CampusLearn
← CSC 416 Machine Learning and Data Science

Supervised Learning and Credit Risk Evaluation

Notes

Introduction to Supervised Learning Supervised learning entails training algorithms on labeled datasets where each observation comprises input features alongside an associated target ground truth. In Year 4 Computer Science curricula, emphasis transitions from basic algorithm syntax to mathematical formulation, generalization bounds, bias-variance tradeoffs, and production evaluation. Supervised tasks are broadly divided into regression, where target variables are continuous real values, and classification, where predictions map to discrete category labels. Core Classification Paradigms Classification relies on defining decision boundaries in feature space. Logistic regression establishes linear decision surfaces using the sigmoid link function to output calibrated probabilities between zero and one. Non-parametric methods like Random Forests build ensembles of decorrelated decision trees using bootstrap aggregation and random feature sub-sampling, minimizing variance while managing non-linear feature interactions without requiring extensive feature scaling. Evaluation Metrics Accuracy alone is inadequate for imbalanced real-world distributions. Machine learning practitioners instead rely on precision, recall, the F1-score, and the Area Under the Receiver Operating Characteristic Curve (AUC-ROC). In risk-averse environments, prioritizing recall minimizes false negatives, while prioritizing precision minimizes false alarms. Practical example Consider an automated credit appraisal engine developed for an app-based micro-lender operating in Kenya, such as an M-Shwari or Tala competitor. The objective is predicting whether an applicant will default on a thirty-day micro-loan. The feature matrix includes historical thirty-day M-Pesa transaction velocity in Kenya Shillings, bill payment consistency through Paybill and Till numbers, airtime top-up frequency, and geographic roaming volatility. The target variable is binary: zero denotes timely repayment, and one denotes default. Because ninety-two percent of applicants successfully repay their loans, a severe class imbalance exists. A standard classifier achieving ninety-two percent accuracy by universally predicting zero would fail completely in identifying bad loans. The data science team balances the training split using the Synthetic Minority Over-sampling Technique (SMOTE) and trains an Extreme Gradient Boosting (XGBoost) classifier. During hyperparameter optimization, the decision threshold is calibrated to optimize the F2-score rather than accuracy. This explicitly penalizes false negatives, where a defaulting borrower is misclassified as creditworthy. If the model outputs a predicted default probability of zero point three eight, and the calibrated cutoff threshold is zero point three five, the system automatically declines the disbursement, preserving capital for the digital credit provider.

Quiz

1. Why is accuracy an inadequate evaluation metric for digital micro-lending default prediction in Kenya?

2. Which evaluation metric places greater weight on reducing false negatives over false positives?

3. What primary mechanism allows Random Forests to prevent overfitting compared to individual decision trees?

Sign in to submit and securely grade your quiz.