DAUS

Data Asset Utility Shapley

DAUS is a Shapley-value variant for auditable data-asset contribution attribution.

Core idea:

Traditional Data Shapley:
v(S) = model_performance(S)

DAUS:
v(S) = utility_score(S)

For participant i:

marginal_i(S) = v(S union {i}) - v(S)

DAUS preserves Shapley-style marginal contribution attribution. The innovation is replacing model-performance utility with auditable data-asset utility.

What DAUS Is

DAUS is:

• a data attribution algorithm
• a Shapley-value variant
• a coalition utility framework
• a contribution attribution layer

What DAUS Is Not

DAUS is not:

• a pricing engine
• a contractual settlement rule
• a distributed-ledger protocol
• a crypto-asset or coin system
• a machine-learning framework
• a commercial exchange platform

Why DAUS?

Traditional Data Shapley is powerful when v(S) can be evaluated through repeated model training or model-performance experiments. In many data-asset settings, that performance signal is unavailable, expensive, incomplete, or not the only auditable basis for attribution.

DAUS keeps the Shapley attribution structure but lets v(S) be built from auditable data-asset utility evidence such as:

• measured contribution units
• quality
• coverage
• scarcity
• sample scale
• scenario fit
• compliance usability
• measured contribution evidence
• expert-estimated evidence
• contract-agreed evidence
• simulation evidence

Model contribution evidence can be one input to the utility function when available. It is not required for DAUS to run.

Input

The canonical input is DataAssetUtilityInput:

• participant_id
• role
• measured_contribution_units
• quality_score
• coverage_score
• scarcity_score
• sample_score
• scenario_fit_score
• compliance_usability_score
• model_contribution_score optional
• expert_score optional
• contribution_source_type: measured_data, expert_estimate, contract_agreement, or simulation
• confidence_level
• evidence
• assumptions

Output

calculate_daus(...) returns DAUSShapleyResult, including:

• evaluated coalitions S
• coalition utility scores v(S)
• participant Shapley values
• contribution shares
• source types and confidence levels
• audit records
• assumptions

Mathematical Definition

For participants N, DAUS defines a coalition utility function:

v_DAUS(S) = UtilityScoreFunction(S)

The default MVP utility function is additive:

u_i = measured_contribution_units_i
    * quality_factor_i
    * coverage_factor_i
    * scarcity_factor_i
    * sample_factor_i
    * scenario_fit_factor_i
    * compliance_usability_factor_i

v_DAUS(S) = sum(u_i for i in S)

This additive form is only the MVP default case. Under additive utility, each participant's DAUS Shapley value equals its standalone utility contribution. That equality is a special case, not the full DAUS definition.

DAUS attribution uses the Shapley formula:

phi_i = sum over S subset N\{i} of
        |S|! * (|N|-|S|-1)! / |N|!
        * (v_DAUS(S union {i}) - v_DAUS(S))

Relationship With Shapley

DAUS is not a replacement for Shapley. It is a data-asset version of Shapley.

• Traditional Data Shapley: v(S) is model performance.
• DAUS: v(S) is auditable data-asset utility.

This distinction matters when model-performance experiments are unavailable, incomplete, expensive, or not the right governance basis.

Quick Start

python3 -m pip install -e .
PYTHONDONTWRITEBYTECODE=1 python3 -m pytest -q

Example

from decimal import Decimal
from daus import DataAssetUtilityInput, calculate_daus

inputs = [
    DataAssetUtilityInput(
        participant_id="source-a",
        role="data_provider",
        measured_contribution_units=Decimal("100"),
        quality_score=Decimal("95"),
        coverage_score=Decimal("90"),
        scarcity_score=Decimal("80"),
        sample_score=Decimal("100"),
        contribution_source_type="measured_data",
        confidence_level=Decimal("0.9"),
        evidence="validated contribution evidence batch",
    ),
    DataAssetUtilityInput(
        participant_id="source-b",
        role="data_provider",
        measured_contribution_units=Decimal("80"),
        quality_score=Decimal("85"),
        coverage_score=Decimal("75"),
        scarcity_score=Decimal("90"),
        sample_score=Decimal("95"),
        contribution_source_type="expert_estimate",
        confidence_level=Decimal("0.7"),
        evidence="expert-reviewed simulation evidence",
    ),
]

result = calculate_daus(inputs)
for attribution in result.participant_attributions:
    print(attribution.participant_id, attribution.shapley_value, attribution.contribution_share)

Roadmap

• Additional coalition-level utility functions with interaction terms.
• Stronger audit serialization helpers.
• Reference examples for measured, expert-estimated, and simulation-based evidence.
• Optional host-project adapters kept outside DAUS Core.