SELECTING TIME SERIES NESTING PARAMETERS TO DETECT DATA ANOMALIES IN BEHAVIORAL ANALYTICS SYSTEMS

Purpose of the article: this article investigates the impact of embedding algorithm selection on the performance of anomaly detection in behavioral analytics systems for scalar time series. Recommendations for algorithm selection are provided. Method: the research methodology is based on Takens-Mane theorems for reconstructing attractors of dynamical systems from scalar observables and methods of topological data analysis. Result: the study experimentally establishes that the effectiveness of anomaly detection in behavioral analytics systems using machine learning and topological data analysis significantly depends on the quality of the reconstructed space from the scalar observable. Recommendations are developed for selecting embedding parameters – the dimension of the reconstructed space and the time delay – depending on the embedding method used. Scientific novelty: unlike previous studies on the optimal selection of time series embedding parameters in predicting the behavior (functioning) of dynamical systems, this is the first scientific work that directly investigates the application of popular algorithms for selecting time series embedding parameters to improve the accuracy of anomaly detection by behavioral analytics systems using topological descriptors.