Skip to content

Model Interpretability in FoodSpec

Questions this page answers: - How can I interpret ML/DL models applied to spectra? - Which tools does FoodSpec provide for feature importance and latent factors? - How should interpretability be reported in food spectroscopy studies?

Why interpretability matters

  • Connects spectral features to chemistry (e.g., unsaturation bands).
  • Builds trust for QA/QC and regulatory contexts.
  • Helps diagnose spurious correlations or leakage.

Tools in FoodSpec

  • PCA/PLS loadings: via plot_pca_loadings and PLS loadings to identify influential wavenumbers/peaks.
  • Random Forest feature importances: inspect feature_importances_; relate top bands to chemistry.
  • Peak/ratio-based features: inherently interpretable; report definitions and effect directions.
  • Confusion matrices and per-class metrics: clarify where models struggle.
  • Residuals and calibration plots: show bias and spread in regression.

Reading loadings (PCA/PLS)

  • Scores vs loadings: scores plot shows samples in latent space; loadings plot shows which wavenumbers drive each component.
  • Example (oils): If oil A vs B separate along PC1 and loadings have strong positive contributions near ~1655 cm⁻¹ (unsaturation), that band is characteristic of oil A. Use Spectroscopy basics for vibrational assignments.
  • Pair qualitative plots with metrics (e.g., silhouette/between-within on scores) to quantify separation.

Practical examples

from foodspec.chemometrics.pca import run_pca
from foodspec.viz import plot_pca_loadings

pca, res = run_pca(X, n_components=2)
plot_pca_loadings(res.loadings, wavenumbers, components=(1, 2))

# RF feature importances
rf = make_classifier("rf", n_estimators=200, random_state=42)
rf.fit(X, y)
importances = rf.feature_importances_

For PLS-DA/PLS regression, examine loading vectors and VIP-like interpretations; ensure preprocessing is consistent across train/test.

Visual examples

Random Forest feature importances Figure: RF feature importances on synthetic spectra. Top bands correspond to simulated peaks; relate them to known chemical bands. Flat or noisy importances may indicate weak signal or overfitting.

PLS loadings example Figure: PLS loadings for the first component, showing influential wavenumbers in a calibration task. Loadings sign shows positive/negative association with the component; peaks in loadings point to bands driving separation or prediction.

Reporting guidance

  • Report top contributing bands/ratios and their chemical meaning.
  • Include loadings/importance plots in supplementary material; summarize key drivers in main text.
  • Avoid over-interpretation of noisy or collinear features; cross-check with domain knowledge.
  • Pair interpretability with metrics: a model that performs poorly but shows plausible bands still needs improved performance.

When Results Cannot Be Trusted

⚠️ Red flags for feature importance and model interpretability:

  1. Feature importance from overfit model (high training accuracy but low test accuracy)
  2. Importance scores reflect noise fit during overfitting, not true signal
  3. Top features may be coincidental collinearities in training data

  4. Fix: Report feature importance from test set or cross-validation; use permutation importance; validate feature relevance independently

  5. Correlations with labels misinterpreted as feature importance

  6. High correlation doesn't mean feature is used by model or is important for prediction
  7. Spurious correlations (batch confounding, data leakage) can rank high

  8. Fix: Use model-specific importance (coefficients, tree splits, SHAP values); don't rely on univariate correlations alone

  9. Loadings/coefficients from collinear features not interpreted cautiously

  10. High collinearity (VIF > 10) makes coefficients unstable; small data changes flip signs
  11. Ranking features by coefficient magnitude misleads when collinearity present

  12. Fix: Check VIF or condition number; use regularization (Ridge/Lasso) to stabilize coefficients; use permutation importance

  13. Confidence intervals not reported for feature importance

  14. Point estimates of importance vary across CV folds; single-fold importance is unstable
  15. Without CI, importance scores appear more certain than they are

  16. Fix: Compute importance across CV folds; report mean ± SD or 95% CI; visualize variability

  17. Statistical significance confused with practical importance (p < 0.05 band ≠ important for prediction)

  18. ANOVA or t-test on feature values tests association, not model utility
  19. Significant association doesn't guarantee predictive power

  20. Fix: Report model-based importance; tie to prediction accuracy; validate with independent data

  21. Interpreting absence of feature as "not important" (feature has importance = 0)

  22. Feature with zero importance may be redundant with other features, not truly unimportant
  23. Or importance measure is insensitive to that feature's contribution

  24. Fix: Use multiple importance measures (permutation, SHAP, coefficients); test feature removal effect on metrics

  25. Chemically implausible features ranked as most important without flagging as suspicious

  26. Food spectra are noisy; random noise peaks can rank highly if model overfit
  27. Chemical bands should make physical/chemical sense

  28. Fix: Cross-check with domain expertise; validate top features in independent data; avoid over-interpreting noisy regions

  29. Feature importance from single model type only (only PLS loadings, no RF permutation importance)

  30. Different models and importance measures can rank features differently
  31. Robust conclusions require agreement across multiple approaches
  32. Fix: Report importance from multiple models/methods; highlight consensus features; note disagreements

See also