Reranking Serendipity Technique in Recommendation System

The problem of over-specialization in recommender systems can limit user exploration. To address this limitation, this study proposes a serendipity-oriented reranking method called Serendipity Reranking based on Genre Dissimilarity (SRGD), which incorporates three key components: relevance, unpopularity, and genre dissimilarity. Experimental results on the MovieLens-100K dataset show that hybrid methods combining SRGD with SVD (SVD-SRGD) and SVD++ (SVD++-SRGD) achieve high average serendipity scores of 0.535 and 0.540, and NDCG scores of 0.626 and 0.633 for n from 1 to 5. However, for n from 6 to 10, average serendipity scores decrease to 0.465 and 0.472, while NDCG scores increase to 0.641 and 0.649. Additionally, the hit-ratio metric shows substantial improvement to produce at least one serendipitous item for each user, rising from 0.799 and 0.881 for n from 1 until 5 to 0.963 and 0.969 for n from 6 to 10. These findings confirm a trade-off between serendipity and NDCG, where increased relevance (NDCG) can reduce serendipity. Nonetheless, the SRGD method demonstrates strong potential to enhance serendipitous recommendations. Future work may further improve performance by incorporating additional features such as user curiosity and item similarity. © 2025 IEEE.