Abstract
Aggregate performance on continuous-time dynamic graphs (CTDGs) combines, in a single score, the portion attributable to known temporal regularities and the additional predictive power of neural models. This study separates the two at the query level. We construct a mechanism-constrained predictor that uses pair recurrence, recency and history position, renewal patterns, and short sequential transitions while learning the compatibility within each mechanism. Across four CTDG datasets, this predictor recovers a substantial portion of the performance of strong neural baselines, and the recovered performance quickly saturates with a small, dataset-specific set of explicit mechanisms. Neural residuals concentrate on queries for which the positive and negative candidates have similar mechanism-execution profiles. Allowing conditional interactions among mechanisms is more effective than simply reweighting their existing contributions. Conditioning the contribution of one mechanism on the execution state of another recovers 54.9-73.2% of the original neural-only queries and improves overall paired accuracy on all four datasets. Although the magnitude of the effect varies across datasets, these results show that the performance gap of neural CTDG models need not be treated solely as an opaque difference in representational capacity. At least part of the gap is localized to queries with similar candidate execution profiles and can be functionally explained by conditional coordination among known, low-dimensional mechanisms.
本页仅发布可检索的书目信息与作者摘要。HTML 正文、PDF 和 LaTeX 源文件是独立阅读资源，不作为搜索引擎落地内容。打开页面后会按浏览器语言自动翻译标题与摘要，并写入缓存。