Biases in Decision Making

Abstract

Decisions are typically taken on the basis of heuristics that are a door opener for di erent types of decision biases. Such biases can be interpreted as a tendency to decide in certain simpli ed ways which can often lead to suboptimal decision outcomes. Recommender systems support users in di erent types of decision making tasks and thus should be aware of such biases. In this paper we provide a short overview of di erent types of decision biases and their impacts on recommender systems. We also discuss some issues for future work.

Keywords

• Decision Making
• Biases
• Recommender Systems

Introduction

Recommender systems [ 5 ] support users in identifying relevant candidates from an item assortment. Collaborative ltering [ 14 ] is based on word-of-mouth promotion where ratings of users with similar preferences are exploited for recommending items. Content-based ltering [ 24 ] recommends items that are similar to those the user has experienced in the past. Knowledge-based approaches [ 4, 7 ] rely on semantic item knowledge that is exploited for determining recommendations. For example, constraint-based recommenders [ 6, 7 ] rely on an explicit set of constraints that support the determination of a recommendation. Finally, group recommenders determine recommendations for groups of users on the basis of group decision heuristics [ 20 ]. In this paper we focus on knowledge-based and also group and collaborative recommendation approaches. An example of a knowledge-based recommendation environment is WeeVis [ 12 ] which is a MediaWiki extension for the de nition and execution of recommender applications. Furthermore, Choicla [ 25, 26 ] is an environment that supports group decision tasks on the basis of group recommendation technologies. In the remainder of this paper we will discuss di erent types of decision biases and their role in recommendation scenarios. In most of the cases when users are interacting with recommender systems, they do not know their preferences beforehand but rather construct and frequently adapt them [ 17, 23 ]. In this context, users do not optimize their decisions but apply decision heuristics which can act as a door opener for di erent cognitive (decision) biases [ 23 ]. In the following we provide an overview of example decision biases (100's of these exist) and their role in recommender systems.

Decoy E ects. A decision is taken depending on the context in which alternatives are presented. Thus, completely inferior alternatives can trigger changes in choice behaviors. An overview of decoy e ects (context e ects) is provided in Figure 1.1 Item T is denoted as target item for which we want to increase the selection share. Item C is the competitor of T and D is assumed to be the decoy item which can be used to increase the selection share of item T. A target T is a compromise to decoy item D if it is less expensive and has a slightly lower quality. Furthermore, the attraction e ect denotes a situation where T is slightly more expensive but has a signi cantly higher quality. Finally, asymmetric dominance denotes a situation where T is cheaper than D and has a higher quality.

An example of an asymmetric dominance e ect in the evaluation of Internet connection alternatives is depicted in Figure 2. In this example, item A (the target item) dominates the decoy item in two dimensions whereas item B (the competitor) dominates the decoy item in only one dimension. The decision heuristic often applied in this context is a pairwise comparison of attributes [ 23 ]. In our example, T is the clear winner since it dominates D in two dimensions.

Impacts of decoy e ects on recommender applications can be summarized as follows. First, decoy items could be exploited for increasing the selection share of speci c target items [ 30 ] (see also the above example) { for sure, this application comes along with ethical issues. An empirical study related to decoy 1 Note that the dimensions cost and quality are examples { other dimensions could be used as well [ 28 ]. e ects in the nancial services domain is presented in [ 29 ]. Further decoy-related studies in the context of recommender systems are reported in [ 27 ] (hotel rooms) and [ 28 ] (game characters). Knowledge about decoy items can also be exploited for de-biasing purposes. Such a scenario is discussed in [ 10 ] where a dominance model is introduced to gure out dominance relationships between di erent items in a candidate set. On the basis of identi ed dominance relationships decoy items can be eliminated from the result set. Finally, decoys can also trigger the construction of explanations that are related to decision heuristics (e.g., attribute-wise comparison ): item A (see Figure 2) is the clear winner since it dominates D in both dimensions.

Primacy and Recency. Primacy/recency e ects describe situations in which items presented at the beginning and the end of a list are evaluated signi cantly more often than others. Since users are not interested in evaluating large lists to identify relevant items, they often focus their evaluations to the beginning and end of a list (interpretation of primacy/recency is a decision phenomenon). Murphy et al. [ 21 ] show this e ect in the context of an analysis of the clicking behavior of users. Primacy/recency has also a cognitive aspect: information units at the beginning (primacy) and at the end (recency) of a list are recalled more often than information units in the middle of a list. Felfernig et al. [ 8, 11 ] show the existence of primacy/recency e ects in the context of recommendation dialogs. The outcome of their analysis is that product properties at the beginning and the end of a recommendation dialog are recalled more often and are then also preferred as selection criteria when selecting items from a consideration set. This also holds for unfamiliar product properties (see Figure 3).

