23 — Automated Smell Detection

Driving question: How should a detector act when the smell concept itself has fuzzy boundaries?

Learning objectives

• Implement threshold/rule-based smell detection.
• Compare supervised and unsupervised approaches.
• Handle class imbalance and project-specific thresholds.
• Evaluate agreement with human judgments and downstream outcomes.

From rule to prediction

A detector can be framed as:

• rule evaluation;
• anomaly detection;
• binary/multilabel classification;
• ranking/prioritization;
• change-risk prediction.

Ranking may be more useful than binary labeling when developers have limited remediation capacity.

Evaluation beyond classification

Ask whether the detector:

• finds actionable issues;
• ranks important instances early;
• generalizes across projects;
• remains stable across versions;
• helps developers make better decisions;
• produces explanations that correspond to the actual rule/evidence.

Design / research exercise

Build a smell-ranking function that combines smell intensity and change frequency. Compare the top-10 ranking with a pure threshold detector and inspect which list seems more actionable.

Suggested reading

• Empirical literature on code smell detection and prioritization.