DoorDash Delivery Duration Prediction Take-Home

Overview

In this exercise, you will use your machine learning experience to solve a straightforward but challenging prediction problem. The exercise contains two parts:

1. Building a machine learning model for a prediction task
2. Writing an application to make predictions using that model

In Part 1, you will be evaluated on:

• Performance on the test set
• Feature engineering choices, including features used and encoding of features
• Data processing
• Choice of models used
• Description of model performance and insights and observations from the model

In Part 2, you will be evaluated on:

• Performance of the application
• Adherence to common software engineering patterns (unit tests, modular code, etc.)
• Ability to make educated trade-offs based on the given constraints

Note: For this part, you must use a production-ready language like Python, Java, C++, Scala, Ruby, etc.

Problem Description

When a consumer places an order on DoorDash, the expected time of delivery is shown. It is very important for DoorDash to get this right, as it has a big impact on consumer experience. In this exercise, you will build a model to predict the estimated time taken for a delivery and write an application that can make these predictions.

Concretely, for a given delivery you must predict the total delivery duration in seconds, i.e. the time taken from:

• Start: the time the consumer submits the order (created_at)
• End: when the order will be delivered to the consumer (actual_delivery_time)

To help with this, the following files are provided:

• historical_data.csv: table of historical deliveries
• data_to_predict.json: JSON list of deliveries that you must predict on (for the second part)
• data_description.txt: description of all columns in historical_data.csv and details of data_to_predict.json

Requirements

Part 1

• Build a model to predict the total delivery duration in seconds as defined above. Feel free to generate additional features from the given data to improve model performance.
• Explain:
  • the model(s) used
  • how you evaluated your model performance on the historical data
  • any data processing you performed on the data
  • feature engineering choices you made
  • any other information you would like to share about your modeling approach
• Based on the findings from the model, list recommendations to reduce delivery time.

Part 2

• Write an application that accepts data from the JSON file (data_to_predict.json), uses the model to make a prediction for each delivery in the JSON file, and writes out predictions to a new tab-separated file with columns:
  • delivery_id
  • predicted_delivery_seconds
• Your predictions on this test data set will be evaluated using RMSE (Root Mean Squared Error), and your score must exceed a baseline set for the task.
• Note: For this part, you must use a production-ready language like Python, Java, C++, Scala, Ruby, etc.

Deliverables

Submit exactly one PDF file that includes:

• A write-up explaining your model, choices made, and discussion on the questions above
• The code used for Part 1
• The application code for Part 2, including instructions for running it (dependencies, packages required, etc.)
• The predictions for data_to_predict.json embedded in the PDF as part of the application output

Your application code must be runnable from the command line with data_to_predict.json passed as input. Any code included in the PDF must be executable so it can regenerate the required outputs.

Notes

• We expect the exercise to take 5-6 hours in total, but feel free to spend as much time as you like on it.
• Feel free to use any open source packages for the task.
• Thank you for your hard work! Please let us know if you have any questions. Good luck!