Improving Matrix Completion by Exploiting Rating Ordinality in Graph Neural Networks

TL;DR

A new method is introduced, called ROGMC, to leverage Rating Ordinality in GNN-based Matrix Completion, which uses cumulative preference propagation to directly incorporate rating ordinality in GNN's message passing, allowing for users' stronger preferences to be more emphasized based on inherent orders of rating types.

Abstract

Matrix completion is an important area of research in recommender systems. Recent methods view a rating matrix as a user-item bi-partite graph with labeled edges denoting observed ratings and predict the edges between the user and item nodes by using the graph neural network (GNN). Despite their effectiveness, they treat each rating type as an independent relation type and thus cannot sufficiently consider the ordinal nature of the ratings. In this paper, we explore a new approach to exploit rating ordinality for GNN, which has not been studied well in the literature. We introduce a new method, called ROGMC, to leverage Rating Ordinality in GNN-based Matrix Completion. It uses cumulative preference propagation to directly incorporate rating ordinality in GNN's message passing, allowing for users' stronger preferences to be more emphasized based on inherent orders of rating types. This process is complemented by interest regularization which facilitates preference learning using the underlying interest information. Our extensive experiments show that ROGMC consistently outperforms the existing strategies of using rating types for GNN. We expect that our attempt to explore the feasibility of utilizing rating ordinality for GNN may stimulate further research in this direction.

Figures (2)

• Figure 1: Method overview. 'U' denotes observed interaction with no rating from users. This example uses $\mathcal{T}=\{1,3,5\}$.
• Figure 2: Effects of ROGMC. (a-b) L2 distances between the averaged user representations across various rating types. (a) and (b) corresponds to the results of GCMC+ and ROGMC, respectively. (c) RMSE with varying hyperparameters: $\lambda$ and $\alpha$. Results of Rating-frac=50% on ML-100K.