Anomaly Detection

[JohT]

🚀 Features

✨Anomaly Detection

This update adds a powerful new component to the code analysis pipeline: an unsupervised machine learning module that automatically detects anomalies across your codebase—with zero manual tuning.
This pipeline generates over 10 insightful plots and CSV reports per code unit type (Java Artifacts, Java Types, Java Packages, and TypeScript Modules). Visualizations range from** intuitive 2D feature plots** that highlight unusual code patterns, to advanced statistical analyses and clustering techniques.
One standout addition: an Isolation Forest model that excels at spotting the “needle in the haystack”—uncovering rare and unexpected feature combinations that could indicate deeper issues or innovation opportunities.
This feature not only flags anomalies—it explains them, helping you understand why certain parts of your code stand out.

• Introduce anomaly detection pipeline with tuned clustering
• Use UMAP for node embedding visualization
• Attach domains folder to report compilations to run them automatically.
• Add unsupervised anomaly detection to the fully automated pipeline
• Reuse common visualization functions for more unified plots
• Add code language to anomaly detection reports and plots

📓 Example for anomaly detection with Isolation Forest

• Use Isolation Forest for anomaly detection
• Use SHAP to explain the detected anomalies
• Use optuna to tune isolation forest and random forest proxy
• Add more SHAP explanation plots and embedding anomalies
• Add lists and plots for most normal non-anomalies
• Update anomaly detection exploration notebooks

📓 Example for hyper-parameter tuning of node embeddings

• Add example for hyperparameter tuning of node embeddings.
• Use optuna for hyperparameter tuning
• Add hyper-parameter tuned Leiden community detection

⚙️ Optimization

• Add elementId to every node embedding: Now, all nodes in the node embedding results are clearly identifiable.
• Move property reset queries into distinct folder
• Explore non null properties for a given label: This is a very helpful query for troubleshooting missing data. You can immediately spot missing properties or properties, that are not present in all nodes.

📖 Documentation

• Update documentation