Decision-Tree-Based Distributed Learning for IoT Devices

Abstract Recently, people have known the concept of the Internet of Things (IoT). Moreover, they have widely used smart devices, such as smartphones and wearable devices, because they are getting smaller and easier to wear. These devices have many sensors so that they can collect many data and process these data using machine learning. Most studies processing sensing data use deep neural network (DNN). However, DNN is not suitable for smart sensing devices because of its high computational complexity. Therefore, this paper adopts ensemble learning of decision trees, especially random forest and gradient boost decision trees (GBDT), to obtain prediction models without many computational resources. In addition, machine learning requires a large amount of data to train highly accurate prediction models. An easy way to gather various data is by uploading data obtained by each device to a server. Although this scheme can collect data easily, it requires the server to be maintained. Further, there is the challenge of the concentration of data and models. To solve these problems, a decentralized machine learning algorithm that works without a server should be developed. This paper proposes a decentralized ensemble learning algorithm for decision trees, with which multiple devices communicate directly with each other to share information without a server. The proposed method can be applied to situations where data are non-independent and identically distributed (non-IID).