Milan Ojsteršek

Finding the right decision tree's induction strategy for a hard real world problem

Decision trees have been already successfully used in medicine, but as in traditional statistics, some hard real world problems can not be solved successfully using the traditional way of induction. In our experiments we tested various methods for building univariate decision trees in order to find the best induction strategy. On a hard real world problem of the Orthopaedic fracture data with 2637 cases, described by 23 attributes and a decision with three possible values, we built decision trees with four classical approaches, one hybrid approach where we combined neural networks and decision trees, and with an evolutionary approach. The results show that all approaches had problems with either accuracy, sensitivity, or decision tree size. The comparison shows that the best compromise in hard real world problem decision trees building is the evolutionary approach.

Contrasting temporal trend discovery for large healthcare databases

With the increased acceptance of electronic health records, we can observe the increasing interest in the application of data mining approaches within this field. This study introduces a novel approach for exploring and comparing temporal trends within different in-patient subgroups, which is based on associated rule mining using Apriori algorithm and linear model-based recursive partitioning. The Nationwide Inpatient Sample (NIS), Healthcare Cost and Utilization Project (HCUP), Agency for Healthcare Research and Quality was used to evaluate the proposed approach. This study presents a novel approach where visual analytics on big data is used for trend discovery in form of a regression tree with scatter plots in the leaves of the tree. The trend lines are used for directly comparing linear trends within a specified time frame. Our results demonstrate the existence of opposite trends in relation to age and sex based subgroups that would be impossible to discover using traditional trend-tracking techniques. Such an approach can be employed regarding decision support applications for policy makers when organizing campaigns or by hospital management for observing trends that cannot be directly discovered using traditional analytical techniques.

Improving a time critical task execution time using an IPRESPS

Abstract An integrated programming environment for static partitioning of time critical tasks into concurrent modules, called grains, and scheduling those modules for the shortest execution time possible on a macro dataflow computer are presented in this paper. We have named above environment with an achronym IPRESPS. DFCL compiler, a program for static partitioning and scheduling of time critical task and a simulator of macro dataflow computer with a visual interface are parts of IPRESPS. Time critical tasks are represented with programs, written in DFCL (a relatively low level data flow control language which we have developed in our laboratory). The DFCL compiler transforms a program, written in DFCL, into a fine grained dataflow graph, which is transformed next into a program graph. The program for static partitioning and scheduling partitions a program graph into grains and schedules them on the basis of the latest possible execution time without increasing the execution time of the whole program graph. To investigate the execution of various time critical tasks on the model of a macro dataflow computer a simulation program was developed. It simulates the execution of program graphs on the macro dataflow real-time computer. Its outputs are various types of profiles, Gantt chart and other results of simulation.

Dynamic Scheduling on a Network Heterogeneous Computer System

Dynamic scheduling on a heterogeneous system was implemented to minimize the application program execution time. Our method decomposes the program workload into computationally homogeneous subtasks, which may be of different size, depending on the current load of each machine in the heterogeneous computer system.

MPI applications' performances in native vs. virtualized environments using InfiniBand IPoIB virtualization and live migration

Over recent years we have seen a constant evolution of High-Performance Computing moving into commodity servers cluster types of systems. At the same time we have seen Cloud Computing being a proven, stable and reliable IT paradigm, utilising virtualized clusters of commodity servers as key components. This paper presents a face-to-face native-to-virtual HPC MPI application performance analysis on a range of general-use HPC applications by virtualizing InfiniBand via IPoIB at the guest virtual machine level on VMware ESXi as well as KVM hypervisor with production ready live migration support. Applications running on various small and medium HPC clusters are targeted to address users, considering full IT infrastructure virtualization by merging general purpose with HPC IT infrastructure. Live migration is a key component within a virtualized environment as it enables dynamic virtual IT infrastructure, therefore live migration performance impact is evaluated through a pluggable multi-platform architecture on a proactive fault-tolerance use case.

Controlling industrial processes with a dataflow controller: a way to achieve better performances

Abstract An application of the macro dataflow computer architecture in controlling industrial processes is described in this paper. The DFCL — a relatively low level data-flow language and a complier, which transforms programs written in this language into program graphs with different levels of granularity are presented first. A suitable macro data-flow industrial controller (IC) model architecture, which was theoretically studied and simulated is described next. The inputs in the simulation are: an fined grained dataflow graph, a number of processors, a communication delay times and execution times for simple nodes. The outputs are: a waiting tokens profile, a parallelism profile, a waiting nodes profile in queue, the program graph with optimal grain size and scheduling strategy for nodes of the program graph. The above research work was done simultaneously with the development of a real, but conventional IC so that many useful comparisons have been done. This has shown the enormous increase in performances of the data-flow IC and pointed out that it is a viable candidate for future industrial controllers.

Establishing of a Slovenian open access infrastructure: a technical point of view

Purpose – The purpose of this paper is to present a technical perspective when implementing the Slovenian open access infrastructure that consists of four institutional repositories (IRs) and a national portal (NP) that aggregates content from the repositories in order to provide a common search engine, recommendations of similar documents, and similar text detection. Design/methodology/approach – During the project, the necessary legal background and processes for mandatory submissions of final study works, research publications and research data were established, as well as processes for data exchange between the IRs and the NP, and processes for similar text detection. Findings – The consortium consisted of four Slovenian universities that significantly differ in size, organisation, and workflows. It was anticipated that exactly the same legal background and software would be used for the four repositories. It turned out that complete unification was impossible due to the differences. Practical implica...

How to build an E-city

We have created the framework for building Web-based information systems which consists of a tool for automatic object persistence in relational mapping and of a custom XSLT (extensible stylesheet language transformations) processor that would use regular objects instead of XML documents. The benefits of using a framework are: independent design and content management, cross referencing of objects, structure inheritance, multilingual support, independence from the database management system, integration of different Web services and automatic support for multimedia information. The main benefit is shorter time for development of distributed services. Current implementation of the framework is based on RDB (relational database). The framework has been used to develop a city information system for the town of Maribor. The information is collected from organisations, companies, users of the information system, and other Internet and database servers. The information is currently used for presentation on the Internet, on infoterminals, on cable TV, and on teletext. The information system is accessible on http://www.sraka.com (WWW interface) and wap.sraka.com (WAP interface).

Dynamic scheduling on a PC cluster

The goal of this work is to generate a better scheme for communicating data among subtasks during application program execution. Dynamic scheduling on a heterogeneous system PC cluster was implemented to minimize the application program execution time. Our method decomposes the program workload into computationally homogeneous subtasks, which may be of different size, depending on the current load of each machine in a heterogeneous computer system. We present some experimental results of two practical applications. The first one is called automatic classification of words in language modeling, and the second one is the asymmetric traveling salesman problem.