Finally, it is worth noting that the classi cations were checked automatically. Many Web sites can be classi ed to more than one category; however, only the designated category is considered to be correct here.

11.4.3.2 Comparison with ACO-Based FRFS Table 11.5 shows the resulting degree of dimensionality reduction, performed via selecting informative keywords, by the standard fuzzy-rough method (FRFS) and the ACO-based approach (AntFRFS). AntFRFS is run several times, and the results averaged both for classi cation accuracy and number of features selected. It can be seen that both methods drastically reduce the number of original features. AntFRFS performs the highest degree of reduction, with an average of 14.1 features occurring in the reducts it locates. To see the effect of dimensionality reduction on classi cation accuracy, the system was tested on the original training data and a test dataset. The results are summarized in Table 11.6. Clearly, the fuzzy-rough methods exhibit better resultant accuracies for the test data than the unreduced method for all classi ers. This demonstrates that feature selection using either FRFS or AntFRFS can greatly aid classi cation tasks. It is of additional bene t to rule inducers as the induction time is decreased and the generated rules involve signi cantly fewer features. AntFRFS improves on FRFS in terms of the size of subsets found and resulting testing accuracy for QSBA and PART, but not for C4.5 and JRip. The challenging nature of this particular task can be seen in the overall low accuracies produced by the classi ers (perhaps due to over tting), though improved somewhat after feature selection. Both fuzzy-rough approaches require a reasonable fuzzi cation of the input data, while the fuzzy sets are herein generated by simple statistical analysis of the dataset with no attempt made at optimizing these sets. A ne-tuned fuzzi cation will certainly improve the performance of FRFS-based systems. Finally, it is worth noting that the classi cations were checked automatically. Many Web pages can be classi ed to more than one category; however, only the designated category is considered to be correct here.

Classi cation performance Original FRFS Train Test Train Test 95.89 44.74 86.30 57.89 100.0 39.47 82.19 46.05 72.60 56.58 78.08 60.53 95.89 42.11 86.30 48.68

This chapter has presented a fuzzy-rough method to aid the classi cation of Web content, with promising results. In many text categorization problems, feature weights within datasets are real valued, posing a problem for many feature selection methods. FRFS can handle this type of data without the need for a discretizing step beforehand. In particular, while retaining fewer attributes than the conventional crisp rough set-based technique, the work resulted in an overall higher classi cation accuracy.

The ever-increasing demand for dependable, trustworthy intelligent diagnostic and monitoring systems, as well as knowledge-based systems in general, has focused much of the attention of researchers on the knowledge-acquisition bottleneck. The task of gathering information and extracting general knowledge from it is known to be the most dif cult part of creating a knowledge-based system. Complex application problems, such as reliable monitoring and diagnosis of industrial plants, are likely to present large numbers of features, many of which will be redundant for the task at hand [268,310]. Additionally inaccurate and/or uncertain values cannot be ruled out. Such applications typically require convincing explanations about the inference performed. Therefore a method to allow automated generation of knowledge models of clear semantics is highly desirable. The most common approach to developing expressive and human-readable representations of knowledge is the use of if-then production rules [192]. Yet real-life problem domains usually lack generic and systematic expert rules for mapping feature patterns onto their underlying classes. The present work aims to induce low-dimensionality rulesets from historical descriptions of domain features that are often of high dimensionality. In particular, a recent fuzzy rule induction algorithm (RIA), as rst reported in [51], is taken to act as the starting point for this. The premise attributes of the induced rules are represented by fuzzy variables, facilitating the modeling of the inherent uncertainty of the knowledge domain. It should be noted, however, that the exibility of the system discussed here allows the incorporation of almost any rule induction algorithm that uses

Computational Intelligence and Feature Selection: Rough and Fuzzy Approaches, by Richard Jensen and Qiang Shen Copyright 2008 Institute of Electrical and Electronics Engineers