ScholarWorks@성균관대학교

초록

Previous Click-Through Rate (CTR) prediction models rely on enumerating high-order feature combinations up to a fixed order, limiting expressiveness and scalability. Recent studies explored arbitrary-order interaction modeling through two major paradigms: log-based and graph-based methods. However, both paradigms suffer from inherent weaknesses: log-based methods lack stability, and graph-based methods lack generalizability, as both attempt to model overly diverse combinations of features, many of which may be noisy or redundant. This observation provokes a central question: What if only a small set of core interactions is sufficient? To explore this, we progressively mask feature interactions and find that removing up to 90% of them results in negligible performance degradation. This suggests that most interactions are unnecessary. Motivated by this finding, we propose SCV: Selectively Crossing Vectors, a CTR prediction framework that reformulates feature interaction learning as a sparse edge selection task over a globally shared feature-interaction graph. By modeling feature interactions over a globally learned graph and dynamically fusing expert outputs in an instance-aware manner, the SCV effectively leverages global consistency and local adaptability. We further introduce a label-biased self-distillation objective to mitigate the effects of noisy supervision and stabilize training. Experiments on public CTR benchmarks show that SCV achieves state-of-the-art performance while reducing computational cost by up to 66%, validating the effectiveness of globally sparse yet locally adaptive interaction modeling. All codes are available at: https://github.com/bw-99/scv.

키워드