Equitable Valuation of Crowdsensing for Machine Learning via Game Theory

In the era of mobile Internet, it has became easier to obtain personal data through crowdsensing platforms, which promotes the development of data-driven machine learning. A fundamental challenge is how to quantify the value of data provided by each worker. In this paper, we use the powerful tool of game theory called Shapely value to solve this challenge. Shapley value is a classic concept in game theory and can satisfy the equitable valuation of data. However, the calculation of Shapley value is exponentially related to the number of workers. Worse still, in the deep learning model, the time cost of retraining the model and evaluating the contribution of each worker's data to the model is unacceptable. Therefore, we propose two algorithms based on Monte Carlo and batch gradient descent to approximate Shapley value in machine learning and deep learning. We take K-fold validation as the benchmark, and prove that our proposed algorithms can reduce the time overhead while ensuring lower error in the experiment. Finally, we find that it can provide better insight into the labor value of each worker in specific learning tasks.