![]() | Only 14 pages are availabe for public view |
Abstract Classication problems often have a large number of features, but not all of them are useful for classication. Most of the real-valued datasets are high dimensional datasets which contain noisy (irrelevant and redundant) features. Such datasets may reduce the classication accuracy. Feature selection (FS) is the process of selecting a subset of relevant features, which can decrease the dimensionality, shorten the running time and improve the classication accuracy. Consequently, using FS helps to improve classication systems. There are three types of FS approaches, i.e. wrapper, lter and embedded approaches. Their main dierence is that wrappers use a classication algorithm to evaluate the goodness of the features during the FS process while lters are independent of any classi cation algorithm. The embedded approach relaxed the cost of the wrapper approach by incorporating the FS during the training step of the classier. FS is a dicult task because of feature interactions and the large search space. Existing FS methods suer from dierent problems, such as stagnation in local optima and high computational cost. Bio-inspired Optimization Algorithms (BIAs) are wellknown global search algorithms, computationally less expensive and can converge faster than other methods. BIAs have been successfully applied to many areas, but their potentials for FS has not been fully investigated. Whale Optimization Algorithm (WOA) and Salp Swarm Algorithm (SSA) are BIAs that are computationally less expensive and can converge faster than other methods. WOA and SSA have been successfully applied to many areas, but their potential for FS has not been fully investigated. The overall goal of this thesis is to investigate and improve the capability of BIAs for FS to select a smaller number of features and achieve similar or better classication performance than using all features. To overcome these limitations, this thesis suggests a new four BIAs wrapper FS models namely: Improved Whale Optimization Algorithm (IWOA) , Improved Salp Swarm Algorithm (ISSA), Chaotic Salp Swarm Algorithm (CSSA) and Fuzzy Mutual Information Binary Salp Swarm Algorithm (FMIBSSA). Based on benchmark datasets, the proposed models were evaluated according to the number of selected feature, classication accuracy, tness values and run-time. The experimental results conrm that our proposed models achieve better results, improve the classication performance, reduce the number of features and decrease computational time. |