Vehicle rental services like electric scooters and B-cycles are very popular in Austin and are thus a lucrative and competitive business in this city. Understanding when B-cycles in particular are in highest demand and the motivations as to why people use them is useful information to companies to improve the user experience and gain a competitive edge in this profitable venture.
What it does
We explore daily usage patterns and develop a model that uses very limited data to predict how many bicycles will be checked out at a given kiosk depending on the time of the day.
How we built it
We classified specific rides' purposes roughly by calculating the expected ride time using latlong coordinates and comparing that to the actual ride time. Rides with significantly longer times were inferred to be leisure rides (as shown by some following the path of the river), rides close to the predicted time were inferred to be utility rides (to work/class, etc.), and rides with almost 0 distance (returning to the same stop) were inferred to be short-term errands.
We transformed the provided B-cycle data for a particular kiosk to record the number of checkouts for each hour on particular dates. We then used weekday/weekend and time of day features to predict in which frequency class (low, med, high) the checkout rate at a given hour falls into. We achieved 65% cross-validated accuracy using our very limited feature set.
We also compared the distribution of B-cycle traffic following the introduction of the system to UT Austin's students.
Challenges we ran into
The map visualization tools did not have native support for some of the aspects of the data we wanted to highlight. Additionally, the dataset did not have very many features very relevant to the classification task, but with careful selection of classification models, we achieved a nontrivial accuracy.
Accomplishments that we're proud of
Decent classifier accuracy with limited feature set Identification of anomalies in the data and assessing the causes (e.g. ACL festival)
What we learned
How to collaboratively analyze data using Azure
What's next for Understanding Austin B-Cycle Usage
In the future, we would try to better understand the usage with more robust data and understand how data was collected (random selection, etc.)