A list of potential improvements from a recent projects at work

🥼Evaluation, explainability, experiment:​

1. Business knowledge induced feature selection​

2. Visualize training metrics to test overfit​

3. Metric/Model that balance precision recall​

4. Add model calibration to pipeline Sklearn Pipeline Notes​

5. Split data by time group instead​

6. More and Better plots​

🧹Code clean up + feature:​

1. Incorporate post grid search feature selection into the pipeline. Sklearn Pipeline Notes​

2. MLFlow​ auto experiment and output

3. feature source lineage​

4. Top “K” based on the data sample ratio instead.​

5. Unsupervised cluster labeling​

⚡Runtime Performance:​

1. Polars ​

2. multi metrics param search using Optuna ​

3. Combine near miss, random sampling,​

4. Auto EDA with Sweetviz to detect data drift​