Improving Serendipity for Collaborative Metric Learning Based on Mutual Proximity

Today, in web space, where content is constantly expanding, recommendation systems that enable users to explore information passively have become essential technologies, and their accuracy is significantly improving. However, recent studies have focused not only on enhancing recommendation accuracy by suggesting items that match user preferences but also on increasing the appeal of recommendation systems by offering unexpected discoveries and recommendations that exceed expectations. Studies have been conducted to achieve serendipity, recommendations that are both beneficial to the user and marked by novelty and surprise. However, existing systems alter the embedding space compromising this flexibility. In this study, we propose a recommendation method called mutual proximity collaborative metric learning, which improves serendipity for users in an embedding-based recommendation method called collaborative metric learning. The proposed method improves existing techniques by refining the embedding space search algorithm, reducing the bias toward popular items in recommendations without altering the original embedding space, thereby enabling users to achieve serendipity. Furthermore, experimental results demonstrate that our approach outperforms existing methods across various evaluation metrics, such as unpopularity and serendipity. Our approach can thus achieve serendipity and preserve the applicability of the embedding space to diverse tasks.