Knowledge Graph Random Neural Networks for Recommender Systems

In recent years, knowledge graph networks for recommendation have attracted extensive attention, since these methods can capture structured information by linking items with their attributes instead of using only interaction data between users and items. However, existing Graph Neural Networks Methods on Knowledge Graph suffer from over-smoothing and data sparsity, leading to the inability to build deeper networks. To address these problems, the Knowledge Graph Random Neural Networks for Recommender Systems (KRNN) is proposed. Specifically, a random dropout strategy is designed to generate the perturbed entities feature matrices. Then, a feature propagation method is proposed over the perturbed feature matrices for capturing high-order neighbor information, and locating the novel entities representation. The data augmentation matrices are generated by using the new entity representation from the previous step. The consistency regularization is designed to optimize the prediction across different data augmentation matrices in multiple random dropout. Extensive experiments on real datasets demonstrate the proposed method is superior to state of the baselines in alleviating over-smoothing and predicting user preferences, especially in data sparsity scenarios.

Automated summary

This study proposes a Knowledge Graph Random Neural Networks (KRNN) for recommender systems, which addresses the issues of over-smoothing and data sparsity in existing graph neural network methods on knowledge graph. By utilizing a random dropout strategy and feature propagation method, the proposed KRNN achieves superior performance in predicting user preferences, especially in data sparse scenarios.