Decision-tree-based distributed learning for IoT devices

Recently, the concept of the Internet of Things (IoT) has gained recognition. Moreover, smart devices, such as smartphones and wearable devices, have become widely used because they are getting smaller and more convenient to wear. These devices are equipped with numerous, enabling them to collect large amounts of data and process it using machine learning. Most studies on sensing data processing rely on deep neural networks (DNNs). However, DNNs are not suitable for smart sensing devices due to their high computational complexity. Therefore, this paper adopts ensemble learning of decision trees, specifically random forests and gradient boosting decision trees (GBDT), to build prediction models that require fewer computational resources. Additionally, machine learning typically requires a large amount of data to train highly accurate prediction models. A straightforward way to gather diverse data is to upload the data collected by each device to a server. While this approach simplifies data collection, it introduces the need for server maintenance. Moreover, it raises concerns about the centralization of data and models. To address these issues, a decentralized machine learning algorithm that operates without a server is necessary. This paper proposes a decentralized ensemble learning algorithm for decision trees, where multiple devices communicate directly with each other to share information without relying on a server. The proposed method is applicable to situations where data are non-independent and non-identically distributed (non-IID).