The project aims to optimize the placement of Gokada's delivery drivers (referred to as "pilots") to minimize unfulfilled delivery requests. To achieve this, we will leverage causal inference, a statistical framework that goes beyond correlation to uncover cause-and-effect relationships.
The project utilizes two datasets:
Completed Orders: This dataset provides information on successful deliveries, including the origin, destination, and timestamps of each trip. This data is crucial for understanding the spatial and temporal patterns of completed deliveries, which can serve as a baseline for comparison with unfulfilled requests.
Delivery Requests: This dataset includes both completed and unfulfilled requests, along with details about the driver's location and whether they accepted or rejected the request. This data is essential for identifying factors that might contribute to a driver accepting or rejecting a request, such as distance to the pickup location, time of day, or other contextual factors.
By analyzing these datasets together, we can gain insights into the factors that influence the success or failure of a delivery. For example, we can investigate whether drivers in certain locations are more likely to reject requests, or whether certain times of day are associated with higher rates of unfulfilled deliveries.
Causal learning is a critical component of this project. It involves constructing a causal graph, a visual representation of the hypothesized causal relationships between variables. This graph will be based on domain knowledge and expert opinion, and it will be refined using causal discovery algorithms applied to the observational data.
The causal graph will help us identify potential confounding variables, which are factors that might influence both the driver's decision to accept a request and the outcome of the delivery. By controlling for these confounders, we can isolate the true causal effect of driver location on order fulfillment.
Once the causal graph is established, we will use do-calculus, a set of mathematical rules for manipulating causal expressions, to estimate the causal effects of different interventions. For example, we can simulate the effect of relocating drivers to different areas or changing their working hours to see how these changes would impact the number of unfulfilled requests.
By combining causal inference with machine learning techniques, we can develop predictive models that can forecast the outcome of deliveries under different scenarios. These models can then be used to optimize driver placement and other operational decisions, ultimately leading to a more efficient and reliable delivery service.