Little review
Hi Ramya, I just finished looking through the pipeline and the results you shared. Thank you for this, I really liked the work you did here.
Regarding the cleaning phase, I thought it was great that you went so exhaustively through the cleaning steps for each column. I especially liked the approach on the quantity, since you didn't only look for quantity information in the column itself, but also in the product name. I think it's a very clever idea. Even though it only impacted 3 records in this small dataset, I suspect this case happens a lot in the whole Canadian data. The work on brand_name to create a stable brand list also seemed really valuable to me. It made me think that we might want OFF to have a truly stable canonical_brand as you suggested, to allow for a well managed set of brands and affiliated brands.
My only question for this part is do you think this pipeline might work when applied to the whole database without breaking?
Moving on to the candidate generation phase, I felt the overall architecture was very solid. I really liked how methodically you handled the semantic resolution and tried different retrieval methods. I think it would be great to hear your thoughts on which method you found the most reliable, or if you feel that having all of them is necessary to end up with a good grouping.
I think that for OFF, having this kind of granular visibility into why two products match could be great for manual validation, or a better auto validation.
For the actual entity resolution part with Splink, I spent some time looking at the match weights and the distributions. I thought your observation about the strong negative prior (around negative 13) was super interesting. I correlate it with what I read on Splink, since an exact product name match only gives about +8, the model is essentially forced to over rely on nutritional similarity to push the score into the positive.
I was wondering if we could just plug semantic embedders like Cross Encoders trained to differentiate antonyms directly into Splink to catch words like salted vs unsalted?
Either way, I feel like your workflow of flagging these ambiguous pairs and routing them to an LLM is the best way to handle this, and by sending only the reduced, ambiguous batch you keep the criteria simple and the costs low. I also look for potential poisonning with the chain of similarity but I think LLM will be able to surely help to identify them. I just want to hear your thoughts about the thresholds you chose for the no review zones, though. Did you find them mostly through trial and error?
I also thought it could be great to use a Pydantic model to solidify the pipeline and prevent the LLM from formatting its output badly and breaking things.
Overall, I am really impressed! Thank you very much for such solid work! I have great expectations that this might scale really well.
Hi Yanis,
- Yes, this approach should work for the whole dataset without breaking. The current idea is to create a canonical name for every brand (kirkland, great value, presidents choice) since this creates a simpler pipeline and doesn't complicate things as the dataset size increases.
- I believe each method has contributed in getting better grouping results (at least for compliments). After experimenting with another brand, we can confidently conclude which methods we can utilize in our pipeline.
- I like the idea of using Cross Encoders directly with Splink. It is worth trying. Splink failed to deliver better results, mostly in product-variant cases, because there wasn't much difference in nutriments for Splink to catch and differentiate. Splink probability thresholds were selected empirically by inspecting the score distribution and representative product pairs across probability ranges. Because high Splink probability did not always imply product identity, the thresholds were used to define candidate/review zones rather than as definitive identity cutoffs. Ambiguous pairs were subsequently evaluated using structural evidence and semantic review.
- I absolutely agree on using Pydantic model as a next step to solidify the pipeline:)
Thank you very much for your valuable observations!