Causal Solutions for Recommendation System Bias and Practical Applications

The article introduces the problem of bias in industrial recommendation systems, which consist of multiple stages such as recall, coarse ranking, fine ranking, and re‑ranking, and explains why traditional correlation‑based models cannot fully resolve issues like bias, noise, and distribution shift.

It contrasts correlation modeling with causal modeling, highlighting that causal methods use interventions and counterfactual learning to achieve unbiased estimates, illustrated with an e‑commerce coupon example where random exposure breaks the closed loop of biased data.

Using Rubin's potential‑outcome framework, the authors propose a causal analysis pipeline for recommendation tasks, consisting of (1) defining causal estimands for the recommendation goal, (2) analyzing identifiability of the estimand from observational data, and (3) designing models that provide unbiased estimators.

The article details the selection‑bias problem, where exposure position influences click probability, violating the SUTVA assumption, and discusses existing solutions such as ESMM, IPW, and double‑robust (DR) learning that combine propensity scores with error‑imputation to achieve unbiased predictions.

Building on DR, a generalized double‑robust learning method is introduced that jointly controls error, bias, and variance terms by using separate loss functions for the prediction and imputation models, offering a unified framework for existing bias‑correction methods.

To further enhance robustness, a multi‑robust learning approach is proposed, leveraging ensembles of J propensity‑score models and K error‑imputation models; linear combinations that accurately estimate either component guarantee unbiasedness, addressing model‑specification challenges.

For individual treatment‑effect (ITE) estimation, the paper presents a representation‑learning method that decomposes ITE into factual outcome estimation and distribution alignment, and identifies three open issues: distance metric choice, mini‑batch sampling bias, and unobservable confounders.

To solve these, the authors adopt optimal transport as a symmetric, structure‑aware distance metric and introduce the ESCFR framework, which adds RMPR regularization to mitigate mini‑batch sampling bias and PFOR regularization to handle unobservable confounders.

A practical case study on music recommendation shows severe sample noise due to uncertain user attention; two solutions are offered: adaptive label correction based on model loss dynamics, and attention‑weighted learning to reduce the impact of noisy feedback.

The concluding summary emphasizes that causal inference can address many challenges in recommendation systems, offering robust, interpretable, and intelligent‑marketing‑oriented solutions across bias mitigation, multi‑robust learning, and ITE estimation.