Machine Learning & Explainability for Data Science

Tackle a binary classification problem to predict if a candidate is seeking a job change, using a Kaggle dataset, and launch a Jupyter notebook with conda and pip installs.

Ensure reliable, accurate data collection from SQL databases with QA processes and data dictionaries; load, clean, and merge training and test data, and select meaningful features for a binary target.

Perform exploratory data analysis to validate data quality, ensure a record per person with enrollee ids, remove duplicates, handle missing labels, and examine distributions (training hours, education levels) for explainability.

Focus on feature engineering by expanding features with internal and external data, bucketed training hours, and third-party data like census data and Glassdoor to boost job-search prediction.

Choose the gradient boosted decision tree model Lightgbm for binary classification, leveraging its robust handling of class imbalance and null values without imputation.

Drop nonessential columns, convert categorical fields to pandas dummies, convert the target to 0/1, assess correlations, and set up an 80/20 train-test split to prepare x and y.

Run model training with a stratified shuffle split creating five splits, use gridsearchcv over a param grid, and apply cross-validation to optimize hyperparameters and output a performance metric.

Evaluate model performance on a 20% holdout using the roc auc score, compare training and test scores, and note feature importances for data-centric improvements.

Learn how SHAP values quantify each feature's contribution to predictions, compare with importances, and visualize directional effects to explain model decisions for data science applications.