An Analysis of Attribute Utilization in Side Information-integrated Sequential Recommendation

초록

Side Information-integrated Sequential Recommendation (SISR) incorporates item-dependent side information, such as category or genre, to complement item ID representations; however, despite numerous proposed fusion strategies, it remains unclear to what extent current SISR models actually exploit such information in practice. In this work, we present a shuffle-based analysis that examines the functional contribution of item attributes across datasets, models, and fusion strategies by randomly breaking item-attribute alignments and observing the resulting performance changes. Extensive experiments on five real-world datasets show that sensitivity to shuffling is highly heterogeneous and jointly determined by dataset characteristics, attribute types, and model design, with no fusion strategy consistently exhibiting stronger reliance on attributes. Overall, our findings indicate that the contribution of item attributes is conditional and dataset-specific rather than universal, highlighting the need for deeper theoretical and empirical analyses of the conditions under which attributes contribute to performance in SISR. © 2026 Owner/Author.

키워드