Big Data Engineering

Lead
Prof. Dr. Matthias Böhm

Technische Universität Berlin
Ernst-Reuter-Platz 7, 10587 Berlin

Data Science Abstractions and Systems, Performance-Accuracy Tradeoffs in Data Science, Data Cleaning Pipelines and Optimization

The mission of the Big Data Engineering group, led by Prof. Dr. Matthias Böhm, is to simplify data science by providing high-level, data-science-centric abstractions and building systems and tools to execute these tasks in an efficient and scalable manner. The general research interests include the exploration of performance-accuracy tradeoffs, tooling (script generators, label generation, advisors, etc.), seamless data augmentation, cleaning, feature engineering, model debugging and deployment, cost-effective cloud deployments, advanced optimization techniques, adaptive data storage and indexing, and the exploitation of modern hardware.

Current research focuses on:
•   Data Cleaning Pipelines: Automatic enumeration of data cleaning pipelines for target ML application, hyperparameter optimization of cleaning primitives.
•   Model Debugging: Finding the top-k data slices where a trained model underperforms, linear-algebra-based enumeration and pruning algorithms.
•   Fine-grained Lineage Tracing and Reuse: Fine-grained, multi-level lineage tracing for versioning and reuse, lineage deduplication, full and partial reuse of intermediates.
•   Federated Linear Algebra and Parameter Servers: ML model training on federated raw data without central data consolidation, plan generation under awareness of privacy constraints, federated linear algebra programs and parameter servers.
•   Workload-aware Data Reorganization: Compression under awareness of data and workload (linear algebra program) characteristics, asynchronous data reorganization in standing executors (e.g., at standing federated workers).
•   Code Generation for Heterogeneous HW: Extended operator fusion and code generation for GPUs and heterogeneous devices, including sparsity exploitation across operations.