ClassTrim

ClassTrim is a framework to recommend move method refactoring using Non-dominated Sorting Genetic Algorithm-II/III (NSGA-II / NSGA-III) to minimize the number of classes that exceed predefined metric thresholds. After the algorithm produces a Pareto front of candidate refactorings, a single knee solution (the one with the best overall trade-off across objectives) is automatically selected as the recommended refactoring.

Raw experiment data doi: 10.17632/8b3fd45kp6.1

Demo

Run ClassTrim plugin on smelly-demo

Quick start

1. Prerequisites

• JDK 17+
• Maven 3.8+

2. Configure environment

• Copy classtrim-core/src/main/resources/config.properties.example to classtrim-core/src/main/resources/config.properties.
• Edit the new config.properties and set values for your environment:
  • mavenRepo: absolute path to your local Maven repository
  • datasetRoot: absolute path to the datasets root folder
  • outputFolder: output folder for results (optional; default output)
  • projectCacheFolder: folder for serialized project cache (optional; default .project)
  • notifyUrl: optional webhook for notifications

3. Entry point

   • org.classtrim.Main run the move method refactoring algorithm on each project.
   • org.classtrim.NSGAII, org.classtrim.NSGAIII, or org.classtrim.MOEAD run it once.
   • org.classtrim.baseline.BaselineRefactor run baseline comparison experiment.

4. Build and test modules

• Run all modules: mvn test
• Run core only: mvn -pl classtrim-core test
• Run CLI only: mvn -pl classtrim-cli test

5. Build IntelliJ plugin

• Package plugin zip: mvn -pl classtrim-plugin -am package
• Output: classtrim-plugin/target/classtrim-plugin-1.0-SNAPSHOT-idea-plugin.zip

Module extraction (in progress)

This repository now includes extraction scaffolding for library-style usage:

• classtrim-core/: Maven module packaging the current core code and new injectable APIs under org.classtrim.core.*
• classtrim-cli/: Maven module with org.classtrim.cli.RunNsgaiii example entrypoint using the extracted service API
• classtrim-plugin/: IntelliJ IDEA plugin scaffold wired to run core analysis and show notifications

New core API entrypoints:

• org.classtrim.core.analyzer.ProjectAnalyzer / StandardProjectAnalyzer
• org.classtrim.core.engine.RefactoringEngine / NSGAIIIRefactoringEngine
• org.classtrim.core.service.ClassTrimService
• org.classtrim.core.model.ProjectSource / BinaryPathProjectSource
• org.classtrim.core.repository.ProjectRepository (in-memory and filesystem implementations)

How I calculate metrics and some common question are in wiki.

Baseline comparison (baseline/)

This folder contains results from baseline tools for comparison, including:

• Runtime screenshots produced during baseline runs
• Suggested outputs generated by each baseline tool (e.g., *.tsv summaries)

Typical subfolders include JDeodorant/, JMove/, HMove/ and REsolution/, each organized by project-version. These provide reference suggestions and artifacts to compare against this project's NSGA-based refactoring outputs.

Code features

1. Follow unix philosophy.
2. Uses Lombok to simplify data classes (getters/setters, constructors)
3. Uses Java 8 Optional to handle nulls elegantly
4. Emphasizes immutability; returns unmodified collections
5. Caches computed objects to improve performance
6. Keeps dependencies minimal
7. Automatically records experiment results and summaries (summary may be improved by using a database)
8. Includes unit tests

Knee-point selection

NSGA only outputs a set of non-dominated (Pareto) solutions; ClassTrim's RefactorOutput then selects the knee of that front as the single recommended refactoring, following

Das, I., "On characterizing the 'knee' of the Pareto curve based on Normal-Boundary Intersection", Structural Optimization 18, 107–115 (1999). doi:10.1007/BF01195985

How the knee is chosen (org.classtrim.core.util.KneeSelection):

1. Compute the ideal (utopian) point — the per-objective minimum across the front.
2. Find the anchor solutions — the front member minimizing each objective alone (the pay-off rows).
3. Build the CHIM hyperplane (Convex Hull of Individual Minima) through the anchors.
4. Pick the front point of maximum bulge — farthest from the CHIM on the side of the ideal point. This is the point where each objective trade-off is most balanced. The construction is invariant to (global) rescaling of the objectives and generalizes to any number of objectives.

Selection uses only the 3 real objectives (WMC, CBO, RFC); the NSGA-III guiding objectives are search heuristics and are intentionally excluded. Degenerate fronts whose anchors collapse to a point fall back to choosing the solution closest to the ideal point.

Output: the knee is placed first, so diff-01 / metrics-01 are the recommended refactoring, and an explicit KNEE.csv marks its objective values. The full front is still preserved in FUN.csv / VAR.csv.

Unit tests: KneeSelectionTest

Human evaluation

The quality of ClassTrim's Move-Method suggestions is assessed through a blinded expert questionnaire (full plan in human-evaluation-plan.md, results in human-eval/ and human-eval/report.md). Executive summary:

• 5 tools compared — ours (ClassTrim) vs HMove, JDeodorant, JMove, REsolution.
• 100 suggestions (20 per tool) sampled with a fixed seed; ClassTrim is deliberately sampled at an average level (standout feature-envy cases excluded) to avoid cherry-picking.
• Blinding: suggestions are shuffled into neutral IDs S001–S100; reviewers never see the tool, metrics, or internal bookkeeping.
• 20 expert reviewers (senior Java developer personas) judge items on a fixed questionnaire (RQ1-Q5): acceptance, usefulness (1-5), diagnostic value (worth inspecting?), rejection reasons, and free-text comments.
• Each item is reviewed by exactly 2 experts (balanced 20x10 assignment) to measure inter-rater agreement / Cohen's kappa.
• Analysis: acceptance rates and Q2/Q3 means per tool, ClassTrim-vs-baseline deltas, the "detects what not where" hypothesis (Q3 vs Q1), and cross-checks vs output/baseline-overlap.md.

Thanks

• ASM
• JMetal

License

MIT