A Machine Learning Approach for Estimating Missing Data in Nonstationary Environments

The assumption with most learning techniques and algorithms developed this far is that data is complete and will continuously be available or that data conforms to a stationary distribution. Real world applications are often streaming and their measurements are often sampled after an extended period of time thereby giving rise to the formation of a time series. Measurement devices assigned to measure nonstationary quantities are subject to failure. When a failure occurs, the process of approximating missing values becomes difficulty. The process of approximating missing values in such dynamic environments is further exacerbated by the chaotic and unpredictable nature of the evolving data. Typical examples include stock market, network intrusion detection systems and seismic waves. To estimate missing values with traditional statistical methods can lead to bias when applied to environments that evolve with time. This paper introduces an ensemble of regressors approach to approximate missing data in online nonstationary data. The approach learns new concepts incrementally and the current learnt concept is then used to approximate the missing values.