An experimental laboratory for understanding how data properties cause machine-learning algorithms to succeed, fail, and become unreliable.
Most machine-learning tutorials teach:
Dataset → Model → .fit() → Accuracy
ML-XRay asks a different question:
Why did the model behave this way?
The project systematically changes properties of data and observes how different algorithms respond.
Every investigation follows:
Understand → Derive → Implement → Predict → Experiment → Explain
Before running a major experiment, a hypothesis is recorded.
Experiments are repeated across multiple random seeds to distinguish real behavior from lucky results.
ML-XRay investigates four major dimensions:
- Geometry
- Data Quality
- Distribution
- Model Complexity
- Linear Regression
- Logistic Regression
- KNN
- Decision Tree
- K-Means
- PCA
- Random Forest
- SVM
- Gradient Boosting / XGBoost
A systematic map showing when common machine-learning algorithms become
unreliable and why. See failure_atlas/README.md.
A condensed writeup of the experiments and what they show:
report/ml_xray_report.pdf (source: report/ml_xray_report.tex).
Ten experiments so far, one card per experiment under
experiment_cards/, one notebook per experiment under
experiments/:
| ID | Algorithm(s) | Question |
|---|---|---|
| XR-001 | Linear Regression | Does noise hurt accuracy and parameter stability? |
| XR-002 | Logistic Regression | Does class overlap hurt performance and confidence? |
| XR-003 | KNN | Do irrelevant dimensions hurt distance-based classification? |
| XR-004 | Decision Tree | How sensitive is a tree to small changes in training data? |
| XR-005 | Random Forest | Does averaging many unstable trees reduce variance? |
| XR-006 | SVM | Does feature scaling change a linear SVM's boundary? |
| XR-007 | Gradient Boosting | Does label noise make more estimators counterproductive? |
| XR-008 | K-Means | Does cluster geometry (non-convex, unequal density) break it? |
| XR-009 | PCA | Does PCA discard low-variance features that are actually predictive? |
| XR-010 | All seven above | Capstone -- same stressors (noise, irrelevant features, overlap, imbalance, sample size) across every algorithm side by side. |
🚧 Experimentation in progress -- 10 experiments done, first report published, more failure conditions and algorithms (SVM kernels, XGBoost) still to come.