Debiased Contrastive Representation Learning for Mitigating Dual Biases in Recommender Systems

TL;DR

This study tackles dual biases—popularity bias ($Z$) and conformity bias ($W$)—in recommender systems by formulating a causal graph and adopting do-calculus to guide debiasing. It introduces Debiased Contrastive Learning for Mitigating Dual Biases (DCLMDB), which learns two latent embeddings $Z$ and $W$ through contrastive losses $\mathcal{L}_{u}$ and $\mathcal{L}_{i}$ and optimizes with $\mathcal{L}_{DCLMDB} = \alpha \cdot \mathcal{L}_{BPR} + \beta \cdot (\mathcal{L}_{u} + \mathcal{L}_{i})$, aiming to decorrelate $Z$ from $I$ and $W$ from $U$. A manipulated graph $\\mathcal{G}_{\\overline{U,I}}$ removing edges $Z \rightarrow I$ and $W \rightarrow U$ under the intervention do$(I,U)$ supports unbiased learning. Empirical evaluation on Movielens-10M and Netflix shows DCLMDB achieving significant improvements in Recall, HR, and NDCG across MF and LightGCN backbones, demonstrating robust debiasing and better diversity. The proposed framework is model-agnostic and offers a principled, causally grounded path to fairer, more accurate recommender systems.

Abstract

In recommender systems, popularity and conformity biases undermine recommender effectiveness by disproportionately favouring popular items, leading to their over-representation in recommendation lists and causing an unbalanced distribution of user-item historical data. We construct a causal graph to address both biases and describe the abstract data generation mechanism. Then, we use it as a guide to develop a novel Debiased Contrastive Learning framework for Mitigating Dual Biases, called DCLMDB. In DCLMDB, both popularity bias and conformity bias are handled in the model training process by contrastive learning to ensure that user choices and recommended items are not unduly influenced by conformity and popularity. Extensive experiments on two real-world datasets, Movielens-10M and Netflix, show that DCLMDB can effectively reduce the dual biases, as well as significantly enhance the accuracy and diversity of recommendations.

Figures (10)

• Figure 1: Two causal graphs are used to show the recommendation process. $U$: user preference, $I$: exposed item, $C$: choice, $Z$: item popularity. (a) Traditional recommender methods; (b) Popularity bias caused by item popularity $Z$.
• Figure 2: Causal graphs showing both popularity bias and conformity bias and an illustration of our solution. $U$: user preference, $I$: exposed item, $C$: choice, $Z$: item popularity, $W$: conformity influence. (a) The causal graph considers the effects of popularity items and conformity influence; (b) we cut off the edges $Z\rightarrow I$ and $W\rightarrow U$ in the training model.
• Figure 3: Overall structure of the proposed DCLMDB. First, we use the backbone to represent the input data as base embeddings ($U$ and $I$) and debiased embeddings ($Z$ and $W$). Both sets of embeddings encompass those of users and items. Subsequently, in the debiasing learning phase (i.e., the dotted box portion of the figure), we employ contrastive learning to steer the debiased embeddings away from biases inherent in the base embeddings. The specific realisations of $\mathcal{L}_{u}$, $\mathcal{L}_{i}$ and $\mathcal{L}_{BPR}$ are in Eq. (\ref{['eq008']}), Eq. (\ref{['eq009']}) and Eq. (\ref{['eq0010']}) respectively. Finally, $\mathcal{L}_{u}$, $\mathcal{L}_{i}$ and $\mathcal{L}_{BPR}$ are summed as in Eq. (\ref{['eq0011']}) to obtain the final loss function $\mathcal{L}_{DCLMDB}$.
• Figure 4: Overlapped items with popular items. A higher IOU indicates that the recommendation result is more similar to the recommended top popular items.
• Figure 5: The IOU ratios of DCLMDB and its variants.

Theorems & Definitions (1)

• Theorem 1: Rules of do-calculus pearl2009causality