Topic-Based Semantic Enhancement for Recommendations
Dongjiao Liu, Ting Jiang · 2024
The massive natural language text produced in the information age contains rich information, how to use text information for recommendation is a hot content in the recommendation field. With the development of deep learning, preprocessing models such as BERT have shown very good results in processing text features. However, BERT is difficult to perform structured semantic extraction and performs poorly on structured text data. For this reason, this paper proposes a model T-DGCF based on topic semantic enhancement, which convolves the structured information and content information of the text through the topic model, and takes the obtained text topic vector as the initial embedding of the products. Then, we obtain the user's embedding about the topic through the interaction relationship. The two-part embedding is used as the initial embedding of the recommender system network for recommendation. We conducted experiments on two public datasets of different sizes and compared them with some competing baseline models. The experimental results show that T-DGCF outperforms several advanced recommendation models in two commonly used evaluation metrics.