Integrated Approaches to Dimensionality Reduction and System Identification using Fuzzy Rule Based Framework

• Main Authors: Yi-Cheng Chen, 陳毅誠
• Other Authors: I-Fang Chung
• Format: Others
• Language: en_US
• Published: 2012
• Online Access: http://ndltd.ncl.edu.tw/handle/16418304903983943438

博士 === 國立陽明大學 === 生物醫學資訊研究所 === 100 === Abstract The present era is an era of biological sciences which has been generating a huge amount of data. Those data usually requires specialized modeling and analysis tools. Computational intelligence tools, such as fuzzy logic, neural networks, and evolutionary computing, have been increasingly used in modeling and analysis of such data. In addition, such biological data almost always have many features, which may lead to enhanced data acquisition time and cost, more design time, more decision making time, and other increased expenditures in cost, time and effort. Hence reducing the dimensionality, if possible, is always desirable through feature selection. In this study, we have developed two kinds of integrated mechanisms for simultaneous construction of a fuzzy rule based system and selection of useful features. In our first system, called Fuzzy Systems - Feature Attenuating Gates (FS-FAG), we have introduced the concept of feature modulator to help select useful features and drop indifferent and derogatory features. For each feature in such a system, an associated feature modulator (or a gate function) is used to act like a gate to prevent bad features from influencing the output results. However, the concept of a feature modulator can only help to drop indifferent and derogatory features, but cannot be used to further detect the redundancy of good features. As we know, if high correlation exists among some good features, we may also be able to achieve good results even though we only include a few of those good features as the input to the system. Thus, in our second study, Fuzzy Systems - Feature Attenuating Gates with Controlled Redundancy (FS-FAGCoR), by integrating a penalty concept into our original feature selection process, whereby a penalty is imposed on features which have high correlation or dependencies to others, the good features are able to be further winnowed down to a small subset consisting of the most effective features. Both for classification problems and regression problems, we have tried our methods on several commonly used data sets as well as on a synthetic data set. Using a 10-fold cross validation setup we have demonstrated the effectiveness of our methods. Here we have also shown the successful selection of important and relevant features. Note that, since fuzzy rule based systems have better interpretability, our constructed fuzzy systems can play a significant role in further explaining the relationships between features and the problem under consideration.