Hierarchical Anomaly Localization in xDSL Internet Access Networks

Abstract

Copper digital subscriber line (xDSL) internet access networks remain widely deployed for the last-mile broadband service, including Fiber-to-the-CABINET (FTTC) configurations whose final segment is copper. Locating the topological origin of degradation in such networks is operationally critical but difficult in the absence of curated fault tickets. We study hierarchical anomaly localization in a real internet service provider (ISP) dataset of <inline-formula> <tex-math notation="LaTeX">$328{,}484$ </tex-math></inline-formula> active xDSL customers (ADSL and VDSL combined) across a multi-province region, organized along the DSLAM, line CARD, street CABINET, and BUILDING hierarchy. We compare three paradigms under a unified copper-pool design: (A) a rule-based hierarchical binomial peer test using vendor-aware robust z-scores; (B) an XGBoost classifier trained on customer-level features; and (C) a heterogeneous graph neural network (HeteroGNN) operating on a multi-relational topology graph with four node types. In the absence of a ground-truth fault registry, we adopt a rigorous label-free evaluation protocol comprising cross-method agreement (Cohen’s <inline-formula> <tex-math notation="LaTeX">$\kappa $ </tex-math></inline-formula>), bootstrap stability of top-<inline-formula> <tex-math notation="LaTeX">$K$ </tex-math></inline-formula> suspect lists, a permutation test for topological non-randomness of anomaly concentration, an edge-type ablation, and an embedding-quality measurement via silhouette score. Method A produces an attribution distribution dominated by clean and individual cases (89.99% and 5.40%) with 4.12% CABINET-, 0.33% CARD-, and 0.17% DSLAM-level attributions. HeteroGNN achieves a pseudo-label fidelity of <inline-formula> <tex-math notation="LaTeX">$F_{1}^{\text {macro}}=0.687$ </tex-math></inline-formula> (5-seed mean ± s.d.: <inline-formula> <tex-math notation="LaTeX">$0.693\pm 0.014$ </tex-math></inline-formula>), with markedly higher pseudo-label recall on minority topology classes than XGBoost (pseudo-label fidelity <inline-formula> <tex-math notation="LaTeX">$F_{1}^{\text {macro}}=0.529$ </tex-math></inline-formula>, 5-seed mean ± s.d.: <inline-formula> <tex-math notation="LaTeX">$0.529\pm 0.000$ </tex-math></inline-formula>), and an ablation indicates that removing the street-CABINET edge causes the largest performance drop (<inline-formula> <tex-math notation="LaTeX">$\Delta F_{1} = -0.230$ </tex-math></inline-formula>). A fourth control is additionally evaluated, in which the same neighbor-aggregate information is supplied to XGBoost directly as three peer-anomaly-fraction features; a pseudo-label fidelity of <inline-formula> <tex-math notation="LaTeX">$F_{1}^{\text {macro}}=0.954$ </tex-math></inline-formula> is attained by this variant, substantially exceeding the heterogeneous GNN. This result indicates that the practical value of the heterogeneous graph representation in this setting lies primarily in demonstrating that neighbor-aggregate information is highly predictive of hierarchical fault attribution, rather than in the relational architecture itself outperforming an explicit feature-engineering treatment of the same information. Permutation tests confirm that topology-level anomaly concentration is statistically non-random at all valid hierarchical levels (<inline-formula> <tex-math notation="LaTeX">$p \lt 0.002$ </tex-math></inline-formula>). The work is positioned as an observational methodological comparison and its limitations are made explicit.