Causal Incremental Graph Convolution for Recommender System Retraining.

Area of Science:

• Computer Science
• Artificial Intelligence
• Recommender Systems

Background:

• Real-world recommender systems require frequent retraining with new data to maintain performance.
• Graph Convolutional Network (GCN)-based models are state-of-the-art for collaborative recommendation but retraining is computationally expensive.
• Existing retraining methods often require full model retraining, which is inefficient and time-consuming.

Purpose of the Study:

• To develop an efficient retraining mechanism for GCN-based recommender models.
• To enable model updates using only new interaction data without sacrificing recommendation accuracy.
• To address the challenge of incorporating new data while excluding old graph structures in model updates.

Main Methods:

• Proposed a causal incremental graph convolution (IGC) approach for efficient GCN retraining.
• Introduced two novel operators: IGC for combining old representations with incremental graphs and fusing preference signals.
• Developed colliding effect distillation (CED) to address the out-of-date issue of inactive nodes using causal inference.

Main Results:

• The IGC approach successfully combines historical and new data representations, effectively fusing long- and short-term user preferences.
• CED mitigates the representation staleness of inactive nodes by estimating the causal effect of new data.
• Experiments on three real-world datasets showed significant speed-ups and accuracy gains compared to existing retraining methods.

Conclusions:

• The proposed causal IGC approach offers an efficient and accurate solution for retraining GCN-based recommender systems.
• This method enables effective model updates using only incremental data, overcoming the limitations of full retraining.
• The approach demonstrates practical applicability for maintaining high-performance recommender systems in dynamic environments.