Since context describes the current status of a user in terms of location, activity, and surroundings, there is a need to model the context’s attributes efficiently by reducing the dimensionality space. Li et al. proposed a hybrid process combining activity classification with a subspace clustering approach in order to analyze high dimensional and heterogeneous synthetic context data. Subspace clustering merges clusters from different dimensions based on the data objects they share; thus, complexity time can grow significantly when the number of dimensions is much higher. In our approach, which was empirically tested with realworld sensor data, we aimed to learn these relations by unsupervised techniques such as PCA and deep learning, which is linear to the amount of contextual features and allows diversity in the learned patterns.