Wrapper-Based Federated Feature Selection for IoT Environments

Novel Internet of Things (IoT) applications have emerged as enabling technologies for the smart city initiative. IoT devices collect or produce huge multi-modal data that is either processed on the edge or sent to a central cloud for processing. The collected data sets are pre-processed by methods known as "feature selection", to remove redundant, irrelevant, or noisy features. Feature selection will help with improving the results achieved by the learning method as well as reducing the computational complexity of the model. The goal is to select the most informative features of data and only transmit the selected features to the edge/cloud servers for further processing. This leads to smaller costs for data transmission to the servers. In this paper, a novel wrapper-based federated feature selection (FFS) algorithm is proposed, where IoT devices collaborate to select the most informative features without sharing their local data sets. The proposed FFS algorithm uses binary gravitational search algorithm (BGSA) in a federated and collaborative manner to select a small enough subset of informative attributes and provide an improved trade-off between communication cost and learning accuracy. Our experimental results on three data sets including MNIST, Fashion-MNIST, and MAV demonstrate that the proposed BGSAFFS method can in average remove more than 50% of features without losing information. The obtained results prove the effectiveness of the proposed method in achieving a good trade-off between accuracy and communication cost in comparison to other state-of-the-art feature selection methods as well as a no-feature selection baseline.