A New Symbolic Time Series Representation Method Based on Data Fuzzification

Time series classification is an essential data processing task that relies on assigning class labels to sequences of temporal data. A fundamental component of any time series classification method is data representation. There exist several approaches to that task ranging from straightforward sequence distance-based methods to neural networks. We focus on symbolic time series representation-based methods. The literature of the domain repeatedly underlines their flexibility and good classification quality. We propose a new approach to convert numeric time series into symbolic ones based on fuzzy clustering. The goal is to reduce noise in the data. The proposed method utilizes cluster membership values to determine symbols that characterize the time series. The new approach was tested in an empirical procedure to validate its correctness while achieving satisfying results.