Felix Jungermann

Enhanced Services for Targeted Information Retrieval by Event Extraction and Data Mining

We present a framework combining information retrieval with machine learning and (pre-)processing for named entity recognition in order to extract events from a large document collection. The extracted events become input to a data mining component which delivers the final output to specific users questions. Our case study is the public collection of minutes of plenary sessions of the German parliament and of petitions to the German parliament.

Towards adjusting mobile devices to user's behaviour

Mobile devices are a special class of resource-constrained embedded devices. Computing power, memory, the available energy, and network bandwidth are often severely limited. These constrained resources require extensive optimization of a mobile system compared to larger systems. Any needless operation has to be avoided. Timeconsuming operations have to be started early on. For instance, loading files ideally starts before the user wants to access the file. So-called prefetching strategies optimize systems operation. Our goal is to adjust such strategies on the basis of logged system data. Optimization is then achieved by predicting an applications behavior based on facts learned from earlier runs on the same system. In this paper, we analyze system-calls on operating system level and compare two paradigms, namely server-based and device-based learning. The results could be used to optimize the runtime behaviour of mobile devices.

Enhancing Ubiquitous Systems through System Call Mining

Collecting, monitoring, and analyzing data automatically by well instrumented systems is frequently motivated by human decision-making. However, the same need occurs when system software decisions are to be justified. Compiler optimization or storage management requires several decisions which result in more or less resource consumption, be it energy, memory, or runtime. A magnitude of system data can be collected in order to base decisions of compilers or the operating system on empirical analysis. The challenge of large-scale data is aggravated if system data of small and often mobile systems are collected and analyzed. In contrast to the large data volume, the mobile devices offer only very limited storage and computing capacity. Moreover, if analysis results are put to use at the operating system, the real-time response is at the system level, not on the level of human reaction time. In this paper, small and most often mobile systems (i.e., ubiquitous systems) are instrumented for the collection of system call data. It is investigated whether the sequence and the structure of system calls are to be taken into account by the learning method, or not. A structural learning method, Conditional Random Fields (CRF), is applied using different internal optimization algorithms and feature mappings. Implementing CRF in a massively parallel way using general purpose graphic processor units (GPGPU) points at future ubiquitous systems.

Relation extraction for monitoring economic networks

Relation extraction from texts is a research topic since the message understanding conferences. Most investigations dealt with English texts. However, the heuristics found for these do not perform well when applied to a language with free word order, as is, e.g., German. In this paper, we present a German annotated corpus for relation extraction. We have implemented the state of the art methods of relation extraction using kernel methods and evaluate them on this corpus. The poor results led to a feature set which focusses on all words of the sentence and a tree kernel which includes words, in addition to the syntactic structure. The relation extraction is applied to monitoring a graph of economic company-directors network.

Handling Tree-Structured Values in RapidMiner

Attribute value types play an important role in mostly every datamining task. Most learners, for instance, are restricted to particular value types. The usage of such learners is just possible after special forms of preprocessing. RapidMiner most commonly distinguishes between nominal and numerical values which are well-known to every RapidMineruser. Although, covering a great fraction of attribute types being present in nowadays datamining tasks, nominal and numerical attribute values are not sufficient for every type of feature. In this work we are focusing on attribute values containing a tree-structure. We are presenting the handling and especially the possibilities to use tree-structured data for modelling. Additionally, we are introducing particular tasks which are offering tree-structured data and might benefit from using those structures for modelling. All methods presented in this paper are contained in the Information Extraction Plugin for RapidMiner.

Tree Kernel Usage in Naive Bayes Classifiers

We present a novel approach in machine learning by combining naBayes classifiers with tree kernels. Tree kernel methods produce promising results in machine learning tasks containing tree- structured attribute values. These kernel methods are used to compare two tree-structured attribute values recursively. Up to now tree kernels are only used in kernel machines like Support Vector Machines or Perceptrons. In this paper, we show that tree kernels can be utilized in a naBayes classifier enabling the classifier to handle tree-structured values. We evaluate our approach on three datasets contain- ing tree-structured values. We show that our approach using tree-structures delivers signifi- cantly better results in contrast to approaches us- ing non-structured (flat) features extracted from the tree. Additionally, we show that our approach is significantly faster than comparable kernel ma- chines in several settings which makes it more useful in resource-aware settings like mobile de- vices.