Logo Logo
Help
Contact
Switch Language to German
Baier, Stephan; Krompass, Denis; Tresp, Volker (2016): Learning Representations for Discrete Sensor Networks using Tensor Decompositions. 2016 IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI), 19-21 September 2016, Baden-Baden, Germany.
Full text not available from 'Open Access LMU'.

Abstract

With the rising number of sensing devices installed in today's and future sensor networks, there is an increasing demand for machine learning solutions performing tasks like automatic behavior detection and decision making. In particular, to classify the state of the complete sensor network, machine learning models are needed, which are capable of fusing the information from multiple sensors. In this paper we examine the use of tensor models to describe the relationship between multiple discrete sensor outputs and attendant class labels describing the overall system state. Tensor decompositions can be considered as a form of representation learning and they have been used for a variety of tasks, e. g. knowledge graph modeling and EEG data analysis. We propose a new approach for multiclass classification using tensor decompositions. As the dimensions of the tensors used in the multi-sensor classification are much higher than in traditional tasks, not all standard decomposition approaches are applicable due to scaling problems. In our experiments on real data, we show that the PARAFAC and Tensor Train decompositions work well for discrete multi-sensor fusion tasks and outperform other stateof- the-art machine learning algorithms.