Modeling Churn in Recommender Systems with Aggregated Preferences
Gur Keinan, Omer Ben-Porat
TL;DR
This work introduces Rec-APC, a recommender-system model that operates with aggregated user information and accounts for churn risk. The RS maintains a Bayesian belief over a finite set of user types $M$ and selects content categories from $K$ to maximize the expected number of likes, formalized as $V^\pi$ under a known $\mathbf{P}$ and prior $\mathbf{q}$. A key theoretical contribution is proving that optimal policies converge to pure exploitation after finite time via belief-walk dynamics, aided by a positive gap parameter and structural assumptions; it also provides a practical branch-and-bound algorithm for computing policies and analyzes its complexity relative to a POMDP baseline like SARSOP. The approach is validated through synthetic experiments and MovieLens-derived parameterizations, demonstrating that the Branch-and-Bound method yields competitive or superior performance, especially as the number of user types grows, thereby enabling scalable, privacy-conscious sequential recommendations with churn considerations.
Abstract
While recommender systems (RSs) traditionally rely on extensive individual user data, regulatory and technological shifts necessitate reliance on aggregated user information. This shift significantly impacts the recommendation process, requiring RSs to engage in intensive exploration to identify user preferences. However, this approach risks user churn due to potentially unsatisfactory recommendations. In this paper, we propose a model that addresses the dual challenges of leveraging aggregated user information and mitigating churn risk. Our model assumes that the RS operates with a probabilistic prior over user types and aggregated satisfaction levels for various content types. We demonstrate that optimal policies naturally transition from exploration to exploitation in finite time, develop a branch-and-bound algorithm for computing these policies, and empirically validate its effectiveness.