Tutorials

The tutorials move from SVM fundamentals to a complete, defensible MISTIC analysis. Code fragments use the same public API and workflows as the notebooks distributed with the package.

• Support vector machine foundations
  • From a separating plane to maximum margin
  • Lagrange multipliers and the classification dual
  • The kernel trick
  • Support vector regression
  • One-class SVM
  • Connecting the three models
  • References
• Feature selection and perturbation analysis
  • Why perturb features?
  • Feature groups
  • Choosing the sample set
  • Reading perturbation values
• Deriving MISTIC’s feature-ranking metrics
  • Notation and a frozen-model perturbation
  • Decision perturbation
  • Probability perturbation
  • The dual-objective criterion
  • Why frozen coefficients?
  • Combined ranking
  • One-class contribution ranking
  • Choosing and reporting a weight
  • References
• Forward and backward feature selection
  • Shared setup
  • Forward selection
  • Backward selection
  • Knee selection and final fitting
  • Comparing the directions
• Interpretations and explanations
  • Three complementary questions
  • Feature rank
  • Support vectors
  • Perturbation explanations
  • Gradients
  • Boundary counterfactuals
  • Integrated gradients
  • Triangulating evidence
  • Reference
• Visualizing results and explanations
  • Selection curves
  • Boundary-counterfactual views
  • Integrated-gradient heatmaps
  • Summary plots
  • Interaction views
  • Reporting checklist
• Blind predictions: best practices
  • Freeze the workflow
  • Generate predictions once
  • Create an auditable export
  • Evaluate the right quantities
  • Explain without tuning
  • Check for drift
• Synthetic benchmark against scikit-learn models
  • Experimental design
  • Shared metrics
  • MISTIC forward selection
  • Plain RBF SVC
  • SVM-RFE followed by RBF SVC
  • Random forest and boosted trees
  • Compare the results
  • Measured paired comparison
  • Repeated benchmark

Suggested paths

New to support vector machines

Read the tutorials in order. The first two explain the geometry and why perturbation supplies useful model-specific evidence.

Experienced SVM user

Begin with Deriving MISTIC’s feature-ranking metrics and Forward and backward feature selection, then compare the explanation types in Interpretations and explanations.

Preparing a final analysis

Review Visualizing results and explanations and Blind predictions: best practices before reporting results.