For LMSR, we compute a cost C(q) and a pricing function p(q) as follows: 12 q a C( ) a Ln j e i / (3) q = ( ) q e i / a pi ( q ) = q a j e i/ (4) where q is a vector that represents the number of orders (future contracts) on each state that have already been accepted by the market maker and a (referred as market liquidity factor) is a parameter that must be chosen by the market maker. a represents the risk that the organizer is willing to accept. The greater the value of a, the more orders the organizer is likely to accept and more liquidity or depth is added to the market, meaning that traders can buy more shares at or near the current price without causing massive price swings. We use the cost function to calculate how much money an agent should pay (receive) when buying (selling) an amount of future contracts. The price function provides the current market prices considering that q future contracts have been accepted. In human-based PMs that measure event probabilities, the sum of the future contracts prices is typically equal to 1. These prices correspond to actual probabilities of the event occurring. Although in our case we cannot allege that prices reflect probabilities, they provide a means to select the winning movie (the highest priced movie) and reveal a relative confidence, estimated by the price differences the higher the differences, the more confident the result of the market. Experiments and Results The primary purpose of our experiments was to compare our approach s predictive accuracy against human-based PMs and polls that attempt to predict the best movie Academy Award winners for the years 2004 to 2010. We could not provide trustworthy analysis prior to 2004 because of the scarcity of social media content. We compared our results with the Hollywood Stock Exchange, the IMDb Oscar polls, and the predictions obtained by ranking movies based on the average of user ratings. Table 1 presents the results of our experiments. For reasons of brevity, this table presents only the case in which market liquidity is set to 100, the time window is set to 60 days, and the BS trading method is used. As Table 1 shows, our approach 54 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS

Method Table 1. Comparison of the predictive accuracy of each social medium, the IMDb polls, HSX, and our approach.* Year 2004 2005 2006 2007 2008 2009 2010 IMDb poll prediction X X X X HSX prediction X Average of user ratings from social media X X X X X X Single social media agent-based PM predictions IMDb.com X X X X Flixster.com X X X RottenTomatoes.com X X X YahooMovies.com (user ratings) X X X X X YahooMovies.com (expert ratings) X X X Twitter Query analysis X X X Our multiple social media agent-based PM predictions * indicates a successful prediction, and X indicates a failed prediction. X outperforms the IMDb polls and the average of ratings, and it can predict the Oscar winner as accurately as HSX. Neither our approach nor HSX or IMDb polls predicted the 2006 award; that year the outsider movie Crash pulled off one of the biggest upsets in Academy Awards history, winning best picture over the front-runner Brokeback Mountain. Table 1 also depicts the predictions achieved when using a single social medium as a source for the agent beliefs in the PM. When single social media are used, predictive accuracy decreases. By incorporating multiple social media, the possibility that the PM satisfies Surowiecki s conditions 13 (opinion diversity, independence, and decentralization) increases. We studied the sensitivity of our results by evaluating the effects of changes in the parameters of our approach. In this respect, we ran experiments for three levels of market liquidity by modifying the market maker s market liquidity factor at the 10, 100 and 1,000 levels for the B and BS trading methods, and 1,000, 1,500 and 2,000 for the L trading method. (We had to increase the market liquidity factor for the latter case as our agents invested more money, which caused a high increase of the market prices.) For each liquidity level, we set the market window at 30 to 90 days with an interval of 10 days because we observed that more than 90 percent of user content is submitted within this timeframe. We examined the effects of market liquidity and market window for each of the three trading methods that is, the use of each method results in up to 147 predictions (seven years, three liquidity factors, and seven time windows). Our results exhibit rather limited variations with respect to the parameters examined. Specifically, compared to HSX, we observed that our results are always successful for market windows greater than 60 days, with the exception of the results for 2010 when the BS trading model is used and market liquidity is set to 10. For lower market windows, we observed 18 failures. Hence, once the time window is big enough to allow adequate user-generated content to be processed, the PM can properly predict the event outcome. When employing limited liquidity in our market (values of the market liquidity factor 1,000 for the B and BS trading models and 2,000 for the L trading model), we observed low variations in market prices. The mean differences between the first-ranked movie contract (highest priced) and the second ranked (second highest priced) were low (0.01 monetary units) for all trading strategies. For moderate and high market liquidity (values of the market liquidity factor 10 and 100 for B and BS or 1,000, and 1,500 for L), the average difference between the winning movie and the second-ranked movie was more than 0.2 monetary units, providing clear evidence of the winner. Concerning the L trading method, we investigated the impact of a continuous PM (a market spanning multiple years) on the wealth of our agents. In this case, we achieved good predictors every year based on their performance and consequently their wealth increases. In this way, good predictors potentially have more money to invest and might greatly affect the outcome of the market. NovEMber/DecEMber 2010 www.computer.org/intelligent 55

