Causal Disentanglement for Regulating Social Influence Bias in Social Recommendation

TL;DR

A Causal Disentanglement-based framework for Regulating Social influence Bias in social recommendation, named CDRSB, is proposed to improve recommendation performance and a disentangled encoder that focuses on disentangling user and item embeddings into interest and social influence embeddings is proposed.

Abstract

Social recommendation systems face the problem of social influence bias, which can lead to an overemphasis on recommending items that friends have interacted with. Addressing this problem is crucial, and existing methods often rely on techniques such as weight adjustment or leveraging unbiased data to eliminate this bias. However, we argue that not all biases are detrimental, i.e., some items recommended by friends may align with the user's interests. Blindly eliminating such biases could undermine these positive effects, potentially diminishing recommendation accuracy. In this paper, we propose a Causal Disentanglement-based framework for Regulating Social influence Bias in social recommendation, named CDRSB, to improve recommendation performance. From the perspective of causal inference, we find that the user social network could be regarded as a confounder between the user and item embeddings (treatment) and ratings (outcome). Due to the presence of this social network confounder, two paths exist from user and item embeddings to ratings: a non-causal social influence path and a causal interest path. Building upon this insight, we propose a disentangled encoder that focuses on disentangling user and item embeddings into interest and social influence embeddings. Mutual information-based objectives are designed to enhance the distinctiveness of these disentangled embeddings, eliminating redundant information. Additionally, a regulatory decoder that employs a weight calculation module to dynamically learn the weights of social influence embeddings for effectively regulating social influence bias has been designed. Experimental results on four large-scale real-world datasets Ciao, Epinions, Dianping, and Douban book demonstrate the effectiveness of CDRSB compared to state-of-the-art baselines.

Figures (8)

• Figure 1: (a) Fork structure of causal graph: confounder affects both the treatment and the outcome. (b) In social recommendations, we treat the social network as a confounder, with user and item embeddings as the treatment and user-item ratings as the outcome. Due to the existence of the social network confounder, there are two paths between user and item embeddings and user-item ratings: a non-causal social influence path and a causal interest path. (c) CDRSB disentangles the user and item embeddings into social influence and interest components and learns dynamic weights of social influence embeddings to fuse them, thereby effectively regulating social influence bias. $C$: social network, $T$: user and item embeddings, $R$: user-item ratings, $T_S$: social influence embeddings, $T_I$: interest embeddings, $T'$: reconstructed user and item representations, $\alpha$: the weight of social influence embeddings.
• Figure 2: The overall framework of CDRSB. It contains two modules: (1) a disentangled encoder that disentangles user and item embeddings learned from a GNN-based network into interest and social influence embeddings. We minimize mutual information-based objectives to reduce redundancy and ensure the separation of these disentangled embeddings. (2) a regulatory decoder that learns dynamic weights to combine interest and social influence embeddings into final user and item representations, achieving reasonable utilization of social influence bias.
• Figure 3: Visualization of user's interest embedding (red points) and social influence embedding (blue points) for different stages: (a) initialization, (b) convergence on the dataset Epinions.
• Figure 4: Visualization of item's interest embedding (red points) and social influence embedding (blue points) for different stages: (a) initialization, (b) convergence on the dataset Epinions.
• Figure 7: RMSE and MAE on different $\lambda$: (a) RMSE (b) MAE.