Several methods exist in classification literature to quantify the similarity between two time series data sets. Applications of these methods range from the traditional Euclidean type metric to the more advanced Dynamic Time Warping metric. Most of these adequately address structural similarity but fail in meeting goals outside it. For example, a tool that could be excellent to identify the seasonal similarity between two time series vectors might prove inadequate in the presence of outliers. In this paper, we have proposed a unifying measure for binary classification that performed well while embracing several aspects of dissimilarity. This statistic is gaining prominence in various fields, such as geology and finance, and is crucial in time series database formation and clustering studies.
A sizable amount of current literature on online drift detection tools thrive on unrealistic parametric strictures such as normality or on non-parametric methods whose power performance is questionable. Using minimal realistic assumptions such as unimodality, we have strived to proffer an alternative, through a novel application of Bernstein's inequality. Simulations from such parametric densities as Beta and Logitnormal as well as real-data analyses demonstrate this new method's superiority over similar techniques relying on bounds, such as Hoeffding's. Improvements are apparent in terms of higher power, efficient sample sizes, and sensitivity to parameter values.
In this paper, we propose a novel approach for classifying incoming continuous data under a non-stationary environment. A class of estimators termed ”Stochastic Learning Weak Estimators” has been generalized to include continuous time sampling and countable state categories. The method is founded on non-stationary Markov chain techniques and is useful in diverse applications such as consumer behavior analysis, email spam classification or understanding drug effectiveness. In terms of tracking the true state probabilities, these weak estimators consistently outperform traditional competitors like maximum likelihood estimates. Only one user defined parameter is necessary and the method is free of subjective ”moving window” type algorithms. We have conducted extensive simulations and real data analyses for classification purposes.
In computer science, stochastic processes, and industrial engineering, stationarity is often taken to imply a stable, predictable flow of events and non-stationarity, consequently, a departure from such a flow. Efficient detection and accurate estimation of non-stationarity are crucial in understanding the evolution of the governing dynamics. Pragmatic considerations include protecting human lives and property in the context of devastating processes such as earthquakes or hurricanes. Cumulative Sum (CUSUM) charting, the prevalent technique to weed out such non-stationarities, suffers from assumptions on a priori knowledge of the pre and post-change process parameters and constructs such as time discretization. In this paper, we have proposed two new ways in which non-stationarity may enter an evolving system -an easily detectable way, which we term strong corruption, where the post-change probability distribution is deterministically governed, and an imperceptible way which we term hard detection, where the post-change distribution is a probabilistic mixture of several densities. In addition, by combining the ordinary and switched trend of incoming observations, we develop a new trend ratio statistic in order to detect whether a stationary environment has changed. Surveying a variety of distance metrics, we examine several parametric and non-parametric options in addition to the established CUSUM and find that the trend ratio statistic performs better under the especially difficult scenarios of hard detection. Simulations (both from deterministic and mixed inter-event time densities), sensitivity-specificity type analyses, and estimated time of change distributions enable us to track the ideal detection candidate under various non-stationarities. Applications on two real data sets sampled from volcanology and weather science demonstrate how the estimated change points are in agreement with those obtained in some of our previous works, using different methods. Incidentally, this study sheds light on the inverse nature of dependence between the Hawaiian volcanoes Kilauea and Mauna Loa and demonstrates how inhabitants of the now-restless Kilauea may be relocated to Mauna Loa to minimize the loss of lives and moving costs.