Article Preview
TopIntroduction
With the advancements in technology in the past few years, a tremendous growth in dimensionality and sizes of the collected data has been observed. Many dimensions or features are found out to be redundant and irrelevant thereby making the process of prediction or classification tedious.
Therefore, for the selection of an appropriate feature subset, researchers have been exploiting various feature selection techniques and their relevance with the output class or labels. The simplest method to implement feature selection is generating and evaluating all the subsets (2^n) i.e.; Exhaustive search method (Kokash, 2005). But the direct evaluation becomes extremely difficult as the number of features increases. Greedy algorithms work by constructing successively, space of solutions. For finding the global optimum of the objective function, it chooses the best available option at each stage of the algorithm (Kokash, 2005). However, heuristic algorithms usually work better than greedy when we are to reach a global optimum (Kokash, 2005).
Because of their unparalleled benefits like effective and inexpensive feedback and faster speed of arriving at good solutions, the heuristic search optimization techniques like Particle Swarm Optimization (PSO), Genetic Algorithm (GA) etc. have been popularly used for feature selection. Algorithms like GA (J.Yang and Honavar, 1998), PSO (Xue et al., 2013) and Firefly (Banati and Bajaj, 2011) have gained admiration due to their high performance (Fister et al., 2013).
However, it is observed that the Gravitational Search Algorithm (GSA) (Rashedi et al., 2009) has been explored comparatively less for selecting relevant features. The researchers (Sombra et al., 2013); Nagpal et al. (2017); Khajehzadeh et al. (2012) have reported that GSA is able to generate promising results and has the following advantages over other evolutionary techniques:
- 1.
GSA gives a better performance by finding a global optimum with a higher convergence rate;
- 2.
It uses less memory than PSO;
- 3.
It allows greater exploration and diversity over a single population than GA.
In most of the prior works, trial and error methods have been used for the initialization of certain parameters of GSA (Nagpal et al., 2017; Papa et al., 2011). A noteworthy speed advantage could be gained if these parameters could be varied during the optimization to demonstrate the local characteristics of the function being optimized (Almeida et al., 1999). Alberto Sombra et al. (Sombra et al., 2013) tried using fuzzy logic for tuning of parameter in GSA and was able to obtain a better convergence rate when the new algorithm was tested on a mathematical function. Use of fuzzy logic for dynamic adaptation of various parameters in GSA is yet to be explored in medical domains.
In this paper, use of Fuzzified GSA (FGSA) has been proposed and tested on few selected medical datasets. In FGSA, GSA is used for selecting the most relevant subset of features whereas Fuzzy logic is used for parameter tuning. K-Nearest Neighbor (KNN) has been used for the purpose of classification. To analyze the performance of the approach proposed in this work, openly available data sets pertaining to the medical domain have been used. The results of FGSA+KNN have been compared with the previous works wherein researchers have used PSO+KNN (Ghasemi et al., 2013) and GSA+KNN (Nagpal et al., 2017) for feature selection in the domains explored.
The results show that with the technique used in this paper, the selected features can be reduced to a maximum of 77.77% in the best case. Also, there was a noteworthy increase in accuracy with the best case increase in accuracy to be 27.3%, 20.8% and 1.5% as compared to KNN, PSO+KNN, and GSA+KNN, respectively.