Cindy Perscheid

Biomarker detection on gene expression data is an important field of research, as the detected biomarkers, i.e. genes, allow for diagnosing diseases and predicting treatment outcomes or survival. Prior knowledge approaches incorporate existing biological findings directly into biomarker detection and are expected to improve biomarker robustness. However, existing prior knowledge approaches are not flexibly applicable to different use cases and their actual effectiveness is not well researched on.

In this talk, we present our research on addressing the aforementioned issues. We have reviewed existing knowledge bases and prior knowledge approaches to subsequently identify different types of prior knowledge and integration strategies that can be flexibly applied to gene expression data. We then have conducted a large case study comparing the performance of multiple prior knowledge approaches, classical feature selection strategies, and knowledge bases, of which we will show first results.

Jan Kossman

Database management systems offer a multitude of configuration options whose proper adjustment is essential for the system's performance. Usually, human database administrators (DBA) tune these configuration options. However, with an ever-increasing number of options, dynamic workloads, and more and more systems to handle, the complexity of such configuration tasks has surpassed the capabilities of human DBAs.

In the future, database systems will optimize their configuration in an unsupervised fashion. Advancements are required in many research areas to achieve this goal. We investigated two aspects that approach unsupervised database optimization from different perspectives. First, we focus on efficient physical database design by analyzing existing index selection algorithms and proposing new, improved approaches. Second, we examine how DBMSs can identify data dependencies in an unsupervised fashion and utilize them for query optimization.

Aikaterini Niklanovits

Given a simple connected weighted graph G=(V,E) and an integer k, the k-Balanced Connected Partition problem (k-BCP) is the problem of partitioning G into k connected components of rougly equal weight. In this talk we are going to introduce this problem, and present some very efficient algorithmic results for the class of c-claw free graphs. In particular, we are going to see approximation algorithms for the two classical objectives of k-BCP, that is maximizing the size of the smallest component and minimizing the size of the larger one, regarding c-claw free graphs. This also results in a 2-approximation algorithm for the balanced edge partition case.

Gerard de Melo

Promoting Your Research Scientific research is often viewed as a noble quest for knowledge and truth. However, upon closer inspection, current scientific practices are highly dependent on what channels scientists pay attention to and how novel research results are disseminated in these channels. Thus, just engaging in research is not enough and we arguably also need to consider how to promote our research. To this end, this talk will present a range of measures that can help in informing our community about our research, including but not limited to the use of social media platforms. The talk will discuss both laudable and questionable recent developments.

Mazhar Hameed

Every day, data are produced from different sources with varying structures for multiple purposes. With increasing heterogeneity, the ability to manage it accurately and according to user requirements has become more challenging for downstream applications. Among other challenges, detecting ill-formed records in files is a difficult problem: Ill-formed records occur in raw data due to loosely defined schemata, incorrect formatting of values, record structure discrepancy, etc. They can lead to aborted loading processes, incorrectly parsed data, and can interfere with the training process of machine learning algorithms.

We propose  SURAGH, a method to detect ill-formed records without relying on external information, such as data types, record structure, or schemata. It parses an input file and generates syntax-based patterns for column values using order-dependent abstractions. We combine these to patterns for entire rows and use them to identify ill-formed records.

Keven Richly

The wide distribution of location-acquisition technologies has led to large volumes of spatio-temporal data, which are the foundation for a broad spectrum of applications. Based on the performance requirements of these applications, in-memory database systems are often used to store and process the data. To efficiently utilize the available DRAM capacities, modern database systems apply various optimizations (e.g., data compression or secondary indexes) to reduce the memory footprint or increase performance. However, the selection of cost and performance balancing configurations is challenging due to the vast number of possible setups consisting of mutually dependent individual decisions. We present a linear programming approach to determine fine-grained configuration optimizations for spatio-temporal workloads. By dividing the data into partitions of fixed size, we can apply different database optimizations for data partitions independently. Our approach jointly optimizes multiple configuration options to optimize the runtime performance for a given workload, data characteristics, and memory budget. To yield maintainable and robust configurations, we extend our linear programming-based approach to incorporate reconfiguration costs as well as a worst-case optimization for potential workload scenarios. We demonstrate on a real-world dataset that our models allow us to reduce the memory footprint (up to 60% with equal performance) or increase the performance (up to 80% with equal memory consumption) compared to established rule-based heuristics.

Marta Lemanczyk

Interpretability methods are used in deep learning to identify individual contributions of input features to neural network predictions. These methods can be also applied on genomic convolutional neural networks to find relevant patterns within genomic sequences. In this way, it is possible to obtain new insights into biological mechanisms or to validate model predictions.
In this talk, I will provide an overview on how to train convolutional neural networks on genomic data and analyze them with common post-hoc interpretability methods.
Additionally, I will give you insights into our current work: We investigate the contribution scores from attribution methods applied on sequence data containing interactions. This includes defining interactions in biological sequences, generating data sets that capture these concepts and evaluating how current interpretability methods behave on CNNs trained on that data.

