The DEEM Lab is a cross-organisational research group uniting the chair for the management of data science processes at Technische Universität Berlin with external members from multiple universities and industry. The lab is led by Prof. Dr.-Ing. Sebastian Schelter and is part of the Berlin Institute for the Foundations of Learning and Data (BIFOLD).

Our lab conducts fundamental research at the intersection of data management and machine learning, which addresses data-related problems in ML applications that cause negative economic, societal or scientific impact. Our goal is to foster the responsible management of data and to lower the technical bar for working with data science technologies.

Our research is accompanied by efficient and scalable open source implementations, many of which are applied in real world use cases, for example in the Amazon Web Services cloud and in large European e-commerce platforms. The focus areas of the lab are:

• Data-centric debugging and testing of machine learning pipelines
• Data processing in compliance with legal regulations, such as the “right-to-be-forgotten”
• Automated validation of data at scale

Our research contributions have been recognized with an ACM SIGMOD Systems Award, an ACM SIGMOD Best Demo Runner Up Award and a Best Paper Runner Up Award from the Table Representation Learning workshop at NeurIPS. We have ongoing collaborations with the University of Amsterdam and CWI in the Netherlands, as well as with the Center for Responsible AI at New York University.

• Jan 11, 2025 ‐ Stefan will be co-organising the workshop on Data Management for End-to-End Machine Learning (DEEM) at SIGMOD 2025 in Berlin.
  
  
• Dec 3, 2024 ‐ We have been co-organising a workshop on ‘The EU AI Act – Developing a technical perspective’ together with our colleagues from machine learning and law as well as industry practitioners.
  
  
• Oct 10, 2024 ‐ Our research group has been covered in an interview on the #ai_berlin website.
  
  
• Jun 1, 2024 ‐ We will be present at the upcoming VLDB conference in China with several contributions:

  • How Data Management Research Helps to Improve Real World ML Applications (Keynote at the workshop on Quality in Databases)
  • A Flexible Forecasting Stack (Industry paper with AWS)
  • Snapcase - Regain Control over Your Predictions with Low-Latency Machine Unlearning (Demonstration)
  • Instrumentation and Analysis of Native ML Pipelines via Logical Query Plans (Paper at the PhD workshop)
  • Assisted design of data science pipelines (Poster of a VLDBJ publication)

  
  
  
• Jun 1, 2024 ‐ At the upcoming SIGMOD conference in Chile, Till will present our paper on SchemaPile: A Large Collection of Relational Database Schemas. SchemaPile is a corpus of more than 200 thousand database schemas, which we envision to be a great resource for ML models dealing with structured data, e.g., in data integration tasks. Furthermore, Stefan will present his initial ideas on Interactively Improving ML Data Preparation Code via ‘Shadow Pipelines’ at the DEEM workshop.
  
  

Faculty, PhD Students & Staff
Prof. Dr. Sebastian Schelter	Stefan Grafberger	Hao Chen
Olga Ovcharenko	Pierre Lubitzsch
External PhD Students & Guests
Zeyu Zhang (University of Amsterdam)	Shubha Guha (University of Amsterdam)	Till Doehmen (Motherduck)
Yichun Wang (University of Amsterdam)	David Campos (Aalborg University)
(External) Master Students
Aynaz Abdollahzadeh (University of Amsterdam)	Leonardo Dominci (University of Amsterdam)

• Barrie Kersbergen, PhD student, Staff scientist at bol.com
• Dr. Olivier Sprangers, PhD student, Applied scientist at Nixtla
• Dr. Mozhdeh Ariannezhad, PhD student, ML scientist at booking.com
• Arezoo Sarvi, PhD student, Data scientist search at Albert Heijn
• Radu Geacu, Master student, Data engineer at the Royal Schiphol Group
• David Vos, Master student, PhD student at the University of Amsterdam
• Benjamin Wang, Master student, ML scientist at booking.com
• Dr. Ji Zhang, Guest researcher, Research scientist at Huawei

Sebastian Schelter is a Full Professor at the Berlin Institute on the Foundations of Learning and Data (BIFOLD) and Technische Universität Berlin. His research is focused on the intersection of data management and machine learning with the goal to foster the responsible management of data and to democratise data science technologies.

The research of his group is accompanied by efficient and scalable open source implementations, many of which are applied in real world use cases, for example in the Amazon Web Services cloud and in large European e-commerce platforms.

