· Xiaojing Yang · Explainability and Responsible AI · 4 min read
中文Ablation vs Shapley: Why Coalition Context Matters
Ablation is simple and useful, but Shapley values ask a broader question by averaging marginal contribution across many coalition contexts.
Core idea
Ablation usually measures what happens when a unit is removed from one reference dataset. Shapley asks what happens when the unit is added across many possible contexts.
This note is part of my series Training Data Attribution for NLP and LLM Research. The series is written as both a research notebook and an interview preparation path: each article should help me explain the idea clearly, connect it to my thesis, and identify what would become future PhD work.
Guiding question: Why is a single leave-one-out ablation not the same as Shapley attribution?
Intuition
Leave-one-out is like asking how important an ingredient is after the soup is already cooked with everything. Shapley asks how much the ingredient contributes across many recipes. Both are useful, but they answer different questions.
In NLP and LLM research, this matters because model behaviour is deeply shaped by data mixture. A model may be fluent because of broad web text, domain-accurate because of specialised documents, safer because of curated instruction data, or biased because of repeated patterns in a subset of the corpus. Training-data attribution gives us language for asking these questions systematically instead of only saying “the data matters”.
Formal lens
Leave-one-out often estimates v(N) - v(N \ {i}). Data addition estimates v({i}) - v(empty) or v(B union {i}) - v(B). Shapley averages v(S union {i}) - v(S) over many S. The difference matters whenever interactions are non-additive.
The important discipline is to define the attribution setup before interpreting the score:
| Design choice | Question to answer |
|---|---|
| Attribution unit | What receives credit: source, group, document, example, or token? |
| Utility function | Which behaviour is being explained: quality, terminology, style, factuality, or safety? |
| Intervention | Are we adding, deleting, reweighting, correcting, or retraining? |
| Estimator | Is the score exact, sampled, gradient-based, surrogate-based, or heuristic? |
| Uncertainty | How stable is the score across seeds, samples, metrics, and evaluation sets? |
NLP / LLM example
A Bokmål-like group might look unimportant in leave-one-out if the full dataset already contains many similar examples. But it may have a large Shapley value if it strongly improves small coalitions where stylistic evidence is otherwise missing.
This is why I do not want to treat attribution as a generic interpretability topic. For my profile, the natural connection is multilingual and domain-specific NLP: low-resource settings, technical terminology, written-standard variation, and evaluation beyond one headline metric.
Connection to my thesis
In my thesis narrative, training-data attribution is useful because it turns a vague data question into an experimental design:
- define interpretable data units;
- define the model behaviour to explain;
- compare controlled data coalitions or interventions;
- estimate contribution;
- report uncertainty and limitations;
- decide what evidence is strong enough to support a causal-style claim.
That structure helps me avoid overclaiming. A score is not automatically a causal explanation. It is a measurement produced by a specific setup.
What I have done, understand, and would extend
| Level | Status |
|---|---|
| Already completed / thesis-ready | Group-level attribution, coalition thinking, metric-based utilities, cautious interpretation, random baselines, bootstrap-style reliability checks. |
| I understand but may not fully implement yet | Instance-level gradient attribution, influence functions, TracIn, Monte Carlo Shapley, surrogate/datamodel approximations. |
| Strong PhD extension | Hierarchical attribution, intervention-based validation, factuality/style-specific utilities, scalable attribution for LLM data mixtures. |
Interview answer
I would say ablation is a useful baseline and sanity check, but Shapley gives a more context-aware estimate of contribution. In my project, this distinction matters because data groups can complement or substitute for one another, especially in low-resource domain MT.
References and reading path
- Lloyd Shapley, A Value for n-Person Games.
- Ghorbani and Zou, Data Shapley: Equitable Valuation of Data for Machine Learning.
- Koh and Liang, Understanding Black-box Predictions via Influence Functions.
- Pruthi et al., Estimating Training Data Influence by Tracing Gradient Descent.
- Ilyas et al., Datamodels: Predicting Predictions from Training Data.
- Rei et al., COMET: A Neural Framework for MT Evaluation.