Temporal-Aware Edge Embeddings via Cross-Dimensional Attention Networks for Fraudulent Transaction Detection in Evolving Graphs
Lists
Only you see this
Abstract
Real-world networks such as financial transaction systems and social networks exhibit highly non-stationary behavior, where noisy pseudo-labels and class imbalance often materialize as subtle edge patterns distributed across time. Conventional approaches either treat structural proximity and temporal evolution as orthogonal factors or compress the entire history into static snapshots, both of which blur fine-grained interaction signatures. We argue that effective dynamic graph anomaly detection demands embeddings that live natively in the spatio-temporal joint space rather than concatenated post-hoc. To this end, we introduce a framework that constructs hybrid edge representations by spatio-temporal unified graph representation n with relative-time positional encoding. Our architecture couples spatio-temporal encoding strategy with a learnable attention gate that weights observed connections according to their chronological distance, allowing the model to soften stale interactions while amplifying recent anomalies. Instead of relying on a single discriminative boundary, we deploy twin position-aware attention transaction manifolds and another exploring rare deviation regions. To cope with pseudo-anomalous labels reports that pollute the training signal, we add a calibrated diffusion anomaly scoring that effectively retracts gradient contributions from likely mislabeled examples. On datasets spanning cryptocurrency flows and e-commerce ratings, our method consistently narrows the detection gap compared to prior diffusion and recurrent bayesian inference baselines, particularly in low-label regimes.
BibTeX
@misc{ding2026temporal,
author = {Xinghao Ding},
title = {{Temporal-Aware Edge Embeddings via Cross-Dimensional Attention Networks for Fraudulent Transaction Detection in Evolving Graphs}},
year = {2026},
url = {https://openreview.net/forum?id=bwBPqx4cwW},
howpublished = {OpenReview},
}Loading related content...