23.10.2015

A. Borusan

## 13.11.2015, 14 Uhr s.t. TU Berlin, EN building, seminar room EN 719 (7th floor), Einsteinufer 17, 10587 Berlin: "Scaling Data Mining in Massively Parallel Dataflow Systems" (Sebastian Schelter PhD Thesis Defense)

This thesis lays the ground work for enabling scalable data mining in massively parallel dataflow systems, using large datasets. Such datasets have become ubiquitous. We illustrate common fallacies with respect to scalable data mining: It is in no way sufficient to naively implement textbook algorithms on parallel systems; bottlenecks on all layers of the stack prevent the scalability of such naive implementations. We argue that scalability in data mining is a multi-leveled problem and must therefore be approached on the interplay of algorithms, systems, and applications. We therefore discuss a selection of scalability problems on these different levels.

We investigate algorithm-specific scalability aspects of collaborative filtering algorithms for computing recommendations, a popular data mining use case with many industry deployments. We show how to efficiently execute the two most common approaches, namely neighborhood methods and latent factor models on MapReduce, and describe a specialized architecture for scaling collaborative filtering to extremely large datasets which we implemented at Twitter. We turn to \textit{system-specific scalability aspects}, where we improve system performance during the distributed execution of a special class of iterative algorithms by drastically reducing the overhead required for guaranteeing fault tolerance. Therefore we propose a novel optimistic approach to fault-tolerance which exploits the robust convergence properties of a large class of fixpoint algorithms and does not incur measurable overhead in failure-free cases. Finally, we present work on an \textit{application-specific scalability aspect} of scalable data mining. A common problem when deploying machine learning applications in real-world scenarios is that the prediction quality of ML models heavily depends on hyperparameters that have to be chosen in advance. We propose an algorithmic framework for an important subproblem occuring during hyperparameter search at scale: efficiently generating samples from block-partitioned matrices in a shared-nothing environment.

For every selected problem, we show how to execute the resulting computation automatically in a parallel and scalable manner, and evaluate our proposed solution on large datasets with billions of datapoints.