Game-Theoretic Perspective in Recommender System
Adversarial Machine Learning in Recommender Systems: State of the art and Challenges
I did some review in this field and there is a nice review (Rethinking search engines and recommendation systems: a game theoretic perspective) This line of research emerges in recent years and most seminal works come from Moshe Tennenholtz’s group. They aim to establish a new paradim for search and recommendation systems by modeling the system as a mediator among selfish participants, i.e., the users and the content providers. They employ a game-theoretic perspective and devise new algorithms via mechanism design.
Since there are three parties in the game, recent works cover three settings in terms of who is the player.
Strategic publishers. The content publishers tend to manipulate the feature of the content to gain more exposure. In this setting, new metrics need to be deployed, like social welfare/fairness that account for the utility of publishers. Recent works studied the corpus dynamics that results from the strategic behavior of publishers in response to rankings, and showed that introducing randomness to ranking functions can increase social welfare by promoting the content diversity in the corpus.
Competing mechanisms. The systems themselves are viewed as players, and their goal is to compete against other systems. This setting is interesting because the goal of an algorithm is no longer the highest accuracy, but the proportion of the data points she predicts more accurately than other players. There is a work considering two competing regression task.
Strategic users. An adversary can create a number of fake users to bias the recommendation system. This profile injection attack, a.k.a. Shilling attack has been extensively studied. Another setting is assuming that users are able to differentiate between exploitation/exploration content, and how to devise a system that will behave close to optimal but will be also incentive compatible, that is, rational users who care only for their expected utility will accept the system’s recommendations.
Indeed, we showed, using a game theoretical analysis, that the PRP is suboptimal as it does not promote content breadth in the corpus in terms of topical coverage.2 In other words, retrieval methods based on the PRP do not provide a strong enough incentive for publishers to produce diversified content. It also turns out that introducing randomness to ranking functions can help to promote content breadth in the corpus, and hence, increase the overall attainable utility for search engine users. We will discuss the sub-optimality of the PRP and the merits of using nondeterministic ranking functions later. The Probability Ranking Principle is Not Optimal in Adversarial Retrieval Settings
RS serves as a mediator among selfish participants social welfare
Strategic publishers
- Study of corpus dynamics that results from the strategic behavior of publishers in response to rankings Information Retrieval Meets Game Theory: The Ranking Competition Between Documents’ Authors
- Ranking robustness and explainable IR Ranking Robustness Under Adversarial Document Manipulations
- Non-Optimality of Classical Search and Recommendation Methods, and Randomness increases social welfare in search A game theoretic analysis of the adversarial retrieval setting
- Fairness in RS A Game-Theoretic Approach to Recommendation Systems with Strategic Content Providers
Competing mechanisms Best Response Regression
Strategic users
Optimizing Long-term Social Welfare in Recommender Systems: A Constrained Matching Approach
Convergence of Learning Dynamics in Information Retrieval Games
Robust Federated Recommendation System
Adversarial Attacks on an oblivious recommender An unscrupulous competitor can easily create a small number of malicious clients to bias the recommendation system. Shilling attacks, a.k.a. profile injection attacks