AutoGCF: Personalized Aggregation on Neural Graph Collaborative Filtering

Graph neural networks have achieved state-of-the-art performance on collaborative filtering (NeuGCFs). The success of NeuGCFs is mainly attributed to the stacking of message aggregation layers. Recent studies have shown that the effectiveness of existing NeuGCFs largely relies on the selection of optimal aggregation steps, which makes the performance on various recommendation scenarios unsatisfactory. To tackle this, we for the first time propose a framework to achieve personalized aggregation step assignment on NeuGCF. First, each user is endowed with a learnable unit to measure the aggregation degree at each aggregation stage. Second, the aggregation degree is used to determine whether the aggregation process should stop after the current stage. Third, the learnable unit is jointly trained with the NeuGCF by bi-level optimization. Finally, we term this new framework AutoGCF, implementing it on the state-of-the-art NeuGCF model - Light-GCN. Empirical experiments on five datasets demonstrate the effectiveness of AutoGCF, and highlight that AutoGCF can adaptively choose optimal steps on different datasets.