Article Preview
TopIntroduction
Various data mining methods have been presented in our world for many years. They play an important role in our lives, in both the business and scientific domains. The continuous growth of data collected in the world increases every day, even reaching a size on the order of zettabytes (where a ZB is 1021 B) (Chen et al., 2014). It is estimated to reach from 168 (Khattak et al., 2019) to 175 (Alnoukari, 2020) ZB by 2025. Moreover, countless devices that form the Internet of Things (IoT) produce quintillions of bytes (1018 B) (Dobre & Xhafa, 2014) of data every day as well. In fact, 500 ZB of data was generated from such IoT sources in 2020 (Mirza et al., 2021) due to their number. The number of such devices increases, and the data cannot be stored in one place because of hardware and network limitations. On the other hand, multiple independent data centers collect such data for different institutions and companies. Therefore, new algorithms and methods are required to enable new findings in this large and distributed amount of data. Distributed data mining (DDM) has been gaining importance in recent years, with various methods implemented (Gan et al., 2017). Depending on the domain of study, these approaches are called Collaborative Systems (e.g., Zhou et al., 2010) or DDM methods and frameworks. Such processing methods should provide both quality results and processing performance while preserving the privacy of transferred data.
Much research on applications implementing various algorithms working in a Spark (Zaharia et al., 2012) cluster has been done. However, such processing requires a specific cluster environment and does not provide control over the processing and data transfer between nodes. Because of this lack of control, which will be discussed later, custom approaches and frameworks are usually implemented utilizing well-known worker-node communication architectures, such as centralized or peer-to-peer (P2P). On account of this expansion, for several years a great effort has been devoted to the study of distributed clustering, classification, and other data mining methods. One of the first examples is presented by Aouad et al. (2007), and a newer solution is presented by Bendechache and Kechadi (2015). Distributed classification methods have been studied by Navia-Vázquez et al. (2006) and Forero et al. (2010) and have even been studied explicitly in the network security domain (Hu et al., 2013). It is worth noting that recently, several authors (Forero et al., 2010; Xu et al., 2015; Jia et al., 2016) have pointed out the importance of preserving privacy in DDM processing. This processing concept, however, is not limited to the previously mentioned types of algorithms; it also applies to other methods, like frequent itemset mining, etc. Nevertheless, for the purposes of this work, we do not focus on these topics.