Enhancing Deep Multimedia Recommendations Using Graph Embeddings

Abstract

Recently, deep learning techniques are widely used inthe field of multimedia recommender systems, like movierecommendation, news recommendation, music recommen-dation, and short video recommendation. With its power-ful ability in extracting representations and learning non-linear relations, deep learning-based collaborative filteringmodels outperform traditional collaborative filtering meth-ods (e.g., matrix factorization) and brings an immense liftin recommendation performances. However, these existing deep-learning-based recommen-dation methods only learn from binary user-item relation-ships - that is to say, deep models learn from low-level user-item interactions and the collaborative information hiddenin more complex user-item relationships in the data is notsufficiently exploited and thus will lead to inferior perfor-mance. To solve this problem, in this work, we proposea graph embedding enhanced method to exploit collabo-rative information by mining deep-level user-item interactions. Instead of learning from binary user-item relations, our method explicitly offers collaborative signals, which areencoded in embedding, to deep models. We conduct exper-iments on four real datasets from different multimedia rec-ommendation scenarios with different sparsity. We find outthat there are significant benefits using the graph enhancedmethod: converging faster and producing a better recom-mendation result. Results show that the proposed methodoutperforms traditional collaborative filtering methods andstate-of-the-art deep learning-based methods