Citibike Classification Challenge
Malte Bonart
October 16, 2019
This work and the underlying source code is available on GitHub.
This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
moderately imbalanced data
| Day Pass/ Three Day Pass | Annual Membership | |
|---|---|---|
| Customer (11%) | Subscriber (89%) | |
| 12.00$ / day | 169.00$ / year | |
| max 30 min | max 45 min | |
| 4.00$ / 15 min | 2.50$ / 15 min |
customers bike for a longer period
Trips with a duration > 2 hours and < 20 seconds have been removed from the analysis (~0.3%). Customer ride on average 1441, Subscribers 721 seconds.
anomaly in the age distribution for customers
6% of all trips from customers have an age value of 49.
gender is mostly unknown for customers
more customers on the weekend
0:Sunday - 6:Saturday.
more clients during summer months
more subscribers during rush hour
model features
- tripduration
- starttime, stoptime
- hour, weekday, month
- start (lat,lon), end (lat,lon)
- age, close_to_fifty
- gender
- NYC neighbourhoods (via geocoding API)
the baseline performs well - nearest neighbour classifier is worse
The baseline was constructed by classifying all trips with unknown gender as customers. Nearest neighbour classification is based on time, start location, end location and tripduration.
classification does not outperform baseline
| features | dimensions | logistic regression (f-score) |
|---|---|---|
| tripduration | 1 | 0.15 |
| + gender | 3 | 0.70 |
| + age | 5 | 0.71 |
| + time | 45 | 0.71 |
| + area | 173 | 0.72 |
The training is based on a random sample of n=5000000 trips, due to resource and time constraints.
random forrest does not improve the classification
comparison of trip duration with biking
| biking vs. | driving | driving (traffic) | transit |
|---|---|---|---|
| Average differences | -107 | -105 | 209 |
| Median differences | 2 | -9 | 263 |
| Biking faster | 50% | 48% | 83% |
Based on a n=2000 random sample of trips, collected with the GoogleMaps Directions API. Wilcoxon signed-rank test and t-test for pairs are both significant. biking faster | biking slower
4726 injured bikers, 10 deaths in NYC car accidents 2018
driver inattenion
failure to yield right of way
confusion of bicyclist
traffic control disregarded
passing or lane usage improber
Top reasons for NYC motor vehicle collisions where at least one biker was injured.
Source: http://tiny.cc/af3kez