Seminar - Machine Learning for Time Series Data: An Applied Perspective
On moodle and TUMonline this seminar is called Seminar - Efficient Inference and Large-Scale Machine Learning (IN2107, IN4874).
The pre-course meeting with information regarding the course format, possible topics etc. is scheduled for Feb 7, 2022 4pm on zoom (Passcode: 281710).
- Pre-course meeting: Feb 7, 2022 4pm - slides
- Kick-off meeting: Apr 28, 2022 2pm - slides
- Final presentations: Jul 26 & 27, 2022
This seminar is intended for Master's students only. You should have attended (and passed) the Machine Learning lecture (IN2064). Having attended Machine Learning for Graphs and Sequential Data (IN2323) or other advanced ML/DL lectures (IN2332, IN2346, etc.) is a plus.
Machine Learning is used in many companies for core business functions where numerical time series (e.g., demand for an online retailer, staff attendance, telemetry data from cloud resources) are ubiquitous. Accordingly, the interest in adapting and innovating machine learning methods that handle time series (i.e. data that exhibit temporal dependencies, going beyond the ubiquitous IID assumption) natively is growing. Common ML tasks that are addressed using time series data include forecasting, anomaly detection, time series classification, representation learning, and missing value imputation. Furthermore, challenges common to the entire ML community like causality or interpretability often need non-trivial adaptations to the time series setting.
In this seminar we will select and discuss topics of current research in machine learning for time series. This seminar will let students get acquainted with current machine learning research, let them explore new fields and ideas and let them analyze and criticize recent publications.
To do so, we will offer students to either go broad or deep: for the broad option, students will select a group of research papers (2-5) from a curated list, covering different approaches and perspectives on a particular task, which they should carefully read, analyze, and critically evaluate. Starting from these they should explore the surrounding literature and summarize their findings, criticism, and research ideas in a 4-page paper (double column). For the deep option, students will select a single paper which has code available and dive more deeply into the particular method, and explore it both theoretically as well as practically, e.g., by comparing it to another, potentially missing baseline or making modifications (e.g., making a point estimator probabilistic). For either the broad or the deep scenario, all students will prepare 25-minute presentations and present their work during a block seminar at the end of the semester.
This seminar will be organized by two external researchers, Jan Gasthaus and Tim Januschowski.
Jan Gasthaus is a Principal Machine Learning Scientist in the Amazon Web Services AI Labs, working mainly on time series forecasting and large-scale probabilistic machine learning. He is passionate about developing novel machine learning solutions for addressing challenging business problems with scalable machine learning systems, all the way from scientific ideation to productization. Prior to joining Amazon, Jan obtained a BS in Cognitive Science from the University of Osnabrueck, an MS in Intelligent Systems from UCL, and a PhD from the Gatsby Unit, UCL, focusing on Nonparametric Bayesian methods for sequence data.
Tim Januschowski is a Sr. Machine Learning Science Manager at Zalando where he leads the article pricing group that owns the algorithmic discount component for Zalando SE. Prior to this, he worked at AWS AI where he helped launch 5 AI services together with his teams, including forecasting and anomaly detection. Tim’s personal interests in forecasting span applications, system, algorithm and modeling aspects and the downstream mathematical programming problems. He studied Mathematics at TU Berlin, IMPA, Rio de Janeiro, and Zuse-Institute Berlin and holds a PhD from University College Cork.
- Point or Probabilistic Forecasting
- Multivariate Methods (for forecasting or anomaly detection)
- Hierarchical Forecasting
- Architectures for sequence data
- Anomaly Detection
- Time Series Representation Learning
- Causality for Time Series Analysis
- Interpretable Machine Learning Methods for Time Series
- Time series classification
- Classical vs. deep learning based methods
- Graph-based methods for time series data