Learning instrumental variable representation for debiasing in recommender systems.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Recommender systems face challenges with data biases, particularly popularity bias and latent confounders, leading to inaccurate and less diverse suggestions.
• Existing debiasing techniques often fail to address latent confounders or require predefined instrumental variables (IVs).

Purpose of the Study:

• To propose a novel causality-based recommendation algorithm, DIVRS, that learns instrumental variable representations directly from user-item interaction data.
• To address the bias amplification issue in Graph Convolutional Networks (GCNs) used in recommender systems.

Main Methods:

• Developed Data-driven IV representation learning for debiasing in Recommender System (DIVRS) to decompose user behavior into causal and confounding relationships.
• Introduced Orthogonal Promotion Regularisation (OPR) and a DIVRS-specific GCN variant (DIVRS-GCN) to mitigate bias amplification.

Main Results:

• DIVRS and DIVRS-GCN effectively mitigate confounding bias in recommender systems.
• Both algorithms demonstrated superior performance over state-of-the-art methods on Douban-Movie and Movielens-10M datasets, improving Recall@20 by up to 10.98%.

Conclusions:

• The proposed DIVRS and DIVRS-GCN approaches offer robust and effective solutions for debiasing recommender systems.
• These methods enhance recommendation accuracy, diversity, and balance, overcoming limitations of existing IV-based systems.