A comparative trade-off analysis on accuracy and efficiency for federated learning in demand forecasting

Abstract

Federated Learning (FL) is an emerging learning mechanism that can achieve accuracy, efficiency, and privacy concurrently, which could be particularly useful in scenarios where large volumes of sensitive data are involved, such as demand forecasting in inventory management. In this paper, we consider a dataset composing of sales records of multiple products across ten different Walmart stores in the USA, and conduct a comparative study of centralized learning, distributed learning, and FL, focusing on the accuracy of predicting future demand of certain products and the required volume of data for transmission by the multi-layer perceptron model. Our results demonstrate that, with the same number of training rounds, FL achieves competitive accuracy in predicting future demands across the selected product categories while significantly reducing data transmission compared to other learning approaches, highlighting the efficiency and practicality of FL. In addition, we compare the performances of FL approaches with different combinations of store-level data from various regions, and examine the trade-off analysis in terms of accuracy and transmission efficiency under different scenarios.