Biases in Decision Making Alexander Felfernig - PowerPoint PPT Presentation

Institute for Software Technology International Workshop on Decision Making and Recommender Systems, Bolzano, 2014 Biases in Decision Making Alexander Felfernig alexander.felfernig@ist.tugraz.at Decision Biases & Recommender Systems 1

Institute for Software Technology Agenda • Recommendation Approaches • Decision Biases • Conclusions & Research Issues Decision Biases & Recommender Systems 2

Institute for Software Technology Applied Software Engineering Research Human Decision Making & Knowledge-based Recommender Systems Recommender Systems • Group Decision Making • Constraint-based Recommenders • Group Recommender Systems • Speech Recognition Rec. • Cognitive (Decision) • Houska Award Nom. (FS) Biases • WeeVis Environment • Choicla Environment Knowledge Software Engineering (KE) Engineering (SE) • Recommenders for • Direct & Anytime Diag. Requirements Eng. (requirements and KBs) • Group Recommendation • Knowledge Understanding for RE (Eye-tracking, studies) • Dependency Detection • Game-based KE, AIGames • IntelliReq Environment • WeeVis Environment Decision Biases & Recommender Systems 3

Institute for Software Technology WeeVis Environment • Provides technologies for the inclusion of recommender applications into Wiki pages. • Suitable for complex item domains such as computers, financial services, and sports equipment. • Includes diagnosis and repair functionalities. • Currently applied by three Austrian universities. • Freely available: weevis.org Decision Biases & Recommender Systems 4

Institute for Software Technology Recommendation Task in WeeVis Customer Properties  V Product Properties  V • Represented as CSP (V, D, C). • Variables V describe customer properties ( ) and product properties ( ). • Compatibility constraints COMP ( ) define relationships between customer properties. • Filter constraints FILT ( ) describe relationships between customer properties and product properties. • Customer requirements R ( ) are unary • Product constraints PROD ( ) constraints on customer properties. describe the item assortment. • Item rankings are based on utility functions. Decision Biases & Recommender Systems 5

Institute for Software Technology Example: Direct Diagnosis of Inconsistent Requirements R: PROD: FILT: COMP: 1 1 1 1 should be consistent! 2 2 2 2 … … … … but: inconsistent! l m k n Diagnosis   R: -  consistent with COMP  FILT  PROD 1 .. k/2+1 .. k/2 k …  COMP  FILT  PROD consistent  1 k/2   … „direct diagnosis“ (increase of domain knowledge) k/2+1 k A. Felfernig and M. Schubert. FastDiag: A. Felfernig, M. Schubert, and C. A. Felfernig, M. Schubert, M. A Diagnosis Algorithm for Inconsistent Mandl, G. Friedrich, and E. Zehentner. An Efficient Diagnosis Constraint Sets, 21st International Algorithm for Inconsistent Con- Teppan. Efficient Explanations for Workshop on the Principles of Diagnosis, Inconsistent Constraint Sets, ECAI straint Sets, AIEDAM, 26(1):53-62, Portland, USA, pp. 31-38, 2010. 2010, pp. 1043-1044, 2010. 2012. Decision Biases & Recommender Systems 6

7 Decision Biases & Recommender Systems WeeVis MediaWiki Environment Institute for Software Technology

8 Decision Biases & Recommender Systems WeeVis MediaWiki Environment Institute for Software Technology

9 Decision Biases & Recommender Systems WeeVis MediaWiki Environment Institute for Software Technology

Institute for Software Technology WeeVis Recommender Applications • >50 Knowledge Engineers. • >70 developed Recommenders. • Interaction logs collected in an anonymous fashion. • Will be exploited for preference learning. Decision Biases & Recommender Systems 10

Institute for Software Technology Heatmap Visualization of Modeling Sessions • Overview of areas, knowledge engineers looked at. • Can be used, for example, for constraint ranking. Decision Biases & Recommender Systems 11

Institute for Software Technology Choicla Environment • Decision about new employees, investment decisions, new cars, choosing a restaurant, … • Modeling environment for decision apps Decision Biases & Recommender Systems 12

