Manual ad filter lists are slow and resource-intensive, prompting the "Project Moonshot" initiative to automate ad detection using AI and machine learning.
The team pivoted from inefficient computer vision models for perceptual ad detection to analyzing HTML structure, which provided richer data for machine learning.
A custom crawler combined with a modified Adblock Plus was used to automatically label HTML nodes on 250,000 web pages, creating a large-scale ground truth dataset.
The data pipeline converted raw HTML into adjacency and feature matrices while solving challenges like severely unbalanced data and slow processing speeds.
Several models were tested for ad classification, including graph neural networks, traditional classifiers with node embeddings, and tree-based models like XGBoost.
Tree-based models significantly outperformed graph neural networks in F1 score, and future work will explore self-supervised learning and more diverse data.
The team tackled deployment challenges by converting Python models to JavaScript, optimizing for latency by moving the model to a background script, and using TensorFlow.js.
The speakers address audience questions regarding how ad companies might circumvent the model and the rationale behind their model experimentation process.
23:50Aditi Godbole
19:27Kapil Gupta
27:23Nico Martin
57:47Fabien Vauchelles
56:55Ekaterina Sirazitdinova
27:23Dora Petrella
27:34David vonThenen
27:02Mirko Ross
Jobs that call for the skills explored in this talk.
Peter Park System GmbH
München, Germany
