Anomaly Detection - as with all Machine Learning terms, many people have heard of it and would like to use it to solve diverse problems in their companies, from system monitoring to fraud detection. But is it always the right approach? And how easy it is to actually set up the entire anomaly detection pipeline in practice? Most importantly, how do you ensure that the insights you gain can be forwarded to the users in real-time? In this talk we will take you with us through the journey we had at Swisscom, establishing an anomaly detection service that is used for various data sources - thousands of time series spanning many topics, from telecommunication to business processes. We will present some pitfalls we faced, how we resolved them and what we found to be the right balance between ML and heuristics. Additionally, we will demonstrate the importance of Root Cause Analysis to achieve actionable insights and make sense of anomalies in the context of complex systems. You will also hear details of our ML pipeline based on open-source technologies, best practices of MLOps and AutoML for operating in a production setting, and the mysterious forecasting ML algorithms we apply. In the end, how soon can you get started with Anomaly Detection yourself and which problems to use it for; that is the question.

Download the slides for this talk.Download ( PDF, 5533.94 MB)