Approximate Quantile Computation over Sensor Networks

Sensor networks have been deployed in various environments, from battle field surveillance to weather monitoring. The amount of data generated by the sensors can be large. One way to analyze such large data set is to capture the essential statistics of the data. Thus the quantile computation in the large scale sensor network becomes an important but challenging problem. The data may be widely distributed, e.g., there may be thousands of sensors. In addition, the memory and bandwidth among sensors could be quite limited. Most previous quantile computation methods assume that the data is either stored or streaming in a centralized site, which could not be directly applied in the sensor environment. In this paper, we propose a novel algorithm to compute the quantile for sensor network data, which dynamically adapts to the memory limitations. Moreover, since sensors may update their values at any time, an incremental maintenance algorithm is developed to reduce the number of times that a global recomputation is needed upon updates. The performance and complexity of our algorithms are analyzed both theoretically and empirically on various large data sets, which demonstrate the high promise of our method.