A Method to Improve Serendipity of Recommendation Lists Based on Collaborative Metric Learning

Abstract

Collaborative Metric Learning (CML) is a recommendation model based on an embedded representation trained by implicit data, i.e., the behavioral histories of clicking and browsing. CML learns a metric space to embed users and items considering not only the relationship between users and items, but also item–item similarity and user–user similarity. Moreover, CML can recommend items close to each user in the trained space that match the preferences of the user. However, CML tends to be influenced strongly by popular items among many users, and the accuracy of embedded representations of other minor items is often neglected. It is necessary to learn the embedded representations of minor items that match user preferences with higher accuracy to provide unexpected recommendations of items that users may not be aware of in advance. In this article, a method is proposed to learn the embedded representations that capture user preferences by weighting loss functions according to the number of observations of implicit data and to make unexpected and effective recommendations, including minor items. Finally, the proposed method is applied to an actual movie evaluation dataset, and the usefulness of the proposed method in making unexpected recommendations based on user preferences is demonstrated.