Learning Knowledge Embeddings with Prior Weights for Sparse Interaction Recommendation

Knowledge-based recommendation models have exhibited their excellent performance in recent years. Most of these models encode knowledge into item embeddings through a graph embedding algorithm, which are useful for uncovering correlations between users and items. However, the graph embedding algorithms in these models neglect the different weights of various relations between items (entities), thus imprecise embeddings are learned resulting in unsatisfactory recommendation results. To address this problem, we propose a deep knowledge-based recommendation model which incorporates a novel graph embedding algorithm with prior relation weights, to learn precise item embeddings. Specifically, an HIN is first constructed based on the entities and relations from open knowledge graphs (KGs). Then, the embeddings of item vertices in the HIN are learned through seeking similar items in terms of various attributes (relations) with different prior weights. Next, the user representations are learned through user-tag-item relationships, based on which recommendation results are obtained by a multi-layer perceptron (MLP) fed with user presentations and item representations (embeddings). All the embeddings learned in our model are regarded as knowledge embeddings. The extensive experiments show that, our model outperforms the previous KG-based recommendation models with help of precise knowledge embeddings. Furthermore, it owns robust performance in the scenario of sparse user-item interactions, since it captures user preferences mainly based on the knowledge rather than observed user-item interactions.