Research On Tabu Search Algorithm-based Feature Selection

With the continuous expansion of the scope of computer applications and the Internet's global popularity,more and more data has accumulated in a variety of application systems,and this creates the "data explosion" problem. It really has come to our side, even more serious "data avalanche" is also moving toward us, which need to develop better countermeasures to avoid being buried alive in the data. In recent years, data mining has aroused great concern in the IT industry, it is found that to use data mining effectively, data pre-processing is essential. And the feature selection is the method often used in data pre-processing.In recent years, a variety of intelligent optimization algorithms continue to emerge,which includes Tabu Search algorithm, the researchers found that when it was applied to Feature Selection, the result would get satisfactory. Then Tabu Search-based Feature Selection was proposed, but the researchers did not study it very well and paid more attention to research of its application. About how to improve the function of the algorithm itself was less than study. This paper is to research how to further improve the method and thus to expand its range of applications.I get three improvements on this algorithm that is on the basis of the study of Tabu Search:(1)As we all know, Tabu Search algorithm has very strong dependence on the initial solution. Good quality initial solution can help complete the search tasks more quickly. Here I have a mixture of Genetic algorithms to generate high-quality materials related to the initial solution. On this basis,we use Tabu Search algortihm for finding the optimal solution.(2)As the objective function have a great impact on the search process of Tabu Search algortihm, so set the appropriate objective function is extremely critical. Here I put the function that is made by Muhammad Atif Tahir to use in this algortihm. It is taking two elements into account that are the classification accuracy rate and characteristics of the dimension. This will not only improve the quality of the characteristics of the final solution, but also conpress feature dimension, then reduce the time consumed by classifier.(3)Finally, I put the adaptive search strategy proposed by He Yi to use in the practice of the algortihm. First, the candidate solutions are divided into half of the concentrated search element of K and half of the diversity search element of K', then adjust the size of K basing on whether the current solution is good. K=K-1or K=K+1. So that wo can adaptively balance the concentrated search and the diversity search.This article use the final design of the KNN classifier's classification accuracy and time-consuming to determine that whether the result of the improved algortihm is better,long as it can improve the function of the algorithm we should accept.Figure 16 table 2 references 53...