S o c i a l M e d i a A n a l y t i c s a n d I n t e l l i g e n c e 4,500 Agent value 4,000 3,500 3,000 2,500 2,000 1,500 1,000 500 Figure 2. Agents wealth per year. Good predictors (such as the query analysis) gain money and thus influence the market more. Bad predictors (such as simple user ratings) lose money and cannot influence the market. The annual total gain of all agents is positive. Figure 2 presents a graph of the agents wealth fluctuation when we initially allocated 1,000 monetary units. Good predictors, such as the Google queries, gain money and thus influence the market. Bad predictors, such as Yahoo Movies (user ratings), lose money and cannot influence the market. The annual total gain of all the agents is positive. Discussion Our results suggest that feeding social media information to an agentbased PM can provide insight on the realization of future events. Table 2 summarizes our choices, provides guidelines for applying our approach in other domains, and lists areas for future research. Our experiments show that we can obtain better results by adding more information sources. The exploration began with using ratings from IMDb. com only; we increasingly added 2004 2005 2006 2007 2008 2009 2010 Year Per year sum of agents gain YahooMovies.com (user ratings) YahooMovies.com (expert ratings) Imdb.com Flixster.com Rotten Tomatoes.com Query analysis more information sources until we obtained predictions as good as those of the human-based prediction market HSX. This demonstrates the need for a dry run to identify the adequate number of information sources for the domain of interest. Regarding our system s self-adjustment by allocating more resources to good predictors, the wealth of agents who invest on the correct outcome increases and subsequently their influence on the market is amplified in a setting with repeated events. The effect on the correct outcome s price was noticeable. Considering the limitations of our work, it can be generalized only for discrete event outcomes and not for event outcomes drawn from a continuous variable. Furthermore, due to the limited access to social media information before 2004, we could not evaluate our approach s predictive accuracy over a longer period. Concerning the processing of social media content, we introduced an arbitrary cut-off between positive and negative ratings, set to the middle of the relevant scales. Although this is a logical assumption, such a selection does not account for ratings by indifferent or indecisive users. In our approach, event prediction is based on PMs. An alternative approach would be to fit a regression model to user-generated content. We investigated various correlations between user ratings and the Academy Award winner but could not identify statistically significant predictors. To test the applicability of our approach in other domains, we set up an experiment to predict the 2010 soccer World Cup winner (see http://imu. iccs.gr/software/socialmediapredict). Our agents traded using information from sports discussion forums, and Spain the actual world cup winner was the highest-priced contract at the market s end. The advantages of using computational agents to abstract and model human behavior in PMs are many-fold. Agents do not require incentives and recruitment, so the markets scale up fairly easy. The agents valuation of probabilities of events depends solely on their belief function, which does not exhibit abnormal behavior at the two ends of the probability spectrum. Lastly, agents are not programmed to manipulate the market. With the growing amount of usergenerated content and the emergence of domain-specific social websites, we expect that approaches such as agent-based PMs will be able to provide useful and reliable predictions about a plethora of future events. 56 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS

Table 2. The steps we followed, the choices we made in our application, a set of guidelines for applications in other domains, and suggestions for further research. Other application Design issue Application choice Justifications Guidelines Further research Social media content extraction Selecting information sources Prominent movie discussion forums Search engine results Twitter posts The selected information sources contain a wealth of diverse, independent, and decentralized social content about the target event Dry runs are necessary to assess whether enough information sources are utilized; check the system capability to predict similar past events with the selected sources and add more sources if needed Investigate types of information sources, quantity of content required for other application domains Social information processing Processing user ratings Processing user comments Query analysis Daily averaging of ratings and separating positive and negative comments at the middle of the scale Jane16 open source tool Phrases expressing positivity 6 Averaging ratings and detecting sentiment polarity on a daily basis provide simple, but sound aggregate measures of users opinions and their evolution Tool is optimized for movie reviews