Introduction to Vector search - Argmx talk
0 likes62 views
The document provides an overview of vector search, its applications in recommendation systems and personalized search solutions, and discusses various algorithms and approaches, such as collaborative and content filtering. It highlights the importance of controlling trade-offs between accuracy and performance based on specific use cases, including considerations for vector density, query types, and filter implementations. Lastly, it emphasizes the standardization represented by the vecsim API for improved item querying in search and recommendation systems.
Introduction to Vector search - Argmx talk
- 1. Intro to Vec Search Uri Goren Chief Data Scientist @ Argmax
- 2. Agenda 1 2 3 4 5 6 About us RecSys & Search Approx Neighbor search Similarity Search stores A new standard Q&A
- 3. About Uri BestPractix ■ Leading argmax.ml ■ Our expertise: ○ Natural Language ○ Recommendation Sys ■ ML Community ○ Podcast ○ Meetup ○ Conferences
- 4. e-Commerce Why argmax cares about search ? We develop recommendation systems and personalized search solutions Creator Economy Ad Tech Direct Insurance
- 8. 3% 42% 49% 23% 47% 2% 13% 48% 89% 99% 51% 71% 77% 87% 99% 63% 97% 51%
- 10. (0, 1, 0, 0, 0 ,0 ,1 ,1, 1) (0, 0, 1, 0 ,0 ,1 ,0, 0, 0) (1, 0, 0, 0, 1 ,0 ,1 ,0, 1)
- 11. 0.03 1 0.9 -0.7 0.88 -0.8 Item vector and User vectors User2Item recommendation Item2Item recommendation
- 12. How to learn the Vectors ? Collaborative filtering Content filtering 0 1 0 0 1 0 1 0 1 1 0 0 1 0 1 0 1 1 Vegan? Keto? Low fat Dairy? Pastry? Snack? …. Shared Features
- 13. Search is essentially the same thing Query Query Encoder Different Encoder 0.8 score
- 14. Personalized search is almost the same thing Query Query Query Encoder Item Encoder 0.8 score User Encoder Query
- 17. Isn’t it just KNN ? Problem solved !?
- 18. How does Naive KNN work ? m Users / Queries n Items m times n times d2 (vector dim) Sort and take top k Sort complexity
- 19. Flat search (Naive) pros and cons Pros ✓ Extremely simple to implement ✓ Does not depend on metric ✓ Accurate results ✓ Best when search space is small Cons ✘ Holds vectors in memory ✘ Not scalable ✘ Does not take advantage of the search space topology
- 20. Hierarchical search John is from Seattle, find his k nearest neighbors No need to compare to the entire world What if John is near the border of 2 states ?
- 21. Inverted Indices 1 0 0 1 0 0 0 100 1 0 1 0 1 0 0 a b c d e When looking for the nearest neighbors of We can filter only records in which the c component is positive (and maybe above a certain threshold)
- 22. Locality sensitive hashing (-0.55,+0.73,+1.7) (-0.7,-0.52,+1.7) (-2.05,-0.1,-1.11) (-,+,+) (-,-,+) (-,-,-) Key Ordered list h(-,+,+) h(-,-,+) h(-,-,-)
- 23. Tradeoffs of the various algorithms Flat HNSW IVF LSH Accuracy Most Least Memory footprint High Medium High Depends on sparsity Low Metric Any L2/Cosine L2/Cosine Any Vector Size Small Small Large if sparse Large
- 26. How pre-filters are implemented US EU
- 27. How post-filters are implemented US EU Is item available for sale? Is item available for sale?
- 28. Vector Similarity Providers Google ScaNN And many more …
- 30. Consideration when choosing a VecSim server ● Is the vector sparse or dense ? ○ Typically, one hot embeddings work best with IVF and servers such as Elastic,Solr,lucene ○ Hybrid queries? ● Should the vectors be stored ? or just the IDs ● Is replication needed ? ● Is GPU acceleration needed ? ● Filtering in Query time ? ● Are items dynamically added and removed often? ● Batch vs online queries ○ E.g. building a mailing list vs live recommendation
- 31. Is vector search a product or a feature?
- 32. The VecSim API Essentially, regardless of the implementation, the API is pretty much the same Taken from the Vecsim pypi module
- 34. Suggestion
- 35. Summary VecSim A new standard for querying items for search and recsys. Control the Tradeoff Between accuracy and performance Depends on use-case Sparse or dense, batch or online, static vs dynamic items
- 36. Ready to get started ? ● Better experimentation, with faster convergence. ● Take context into account and personalize
- 37. THANK YOU! Uri Goren Chief Data Scientist @ Argmax