Preprocessing methods for context extraction from multivariate wireless sensors data — An evaluation

Real time extraction of user context like current location, position, physiological parameters, actions etc. from sensor data is a challenging problem. Probabilistic machine learning algorithms or their variations are used as classifier for accomplishing this task. Sensor data in raw form is voluminous, redundant, noisy and real valued. Most classification schemes cannot perform well on such data. Raw data is preprocessed to counter these inherent problems and transform the data to become more suitable for context extraction algorithms. Horizontal and vertical data reductions are main preprocessing methods. In this work, these methods of data reduction are emphasized and studied. These include imputation, smoothing, feature extraction and data reduction in sequential order. Representative methods applied in conventional data mining are evaluated. A benchmark sensor dataset of ambient, object and wearable sensors is used for the study. Data apart from being real valued also has scores of missing values. The “Mode of Locomotion” context of observed person is extracted by applying Bayesian Belief Network based classifier post processing. Horizontal data reduction has not found to be suitable for sensor data.