Chrome-based shopping classifier

Here's a nice analysis of the recent classifier model added to Chrome to determine if the page you are visiting is an ecommerce page or not: https://dejan.ai/blog/google-shopping-classifier/

Once more Google is adding bloat to it's browser and leaking more data back to their own servers. Use a de-Googled Chrome for your own stuff: https://github.com/ungoogled-software/ungoogled-chromium

Anyway, basically the model splits the content of a page into chunks of around 100 words, truncates them to 64 tokens, and limits the number of chunks it looks at to 10. This means that the model only really sees a maximum of 450 words from the page. From the analysis:

Rendered Page
→ Blink AnnotatedPageContent extraction (5 seconds after load)
→ Text nodes collected from content tree
→ Greedy word-count chunking into passages
→ SentencePiece tokenization (64 tokens per passage)
→ Passage Embedder (TFLite) → 768-dim vectors
→ Mean pooling + title/URL embedding concatenation → 1536-dim input
→ Shopping Classifier (TFLite) → probability score (0 to 1)

Click to expand...

The takeaway from this is that on your shopping pages, you need to be explicit in the first 450 words of hierarchical content in your HTML that your page is an ecommerce page, and that a certain product can be bought from it. Otherwise you won't get all the bells and whistles of price tracking and shopping insights.

Great share.So if your boilerplate or nav menus are too heavy, you're essentially invisible to this local classifier.