Institute for Software Technology This Talk …  Basic introduction to example cognitive biases in the recommender context (100’s exist …)  Cognitive (decision) biases: – “tendency to decide in certain (simplified) ways” – can lead to suboptimal decision outcomes  Bottum-up approach (testing individual biases) Decision Biases & Recommender Systems 13

Institute for Software Technology Why Cognitive Biases? risk [1..10]? fun[1..10]? food [1..10]? credit[1..10]? … Human brains were not primarily designed for the present time but rather for stone-age conditions Also: tradeoff between effort and accuracy, maximizers vs. satisficers Decision Biases & Recommender Systems 14

Institute for Software Technology Frequent Assumptions … • Preferences are known/defined beforehand full HD films • Preferences are stable, 5 pics per sec. users don’t change them WLAN data transfer • Users have an optimization maxprice 1.500€ function in mind waterproof  However, preference max resolution 20MPix stability does not exist! Decision Biases & Recommender Systems 15

Institute for Software Technology Preferences Are Constructed … • Not known beforehand • Often changed • No optimization function used • Decision heuristics applied (e.g., elimination by aspects)  “Door opener” for cognitive biases (tendency to decide in certain ways)! J. Payne, J. Bettman, and E. Johnson. The Adaptive Decision Maker, Cambridge University Press, 1993. Decision Biases & Recommender Systems 16

Institute for Software Technology Example Influence Factors for Decisions with Recommender Systems ordering ordering of of items attributes/ configuration of questions result sets social Decision context explanation of items presentation context Decision Biases & Recommender Systems 17

Institute for Software Technology Examples of Cognitive Biases Theory Description Context effects Additional irrelevant (inferior) items in an item set (decoy effects) significantly influence the selection behavior Primacy/recency Items at the beginning and the end of a list are analyzed effects significantly more often than items in the middle of a list Framing effects The way in which different decision alternatives are presented influences the final decision taken Priming If specific decision properties are made more available in memory, this influences a consumer's item evaluations Defaults Preset options bias the decision process Decision Biases & Recommender Systems 18

Institute for Software Technology Context Effects Decision Biases & Recommender Systems 19

Institute for Software Technology Context Effects • A decision is always made depending on the context in which item alternatives are presented • For example, completely inferior item alternatives can trigger significant changes in choice behaviors • Example context effects are discussed in the following Decision Biases & Recommender Systems 20

Institute for Software Technology Short Note: Ebbinghaus Effect • Illusion of relative size perception • Triggered by context in which objects are shown • Commonalities with context effects Decision Biases & Recommender Systems 21

Institute for Software Technology Context Effects: Overview • Compromise : Target (T) is a compromise to decoy item D (T is less expensive and has slightly lower quality ) • Asymmetric Dominance : T dominates D (T is cheaper and has a higher quality ) • Attraction : T is more attractive than D (T is slightly more expensive but has a higher quality ) Decision Biases & Recommender Systems 22

Institute for Software Technology Compromise Effect Product A (T) B D price per month 30 15 50 download limit 10GB 5GB 12GB The addition of alternative D (the decoy alternative) increases the attractiveness of alternative A because, compared with product D , A has only a slightly lower download limit but a significantly lower price D is a so-called decoy product, which represents a solution alternative with the lowest attractiveness Decision Biases & Recommender Systems 23

Institute for Software Technology Compromise Effect in Financial Services Domain Study performed with real-world products (konsument.at). A. Felfernig, E. Teppan, and K. Isak. Decoy Effects in Financial Service e-Sales Systems, ACM Recommender Systems Workshop on Human Decision Making and Recommender Systems (Decisions@RecSys), Chicago, IL, 2011. Decision Biases & Recommender Systems 24

Institute for Software Technology Asymmetric Dominance Effect Product A (T) B D price per month 30 15 50 download limit 10GB 5GB 9GB Product A dominates D in both dimensions (price and download limit) Product B dominates alternative D in only one dimension (price) The additional inclusion of D into the choice set could trigger an increase of the selection probability of A Decision Biases & Recommender Systems 25