In the past, he has been an assistant professor at the University of Amsterdam, a faculty fellow at New York University, a senior applied scientist at Amazon Research and a research intern at Twitter and IBM Almaden in California. His research contributions have been recognized with an ACM SIGMOD Systems Award, an ACM SIGMOD Best Demo Runner Up Award, and a Best Paper Runner Up Award from the Table Representation Learning workshop at NeurIPS.

• Email: schelter[at]tu-berlin[dot]de
• Profiles: Google Scholar, Twitter, Linkedin, DBLP, ORCID
  
  

Winter semester 24-25:

Details on registration and seat assignment will be discussed in the first in-person class of each course.

EDML – Engineering Data for Machine Learning

This highly technical course focuses on the engineering and life-cycle management of data for production machine learning deployments. The course starts by recapping fundamentals about relational data processing and dataflow systems. Subsequently, students learn about encoding, storing and managing vectorised feature representations of heterogeneous input data sources for machine learning applications, and the architecture of current state-of-the- art systems for this task such as Google’s Tensorflow Extended Platform. Concurrently, the students will be exposed to foundational theory for this problem space, such as incremental view maintenance for relational data, fine-grained data provenance tracking via provenance semi-rings and differential computation.

In addition, students will learn to identify, quantify and address common quality issues with respect to the completeness and consistency of the data. Furthermore, students will be exposed to ongoing research efforts in this space such as ML pipeline debugging or error detection techniques from data-centric AI. In addition, they will have the opportunity to discuss the practical implications of the covered technologies with invited practitioners.

Course page available at https://isis.tu-berlin.de/course/view.php?id=40407

RDSEM – Seminar on Responsible Data Engineering

In this seminar, students will learn how to: (a) critically read and interpret scientific papers drawn from literature on responsible data management and responsible data science, (b) give a good scientific presentation that is technically precise, concentrated on the relevant topics, and also enjoyable; © write a scientific survey based on papers drawn from varying sources, such as contemporary computer science journals and conference proceedings. In addition, students will learn about state-of-the art and current research topics in responsible data science and data management.

Course page available at https://isis.tu-berlin.de/course/view.php?id=40406

Programming Practical – Machine Learning Pipelines

At the beginning of every semester, we will define a new project (or multiple) related to the implementation of machine learning pipelines (which prepare data from heterogeneous sources, encode it as features and train one or more ML models afterwards). Students receive specifications of these components as well as selected implementation goals. The task is then to create - in self-organised teams of 4 persons - correct implementations of these components in Python, Scala or Rust. Apart from developing the source code of the prototypes, other important aspects include the use of version control tools, test-driven development, design documentation, as well as runtime experiments and improvements. At the same time, this project allows for a deeper understanding of methods related to machine learning pipelines and data analysis, as well as algorithms and data structures. The focus is, however, on a problem-oriented utilisation of programming skills so solve a practical problem, not a holistic coverage of functionality of machine learning applications.

Course page available at https://isis.tu-berlin.de/enrol/index.php?id=39392

PhD Positions in Data Engineering (via the BIFOLD Graduate School)

The “Berlin Institute for the Foundations of Learning and Data” (BIFOLD) offers several positions for doctoral students as part of its graduate school. Join the graduate school allows you to conduct research with the DEEM Lab as well.

If you are interested in this opportunity, please send your application via email (quoting the job reference number IV-63124) as one file in PDF format to Prof. Dr. Volker Markl and Prof. Dr. Klaus-Robert Müller, at gsapplication@bifold.tu-berlin.de until the 3rd of February.

Your application for the DEEM Lab has to include the following documents:

• Filled-in application form, in order to apply for our lab, please enter the following information under “2.5 Research interests”:
  
  • Research group leads name: Prof. Dr. Sebastian Schelter
  • Research areas of interest: Pick one from “Data-centric debugging and testing of machine learning pipelines”, “Data processing in compliance with legal regulations”, “Automated validation of data at scale”
• Letter of motivation and CV
• Copies of your degree certificates, academic transcripts, and a list of publications
• Names and contact details of 2 referees (whose letters should be available by the deadline of this call)

Email: schelter [at] tu-berlin [dot] de

Technische Universität Berlin
FG Management of Data Science Processes
Sekr. TEL 9-2
Ernst-Reuter Platz 7
10587 Berlin
Germany

Responsibility under the German Press Law §55 Sect. 2 RStV:
Prof. Dr.-Ing. Sebastian Schelter