Ablation

Ablation answers one question about a prediction: how much did this part of the context matter? The workflow is explicit. Choose cells, create a modified batch, and compare the same model’s outputs.

without_comments = batch.ablate(comment_positions)
full, ablated = model.predict([batch, without_comments], activation="identity")
delta = ablated - full

A large delta means the removed cells were load-bearing context for this prediction. A delta near zero means the model reached the same score without them. Compare logits; sigmoid saturation can hide real movement in probability space.

How Ablation Works

RelationalBatch.ablate marks the selected positions as padding and leaves them in place. Positions stay stable, node identities and foreign-key references need no remap, and the returned batch is a copy, so the original stays usable. Attention masks exclude padded cells entirely, which makes an ablated cell invisible to the model.

Target cells cannot be ablated; attempting it raises ValueError.

Choosing What to Ablate

Positions are yours to group. Applications typically group by table (all support tickets), by column (every amount), by entity (one specific order), or by time range. The model never guesses which ablation is meaningful, so the grouping should map to a question someone actually asked.

support_positions = [i for i, table in enumerate(tables) if table == "support_tickets"]
without_support = batch.ablate(support_positions)

Measuring Over a Dataset

AblationEvaluator runs named ablations across a set of examples and reports the mean and mean absolute prediction delta per group:

from relational_transformers import AblationEvaluator

evaluator = AblationEvaluator(
    examples,
    ablations={"support_history": support_positions, "orders": order_positions},
)
metrics = evaluator(model)
# => {'support_history_mean_delta': ..., 'support_history_mean_absolute_delta': ..., ...}

Deltas are computed on identity-activation outputs, so classification results are in logit space. The support-history ablation example runs this end to end on typed customer contexts.