Bernhard Renard

The German constitution explicitly protects freedom of research. At the same time, science is not free of outside influences and scientific discoveries can change daily life in a society. Within the discussion, we want to focus on exploring the potential interplay between science and politics. The goal is on a pragmatic view to identify areas of our research potentially affected rather than free floating philosophical discussions.

Martin Boissier

Data encoding has been applied to database systems for decades as it mitigates bandwidth bottlenecks and reduces storage requirements and thus costs. But even in the presence of these advantages, most in-memory systems only defensively employ compression as the negative impact on runtime performance can be severe. Nevertheless, real-world systems with large parts being infrequently accessed and cost efficiency constraints in cloud installations necessitate solutions that automatically and efficiently select encoding techniques, including heavy ones. In this paper, we study approaches to determine memory budget-constrained encoding configurations using greedy heuristics and linear programming. We show that optimal configurations can reduce the main memory footprint by 60 % for TPC-H without a loss in runtime performance over state-of-the-art dictionary encoding. To yield robust selections, we extend the linear programming-based approach to incorporate query runtime constraints and mitigate unexpected performance regressions.

Stefan Halfpap

Database replication and query load-balancing are mechanisms to scale query throughput. Using workload knowledge allows load-balancing queries to replica nodes according to their accessed data. As a result, replica nodes must only store and synchronize subsets of the data. However, balancing the load of large-scale workloads evenly while minimizing the memory footprint is challenging, particularly if robustness against node failures and uncertain workloads is required: Optimal solutions take too much time. Solutions of existing heuristics are either not memory-efficient or cause load imbalances, both degrading cost/performance. We present allocation approaches based on linear programming to quickly calculate allocations with a better mix of robust performance and memory efficiency than state-of-the-art approaches. We demonstrate the applicability and the effectiveness of our approaches for the standardized benchmarks TPC-H and TPC-DS as well as a real-world accounting workload.

Frank Pawlitschek

The HPI School of Entrepreneurship here for you when it comes to founding a business.

We inspire, motivate and support you. Every day. From finding suitable business ideas, through the first tentative thoughts of a start-up, to successful company growth.

We offer an inspiring ecosystem and a lot of formats that individually help entrepreneurs to bring together all the necessary ingredients: Mentoring and coaching, methods around ideation and business ideas testing, business know-how and tool sets, finding co-founders, funding, network access, collaboration partners and much more.

We are looking to reflect in this session on thinking about the dissertation topic from an entrepreneurial perspective.

https://hpi.de/en/entrepreneurship/hpi-e.html

Johannes Maximilian Kroschewski

Single Sign-On (SSO) is a method allowing users to log in to an application through a third-party account, such as their Google account. Today, the most widely used SSO protocol is OpenID Connect (OIDC), which poses a significant privacy problem: The SSO provider can track users by learning which applications are being accessed. My thesis addresses this problem and makes three major contributions. First, it provides a critical discussion of proposed approaches for privacy- and security-enhancing SSO protocols. We find that the dominant approaches are incompatible with the common OIDC standard. Such compatibility is essential to enable the application in practice. Our second contribution is to propose two protocol extensions, which are (1) compatible with OIDC and (2) enhance user privacy. Lastly, we evaluate if the proposed protocols cause disruption with other functions in OIDC and assess the necessary adaptations to the protocol flow. The goal is to establish usable protocol extensions which function smoothly in practice and resolve long-standing privacy issues in SSO protocols.

Stefan Neubert

Among the most important skills of scientists is the capability to communicate their work to colleagues. This tends to be relatively easy when interacting with other experts from the same field: One tackles related questions, uses similar techniques, builds on the same (or at least overlapping) prior work and -- most importantly -- knows the same technical terms.
Even within a field, the definition of such terms can be fuzzy and their interpretation by different people might differ ever so slightly. However, once we target a broader audience, communication becomes a much bigger issue: Other scientists might not know some of our terms and phrases or (even worse) know and use them, but with entirely different semantics.
For all posters we design, talks we give and texts we write, we have to find an audience-specific balance between 'defining everything' and 'expecting perfect prior knowledge'. The research school provides an ideal stage to sharpen our skills in such inter-group communication. In teams of ~3 people from different research groups we will revisit our research symposium posters to track down and do away with possible misunderstandings and language barriers.

Felix Naumann

Conflicts of interest in research occur when a person reviewing another person's work (paper, thesis, funding proposal, application, nomination, etc.) is potentially biased or can be perceived to be biased. Together, we will discuss different definitions of Col, the need and difficulty to disclose conflicts, how unbiased reviewing is subverted and what we can do about it. Please prepare for this meeting by searching and noting the Col-policies of the top conference in your particular field.

Karen Seidel

This talk is about learning from informant, a formal model for binary classification. Illustrating examples are linear separators and other uniformly decidable sets of formal languages. Due to the learning by enumeration technique by Gold the learning process can be assumed consistent when full-information is available.
The original model is then adjusted towards the setting of deep learning. We investigate the learnability of the set of half-spaces by this incremental learners. Moreover, they have less learning power than the full-information variant by a fundamental proof technique due to Blum and Blum. This technique can also be used to separate consistency.
Finally, we present very recent results towards a better understanding of (strong) non-U-shaped learning from binary labeled input data.