TrackerML - Web Tracking Intelligence

About this project

This visualization maps 3,000 third-party domains from DuckDuckGo's open-source Tracker Radar dataset, which catalogs domains observed loading on popular websites along with their behavioral metadata.

Click any node to run three ML models against it:

Tracking Detection

Binary classifier (F1 0.893) trained on 295 behavioral features. Detects tracking across all vectors, cookies, pixels, fingerprinting, not just browser API abuse. Identifies 338 trackers the fingerprinting heuristic misses, including facebook.com and adnxs.com.

Entity Type

Classifies domains into four behavioral clusters: ad tech, CDN/infrastructure, platform, or ad management. 75% accuracy. Useful because 57% of domains have no ownership information.

Entity Attribution

For ad tech domains, predicts the likely owning company (Google, Microsoft, Adobe, etc.) from behavioral patterns alone. Shown only when the cluster model identifies ad tech. Works well for distinctive entities; less reliable for smaller companies.

Limitations. The entity cluster model can over-predict ad tech for domains with limited metadata. Entity attribution is a probabilistic guess, not a definitive answer, domains with unusual behavioral profiles may be misattributed to larger entities that dominate the training data. The tracker classifier's main failure mode is false positives on CDN domains owned by large tech companies, where ownership metadata mimics tracking signals.