Dissimilarity Metric Score Estimation for Time Series with Missing Values

Abstract

Missing values in time series data pose significant challenges for data modeling and further analyses. Interpolation methods are often used to fill in the missing values in the data, however, they may cause extra computational complexities and may make the analysis not suitable for real-time operations. Hereby, considering this, this paper focuses on the problem of estimating the dissimilarity metric score for time series data with missing values without interpolating the data. Hereby, we propose an approach to estimate the dissimilarity metric scores without utilizing the imputation methods. Our pro- posed algorithm utilizes a basic, but effective statistical model composed of statistical moments of a time series window to estimate the dissimilarity score of the respective window without applying the interpolation methods. Correlation between the proposed approach scores and the Euclidean dissimilarity metric scores on a benchmark dataset is computed for the most commonly used interpolation methods. To observe the dissimilarity values, several different missing value rates were selected to randomly erase the samples with that ratio from the data. The experimental results show that our proposed method provides comparable correlation results with some dissimilarity measures especially with spline interpolation by creating a correlation coefficient value of 0.819. Hence, the application of such a basic approach to estimating the dissimilarity values without applying interpolation or dissimilarity calculations to observe the time-varying data behavior can be used to reduce the computational complexity in real-time applications.