Krzysztof Grąbczewski

Transformations of symbolic data for continuous data oriented models

Most of Computational Intelligence models (e.g. neural networks or distance based methods) are designed to operate on continuous data and provide no tools to adapt their parameters to data described by symbolic values. Two new conversion methods which replace symbolic by continuous attributes are presented and compared to two commonly known ones. The advantages of the continuousification are illustrated with the results obtained with a neural network, SVM and a kNN systems for the converted data.

Universal Meta-Learning Architecture and Algorithms

There are hundreds of algorithms within data mining. Some of them are used to transform data, some to build classifiers, others for prediction, etc. Nobody knows well all these algorithms and nobody can know all the arcana of their behavior in all possible applications. How to find the best combination of transformation and final machine which solves given problem?

Meta-learning with Machine Generators and Complexity Controlled Exploration

We present a novel approach to meta-learning, which is not just a ranking of methods, not just a strategy for building model committees, but an algorithm performing a search similar to what human experts do when analyzing data, solving full scope of data mining problems. The search through the space of possible solutions is driven by special mechanisms of machine generators based on meta-schemes. The approach facilitates using human experts knowledge to restrict the search space and gaining meta-knowledge in an automated manner. The conclusions help in further search and may also be passed to other meta-learners. All the functionality is included in our new general architecture for data mining, especially eligible for meta-learning tasks.

Techniques of Decision Tree Induction

Finding optimal DT for given data is not easy (with exceptions of some trivial cases). The hierarchical structure of DT models could suggest that the optimization process is also nicely reduced with subsequent splits, but it is not so.

Validated decision trees versus collective decisions

In the most common decision tree (DT) induction approaches, crossvalidation based processes validate the final DT model. This article answers many questions about advantages of using different types of committees constructed from the DTs generated within the validation process, over single validated DTs. Some new techniques of providing committee members and their collective decisions are introduced and evaluated among other methods. The conclusions presented here, are useful both for human experts and automated meta-learning approaches.

Increasing efficiency of data mining systems by machine unification and double machine cache

In advanced meta-learning algorithms and in general data mining systems, we need to search through huge spaces of machine learning algorithms. Meta-learning and other complex data mining approaches need to train and test thousands of learning machines while searching for the best solution (model), which often is quite complex. To facilitate working with projects of any scale, we propose intelligent mechanism of machine unification and cooperating mechanism of machine cache. Data mining system equipped with the mechanisms can deal with projects many times bigger than systems devoid of machine unification and cache. Presented solutions also reduce computational time needed for learning and save memory.

Task management in advanced computational intelligence system

Computational intelligence (CI) comes up with more and more sophisticated, hierarchical learning machines. Running advanced techniques, including meta-learning, requires general data mining systems, capable of efficient management of very complex machines. Requirements for running complex learning tasks, within such systems, are significantly different than those of running processes by operating systems. We address major requirements that should be met by CI systems and present corresponding solutions tested and implemented in our system. The main focus are the aspects of task spooling and multitasking.

Future Perspectives of Meta-Learning

The preceding chapter shows various approaches to learning at meta-level. It emphasizes that although the goals may be defined in many different ways, the ultimate goal should always be an improvement in learning at base-level. Even the most attractive form of meta-knowledge is not a value for itself, but only if it can help improve learning processes, so that learning at object-level gets faster or more accurate.

Intemi: Advanced Meta-Learning Framework

Serious meta-learning applications may require running huge counts of learning processes. The results obtained from the calculations must be reliably analyzed by many comparisons and statistical tests. To gain valuable knowledge about data at hand, it does not suffice to run a couple of methods and show the results. Also in scientific experiments, it should no longer be accepted, that several algorithms are tried and new approaches are claimed to be advantageous, on the basis of several simple tests and comparisons to several other methods.

Meta-Level Analysis of Decision Tree Induction

Object oriented design divides complex algorithms and data structures into smaller and simpler components, specializing in solving extracted subproblems. As a result, also in the approach to a general framework for DT induction, the algorithms can be composed by a number of compatible components. In the framework described in Chap. 3, even the simplest DT induction algorithms are composed of several components responsible for such tasks as performing search process, estimating split quality, pruning and so on.