Attributed Graph Convolutional Network for Enhanced Social Recommendation Through Hybrid Feedback Integration

Abstract

Social recommendation, a technique aimed at predicting user preferences by harnessing social ties, has frequently employed collaborative filtering (CF) due to its demonstrated efficiency and scalability. Nonetheless, a decline in performance of most extant CF techniques has been observed when confronted with extreme sparsity in explicit feedback. Past investigations predominantly merged both explicit and implicit feedback to mitigate the data scarcity issue, embedding based solely on explicit characteristics and formulating objective functions founded on user-item associations. Such a paradigm signifies a dependency on these interactions to compensate for deficient embeddings. Notably, a considerable discrepancy exists between implicit feedback and genuine user satisfaction in social recommendations, attributed to pervasive false positive interactions devoid of detailed user/item attributes. Furthermore, the establishment of connectivity between users/items has been partially dependent on users' inclinations, suggesting that the aggregation procedure might overlook certain neighbourhood preferences. In response to these challenges, a hybrid neural graph model endowed with attributive features has been introduced. This model amalgamates explicit/implicit feedback, attribute data, and a user-item interaction graph. To counteract data sparsity, a variational graph framework has been devised to extract latent representations from both feedback and attribute data. For the effective and explicit discernment of collaborative signals, the embedding incorporates a user-item interaction graph, which offers a potent modelling of elevated-order connectivities and the detection of latent user-item associations. The user and item embeddings are derived via an attentive propagation method, with the ultimate item embeddings being sourced through a linear weighted sum, eschewing non-linear activation functions. Comparative analyses on four real-world datasets have demonstrated the superior efficacy of the proposed methodology in relation to leading contemporary recommendation systems.