Collaborative Competitive filtering II: Optimal Recommendation and Collaborative Games

TL;DR

This paper introduces a game-theoretic framework for recommender systems that explicitly optimizes strategic goals like revenue and diversity, moving beyond traditional preference recovery methods.

Contribution

It extends the Collaborative-Competitive Filtering model into a game-theoretic approach, enabling explicit optimization of strategic objectives in recommendation systems.

Findings

Optimized for click-through rate, sales revenue, and diversity.

Demonstrated promising results on a commercial system.

Bridged the gap between strategic goals and recommendation proxy.

Abstract

Recommender systems have emerged as a new weapon to help online firms to realize many of their strategic goals (e.g., to improve sales, revenue, customer experience etc.). However, many existing techniques commonly approach these goals by seeking to recover preference (e.g., estimating ratings) in a matrix completion framework. This paper aims to bridge this significant gap between the clearly-defined strategic objectives and the not-so-well-justified proxy. We show it is advantageous to think of a recommender system as an analogy to a monopoly economic market with the system as the sole seller, users as the buyers and items as the goods. This new perspective motivates a game-theoretic formulation for recommendation that enables us to identify the optimal recommendation policy by explicit optimizing certain strategic goals. In this paper, we revisit and extend our prior work, the Collaborative-Competitive Filtering preference model, towards a game-theoretic framework. The proposed framework consists of two components. First, a conditional preference model that characterizes how a user would respond to a recommendation action; Second, knowing in advance how the user would respond, how a recommender system should act (i.e., recommend) strategically to maximize its goals. We show how objectives such as click-through rate, sales revenue and consumption diversity can be optimized explicitly in this framework. Experiments are conducted on a commercial recommender system and demonstrate promising results.