Comparing apples to apple trees: Injecting domain knowledge to similarity searches
summary
Claim: product-level similarity improves by injecting known image→product links into image similarity and aggregating up to products. Implementation overlays an image similarity matrix with a product-link block matrix (ones between images of the same product), then compresses images into product nodes via graph minors and uses the max-pair aggregation (or thresholds/counts) to build product similarity. NetworkX prototyping works but doesn’t scale; custom Python/R packages use sparse matrices for a 1–2 order magnitude memory/speed gain and a TensorFlow version is in progress.
AI-written summary · about
chapters (7)
- 0:00Introduction to domain knowledge
- 1:54Product similarity challenges
- 3:59Domain knowledge in similarity
- 5:23Implementing domain knowledge
- 7:10Graph-based aggregation method
- 8:37Scaling the approach
- 10:00Future applications and conclusion
related talks
36mFinding needles in haystacks with Deep Neural Networks
22mHarnessing image and text data for product matching
37mSimilarity Search for Product Matching @ Semantics3
45mTransformer based clustering: Identifying product clusters f...
17mCalculating recommendations based on product images #bbuzz
21mShopping Recommendations
49mFrom research to production at ebay
41mSearching for Products by Photo