Previously studied, effective approach 6 Market-based collective intelligence with artificial agents Agents utility Log utility 7 Already applied in theoretical modeling of prediction markets (PMs), which proves the reflection of agents beliefs on market prices 7 Selecting trading mechanism Logarithmic market scoring rules 9 Logarithmic market scoring rules assure liquidity because the market maker always announces prices offering to buy and sell future contracts For each information source, identify the appropriate cutoff between positive and negative ratings Sentiment analysis tool should apply to the domain Use expressions relevant to the domain Useful in other domains because it provides a straightforward, proven utility function Logarithmic market scoring rules have been proven in various PM settings and can be used readily in other domains Investigate advanced rating aggregation approaches that account for indifferent or indecisive entries Use other sentiment analysis methods and tools that incorporate lexical analysis 12 Extend the set of expressions and apply natural language processing techniques to automatically extend the vocabulary with synonyms Investigate using other utility functions such as constant absolute and constant relative risk aversion Research using other market mechanisms such as continuous double auction or pari-mutuel betting References 1. F.Y. Wang et al., Social Computing: From Social Informatics to Social Intelligence, IEEE Intelligent Systems, vol. 22, no. 2, 2007, pp. 79 83. 2. L. Tang and H. Liu, Toward Predicting Collective Behavior via Social Dimension Extraction, IEEE Intelligent Systems, vol. 25, no. 4, 2010, pp. 19 25. 3. S. Asur and B.A. Huberman, Predicting the Future with Social Media, HP Labs report, 2010; www.hpl.hp.com/research/ scl/papers/socialmedia/socialmedia.pdf. 4. R. Hanson, Decision Markets, IEEE Intelligent Systems, vol. 14, no. 3, 2000, pp. 16 19. 5. J. Wolfers and E. Zitzewitz, Prediction Markets, J. Economic Perspectives, vol. 18, 2004, pp. 107 126. 6. R. Hanson, Foul Play in Information Markets, Information Markets: A New Way of Making Decisions, R.W. Hahn and P.C. Tetlock, eds., AEI Press, 2006, pp. 126 141. 7. S.-H. Chen, Software-Agent Designs in Economics: An Interdisciplinary Framework, IEEE Computational Intelligence, vol. 3, no. 4, 2008, pp. 18 22. NovEMber/DecEMber 2010 www.computer.org/intelligent 57

S o c i a l M e d i a A n a l y t i c s a n d I n t e l l i g e n c e T h e A u t h o R S Efthimios Bothos is a PhD candidate and researcher in the Information Management Unit at the National Technical University of Athens (NTUA). His research focuses on prediction markets and methods for information aggregation. Bothos has a diploma degree in electrical and computer engineering from NTUA. He is a member of the IEEE Computer Society. Contact him at mpthim@mail.ntua.gr. Dimitris Apostolou is an assistant professor in the Department of Informatics at the University of Piraeus, Greece. His research focuses on group support systems, knowledgebased decision support systems, and knowledge management systems. Apostolou has a PhD in electrical and computer engineering from NTUA. He is a member of the IEEE Computer Society. Contact him at dapost@unipi.gr. Gregoris Mentzas is a full professor in the School of Electrical and Computer Engineering at the National Technical University of Athens and the director of the Information Management Unit at the Institute of Communication and Computer Systems, Athens. His area of expertise is information technology management, and his research concerns e-government, knowledge management, and e-service technologies. Mentzas has his PhD in operations research and information systems from NTUA. Contact him at gmentzas@ mail.ntua.gr. Probabilities, Graduate School of Business, Stanford Univ., 2005. 11. R.D. Hanson, Logarithmic Market Scoring Rules for Modular Combinatorial Information Aggregation, J. Prediction Markets, vol. 1, no. 1, 2007, pp. 1 15. 12. Y. Chen and D. Pennock, A Utility Framework for Bounded-Loss Market Makers, Proc. 23rd Conf. Uncertainty in Artificial Intelligence, 2007, pp. 49 56. 13. J. Surowiecki, The Wisdom of Crowds, Doubleday, 2004. 8. J. Krauss et al., Predicting Movie Success and Academy Awards through Sentiment and Social Network Analysis, Proc. European Conf. Information Systems, 2008, pp. 2026 2037. 9. H. Chen and D. Zimbra, AI and Opinion Mining, IEEE Intelligent Systems, vol. 25, no. 3, 2010, pp. 74 80. 10. J. Wolfers and E. Zitzewitz, Interpreting Prediction Market Prices as Selected CS articles and columns are also available for free at http://computingnow.computer.org. Running in Circles Looking for a Great Computer Job or Hire? Make the Connection - IEEE Computer Society Jobs is the best niche employment source for computer science and engineering jobs, with hundreds of jobs viewed by thousands of the finest scientists each month - in Computer magazine and/or online! > Software Engineer > Member of Technical Staff > Computer Scientist > Dean/Professor/Instructor > Postdoctoral Researcher > Design Engineer > Consultant http://www.computer.org/jobs IEEE Computer Society Jobs is part of the Physics Today Career Network, a niche job board network for the physical sciences and engineering disciplines. Jobs and resumes are shared with four partner job boards - Physics Today Jobs and the American Association of Physics Teachers (AAPT), American Physical Society (APS), and AVS: Science and Technology of Materials, Interfaces and Processing Career Centers. 58 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS