A Novel Approach Based on Matrix Factorization for Recovering Missing Time Series Sensor Data

Missing data are unavoidable in wireless sensor networks, due to weather, equipment failure or some uncontrollable human factors. This results in data sets incompleteness, which decreases the performance of related applications greatly. Although many related researches have been carried out to recover missing data, most of them are not effective enough when dealing with multiple sensor data series simultaneously. Moreover, few studies have made the best of the characteristics of these sensor data. In this paper, we have taken up this challenge by employing an efficient approach based on matrix factorization (MF) with any weights. First, to improve the accuracy of missing data recovery, we make full use of the interior patterns of each data series and the correlation information across multiple sources to constrain the objective function of MF. Then, to reduce the time complexity of solving the MF parameters, an element-wise alternating least squares algorithm is proposed. Further, we apply a series of caching strategies to optimize the learning process of parameters. Simulation results on real-world data sets demonstrate that the approach proposed is a promising approach to ensure the data quality in wireless sensor networks.