Causal Intervention for Leveraging Popularity Bias in Recommendation

Abstract

Recommender system usually faces popularity bias issues: from the data perspective, items exhibit uneven (usually long-tail) distribution on the interaction frequency; from the method perspective, collaborative filtering methods are prone to amplify the bias by over-recommending popular items. It is undoubtedly critical to consider popularity bias in recommender systems, and existing work mainly eliminates the bias effect with propensity-based unbiased learning or causal embeddings. However, we argue that not all biases in the data are bad, i.e., some items demonstrate higher popularity because of their better intrinsic quality. Blindly pursuing unbiased learning may remove the beneficial patterns in the data, degrading the recommendation accuracy and user satisfaction.This work studies an unexplored problem in recommendation -how to leverage popularity bias to improve the recommendation accuracy. The key lies in two aspects: how to remove the bad impact of popularity bias during training, and how to inject the desired popularity bias in the inference stage that generates top-𝐾 recommendations. This questions the causal mechanism of the recommendation generation process. Along this line, we find that item popularity plays the role of confounder between the exposed items and the observed interactions, causing the bad effect of bias amplification. To achieve our goal, we propose a new training and inference paradigm for recommendation named Popularity-bias Deconfounding and Adjusting (PDA). It removes the confounding popularity bias in model training and adjusts the recommendation score with desired popularity bias via causal intervention. We demonstrate the new paradigm on the latent factor model and perform extensive experiments on three real-world datasets from Kwai, Douban, and Tencent. Empirical studies validate that the deconfounded training is helpful to discover user real interests and the inference adjustment with popularity bias could further improve the recommendation accuracy. We release our code at https://github.com/zyang1580/PDA.

A Model-Agnostic Causal Learning Framework for Recommendation using Search Data

“…In recommendation scenario, however, we still focus on promoting the accuracy of preference estimation, rather than just identifying cause-effects. Existing studies also found that the non-causal associations can contribute to the prediction accuracy [44]. The observation motivates us that not all confounders (biases) need to be discarded.…”

A Semi-Synthetic Dataset Generation Framework for Causal Inference in Recommender Systems

“…Causal graphical model [25] is a powerful tool in causal inference and has also attracted much interest in recommender systems [41].…”

“…Besides, they are usually learned based on observed data and are skewed due to the closed feedback loop in recommender systems, resulting in "the rich get richer" Matthew effect [11,32,35]. In particular, the observed data possibly contains many biases including position bias [2,40], item exposure bias [32,39], user self-selection bias [31,36], and popularity bias [1,41], which are likely to result in biased models and affect users' experiences. There are several benchmarked datasets for unbiased recommendation, e.g., Yahoo!R3 [22], Coat [32], MSSD [6], and MovieLens [14], etc., which have been widely used to develop and evaluate recommendation models.…”

Deconfounded Recommendation for Alleviating Bias Amplification

“…Causal inference has been widely used in many machine learning applications, spanning from computer vision [23,34], natural language processing [11,12,43], to information retrieval [4]. In recommendation, most works on causal inference [25] focus on debiasing various biases in user feedback, including position bias [18], clickbait issue [37], and popularity bias [45]. The most representative idea in the existing works is Inverse Propensity Scoring (IPS) [2,28,41], which first estimates the propensity score based on some assumptions, and then uses the inverse propensity score to re-weight the samples.…”