Paper 1（Regular）
■書誌情報
Ryotaro Hamada, Yohei Kobashi, Hiroto Sudo, Yutaka Matsuo: Lightweight Ban Prediction in YouTube Live-Stream Chat: Per-Genre Feature Selection and Automatic Genre Assignment, Proceedings of the 25th IEEE/WIC International Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT 2026), Regular Paper, December 2026
■概要
In YouTube live-stream chat, streamers and moderators sanction viewers who misbehave by prohibiting them from posting comments (hereafter, a ban). Detectors trained on ban records have been proposed, but most channels accumulate only a few such records, and these detectors cannot be applied to streamers who have none. We propose a detector that requires no ban records from the channel it is applied to. The detector ranks, in real time, the comments posted immediately before viewers are banned, using only the degree of semantic deviation from the conversation, measured with a lightweight embedding model, and statistics of text and behavior. Which features to use is determined per streaming genre; only this selection uses ban records, taken from existing channels other than the target, so the detector can be applied as is to new streamers who have no records. In an evaluation on about 6.8 million comments from 38 channels across five YouTube genres, the AUROC of the proposed method exceeded that of the harmfulness detection model Qwen3Guard-0.6B in four of the five genres, reaching 0.840 versus 0.668 for Japanese VTubers (difference +0.171, 95% confidence interval 0.147–0.197). Replacing this model with the four-billion-parameter Qwen3Guard-4B leaves the 95% confidence interval of the AUROC difference containing zero in every genre. The proposed method runs at 20.4 ms per comment on a CPU, about 208 times as fast as Qwen3Guard-0.6B, and for News a single-feature configuration that requires no embedding model runs at 0.014 ms. Furthermore, streams of unknown genre can be assigned to a known genre with an accuracy of 0.846 using only the embeddings of their first 10 comments.
Paper 2（Short）
■書誌情報
Soichi Onozuka, Yohei Kobashi, Takaaki Ohnishi, Yutaka Matsuo: Visual Similarity and Structural Aggregation in Examiner-Citation Networks for GNN Evaluation, Proceedings of the 25th IEEE/WIC International Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT 2026), Short Paper, December 2026
■概要
Design patents protect the ornamental appearance of products and contain little text, so prior-art search depends on visual comparison. The citations that examiners at the United States Patent and Trademark Office (USPTO) add between design patents therefore form a human-curated network of visual prior-art relationships. On the IMPACT corpus we ask what makes a design patent citable, and answer in three layers. Visually, examiner-citation pairs are consistently more similar than non-connected pairs under both image-embedding cosine similarity and zero-shot Multimodal Large Language Model (MLLM) judgments. Corporately, the gap concentrates in same-company pairs across every local citation structure, and the product-name decomposition of closed triangles identifies systematic repeated filing of near-identical design variants. Temporally, a nonparametric attachment diagnostic rejects degree-blind attachment while stopping short of the linear kernel, and a pre-registered holdout rejects static node fitness as a calibrated predictor even as it improves ranking. This layered answer makes a second question precise: does structural aggregation by a graph neural network (GNN) add anything beyond concatenating the features that already encode these layers? The experiment uses a co-citation representation of the same network, with same-company closed triangles contracted by Union–Find as the only preprocessing. A static graph transformer is compared against its own aggregation-removed ablation (K=0: features only) under a single pre-registered endpoint. Across all four executed cells, feature concatenation exceeds aggregation in test mean reciprocal rank (MRR). The pre-registered endpoint fails on the contracted graph (−0.0064, 95% CI [−0.0092, −0.0036], one-sided p≈1), so by the frozen commitment the main result is that structural aggregation does not exceed feature concatenation—and the interval lying entirely below zero suggests, post hoc, that it subtracts. Contraction significantly lifts both arms (post-hoc p < 10⁻⁴); the 2×2 interaction is not significant (p=0.22). The corporate signal is thus consumed more reliably as features than as structure—the paper’s central lesson restated at the level of model inputs.