DISR: Distributional Item Representations for Sequential Recommendation

Abstract

Sequential recommender systems predict the next item from a chronological interaction history, typically representing users and items as point embeddings scored by dot product. DISR (Distributional Item Representations for Sequential Recommendation) instead represents items and sequence queries as diagonal Gaussian distributions. Item means are initialized from frozen Sentence-BERT text representations, while query and item variances are learned during training. Candidates are scored using closed-form symmetric KL divergence within the standard sampled-softmax cross-entropy framework. Experiments on Amazon Reviews 2023 Beauty and Sports evaluate 50K-user and full-scale settings, with five seeds at 50K. DISR improves NDCG@10 over the strongest non-DISR baseline by +12.8% on Beauty 50K and +15.6% on Sports 50K. At full scale, DISR leads on Sports by +2.6% over DuoRec and remains within 0.8% of the best NDCG@10 on Beauty. The largest gains occur for users with one to three interactions, reaching +14.5% on Beauty and+19.3% on Sports. DISR also converges substantially faster than the strongest contrastive baselines at full scale. Ablation results support the contribution of the overall distributional formulation, while the full-scale evaluation remains single-seed.