Enhancing Long-Tail Exposure in Session-Based Recommendation via Tail-Aware Training

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Recommender systems often concentrate exposure on popular items, limiting catalogue utilization and leaving long-tail items underexposed. This study evaluates TAILPromote under a leakage-controlled temporal protocol in which all eligibility, vocabulary, popularity, and long-tail definitions are derived from training data alone. TAILPromote combines a SASRec backbone with a gated tail-specific residual expert, popularity-aware logit adjustment, and catalogue-coverage regularization. Across five seeds on YooChoose 1/64, it attains MRR@20 of 33.04% ± 0.12, Coverage@20 of 89.93% ± 0.42, and LTP@20 of 39.63% ± 0.51. Paired hierarchical-bootstrap tests with Holm correction show lower accuracy than NISER+ but higher accuracy than SASRec and TailNet, together with significantly greater long-tail exposure than all three baselines. Factorial analyses identify logit adjustment as the main exposure contributor, with a smaller significant effect from Tail Expert gating. These results show that TAILPromote provides a controllable accuracy–exposure trade-off. Further evaluation on additional datasets and online settings is needed to assess generalizability and user-level utility.