Evaluation of Selected Autoencoders in the Context of End-User Experience Management

Empirical research shows that a significant portion of employees regularly faces IT-related challenges in their workplace, resulting in lost productivity, customer dissatisfaction, and increased employee turnover [1]. Although the significant impact of these problems, keeping the IT administration informed about ongoing issues is a major challenge. End-User Experience Management (EUEM) aims to help IT administrators address this problem. For example, in the context of EUEM, telemetry data collected from employees' devices can help IT administrators to identify potential issues [2]. Machine learning algorithms can automatically detect anomalies in the collected telemetry data, providing IT administration with essential insights to optimize the end-user experience [2]. This paper examines the advantages and disadvantages of three different autoencoder-based algorithms identified in the literature as well-suited for detecting anomalies applied in this paper to hardware telemetry: Autoencoder (AE), Variational Autoencoder (VAE), and Deep Autoencoding Gaussian Mixture Model (DAEGMM). The results show that all three models provide anomaly detection in hardware telemetry data, though with significant differences. While the AE is the fastest Algorithm, the VAE offers the most stable results. The DAEGMM provides the best separation of endpoints into outliers and normal data points but has the most extended runtime. For all models, data aggregation has a significant potential for data reduction by aggregating the measurements over a longer time interval.