טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
Ph.D Thesis
Ph.D StudentBen-Porat Omer
SubjectStrategic Behavior in Prediction and Recommendation
DepartmentDepartment of Industrial Engineering and Management
Supervisor Professor Moshe Tennenholtz
Full Thesis textFull thesis text - English Version


Abstract

The 21 century, and especially its second decade, has witnessed the rise of machine learning (ML) algorithms in commercial applications. Search engines, recommender systems, and advertising, among others, have transformed dramatically by embracing tremendous amounts of data to enhance content quality, personalization and offer more accurate predictions and recommendations. Companies use ML algorithms to increase their market value (e.g., profits or customer engagement) by providing better products and services. In sheer contrast, ML algorithms are almost always developed in isolation, optimizing internal criteria (for instance, estimating the relevance of a document to a query, or predicting the selling value of apartments). When utilized in markets, ML algorithms are susceptible to strategic behavior?agents misuse algorithms for their advantage, and competing companies exploit their structure to outperform them.


This thesis bridges the gap between decades of economic studies on market competition and the use of ML algorithms in markets. It enriches the literature by studying challenges that are unique to data-oriented contexts. Its first part investigates competition between learning agents. In such scenarios, agents (i.e., companies) employ algorithms to attract users by providing prediction and segmentation. We develop theory and practical tools for the induced competitions and demonstrate experimentally that current ML algorithms miserably fail if not adjusted to market competition. The second part investigates the inefficiency of recommendation systems that interact with strategic content providers, e.g., search engines. Such systems retrieve web-pages owned by selfish agents who wish to attract as many readers as possible. It provides novel game-theoretic modeling, analyzes the dynamics of the system and its stability points, and offers mechanisms that mitigate strategic behavior. Overall, this thesis contributes to the Game Theory and ML literature, and especially to the emerging area of learning in the presence of strategic behavior.