The Method of Time-Series Classification Based on Spectral Analysis of Self-Organizing Map
Abstract
Purpose: Most time-series classification methods take into account the internal structure of objects, resting only upon Markov assumptions. This leads to a significant loss of the discriminant information contained in the time-series data. The purpose of this work is improving the quality of time-series classification methods by analyzing the information about the structure of SOM (Self-Organizing Map) nodes. Results: A classification method is proposed, based on spectral analysis of graphs built by a supervised learning model. A time-series classification method is developed through analysing the structure of clusters obtained as a result of mapping data on a nonlinear principal manifold by SOM algorithm. This set of nodes (clusters) is represented as a graph. A data graph and a model graph are defined, being specified on a topological ordered SOM structure. The integer score of matching between the data graph and the model graph is calculated using the method of spectral graph theory. It is experimentally proved that classification quality is higher when the method proposed in the paper is combined with the state-ofart methods, such as HMM (Hidden Markov Model) and HCRF (Hidden Conditional Random Fields), or with NPM-PGM (Nonlinear Principal Manifolds - Probabilistic Graphical Model) which we previously developed. Practical relevance: The developed method can be used for the recognition of time-series data, for example, handwriting recognition or human motion primitives recognition.Published
2015-04-01
How to Cite
Yulin, S., & Palamar, I. (2015). The Method of Time-Series Classification Based on Spectral Analysis of Self-Organizing Map. Information and Control Systems, (2), 23-29. https://doi.org/10.15217/issn1684-8853.2015.2.23
Issue
Section
Information processing and control