Recommender Performance for Users with Interest in Several Fields on Collaborative Topic Regression through Vocabulary Completion

A recommender system is a tool that assists users with varying interests in obtaining information. While certain users may have deep interests in single fields, others may exhibit such interests across several fields. This latter group is referred to as “broad perspective users” throughout this paper. Recommendations of overly specialized content may not be appropriate for broad perspective users. Collaborative filtering is one of the most widespread methods for selecting preferable content, wherein user preferences are represented by a rating matrix. However, the typically high percentage of missing ratings, known as sparsity, results in suboptimal performance. To mitigate the degradation of performance in high-sparsity cases, a hybrid method taht combines ratings with other relevant data pertaining to items and users has been proposed. Collaborative topic regression (CTR) is a pioneer model, that extracts latent topics from documents about users or items and employs them alongside the rating matrix. Thus, words contained within a document has a significant impact on model performance. In a highly specialized domain such as scientific articles, we assume that adding less specialized words that fit the original document's content is more effective for broad perspective users. In this study, we conducted to improve the recommendation performance for broad perspective users by complementing vocabulary to the documents. To achieve this, we use a pre-trained language model for sentence generation. We verified that the performance of this approach is better for broad perspective users than for users that are not restricted. Furthermore, we conducted experiments with multiple sparsity levels, as the sparsity of the rating is interlinked with relevant topics in CTR.