Marcin Wojnarski

On the Evolution of Rough Set Exploration System

We present the next version (ver. 2.1) of the Rough Set Ex- ploration System - a software tool featuring a library of methods and a graphical user interface supporting variety of rough-set-based and re- lated computations. Methods, features and abilities of the implemented software are discussed and illustrated with examples in data analysis and decision support.

TunedIT.org: system for automated evaluation of algorithms in repeatable experiments

In this paper we present TUNEDIT system which facilitates evaluation and comparison of machine-learning algorithms. TUNEDIT is composed of three complementary and interconnected components: TunedTester, Repository and Knowledge Base. TunedTester is a stand-alone Java application that runs automated tests (experiments) of algorithms. Repository is a database of algorithms, datasets and evaluation procedures used by TunedTester for setting up a test. Knowledge Base is a database of test results. Repository and Knowledge Base are accessible through TUNEDIT website. TUNEDIT is open and free for use by any researcher. Every registered user can upload new resources to Repository, run experiments with TunedTester, send results to Knowledge Base and browse all collected results, generated either by himself or by others. As a special functionality, built upon the framework of automated tests, TUNEDIT provides a platform for organization of on-line interactive competitions for machine-learning problems. This functionality may be used, for instance, by teachers to launch contests for their students instead of traditional assignment tasks; or by organizers of machine-learning and data-mining conferences to launch competitions for the scientific community, in association with the conference.

IEEE ICDM 2010 Contest: TomTom Traffic Prediction for Intelligent GPS Navigation

In this foreword, we summarize the IEEE ICDM 2010 Contest: “TomTom Traffic Prediction for Intelligent GPS Navigation”. The challenge was held between Jun 22, 2010 and Sep 7, 2010 as an interactive on-line competition, using the TunedIT platform (http://tunedit.org). We present the scope of the ICDM contest series in general, the scope of this year’s contest, description of its tasks, statistics about participation, details about the TunedIT platform and the Traffic Simulation Framework. A detailed description of winning solutions is part of this proceeding series.

Debellor: A Data Mining Platform with Stream Architecture

This paper introduces Debellor (www.debellor.org) --- an open source extensible data mining platform with stream-based architecture, where all data transfers between elementary algorithms take the form of a stream of samples. Data streaming enables implementation of scalable algorithms, which can efficiently process large volumes of data, exceeding available memory. This is very important for data mining research and applications, since the most challenging data mining tasks involve voluminous data, either produced by a data source or generated at some intermediate stage of a complex data processing network. Advantages of data streaming are illustrated by experiments with clustering time series. The experimental results show that even for moderate-size data sets streaming is indispensable for successful execution of algorithms, otherwise the algorithms run hundreds times slower or just crash due to memory shortage. Stream architecture is particularly useful in such application domains as time series analysis, image recognition or mining data streams. It is also the only efficient architecture for implementation of online algorithms. The algorithms currently available on Debellor platform include all classifiers from Rseslib and Weka libraries and all filters from Weka.

Interactive Rough-Granular Computing in Pattern Recognition

We discuss the role of generalized approximation spaces and operations on approximation spaces in searching for relevant patterns. The approach is based on interactive rough-granular computing (IRGC) in the WisTech program. We also present results on approximation of complex vague concepts in real-life projects from different domains using the approach based on ontology approximation. Software projects supporting IRGC are reported.

Absolute contrasts in face detection with adaboost cascade

Object detection using AdaBoost cascade classifier was introduced by Viola and Jones in December 2001. This paper presents a modification of their method which allows to obtain even 4-fold decrease in false rejection rate, keeping false acceptance rate - as well as the classifier size and training time - at the same level. Such an improvement is achieved by extending original family of weak classifiers, which is searched through in every step of AdaBoost algorithm, with classifiers calculating absolute value of contrast. Test results given in the paper come from a face localization problem, but the idea of absolute contrasts can be applied to detection of other types of objects, as well.

Nondeterministic Discretization of Weights Improves Accuracy of Neural Networks

The paper investigates modification of backpropagation algorithm, consisting of discretization of neural network weights after each training cycle. This modification, aimed at overfitting reduction, restricts the set of possible values of weights to a discrete subset of real numbers, leading to much better generalization abilities of the network. This, in turn, leads to higher accuracy and a decrease in error rate by over 50% in extreme cases (when overfitting is high). Discretization is performed nondeterministically, so as to keep expected value of discretized weight equal to original value. In this way, global behavior of original algorithm is preserved. The presented method of discretization is general and may be applied to other machine-learning algorithms. It is also an example of how an algorithm for continuous optimization can be successfully applied to optimization over discrete spaces. The method was evaluated experimentally in WEKA environment using two real-world data sets from UCI repository.

Unsupervised learning of image recognition with neural society for clustering

New algorithm for partitional data clustering is presented, Neural Society for Clustering (NSC). Its creation was inspired by hierarchical image understanding, which requires unsupervised training to build the hierarchy of visual features. Existing clustering algorithms are not well-suited for this task, since they usually split natural groups of patterns into several parts (like k-means) or give crisp clustering. Neurons comprising NSC may be viewed as a society of autonomous individuals, proceeding along the same simple algorithm, based on four principles: of locality, greediness, balance and competition. The same principles govern large groups of entities in economy, sociology, biology and physics. Advantages of NSC are demonstrated in experiment with visual data. The paper presents also a new method for objective and quantitative comparison of clustering algorithms, based on the notions of entropy and mutual information.