Towards Practical Causal Inference for Recommender Systems: Combinatorial Interventions, Complex Evaluations, and Robust Generalization
Abstract
Recommender systems play a central role in connecting users to products on modern digital platforms. To build effective recommendersystems, we analyze past user behaviors to understand what recommendations can improve user experience and increase user engagement. Causal inference is a key tool for answering such questions about recommendation effectiveness; it helps us reason about how userswouldbehave if they were recommended (or were not recommended) different items. While proven useful, existing causal inference approaches are not readily applicable to large-scale recommender systems. Current causal inference approaches do not scale to practical recommender systems with an enormous number of intervention options, e.g. combinatorial in the total number of items. Second, existing causal inferences on recommender systems struggle to designate suitable counterfactual outcomes as evaluation metrics. Many systems rely on user click metrics to measure recommendation effectiveness, but they lead to more click-bait items in practice. Finally, existing causal approaches to recommender systems do not immediately generalize to new populations and new items. This research aimsto address these issues via a generative modeling approach. We will reason about how the counterfactual outcomes of different interventions are generated and what structure this generative process entails. This generative reasoning will lead to efficient algorithms for handling combinatorial interventions, complex evaluations, and robust generalization.The project will adapt generative modeling to practical causal inference on recommender systems in three topics. We will study the structure of how counterfactual outcomes are generated, and how it can enable efficient approximations of causal inference with enormous intervention options. We will also study how the structure of counterfactual outcomes and the corresponding metrics would affect the user behavior we observe. Building generative models for such user behavior will help design suitable outcomemetrics for recommender systems in a data-driven manner. We will finally model the generative process of counterfactual outcomes in new populations, leading to optimal predictors that are generalizable to new populations and items. This project will explore how generative modeling of counterfactual outcomes would help causal inference improve scalability, design metrics, and enhance generalizability.Approved for Public Release.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 29, 2023
- Source ID
- N000142312590
Entities
People
- Zhaoran Wang
Organizations
- Board of Regents of the University of Michigan
- Office of Naval Research
- United States Navy