Building A Predictive Model – An Example of Product Recommendation

Slides presented by Alex to the NYC Predictive Analytics group on April 1, 2010 (no joke!)

view/download pdf

 

Presentation Outline
1. Building a Predictive Model by Alex Lin, Senior Architect, Intelligent Mining
2. Outline
– Predictive modeling methodology
– k-Nearest Neighbor (kNN) algorithm
– Singular value decomposition (SVD) method for dimensionality reduction
– Using a synthetic data set to test and improve your model
– Experiment and results
3. The Business Problem: Design product recommender solution that will increase revenue.
4. How Do We Increase Revenue?
5. Example: Is this recommendation effective?
6. What am I going to do?
7. Predictive Model Framework 1
8. What Data to Use?
9. Predictive Model Framework 2
10. What Features to Use?
11. Data Representation / Features
12. Data Normalization 1
13. Data Normalization 2
14. Predictive Model Framework version 3
15. Which Algorithm?
16. k-Nearest Neighbor (kNN)
17. Similarity Measure for kNN
18. k-Nearest Neighbor (kNN) Item feature space
19. Predictive Model Version 1
20. Cosine Similarity
21. Singular Value Decomposition (SVD)
22. Reduced SVD
23. SVD Factor Interpretation
24. SVD Dimensionality Reduction
25. Missing values
26. Singular Value Decomposition (SVD)
27. Predictive Model version 2
28. Synthetic Data Set – Why do we use it?
29. Synthetic Data Set – 16 latent factors synthetic e-commerce data set
30. Synthetic Data Set – Item property, User preference, Purchase Likelihood
31, Synthetic Data Set – final slide
32. Experiment Setup – Each model (Random / kNN / SVD+kNN) will generate top 20 recommendations for each item
33. Experimental Result – kNN vs. Random (Control)
34. Experimental Result – Precision % of SVD+kNN Recall
35. Experimental Result – Quality of SVD+kNN Quality
36. Experimental Result – The effect of using Cart data Precision
37. Experimental Result – The effect of using Cart data Quality
38. Outline
39. References
40. Thank you

  • Twitter
  • del.icio.us
  • HackerNews
  • Facebook
  • RSS
  • LinkedIn
  • Tumblr
  • email
  • Reddit