Catálogo de publicaciones - libros

Feature selection refers to the problem of selecting those input features that are most predictive of a given outcome; a problem encountered in many areas such as machine learning, pattern recognition and signal processing. In particular, solution to this has found successful application in tasks that involve datasets containing huge numbers of features (in the order of tens of thousands), which would be impossible to process further. Recent examples include text processing and web content classification. Rough set theory has been used as such a dataset pre-processor with much success, but current methods are inadequate at finding reductions, the smallest sets of features possible. This paper proposes a technique that considers this problem from a propositional satisfiability perspective. In this framework, minimal subsets can be located and verified. An initial experimental investigation is conducted, comparing the new method with a standard rough set-based feature selector.

We present a study on a rough set based approach for feature selection. Instead of using significance or support, Parameterized Average Support Heuristic (PASH) considers the overall quality of the potential set of rules. It will produce a set of rules with balanced support distribution over all decision classes. Adjustable parameters of PASH can help users with different levels of approximation needs to extract predictive rules that may be ignored by other methods. This paper finetunes the PASH heuristic and provides experimental results to PASH.

Most existing rough set-based feature selection algorithms suffer from intensive computation of either discernibility functions or positive regions to find attribute reduct. In this paper, we develop a new computation model based on relative attribute dependency that is defined as the proportion of the projection of the decision table on a subset of condition attributes to the projection of the decision table on the union of the subset of condition attributes and the set of decision attributes. To find an optimal reduct, we use information entropy conveyed by the attributes as the heuristic. A novel algorithm to find optimal reducts of condition attributes based on the relative attribute dependency is implemented using Java, and is experimented with 10 data sets from UCI Machine Learning Repository. We conduct the comparison of data classification using C4.5 with the original data sets and their reducts. The experiment results demonstrate the usefulness of our algorithm.

Consistent extensions of information systems have been earlier considered in the literature. Informally, a consistent extension of a given information system includes only such objects corresponding to known attribute values which are consistent with the whole knowledge represented by rules extracted from the information system. This paper presents a new look at consistent extensions of information systems focusing mainly on partially consistent extensions and broadening the approach proposed earlier by Z. Suraj. To this end, a notion of a partially consistent extension of an information system is introduced. The meaning, properties and application examples of such extensions are given. In the approach presented, we admit the situation that some objects in an extension are consistent only with a part of the knowledge extracted from the information system. We show how a factor of consistency with the original knowledge for a given object from an extension can be computed. Some coefficients concerning rules in information systems are defined to compute a factor of consistency. The notions presented are crucial for solving different problems. An example is given in the paper to show an application of the proposed approach in modelling of concurrent systems described by information systems.

According to the sizes of the attribute set and the information table, the information tables are categorized into three types of Rough Set problems, Pattern Recognition/Machine Learning problems, and Statistical Model Identification problems. In the first Rough Set situation, what we have seen is as follows: 1) The ”granularity” should be taken so as to divide equally the unseen tuples out of the information table, 2) The traditional ”Reduction” sense accords with the above insistence, and 3) The ”stable” subsets of tuples, which are defined through a ”Galois connection” between the subset and the corresponding attribute subset, may play an important role to capture some characteristics that can be read from the given information table. We show these with some illustrative examples.

In incomplete data missing attribute values may be universally interpreted in several ways. Four approaches to missing attribute values are discussed in this paper: lost values, ”do not care” conditions, restricted ”do not care” conditions, and attribute-concept values. Rough set ideas, such as attribute-value pair blocks, characteristic sets, characteristic relations and generalization of lower and upper approximations are used in these four approaches. A generalized rough set methodology, achieved in the process, may be used for other applications as well. Additionally, this generalized methodology is compared with other extensions of rough set concepts.

Minimal rule generation in - (), which follows rough sets based rule generation in (), is presented. According to and in , and are defined. are also introduced into for generating minimal certain rules. Like minimal rule generation in , the condition part of a minimal certain rule is given as a solution of an introduced discernibility function. As for generating minimal possible rules, there may be lots of discernibility functions to be solved. So, an algorithm based on an order of attributes is proposed. A tool, which generates minimal certain and minimal possible rules, has also been implemented.

Rough Set Theory is a mathematical tool to deal with vagueness and uncertainty. Rough Set Theory uses a single information table. Relational Learning is the learning from multiple relations or tables. This paper studies the use of Rough Set Theory and Variable Precision Rough Sets in a multi-table information system (MTIS). The notion of approximation regions in the MTIS is defined in terms of those of the individual tables. This is used in classifying an example in the MTIS, based on the elementary sets in the individual tables to which the example belongs. Results of classification experiments in predictive toxicology based on this approach are presented.

The rough relational database model was developed for the management of uncertainty in relational databases. In this paper we discuss rough functional dependencies and the normalization process used with them. Normalization is an important part of the relational database design process and rough normalization provides similar benefits for the rough relational database model.

The article introduces the basic ideas and investigates the probabilistic version of rough set theory. It relies on both classification knowledge and probabilistic knowledge in analysis of rules and attributes. One-way and two-way inter-set dependency measures are proposed and adopted to probabilistic rule evaluation. A probabilistic dependency measure for attributes is also proposed and demonstrated to have the monotonicity property. This property makes it possible for the measure to be used to optimize and evaluate attribute based-representation through computation of attribute reduct, core and significance factors.