CSCI 1109 - M47 - Logistic regression & error metrics

Описание: In this module, we meet our first true “workhorse” classifier: logistic regression. Instead of predicting a number, we now predict the probability of a yes/no outcome—whether a tumour is malignant, a transaction is fraud, or a user will churn. You’ll see how a simple linear score gets squashed through the sigmoid curve into a probability, how that turns into a hard decision via a threshold, and why different error types (false positives vs false negatives) matter so much in real applications. Along the way we move beyond raw accuracy to confusion matrices, precision, recall, and F1, using small but impactful examples to show how metric choice and threshold choice are really about values and tradeoffs.

Explain logistic regression as a model that makes the log-odds of a binary outcome linear in the features, and interpret the role of the sigmoid function.
Train and evaluate logistic regression models in scikit-learn, including basic preprocessing like feature scaling.
Compute and interpret confusion matrices, accuracy, precision, recall, and F1, and recognize when accuracy alone is misleading (e.g., under class imbalance).
Experiment with different decision thresholds and describe the tradeoff between catching more positives (recall) and avoiding false alarms (precision) in concrete domains like medicine and fraud detection.
Reflect on how metric and threshold choices connect to real-world stakes, fairness, and accountability when probabilistic models are deployed on people.o the bias–variance tradeoff and overfitting.
Implement regularized regression models in scikit-learn, using simple validation workflows to see how the regularization strength λ changes coefficients, curves, and error metrics.
Use small, visual examples to diagnose when missing scaling or mis-tuned regularization is the main problem, and articulate how these choices can affect fairness, interpretability, and downstream decisions in domains like health, credit, and infrastructure.

Course module page: https://web.cs.dal.ca/~rudzicz/Teaching/CS...