CatDive: A simple yet effective method for maximizing category diversity in sequential recommendation

How can we offer users items of diverse categories while considering their interests in sequential recommendation? Traditional sequential recommendation methods often prioritize accuracy over diversity, recommending only items similar to those with which users have previously interacted. Recommendation biased toward only the favorite category of each user fails to achieve novelty and serendipity, demoting users' satisfaction. Thus, it is important to maximize category diversity in sequential recommendation. However, the state-of-the-art work on category diversified sequential recommendation tackles it only with a post-processing method which sacrifices accuracy to improve category diversity. In this paper, we propose CatDive, a simple yet effective method for maximizing category diversity without sacrificing accuracy in sequential recommendation. In the training phase, CatDive incorporates category information into the recommendation model by using multi-embedding, composed of item and category embeddings. Also, CatDive boosts category diversity by placing greater emphasis on users' preferred categories when selecting negative samples, while still prioritizing high-confidence negatives based on item popularity to maintain accuracy. Finally, we enhance and control category diversity by prioritizing underrepresented categories in the reranking phase, using a tunable hyper-parameter to balance adjustments. CatDive achieves the state-of-the-art performance on real-world datasets, achieving up to 152.49% higher category diversity in the similar level of accuracy and up to 39.73% higher accuracy in the similar level of category diversity compared to the best competitor.