iST-MRF: Interactive Spatio-Temporal Probabilistic Models for Sensor Networks

Streams of sensor measurements arise from twitter, mobile phone networks, internet traffic, road traffic, home automation systems, seismic motion and sea level to mention just a few. The interactive exploration and modelling of such measurements from multiple sensors induces the need for algorithms that are capable of processing the data as it becomes available and that can quickly provide partial results based on the data seen so far. Beside these requirements, the algorithm should capture the inherent spatio-temporal dependency structure within sensor data and allow a predictive analysis on arbitrary subsets of sensors. Spatio-Temporal Markov Random Fields (ST-MRF) are known to meet the requirements of modelling the dynamics of sensor networks. ST-MRF tracks the empirical distribution of each sensor and concurrently updates a Maximum Likelihood estimate of the underlying distribution. In the first part of this paper, we show how to train such models in an online fashion in order to perform near-instant updates to the model and provide them to the user. In the second part, we present iST-MRF, a free open source software for interactive modelling and analyzing data from sensor networks, which implements and visualizes ST-MRF. It guarantees high performance computations for offline models and concurrent learning and prediction for online models. We present two exemplary applications of iST-MRF to sensor network data, namely modelling a network of temperature sensors and a location prediction task.