Contributions to time series data mining towards the detection of outliers / anomalies
- DIRECTORES: Angel Conde, Usue Mori y Jose A. Lozano
- UNIVERSIDAD: UPV/EHU
Recent advances in technology have brought about major breakthroughs in data collection, enabling a large amount of data to be gathered over time. These data are often presented in the form of time series, where the observations have been recorded in an orderly fashion and are correlated in time. In recent years, a great interest has arisen in extracting meaningful and useful information from such data. The research area that focuses on this task is called time series data mining.
The time series data mining community has been devoted to solving different tasks, including the detection of outliers/anomalies. Outliers or anomalies are those observations that do not follow the expected behavior in a time series. These observations typically represent unwanted data or events of interest, and thus, detecting them is desirable because they may worsen the quality of the data or reflect interesting phenomena that the analyst intends to detect.
This thesis presents several contributions in the field of time series data mining, more specifically, on the detection of outliers or anomalies. Indeed, the contributions we present in this thesis are 1) a comprehensive review and taxonomy for the unsupervised outlier/anomaly detection techniques in time series data; 2) a novel self-supervised anomaly detection technique for whole univariate time series, aimed at detecting water leaks; and 3) a new technique for processing multivariate time series with missing values based on selective imputation, which can be applied in subsequent tasks such as outlier/anomaly