Skip to content

Next Steps and Active Learning

Active Learning Exercise

To solidify your understanding of building and tuning robust modeling pipelines, we recommend you try the following tasks. These exercises are designed to give you hands-on practice with the iterative and exploratory nature of data science.

  1. Incorporate Categorical Features: Our model only used numerical features.

    • Task: Add the zipcode feature to your feature set X. Since zipcode is a categorical variable, you will need to add a preprocessing step for it in your pipeline. Use the OneHotEncoder or TargetEncoder from scikit-learn to handle this. You will need to use ColumnTransformer to apply different preprocessing steps (scaling for numerical, one-hot encoding for categorical) to different columns.
    • Analysis: Does adding this location data improve your model's final performance?

  2. Experiment with a Different Model: We demonstrated how to swap a DecisionTreeRegressor for an XGBRegressor.

    • Task: Build and tune a new pipeline using a KNeighborsRegressor. Research its key hyperparameters (like n_neighbors) and create a new parameter grid for GridSearchCV.
    • Analysis: How does the performance and training time of the K-Nearest Neighbors model compare to the Decision Tree and XGBoost models?

  3. Optimize for a Different Metric: We used Mean Absolute Error as our scoring metric.

    • Task: Re-run one of your GridSearchCV experiments, but this time set scoring='neg_root_mean_squared_error'.
    • Analysis: Does optimizing for RMSE result in a different set of best hyperparameters? Why might a business choose to optimize for RMSE over MAE? (Hint: How does RMSE penalize large errors?)

Lesson Summary

In this lesson, you have significantly upgraded your modeling capabilities, moving from a manual, analytical workflow to a robust, automated process ready for real-world application.

You learned that a Pipeline is the professional standard for encapsulating preprocessing and modeling steps, ensuring consistency and preventing critical errors like data leakage. We replaced the unreliable, single train-test split with k-fold cross-validation to build confidence in our model's performance. You saw how to systematically optimize hyperparameters using GridSearchCV, and witnessed the power of ensemble models like XGBoost to improve predictive accuracy.

Most importantly, you now have a reusable framework for building, evaluating, and tuning models in a systematic and reproducible way—a core competency for any data science practitioner.