Extending the Growing Neural Gas Classifier for Context Recognition
R. Mayrhofer - Extending the Growing Neural Gas Classifier for Context Recognition - Proceedings of EUROCAST 2007: 11th International Conference on Computer Aided Systems Theory, Las Palmas, Spanien, 2007, pp. 920-927
Context awareness is one of the building blocks of many applications in pervasive computing. Recognizing the current context of a user or device, that is, the situation in which some action happens, often requires dealing with data from different sensors, and thus different domains. The Growing Neural Gas algorithm is a classification algorithm
especially designed for un-supervised learning of unknown input distributions; a variation, the Lifelong Growing Neural Gas (LLGNG), is well suited
for arbitrary long periods of learning, as its internal parameters are self-adaptive. These features are ideal for automatically classifying
sensor data to recognize user or device context. However, as most classification algorithms, in its standard form it is only suitable for numerical input data. Many sensors which are available on current information appliances are nominal or ordinal in type, making their use difficult. Additionally, the automatically created clusters are
usually too fine-grained to distinguish user-context on an application level. This paper presents general and heuristic extensions to the LLGNG classifier which allow its direct application for context recognition.
On a real-world data set with two months of heterogeneous data from different sensors, the extended LLGNG classifier compares favorably
to k-means and SOM classifiers.