Towards Efficient, Fair and Interpretable Neural Dynamic Data Valuation

Abstract In the modern data economy, data has emerged as a valuable form of property, driving the growth of data marketplaces and elevating the importance of data valuation. While existing methods have made progress in this field, they often struggle to simultaneously achieve efficiency, fairness, and interpretability. To address these challenges, we introduce a novel approach called Neural Dynamic Data Valuation (NDDV), which leverages the principles of optimal control theory. NDDV offers a comprehensive framework that accounts for the dynamic nature of data value and the intricate interactions between data points. By reformulating classical valuation criteria through the lens of stochastic optimal control, our method provides a fresh perspective on quantifying data worth. At its core, NDDV employs trajectory learning to accurately identify data value, supported by robust theoretical interpretations. We enhance the method's fairness through a data re-weighting strategy that captures unique features of individual data points, facilitating interactions between individual data states and the weighted mean-field state. Our results demonstrate NDDV's ability to unveil the dynamic process of data valuation while extending to an interpretable model. Importantly, NDDV achieves significant computational efficiency by requiring only a single training session to estimate the value of all data points. This advancement represents a substantial improvement over existing methods, potentially revolutionizing how we approach data valuation in complex, real-world scenarios.