An adaptive algorithm for dealing with data stream evolution and singularity

Abstract

The analysis of representative features for multi-source and massive data streams, from which implicit knowledge is mined, has become a research hotspot in the era of rapid development of data industry. In response to the weak adaptability of traditional algorithms, an adaptive algorithm GNG-L is proposed based on growing neural gas (GNG) for monitoring and tracking the drift and singularity of real-time data stream in non-stationary environments, which includes three mechanisms, namely weight adaptation, neuron deletion and generation. Firstly, the mechanism of weight adaptation is proposed by analyzing the changes of the local characteristics for data streams, which ensures the network topology is adjusted accurately and quickly. Secondly, the adaptive deletion mechanism removes neural nodes that are no longer updated due to the evolution of data streams. Finally, the generation mechanism is trigged when the new feature of data stream evolution needs to be described in new regions of the feature space. The proposed model has been validated based on a number of data sets, and the results show that the algorithm proposed in this paper can effectively track changes of data sets in non-stationary environments.