Driver Behavioral Segmentation
Behavioral segmentation of drivers for a ride-hailing platform aimed at reducing incentive burn
Problem Statement
An Asian ride-hailing company was facing high burn due to incentives given to drivers and wanted to develop personalized incentive schemes that are designed to reduce the overall burden on the company while having low impact on driver churn. A segmentation scheme based on driver behavioral patterns was required.
Challenges
- The need to separate out drivers by different business verticals and analyze them separately requiring multiple segmentation schemes each with its own nuances
- Extracting only relevant data from a large dump of data sources
- Obtaining a data-driven understanding of why driver segments behave a certain way in order to derive insights around game-ability of incentives
Solution
Summary
I, along with a team, developed an unsupervised machine learning solution to identify clusters of drivers with common behaviors. Data around trips such as trip frequency, trip length, geographic concentration, etc. and driver activity on the platform such as total active duration, trip acceptance rate, active hours of day, etc. was used to identify patterns. A k-means clustering approach was followed for clustering.
Approach
- The company maintained a large database of thousands of tables on BigQuery which needed scouring to identify relevant data for the project
- Data aggregations, transformation and cleaning was done to create a master modeling database
- Feature engineering and feature selection techniques were used to reduce the dimensionality and convert the data into a purely numerical dataset with categorical encoding
- Drivers were separated by business verticals and k-means clustering was used to identify segments
- Segments were profiled and discussions were held with the marketing team to help design the right experiments for each segment
Key Takeaways
- Behavioral patterns such as infrequent long distance trips, frequent short distance trips, high peak hour activity, etc. were identified
- Segments that were causing the most incentive burn were identified and segment profiles helped in estimating the optimal incentive schemes for such segments
Have a similar challenge?
Let's discuss how we can develop solutions for your specific use case.