Predict last-minute give-up rentals to reduce operating loss for WeMo Scooter

Project Details


Fall 2019


Will Kuan, Beverly Lin, Adam Yu, PeiPei Chen



WeMo Scooter is an e-scooter sharing company that provides urban transportation solutions by convenient, economic, free and interesting riding experiences. WeMo offers an app as an interface for users to find, reserve, rent, and turn on/off an e-scooter easily.

Business Problem: The rental process is done by its mobile app, allowing users to reserve the e-scooter remotely. Once users click the reserve button on a specific e-scooter, there will be a ten-minute buffer time for them to physically find and rent it. During the ten-minute buffer, the e-scooter becomes unavailable for others. If users eventually give up the renting process in any circumstances, the unavailability of e-scooters causes inconvenience for nearby potential users and hence operating loss for WeMo. In fact, in our initial data exploration, we are inferred that 16.5% of users are give-up rentals, which supports the fact that it’s imperative to act on this phenomenon. We expect to predict for WeMo’s operation team to offer a more flexible buffer time.

Data Mining Goal: We would like to predict give-up rentals, those users who reserve but do not rent, immediately after a reservation is made using the behavioral data. We will need a light-weight and cost-efficient model because the predictions need to be made in a short time.

Data: Two weeks of data from Sep.1 to Sep.14 in 2019 is extracted from the database. Data is transformed from a nested structure to an event count table and is later processed through a variable selection and derivation to generate the final dataset for modeling. The final dataset includes 16 variables and has 133684 records.

Method & Evaluation: We applied classification tree, random forest, logistic regression, and stepwise regression as our methods. And we choose sensitivity (among all the actual give-up rentals, the percentage we predict correctly) as our metric of interest since its an imbalanced data. And as a result, the undersampling approach is also adopted. Finally, we come up with an ensemble model of Random Forest and Stepwise Regression considering the stability of the model.

Recommendation: In short, we have three recommendations for WeMo: 1. Adopt undersampling and ensemble approach for model building. 2. Investigate the reasons for dropped rentals via pop-up surveys for high-risk users 3. Improve results by adding meaningful derived variables (such as distance between users and scooters) as well as trying out other approaches such as lasso regression and different ensemble combinations.

Application Area: