Interpretability
Interpretability
Interpretability Analysis
RQ4: What trade-offs exist between predictive performance and interpretability across representations?
Interpretability is scored 1–5 (higher = more interpretable). Score 5 means a domain expert can audit every prediction from named feature coefficients. Score 1 means predictions rely on opaque internal representations.
Performance vs Interpretability Trade-off
Each point represents the best model for that representation
Logistic Regression Coefficients
Feature importances for interpretable representations
DistilBERT Attention Weights
Average last-layer attention weights per class (indicative, not faithful attribution)
Attention ≠ attribution. These show which tokens the model attends to, not necessarily what drives predictions.
