Value functions and normalizers

Coalition utility is distinct from cooperative-game value (Shapley/Banzhaf attributions). Value functions map model outputs to scalar payoffs; normalizers reshape aggregated attributions for display only.

Utility families

Plugin key

Paper utility

Mechanism

tfidf_cosine

U3

Hashed token rows → frozen-corpus TF-IDF → cosine similarity

embedding_cosine

U1 / U2

Cosine similarity of static or contextual embeddings

embedding_euclidean

U4

1 / (1 + L2 distance) between embeddings

logprob

sequence logprob

Sum of per-token log probabilities (stub when unavailable)

Frozen TF-IDF

Unlike MLLM-Shap 0.x, nlp-shap freezes IDF weights from the full run corpus at explain start. Scoring reuses those weights for every coalition comparison.

Normalizers

Applied after ShapleyAggregator (or Banzhaf) aggregation:

  • Identity — raw attributions

  • AbsSum — divide by sum of absolute values

  • PowerShift — shift to non-negative, power transform, scale to sum 1

  • MinMax — scale to [0, 1], then scale to sum 1

Archive rows store pre-normalization utilities and attributions so reanalyze can swap value functions or normalizers without new model calls.