Multi-task convolutional deep neural network for recommendation based on knowledge graphs

Collaborative filtering suffers from data sparseness and cold start heavily in recommendation systems. Although several methods have tried to use auxiliary information in knowledge graphs to discover key representations of items and users to mitigate the issue mentioned above, there is still a certain space for improvement in their performances. This work proposes a multi-task learning scheme that exploits knowledge graphs to enhance recommendations. This scheme mainly includes recommendation tasks and knowledge graph embedding tasks, both of which are associated via convolutional neural networks to obtain richer potential information. Additionally, to better understand users' personalized needs and interests, we have incorporated user preference information into item feature representations to enhance recommendation performance. Furthermore, we have integrated relation importance information into entity feature representations, which enables the head entity to aggregate more crucial neighbor information. Experimental results show the proposed algorithm significantly improves the recommendations of MovieLens-1M, Book-Crossing, and Last.FM, compared with ten competing baselines. Meanwhile, we conduct ablation studies to prove the effectiveness of each module in the proposed scheme on performance improvement. We have made the source code of our proposed model available at https://github.com/MYJiang1102/recommendation-code.