Prof. Dr. Stefan Decker - Theses

Prototypes have been proposed recently as a new way to represent knowledge. In recent years many datasets have been published using RDF. Your task is to find out how the RDF dataset can be converted to prototypes in an efficient manner. For a master thesis, you also have to work on the optimal conversion for different requirements such as updates in the original RDF data.

Semantic technologies and RDF data representation can improve the reusability of scientific data and enable scientist to reproduce the experiments. However there is no tool to support researcher for making their data semantically interpretable by computers. Aim of the thesis is to understand the benefits of semantic web technologies for reproducible research and develop a tool which can convert experimental data to RDF by annotating with selected data models.

Recently, a new approach for computing the personalized pagerank (PPR) for all nodes in a graph was presented by the adviser of this thesis. Personalized pagerank can be used to determine how important specific nodes in a graph are from the viewpoint of a specific node, or set of nodes. For example, it can be used to determine which web pages are relevant for a user, given a set of pages he has visited before. The improvement which the advisor of the thesis devised is useful when the PPR has to be computed for all nodes in the graph. The gain in speed is due to reuse of already computed PPR values for other nodes and a clever ordering of nodes. The student working on this thesis will experiment with this technique and others. First, the student needs to investigate ways to parallelize the algorithms and analyse the algorithms and their parameters experimentally. For a master thesis, the student has to improve the algorithm to optimize loops and experiment whether other improvements in the ordering pay of (i.e., whether other heuristics for finding the order work better and can be computed fast enough to speed up the overall algorithm). Besides, the student would investigate parallel computation of the order and actual PPR. External resources: https://pdfs.semanticscholar.org/c63e/b92a805c3eb636865ecfbd799fda32194753.pdf http://users.jyu.fi/~miselico/papers/GlobalRDFEmbedding.pdf sections 3.2-4

The student will implement several stream sampling algorithms and perform experiments to compare their performance. The implementations are done on top of streaming frameworks like Spark, Apache Flink, and Storm.

Recently, a new approach for finding maximal frequent patterns (MFPAS) was presented by the supervisor and advisers of this thesis. Several further optimizations of the algorithm are possible. The student working on this thesis will experiment with different optimization possibilities and analyse their effect experimentally. For a master thesis, further theoretical analysis of the optimizations and the original algorithm are necessary. Related paper: M. R. Karim, M. Cochez, O. D. Beyan, C. F. Ahmed, and S. Decker. Mining maximal frequent patterns in transactional databases and dynamic data streams: a spark-based approach. Information Sciences, 2017a

Lately several methods for embedding graphs nodes into a vector space have been proposed. These embeddings can then used to train other machine learning models. Learning these embeddings is typically done using CPUs. In this thesis the student would look into the use of other hardware, like GPUs and distributed computation options to speed up the learning process. The challenge is that algorithms working on graphs have typically a bad memory locality. Hence, existing algorithms might need profound modification in order to use them on GPUs or in a distributed fashion.

The objective of this thesis is to analyse breast cancer genomic data with the potential of predicting breast cancer genomic biomarker. Specifically, these analysis will include classification and regression of breast cancer patients based on their genetic information.

Locality-sensitive hashing is used to speed up near-neighbor search in high dimensional space. When the distance of interest is cosine distance, Random hyperplane hashing (RHH) is used. This technique is based on randomly selecting hyperplanes. However, in some cases (when we have more information about the dataset) it seems reasonable to not choose the hyperplanes completely randomly. Further, if normal RRH is performed with a low number of hyperplanes, then the hyperplanes are likely to not cover the space very well. This thesis will be about choosing the hyperplane in a data dependent way and try to sample the hyperplanes such that they cover the space nicely (including a comparison with angular quantization).

In this thesis, the student investigates the use of deep learning techniques (especially computer vision) to perform diagnosis of osteoarthritis. The input to the system are both radiographs (X-RAY) and magnetic resonance images (MRI).

Graph-structured queries provide an efficient means to retrieve desired data from large-scale knowledge graphs. However, it is difficult for non-expert users to write such queries, and users prefer expressing their query intention through natural language questions. Recently, an increasing effort is being exerted to construct graph-structured queries for given natural language questions. At the core of the construction is to deduce the structure of the target query and retrieve vertices/edges of the underlying knowledge graph which constitute the query. Existing query construction methods rely on conventional graph-based algorithms and question understanding techniques, which lead to inefficient and degraded performances facing complicated natural language questions over knowledge graphs with large scales. In this thesis, we focus on this problem and propose novel construction models standing on recent knowledge graph embedding techniques. Extensive experiments were conducted on question answering benchmark datasets, and the results demonstrate that our models outperform baselines in terms of effectiveness and efficiency.