Article Preview
Top1. Introduction
Knowledge extraction plays a crucial role in machine learning and data mining. Many data mining tasks try to exploit the hidden information and knowledge within data for better perception of the underlying system. One of such tasks is classification. It provides a deep insight into the data and helps in better understanding and effective decision-making. However, the enormous data size and its associated noise, are the major obstacle in such a task. The noise in the data can be removed by some feature selection process (Guyon and Elisseeff, 2003; Dash and Liu, 1997). Feature selection is a process of selecting a subset of features from the original set of features without losing the characteristics and identity of an original object. It helps to reduce the unnecessary computation as most of the data collected are redundant and insignificant. Recently knowledge extraction focuses in the direction of interpretability of the knowledge.
Feature selections are done by different techniques (Puch et al., 1993; Inza et al., 2001; Battiti, 1994; Ledesma et al., 2008, De Jesús Rubio, 2017; De Jesús Rubio et.al., 2015). One of the best feature selection process and proper interpretation of classification knowledge is based on neuro-fuzzy technique (Basak, 1998; Chen et al., 2008; Dubois and Prade, 1980; Ghosh et al., 2014; Li, 2002; Toly, 2010; Yang, 2007). Neuro-fuzzy is a hybridization of the artificial neural network (ANN) and fuzzy logic, which exploits the best qualities of these two approaches. ANN is a massively parallel computing, data-driven, fault tolerance, self-adaptive, flexible computational tool with the capability of capturing nonlinear and complex underlying characteristics of any physical process. ANN exhibits excellent behavior in input-output mapping and can resolve complex artificial intelligence problems. However, ANN is back-box in nature i.e. it hides the knowledge within its network. ANN alone cannot interpret the hidden knowledge properly. Fuzzy logic uses linguistic symbols to manipulate the imprecise and ambiguous data and use IF-THEN rules to describe the system. These linguistic symbols are similar to the natural language and the rules using these symbols represent the knowledge in a more understandable way. To deal with the black-box nature of the neural network, it is combined with fuzzy logic.
Many neuro-fuzzy systems (NFS) for feature selection and rule base classification are proposed by different researchers. In the neuro-fuzzy system’s performance is characterized by its transparency and accuracy (Ishibuchi and Nojima, 2009; Wang et al., 2005; Jin, 2004; Jin et al., 1999) but most of them focus on the accuracy rather than transparency. Accuracy refers to the ability of the NFS to conscientiously represent the modeled system. The accuracy depends upon the closeness of the model with the real system. Higher accuracy means better closeness. Transparency refers to the ability of the NFS to express the behavior of the system in an understandable way. Representing discovered information in human understandable form facilitates better data visualization and understanding. Transparency is the main reason for choosing a neuro-fuzzy system for feature selection and classification as it transforms data (information) into human understandable knowledge.