Revisiting explicit recommendation with DC-GCN: Divide-and-Conquer Graph Convolution Network

Abstract

In recent years, Graph Convolutional Networks (GCNs) have primarily been applied to implicit feedback recommendation, with limited exploration in explicit scenarios. Although explicit recommendations can yield promising results, the conflict between the sparsity of data and the data starvation of deep learning hinders its development. Unlike implicit scenarios, explicit recommendation provides less evidence for predictions and requires distinguishing weights of edges (ratings) in the user-item graph.

To exploit high-order relations by GCN in explicit scenarios, we propose dividing the explicit rating graph into sub-graphs, each containing only one type of rating. We then employ GCN to capture user and item representations within each sub-graph, allowing the model to focus on rating-related user-item relations, and aggregate the representations of all subgraphs by MLP for the final recommendation. This approach, named Divide-and-Conquer Graph Convolution Network (DC-GCN), simplifies each model’s mission and highlights the strengths of individual modules. Considering that creating GCNs for each sub-graph may result in over-fitting and faces more serious data sparsity, we propose to share node embeddings for all GCNs to reduce the number of parameters, and create rating-aware embedding for each sub-graph to model rating-related relations. Moreover, to alleviate over-smoothing, we utilize random column mask to randomly select columns of node features to update in GCN layers. This technique can prevent node representations from becoming homogeneous in deep GCN networks. DC-GCN is evaluated on four public datasets and achieves the SOTA experimentally. Furthermore, DC-GCN is analyzed in cold-start and popularity bias scenarios, exhibiting competitive performance in various scenarios.