Pattern Matching in Sequential Data Using Reservoir Projections

Abstract

A relevant problem on data science is to define an efficient and reliable algorithm for finding specific patterns in a given signal. This type of problems often appears in medical applications, biophysical systems, complex systems, financial analysis, and several other domains. Here, we introduce a new model based in the ability of Recurrent Neural Networks (RNNs) for modelling time series. The technique encodes temporal information of the reference signal and the given query in a feature space. This encoding is done using a RNN. In the feature space, we apply similarity techniques for analysing differences among the projected points. The proposed method presents advantages with respect of state of art, it can produce good results using less computational costs. We discuss the proposal over three benchmark datasets.