Topology-Aware Modeling of Relational Anomalies
Open Access DepositedGraph-Based Risk Inference under Weak Supervision
Downloadable Content
item-localized preknowledge, symmetric collusion rings, and hierarchical broker networks. These controlled scenarios permit direct assessment of structural sensitivity under both weakly supervised and fully supervised learning regimes. Empirical results demonstrate that relational modeling advantage is topology-dependent. While flat tabular models perform comparably under independent, item-local anomalies, graph-based models yield substantial improvements under asymmetric and hierarchical coordination structures. Beyond synthetic validation, operational alignment analysis shows that graph-derived rankings significantly enrich real-world flag indicators relative to population base rates. Comparative evaluation against logistic regression and gradient-boosted baselines shows that the proposed weak relational model remains operationally meaningful and competitive, while strong proxy-optimized baselines may align more closely with existing flag logic in some scenarios. Rank overlap analysis further reveals that relational models surface partially distinct candidate sets,suggesting the recovery of higher-order dependency patterns not captured by independent feature-based approaches. Collectively, these findings show that graph-based risk inference extends conventional detection pipelines by incorporating topological structure, thereby bridging representation learning and audit-oriented decision support under measurement uncertainty.
verified ground-truth misconduct labels are typically unavailable. Operational flagging systems rely on statistical heuristics and expert review, producing risk indicators rather than confirmed outcomes. This measurement uncertainty complicates both model evaluation and algorithmic comparison. This thesis introduces a topology-aware risk inference framework based on graph representation learning.Examinee--item interactions are modeled as a bipartite graph, allowing relational dependencies to be encoded through neighborhood structure rather than independent feature aggregation. To enable systematic evaluation in the absence of verified labels, a synthetic injection framework is developed to generate operationally motivated anomaly patterns of increasing structural complexity
Detecting coordinated irregularities in high-stakes assessment data poses a fundamental challenge
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.