Enhanced Multi-Task Learning and Knowledge Graph-Based Recommender System

Abstract

In recent years, the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">m</i> ulti-task learning for <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i> nowledge graph-based <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">r</i> ecommender system, termed MKR, has shown its promising performance and has attracted increasing interest, because a recommendation task and a knowledge graph embedding (KGE) task can help each other to improve the recommendation. However, MKR still has two difficult issues. The first is how fully to capture users' historical behavior pattern in the recommendation task and how fully to utilize deep multi-relation semantic information in the KGE task. The second is how to deal with datasets with different sparsity. Tackling these challenging issues, this paper proposes an enhanced MKR (EMKR) approach with two novelties. First, we propose to utilize the attention mechanism to aggregate users' historical behavior for more accurately mining preferences in the recommendation task, and utilize the relation-aware graph convolutional neural network to fully capture the deep multi-relation neighborhood features in the KGE task, so as to address the first issue. Second, a two-part modeling strategy is proposed for a better representation of users in the recommendation task to expand the expressive ability of the model for adapting to datasets with different sparsity, so as to address the second issue. Extensive experiments are conducted on widely-used datasets and 11 approaches are used for comparison. The results show that the proposed EMKR can achieve substantial gains over the compared state-of-the-art approaches, especially in the situation where user-item interactions are sparse.