Ibraheem Wagih · Fraud & Financial Crime → Analytics
Projects / MSc Dissertation

Clustering WiFi Telemetry into Failure Archetypes

Individual MSc Dissertation · Lifemote partnership UCL MSc Business Analytics

To what extent can unsupervised clustering of WiFi telemetry identify distinct gateway failure archetypes, and how should ISPs prioritise them for proactive intervention? An industry dissertation with Lifemote, a WiFi analytics scaleup whose ISP clients serve roughly 6 million UK and European households.

525,983
active gateways in one weekly telemetry snapshot
2-stage
clustering: healthy vs degraded, then archetypes
50
PCA components carried into clustering
12,000
words, submitted August 2026

Problem

ISPs mostly discover in-home WiFi problems reactively, through helpline calls, engineer dispatches and churn. Existing proactive systems detect that a gateway is degrading but not how: a router hidden behind a television needs repositioning, a home saturated by neighbouring 2.4 GHz networks can be fixed remotely by band steering, a boot-looping gateway needs replacing. Treating those as one "degraded" population wastes intervention budget. The missing step is typing: segmenting degraded gateways into recognisable failure archetypes, each mapped to a feasible remediation route, without any hand-labelled fault data.

Design

A cross-sectional weekly snapshot rather than a pooled panel, because the weekly fields are rolling seven-day windows and pooling daily exports would duplicate each gateway into near-identical rows, inflating the apparent sample. The pipeline: data triage with an audit trail, leakage prevention, feature engineering over mixed telemetry, then PCA into a 50-component space. Stage one clusters the full fleet at k=2 to separate healthy from degraded. Stage two re-fits PCA on the degraded population alone and re-clusters it, letting failure types emerge without the healthy mass as a confound. K-Means (k-means++ initialisation) runs against a Gaussian Mixture Model comparator, with DBSCAN excluded on methodological grounds.

PCA scree plot and cumulative variance curve
Dimensionality selection: scree plot and cumulative explained variance for the PCA space carried into clustering.

Validation Before Interpretation

Internal validity uses Silhouette and Davies-Bouldin scores across candidate cluster counts, computed on a fixed, seeded 25,000-row subsample because Silhouette's quadratic cost is prohibitive at fleet scale. External validity checks cluster assignments against an operational quality-of-experience label that the clustering never saw, using Adjusted Rand Index and Normalised Mutual Information. Temporal stability re-runs the assignment on the following week's snapshot to test whether archetypes persist. Clusters are treated as hypotheses to be validated, not answers.

Silhouette and Davies-Bouldin scores across cluster counts
Model selection: Silhouette (higher better) and Davies-Bouldin (lower better) across k, K-Means vs GMM on the degraded subpopulation.
2D PCA projection coloured by QoE status
External grounding: the PCA space coloured by the held-out QoE label (20,000-gateway sample), checking that structure aligns with an operational signal.

Findings

The clustering recovers a small set of stable, interpretable failure archetypes with distinct remediation routes, validated externally and across consecutive weeks. Detailed findings and fleet statistics are held back from this page pending partner sign-off; the full dissertation is available on request where confidentiality allows.