FAIRTIQ - Machine learning for travel mode detection, at AMLD 2020
0 likes142 views
The document discusses machine learning for travel mode detection. It describes how FAIRTIQ works to detect travel modes using input time windows of user location data and statistical features like speed. It summarizes that initially a random forest model was used which is quick to implement and robust with low training samples, but this could confuse slow movements like walks with funiculars. The document hints that neural network models may provide advantages over the random forest approach for travel mode detection.
FAIRTIQ - Machine learning for travel mode detection, at AMLD 2020
- 1. Machine learning for travel mode detection Roman Prokofyev Co-founder and Chief Scientist AMLD 2020, EPFL. 28.01.2020
- 2. How FAIRTIQ works? 2 R m1 Lausanne Renens VD EPFL
- 3. Travel mode detection? 3 Lausanne Renens VD EPFL TRAIN SUBWAY
- 4. 4 Labelling the data: manually Lausanne Renens VD EPFL
- 5. 5 Input: sliding time windows Time ! Min # of locations in a window ! Min duration of a window ~ 3 min.
- 6. 6 Features: mostly statistical ! Speed min/max/percentiles/std/... ! Distances ! Location accuracy Accelerometer-based transportation mode detection on smartphones, SenSys 13
- 7. 7 Why percentiles/medians? Outliers Renens VD EPFL Lausanne
- 8. 8 Features based on domain data
- 9. 9 First model: random forest ! Quick to implement and to inspect ! Robust with a low # of training samples
- 10. 10 What happened with forgotten checkouts
- 11. 11 Funiculars Jan Jun Nov ! Slow movement can be confused with walks ! Feature that measures distance to the funi route % empty journeys
- 12. Whats next?
- 13. 13 The problem with Random Forest model w1 wnInput w2 w3 Journey RF model Output o1 ono2 o3
- 14. 14 The advantages of NN models w1 wnInput w2 w3 Journey Output o1 ono2 o3 Sequential model
- 15. Thank you for your attention linkedin.com/in/rprokofyev