A distributed evolutionary fuzzy system-based method for the fusion of descriptive emerging patterns in data streams

Abstract

Nowadays the amount of networks of devices and sensors, such as smart homes or smart cities, is rapidly increasing. Each of these devices generates massive amounts of data on a continuous basis where an interpretable description of its state is interesting for the experts. This knowledge can be extracted by means of emerging pattern mining techniques. In fact, it can be extracted locally on each device and joined together afterwards in order to obtain a global vision of the system without transferring any data. However, the traditional massive data processing frameworks are focused on the extraction of this global model, which produces huge amounts of data transfers throughout the network. This paper proposes a distributed method based on evolutionary fuzzy systems for both the extraction and subsequent fusion of descriptive emerging patterns in data streams from different sources of the same kind. First, an evolutionary algorithm following an informed approach for efficient data processing is presented for the extraction of emerging patterns on the data stream generated by each device, in order to obtain a local model for each stream. Then, several fusion methods are proposed for the aggregation of these patterns in order to extract the global model. A wide experimental study has been carried out to analyse the suitability of the evolutionary algorithm for the extraction of high-quality emerging patterns and its capacity to deal with concept drift. Finally, the quality of the proposed fusion methods is also analysed.