On the Usage of Gaussian Process for Efficient Data Valuation

2025-06-04

AI Summary

This paper addresses data valuation—the quantification of individual data points’ influence on model performance in machine learning. We propose an efficient, interpretable valuation framework based on Gaussian processes (GPs). Methodologically, we introduce GPs for the first time to model submodel utility, integrating Bayesian inference with a canonical decomposition of utility functions to enable rapid, incremental estimation of sample contributions. Our key contributions are: (1) the first interpretable decomposition framework for data valuation, explicitly attributing influence to individual samples and their interactions; and (2) leveraging GPs to jointly ensure theoretical rigor—via probabilistic modeling of utility uncertainty—and computational efficiency—through closed-form posterior updates. Experiments demonstrate that our approach accelerates valuation by over an order of magnitude compared to full retraining, while achieving high rank correlation with Leave-One-Out and other baselines in influence ranking. The framework thus enables near real-time, principled assessment of data value.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Relational Probabilistic Models
• Data Mining & Knowledge Management: Representing, Reasoning, and Using Provenance, Trust

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets
• Security and Privacy: Data transparency and provenance

Abstract

In machine learning, knowing the impact of a given datum on model training is a fundamental task referred to as Data Valuation. Building on previous works from the literature, we have designed a novel canonical decomposition allowing practitioners to analyze any data valuation method as the combination of two parts: a utility function that captures characteristics from a given model and an aggregation procedure that merges such information. We also propose to use Gaussian Processes as a means to easily access the utility function on ``sub-models'', which are models trained on a subset of the training set. The strength of our approach stems from both its theoretical grounding in Bayesian theory, and its practical reach, by enabling fast estimation of valuations thanks to efficient update formulae.

Problem

• Efficiently valuing data impact on model training
• Decomposing data valuation into utility and aggregation
• Using Gaussian Processes for fast valuation estimation

Innovation

• Novel canonical decomposition for data valuation
• Gaussian Processes for utility function access
• Efficient update formulae for fast estimation