Investigate Fathom-based page extraction

[Osmose]

As per the latest mockups, we probably want to extract the following info about a product:

• Name of shopping site (e.g. Amazon)
• Price
• Product name
• Representative image of product

In lieu of a better way of verifying whether the Fathom ruleset, let's use this:

• A training corpus of 100 pages from the 5 supported commerce sites.
• Train the ruleset using 50 of the pages, verify it against the other 50
• Aim for 95% accuracy

[Osmose]

@javaun What 5 shopping sites do we want to target for the first round? We can expand later if necessary.

[biancadanforth]

Last night, I was able to get Fathom running in our web extension, in large part thanks to Victor Ng's and Swathi Iyer's prior work.

This currently runs Fathom in the content script for the tab, and I am getting consistent notifications from Firefox on an Amazon product page that it is slowing down Firefox.

@Osmose , would it be possible to run this script in a separate thread/process? In general, what might be some options to work through this? Currently I cannot get Fathom to run reliably without hanging.

[javaun]

Top domains for EN-US:

1. Amazon
2. eBay * +
3. Walmart
4. Best Buy
5. Macy's

• eBay is a distant no. 2 in sales and may be a PITA. It's also a marketplace and a mixed bag. But they're huge (2x the size of Walmart by online sales) AND we have a rev share with them. If they turn into a Fathom pain or don't (or testing says users don't want them), we could drop them at some point.

• Not in this list is Apple.com, who is roughly the size of Walmart in ecommerce. But they're a brand destination and probably not a great us of one of our precious 5 slots for MVP.

[javaun]

My bad: for 5 let's use Homedepot.com instead of Macy's. Home Depot is slightly bigger and less likely to be roadkill like Macy's. Clothes are getting hammered. Hardware/home much more defensible

[brucko]

Dropping these here if you need more inspiration

Amazon, Walmart, Best Buy, Apple, Target, Etsy, Ikea, Gap, Macys, Home Depot, Lowes, Staples, Office Depot, Zappos, B&H, Costco

[biancadanforth]

Re: Perceived Fathom slowness

• This was observed when running Fathom on every page with Swathi Iyer's rulesets.
• We profiled the extension using Fathom with and without Swathi's rulesets, and we discovered the performance issues were largely related to the complexity of her rulesets (principally, some RegExp rules), not Fathom itself.

---

Possible rules to consider for extracting product features (title, price, image) from a product page

Fathom scores elements in the DOM by how well they follow a given ruleset, with modifiers applied based on how good that rule seems to be at finding the correct element.

Here is a list of some rule ideas collected from swathiiyer2, Osmose, erikrose and myself by feature, ranked in order of how well they seem to match a product page for our top 5 sites (listed here and here):

All three features:

• Above the fold
• Clustered "nearby" one another
  • This is an interdependency rule; for example, once we are confident in an image, find title-like and price-like elements nearby
• Has the feature name (e.g. "title" or "price") in its id
• Has the feature name (e.g. "title" or "price") in one of its class names
  • Use includes for better perf than a RegExp.
  • Check ancestor elements for class names and add a smaller modifier.
  • Add a smaller modifier for this rule compared to the id rule above.
• Larger font size (for "title" and "price")
• Optimal string length (for "title" and "price")

For each feature below, these are additional, feature-specific considerations on top of the cross-feature considerations above:

Product title:

• An h1 element

Product price:

• A span or h2 element
  • Ignore ancestors with same innerText value
• Has a $ in the string
• Looks like or contains a price-like pattern, like '$XXXX.XX', with any number of Xs to the left of the decimal and exactly two to the right. X is a digit, 0-9.
  • May need to trim off substring before $
  • Home Depot notably does not match the '.XX' at the end.
    Non-default text color
  • Only seems to be the case for Amazon
• A bunch more rules from Swathi's work

Product image:

• An img element
• Largest image
• Sufficiently large image
• Aspect ratio closest to 1:1 (image is more square in shape than rectangular)
• A bunch more rules from Swathi's work

[Osmose]

Updating this issue to reflect that this is an investigation and we haven't yet committed to using Fathom-based extraction.

[biancadanforth]

Referencing PR #45 , here are the conclusions from my Fathom assessment:

Conclusions:

• My ruleset has an average accuracy of 98.7% across a training set of 50 product pages, 10 each from our top five sites.
  • Note that the accuracy “in the wild” will likely be lower, so ideally these rules need to be further refined and trained across more pages to get us to 100%.
• My recommendation is to use Fathom with a CSS selector fallback. We can switch to CSS selectors as primary if there are performance or other concerns that arise with Fathom as the add-on takes shape.
• For this approach, we will need to build up a test set of product pages to test Fathom “in the wild” and an extensive list of CSS selectors for our top five sites.

Additional information:

Osmose and I met with erikrose last Friday and outlined our needs and aspirations for Fathom. There are two action items for Erik there that he said he could complete in a matter of weeks.