Serendipity into session-based recommendation: Focusing on unexpectedness, relevance, and usefulness of recommendations

Conventional recommender systems have a potential problem of hyper-personalization that produces biased recommendations, decreasing the diversity of items. Concerning this issue, serendipity-oriented recommender systems have been introduced as an alternative to increasing recommendation diversity. However, these systems have a drawback in potentially recommending uninteresting items from a target user's perspective. To overcome this limitation, the present study introduces a novel algorithm to recommend diverse items that are likely to be favored via fusing serendipity into existing algorithms with maintaining accuracy. Specifically, we incorporate elements of serendipity into session- and graph neural network-based recommender systems by utilizing not only unexpected but also relevant data from historical data. Furthermore, our algorithm reflects the personalized balancing for trade-off via a user-controllable parameter. In the application of our algorithm, the experimental results on two real-world datasets show the algorithm's trade-off between accuracy and diversity and its potential from the user-centric perspective. This study contributes to proposing a new algorithm that provides not solely useful but also varied recommendations to broaden users’ horizons and to simple and clear algorithmic improvement by its application.