Insight_Smartan_WorkoutRecommender

Transcript

1. Smartan A Smart Workout Recommender Judith Yue Li Insight Data Science
2. Workout Routines for Fitness My Workout Routine for Back Pain
3. Workout Routines for Fitness Back stretching Core strength My Workout Routine for Back Pain
4. Matching Routines by Keywords Workout Routines Database myfitnesspal
5. Matching Routines by Keywords Workout Routines Database Female beginner “Female Beginner Routines” “Perfect Woman Routines” myfitnesspal
6. Matching Routines by Keywords Workout Routines Database Female beginner “Female Beginner Routines” “Perfect Woman Routines” Arm focus Abs focus myfitnesspal
7. Smartan: Matching Exercise Patterns ____ ____ ____ User current workout
8. ____ ____ ____ ____ ____ ____ ____ ____ ____ Smartan Database 2000 + Routines 1300 + Exercises Smartan: Matching Exercise Patterns ____ ____ ____ User current workout
9. Algorithm: Non-negative Matrix Factorization •  Unsupervised learning for feature extraction •  Similar to topic modeling (LDA) •  Similarity is measured by Pearson correlation
10. Algorithm: Non-negative Matrix Factorization •  Unsupervised learning for feature extraction •  Similar to topic modeling (LDA) •  Similarity is measured by Pearson correlation
11. Algorithm: Non-negative Matrix Factorization •  Unsupervised learning for feature extraction •  Similar to topic modeling (LDA) •  Similarity is measured by Pearson correlation
12. Algorithm Validation Output from Smartan.info 53% Correlation
13. Algorithm Validation
14. About Judith •  PhD Candidate, Stanford University •  Computational Geoscience •  “Real-time Monitoring of CO2 Leakage” •  Signal Processing (Kalman Filter) University Singers Choir Dubs Fan
15. Algorithm Validation 1 1 1 1 1 1 1 1 A total of ~12000 data points 2000 + Routine 1300+ Exercises 1 1 1 1 1 1 1
16. •  30 features •  Project gradient by Scikit-Learn (5.3) •  Alternate least square by nimfa (0.85) 1 1 1 1 1 1 1 Take out 20% as testing data A total of ~12000 data points 2000 + Routine 1300+ Exercises 80% as training data Algorithm Validation 1 1 1 1 RMSE = 1 N (r true −r NMF )2 ∑ 1
17. Comparison of Two Female Beginner Routines Female Beginner Routine Perfect Woman Routine 0 0.5 1 1.5 2 2.5 Cardio/Weight Loss Core/Abs Strength Arm/Bicep Glute Building Score
18. Recommendations based on Workout Themes Walking Elliptical Plank Air Bike Crunches Barbell Curl Drag Curl Lying Leg Curls/Leg Extension Cardio Core strength Arm Glute 30 Workout Themes ____ ____ ____ 2000 Workout Routines ____ ____ ____ ____ ____ ____ •  Unsupervised learning •  Similar to topic modeling (LDA) 17 26 55 43 20 40 45 34 Routine A User A
19. NMF: Sparse representation
20. NMF: Non negative score