Power of Visualizing Embeddings
2 likes514 views
The document outlines a presentation by Pramod Singh on visualizing embeddings and their application in understanding customer journeys across various domains. It discusses the challenges of high-cardinality variables and the use of embeddings for representing user interactions and categorizing journeys. Key takeaways include the advantages of embeddings in machine learning predictions and the ability to visualize relationships between categories.
Power of Visualizing Embeddings
- 2. The Power of Visualizing Embeddings Pramod Singh
- 3. About Me .. ▪ Team Lead – Data Science Bain and Company ▪ Speaker -O’Reilly Strata conference -GIDS ▪ Published Author • Machine Learning using PySpark • Learn PySpark • Learn TensorFlow 2.0 : The easy way • Machine Learning in Production ( WIP) ▪ https://www.linkedin.com/in/pramodchahar/
- 4. Agenda ▪ Inspiration for this session ▪ Conventional Approach ▪ Learning Embeddings ▪ Custom Embeddings ▪ Visualizing Embeddings ▪ FAQs
- 6. Interactions Finance Exteriors Interiors Maintenance Car DealerFeatures
- 7. User Journey – Core Elements Different Pages Categories Time Spent Sequence of Events
- 8. User Representation – I User ID Total Visits Total Time Spent Total Pages Total Sessions Converted 121A 10 25 110 4 0
- 9. User Representation – II User ID Total Pages Finance Specification Dealer … … Finance(sec) Specification(sec) Dealer(sec) … Converted 121A 10 3 4 4 3 5 5 0
- 10. All Users User ID Total Pages Finance Specification Dealer … … Finance(sec) Specification(sec) Dealer(sec) … Converted 121A 10 3 4 4 3 5 5 0 19X2 50 0 21 0 0 350 0 0 GG52 33 8 4 9 45 50 78 1
- 11. Applicable to other domains Finance & Insurance E-Commerce/Retail Real Estate
- 12. Key Questions ▪ Which set of customer journeys are similar to each other ? ▪ Which set of customer journeys indicate broken vs seamless experience ? ▪ Which are those 4-5 major routes that customers takes in order to convert ?
- 13. Category Representation Frequency Based Prediction BasedOne Hot Encoding
- 14. Challenges High-Cardinality Variables Sematic Signal w/o Supervision
- 15. Challenges Specifications Number of columns = Number of unique categories Price Features Specifications Price Features Reviews .. … 1 0 0 0 0 0 0 1 0 0 0 0 0 0 1 0 0 0 Similarity between Specifications and Price = 0 Similarity between Price and Features = 0 Similarity between Features and Specifications = 0
- 16. Gaps • Sequence of events is ignored Can we represent each of these page categories with a vector which captures the underlying semantics ? Using this vector , can we represent each user journey?
- 17. Embeddings
- 18. Embeddings “An embedding is a mapping of a discrete — categorical — variable to a vector of continuous numbers such that the vectors of similar entities are closer to one another in vector space.” king - man + woman = queen
- 19. Category Similarity using Embeddings Price 0.43 0.75 0.98 … …. … 0.55 0.87 Specification 0.23 0.10 0.33 … …. … 0.45 0.20 Features 0.22 0.09 0.30 … …. … 0.44 0.18 Similarity between Specifications and Price = - 0.75 Similarity between Price and Features = - 0.83 Similarity between Features and Specifications = 0.91
- 20. Immediate Advantages Fix Size Representation Similar Categories
- 21. Embeddings Image Text Music User
- 22. Learning Embeddings Without Label With Label Pre-Trained
- 23. — John R. Firth (a dominant figure in 20th century Linguistics) “You shall know a word by the company it keeps.”
- 24. Without Label
- 25. Sequence Based Embedding The earth is round and moves around the sun “All we need is a sequence of categories”
- 26. Sequence Based Embedding* The earth is round and moves around the sun • Context and Target Words • Given a word, which are the neighboring words ? • Given the neighboring words, what's the target word ? *Window size
- 29. Word2Vec Homepage Offers Finance Offers Specification … … Test Drive
- 30. With Label
- 31. Embedding layer in DNN Homepage Offers Finance Specifications … … … … CrossEntropy Loss Observed target Predicted target Embedding Layer
- 33. Category Embeddings Page-Category Embedding Brochure Reviews Finance Test Drive Specification 0.13 0.45 .. 0.21 0.67 Column Length : Embedding Size : 100 0.25 0.23 .. 0.53 0.98 0.98 0.12 .. 0.34 0.76 0.21 0.53 .. 0.23 0.87 0.87 0.24 .. 0.63 0.25
- 34. Embedding Visualization Categories related to services, warranty, review are closer Categories related to test drive activities are closer Categories vehicle information are closer
- 35. User Journey Mapping Page ‘A’ Page ‘B’ Page ‘D’Page ‘C’Visitor 1 Page ‘E’ Page ‘B’ Page ‘D’ Page ‘E’Visitor 2 0.43 0.75 0.98 0.55 0.87 0.54 0.23 0.56 0.35 0.76
- 36. Customize Embeddings Page-Category Embedding Brochure Reviews Finance Test Drive Specification 0.13 0.45 .. 0.21 0.67 Column Length : Embedding Size : 100 0.25 0.23 .. 0.53 0.98 0.98 0.12 .. 0.34 0.76 0.21 0.53 .. 0.23 0.87 0.87 0.24 .. 0.63 0.25 User Journey 0.43 0.75 0.98 0.55 0.87 Brochure Specification Finance Reviews Test Drive Time Spent Brochure 0.13 0.45 .. 0.21 0.67 0.43 Time Spent Specification 0.25 0.23 .. 0.53 0.98 0.75 … … … … … … Test Drive 0.87 0.24 .. 0.63 0.25 0.87 User Journey Embedding 0.53 0.76 0.35 0.65 0.89
- 39. Advantages of Embeddings • Finding nearest neighbours in the low dimensional space • Input features for machine learning prediction • For understanding relations between between categories
- 40. Additional Resources • https://towardsdatascience.com/neural-network-embeddings-explained-4d028e6f0526 • http://jalammar.github.io/illustrated-transformer/ • https://www.youtube.com/results?search_query=sequence+embeddings+pramod
- 41. Feedback Your feedback is important to us. Don’t forget to rate and review the sessions.