Impacts of primacy/recency e ects on recommender applications can be summarized as follows. Similar to decoy e ects, primacy/recency e ects can be exploited for controlling item selection behavior when interacting with a recommender system (on the basis of attribute orderings). Further related studies are a major issue for future work, for example, the impact of di erent attribute orders on product comparison pages, di erent orderings of argumentations in item reviews, and di erent orderings of repair proposals [ 9 ] in knowledge-based recommendation.

Framing. The way a decision alternative is presented in uences the decision behavior of the user. For example, users will prefer meat that is 80% lean compared to meat that is 20% fat. Another example is price framing [ 3 ]: if a user has to choose between two companies selling wood pellets (X,Y) where X sells pellets for e24.50 per 100kg and gives a discount of e2.50 if the customer pays with cash and Y sells pellets for e22.00 per 100kg and charges a e2.50 surcharge if the customer uses a credit card, users will prefer the rst alternative. This selection behavior can, for example, be explained by prospect theory [ 15 ] which suggests that alternatives are evaluated with regard to gains and losses where losses have a higher negative value compared to equal gains. In the price framing example, the loss would be the surcharge, in the rst example the loss is associated with the 20% fat meat.

Impacts of framing e ects on recommender applications can be summarized as follows. Positive framing can increase the selection probability of items. Price framing can trigger a potential focus shift from quality attributes of items to socalled secondary attributes (e.g., payment services) associated with items and { as a consequence { can change the item selection behavior of a user. In this context it must be pointed out that not every item property is equally salient at decision time and this can lead to signi cant shifts in selection behavior [ 3 ].

Further E ects. Priming [ 16, 22 ] represents the idea of making some properties of a decision alternative more accessible in memory such that this setting directly in uences user evaluations. An example is background priming that exploits the fact that di erent page backgrounds can directly in uence the decisionmaking process [ 16 ]. People often tend to favor the status quo compared to other decision alternatives. If defaults are used, users are reluctant to change predened settings due to the fact that they are loss-averse [ 18, 19 ]. Loss in this context can mean, for example, additional costs resulting from inconsistent item settings triggered by de-selecting a default [ 7 ]. Finally, anchoring denotes the e ect that users often heavily rely on the rst information (anchor) when evaluating decision alternatives. For example, item ratings (e.g., in collaborative ltering) of other users manipulated to be higher result in higher ratings of the current user [ 1 ]. A similar phenomenon has been observed in the context of release planning scenarios where initial evaluations manipulated the follow-up evaluations of requirements [ 13 ]. An approach to de-bias ratings in collaborative ltering recommendation scenarios is presented in [ 2 ]. 3

Conclusions and Future Work

Human decisions are often not based on optimization functions but on decision heuristics that are door-openers for di erent decision biases. We discussed a small set of example biases and their (potential) impact on recommender applications. There are a couple of issues for future research which include an in-depth investigation of possibilities for de-biasing recommendations, the development of consensus-fostering recommendations in group decision making, and the general investigation of the properties of decision biases in group decision making.

References

1. G. Adomavicius , J. Bockstedt , S. Curley , and J. Zhang. Recommender systems, consumer preferences, and anchoring e ects . In Decisions@RecSys11 , pages 35 { 42 , Chicago, IL, USA, 2011 .

2. G. Adomavicius , J. Bockstedt , S. Curley , and J. Zhang. De-biasing user preference ratings in recommender systems . In IntRS 2014 Workshop , pages 2 {9 , 2014 .

3. M. Bertini and L. Wathieu. The framing e ect of price format . In Working Paper, Harvard Business School , pages 1 { 24 , 2006 .

4. R. Burke . Knowledge-based recommender systems . Encyclopedia of Library and Information Systems , 69 ( 32 ): 180 { 200 , 2000 .

5. R. Burke , A. Felfernig , and M. Goeker . Recommender systems: An overview . AI Magazine , 32 ( 3 ): 13 { 18 , 2011 .

6. A. Felfernig . Koba4MS: Selling Complex Products and Services Using KnowledgeBased Recommender Technologies . In CEC 2005 , pages 92 { 100 , 2005 .

7. A. Felfernig and R. Burke . Constraint-based Recommender Systems: Technologies and Research Issues. In 10th ACM Intl. Conference on Electronic Commerce (ICEC'08) , pages 17 { 26 , Innsbruck , Austria, 2008 .

8. A. Felfernig , G. Friedrich , B. Gula , M. Hitz , T. Kruggel , G. Leitner, R. Melcher , D. Riepan , S. Strauss , E. Teppan, and O. Vitouch . Persuasive recommendation: Serial position e ects in knowledge-based recommender systems . In Y. Kort, W. IJsselsteijn, C. Midden , B. Eggen , and B. Fogg, editors, Persuasive Technology , volume 4744 of LNCS , pages 283 { 294 . Springer Berlin Heidelberg, 2007 .

