Estimated Delivery Model

3. Model#

Features engineering#

Training data#

• We can use historical deliveries for the last 6 months as training data. Historical deliveries include delivery data and actual total delivery time, store data, order data, customers data, location, and parking data.

Model#

Gradient Boosted Decision Tree#

• Gradient Boosted Decision Tree sample

• How do Gradient Boosted Decision Trees work?

  • Step 1: Given historical delivery, the model first calculates the average delivery time. This value will be used as a baseline.

  • Step 2: The model measures the residual (error) between prediction and actual delivery time.

    $Error = Actual Delivery Time - Estimated Delivery Time$

  • Step 3: Next, we build the decision tree to predict the residuals. In other words, every leaf will contain a prediction for residual values.

  • Step 4: Next we predict using all the trees. The new predictions will be used to construct predictions for delivery time using this formula:

    $Estimated Delivery Time = Average \_delivery\_time + learning\_rate * residuals$

  • Step 5: Given the new estimated delivery time, the model then computes the new residuals. The new values will then be used to build new decision trees in step 3.

  • Step 6: Repeat steps 3-5 until we reach the number of iterations that we defined in our hyperparameter.

• One problem with optimizing RMSE is that it penalizes similarly between under-estimate prediction and over-estimate prediction. Have a look at the table below. Note that both models use boosted decision trees.

• Although Model 1 and Model 2 have the same RMSE error, model1 overestimates delivery time which prevents customers from making orders. Model2 underestimates the delivery time and might cause customers to be unhappy.