Zhibo Jin, Zhiyu Zhu, Jiayu Zhang, Huijie Xu, Biaoliang Chen, Nan Yang, Fang Chen, Jianlong J. Zhou · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3841106
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).
Modern information and knowledge systems increasingly rely on frozen neural content scorers to rank, filter, recommend, moderate, and curate image, audio, and multimedia content. For trustworthy deployment, operators need attribution tools that explain not only salient input regions, but which structured signal components can locally restore or strengthen the scalar scores consumed by downstream ranking, filtering, and human-review workflows. We propose Adaptive-Path Wavelet Interpretation Maps (AP-WIM), a reference-input-free attribution method that computes nonnegative ascent-sensitivity maps in an invertible wavelet coefficient space. AP-WIM starts from the observed sample, advances along the local gradient field under an amplitude-proportional budget, and accumulates score-increasing sensitivity for input-aligned frequency-structured heatmaps. We characterize the origin-neighborhood risk of zero-baseline straight-line paths and, under local smoothness, derive a conditional score-improvement lower bound; the sign update is first-order optimal within its prescribed coordinate-wise budget box. On ESC-50 audio classifiers, paired-bootstrap tests show significant insertion gains under community-standard perturbation protocols, and a five-fold attribution stability check confirms that the insertion advantage is not unique to one audio fold. On a scorer-level Tiny-ImageNet query-gallery ranking audit, AP-WIM achieves the highest mean insertion AUC across target-score, Recall@1/5, MRR, and nDCG@10 endpoints, and significantly improves relevant-item score restoration over the strongest non-AP baseline. Together, the ranking audit, multimodal attribution benchmarks, runtime measurements, and robustness checks position AP-WIM as a resource-aware, frequency-structured explainability interface for frozen neural content-scoring pipelines. Code is available at: https://github.com/Davidjinzb/AP-WIM.
No comments yet — start the discussion below.