9. A. Felfernig , G. Friedrich, M. Schubert , M. Mandl , M. Mairitsch , and E. Teppan . Plausible repairs for inconsistent requirements . In IJCAI'09 , pages 791 { 796 , Pasadena , California, USA, 2009 .

10. A. Felfernig , B. Gula , G. Leitner, M. Maier , R. Melcher , S. Schippel , and E. Teppan . A dominance model for the calculation of decoy products in recommendation environments . In AISB Symposium on Persuasive Technologies , pages 43 { 50 , 2008 .

11. A. Felfernig , B. Gula , G. Leitner, M. Maier , R. Melcher , and E. Teppan. Persuasion in Knowledge-Based Recommendation . In PERSUASIVE 2008 , pages 71 { 82 , 2008 .

12. A. Felfernig , S. Reiterer , M. Stettinger , and M. Jeran . An Overview of Direct Diagnosis and Repair Techniques in the WeeVis Recommendation Environment . In 25th Intl. Workshop on Principles of Diagnosis , pages 1 { 6, Graz , Austria, 2014 .

13. A. Felfernig , C. Zehentner , G. Ninaus, H. Grabner , W. Maaleij , D. Pagano , L. Weninger , and F. Reinfrank . Group decision support for requirements negotiation . In Decisions@RecSys11 , volume 7138 of LNCS , pages 105 { 116 , 2012 .

14. J. L. Herlocker , J. A. Konstan , L. G. Terveen , and J. T. Riedl . Evaluating collaborative ltering recommender systems . ACM Trans. Inf . Syst., 22 ( 1 ):5{ 53 , 2004 .

15. D. Kahneman and A. Tversky . Prospect theory: An analysis of decision under risk . Econometrica , 47 ( 2 ): 263 { 291 , 1979 .

16. N. Mandel and E. Johnson . Constructing preferences online: Can web pages change what you want ? In Association for Consumer Research Conference, pages 1 { 37 , Montreal, Canada, 1998 .

17. M. Mandl , A. Felfernig , E. Teppan, and M. Schubert . Consumer decision making in knowledge-based recommendation . Journal of Intelligent Information Systems (JIIS) , 37 ( 1 ):1{ 22 , 2010 .

18. M. Mandl , A. Felfernig , and J. Tiihonen . Evaluating Design Alternatives for Feature Recommendations in Con guration Systems . In CEC 2011 , pages 34 { 41 , 2011 .

19. M. Mandl , A. Felfernig , J. Tiihonen , and K. Isak . Status quo bias in con guration systems . In 24th International Conference on Industrial, Engineering and Other Applications of Applied Intelligent Systems , pages 105 { 114 , Syracuse , New York, 2011 .

20. J. Mastho . Group recommender systems: Combining individual models . Recommender Systems Handbook , pages 677 { 702 , 2011 .

21. J. Murphy , C. Hofacker , and R. Mizerski . Primacy and recency e ects on clicking behavior . Computer-Mediated Communication , 11 : 522 { 535 , 2012 .

22. A. North, D. Hargreaves , and J. McKendrick . In-store music a ects product choice . Nature , 390 : 132 , 1997 .

23. J. W. Payne , J. R. Bettman , and E. J. Johnson . The Adaptive Decision Maker . Cambridge University Press, Cambridge, UK, 1993 .

24. M. Pazzani and D. Billsus . Learning and revising user pro les: The identi cation of interesting web sites . Machine Learning , 27 : 313 { 331 , 1997 .

25. M. Stettinger and A. Felfernig . Con guring Decision Tasks . In 16th International Workshop on Con guration, pages 17 { 22 , Novi Sad , Serbia, 2014 .

26. M. Stettinger , G. Ninaus, M. Jeran , F. Reinfrank , and S. Reiterer. WE-DECIDE: A Decision Support Environment for Groups of Users . In IEA/AIE'13 , pages 382 { 391 , 2013 .

27. E. Teppan and A. Felfernig . The asymmetric dominance e ect and its role in e-tourism recommender applications . In Wirtschaftsinformatik (WI' 2009 ), pages 791 { 800 , Vienna, Austria, 2009 .

28. E. Teppan and A. Felfernig . Minimization of decoy e ects in recommender result sets . Web Intelligence and Agent Systems , 10 ( 4 ): 385 { 395 , 2012 .

29. E. Teppan , A. Felfernig , and K. Isak . Decoy e ects in nancial service e-sales systems . In RecSys'11 Workshop on Human Decision Making in Recommender Systems (Decisions@RecSys'11) , pages 1 { 8, Chicago, IL, 2011 .

30. A. Tversky and I. Simonson . Context-dependent preferences . Management Science , 39 ( 10 ): 1179 { 1189 , 1993 .