Combining the Self-Organizing Map and K-Means Clustering for On-Line Classification of Sensor Data

Abstract

Many devices, like mobile phones, use contextual profiles like “in the car” or “in a meeting” to quickly switch between behaviors. Achieving automatic context detection, usually by analysis of small hardware sensors, is a fundamental problem in human-computer interaction. However, mapping the sensor data to a context is a difficult problem involving near real-time classification and training of patterns out of noisy sensor signals. This paper proposes an adaptive approach that uses a Kohonen Self-Organizing Map, augmented with on-line k-means clustering for classification of the incoming sensor data. Overwriting of prototypes on the map, especially during the untangling phase of the Self-Organizing Map, is avoided by a refined k-means clustering of labeled input vectors.