Attribute Selection Method Based On Binary Ant Colony Optimization And Fractal Dimension

Feature Selection focus on eliminating the redundant information and extracting the key feature, but keeping the main value of original data. Therefore, it plays an important role in big data preprocessing. Feature Selection includes four aspects: subset evaluation principle, searching strategy, verification process and stopping condition.Obviously, subset evaluation principle and searching strategy are two key factors. When handling the feature selection problem, any feature has two states: be selected as key feature or not. Herein, number 1 is involved to represent that the feature has been selected, while number 0 represents not. On this basis, starting from one-dimensional cell automata model, Binary Ant Colony optimization (BACO) is introduced as searching strategy, Fractal Dimension (FD) is used as subset evaluation principle in this paper. For the shortcomings of easily dropping into local optima, pheromone deficiency at the initial stage and can not handle a batch of tasks at the same time, several kinds of methods are introduced to BACO to improve its performance.The main research work and contribution are summarized as follows:(1) Incorporating the ant life-cycle (including born, forage, breed, migrate, die) to BACO, nutrition threshold is introduced and the migrate operation, breed operation and die operation is performed to expand the searching space and to keep the diversity of population. A life-cycle based Binary Ant Colony Optimization (LCBBACO) is proposed. Then LCBBACO combine with FD is involved to deal with feature selection problem. The experimental results on 6 UCI datasets show that the method has preferable feasibility and effectiveness.(2) For the equal distribution of pheromone and no heuristic information at the initial stage, BACO have to take a long time to form a routine with distinct pheromone.This is directly influence the running time of the whole algorithm. However, Binary Particle Swarm Optimization (BPSO) has been widely used due to its fast convergence.Combining the advantage of BPSO with BACO, a Binary Ant Colony Optimization with particle feature was proposed (PBACO). Then PBACO is used as searching strategy, FD is involved as subset evaluation principle for feature selection problem.Experimental results manifest that the proposed method cannot only address the feature selection problem effectively, but also largely reduce the number of evaluation and shorten the running time of the whole process.(3) Traditional BACO is usually used to solve a single feature seletion problem.For the character of implict parallelism, no research work has been conducted to address multiple distinct feature selections simultaneously by combining them into a single unified search space. In this paper, the mechanics of mutualistic coevolution is introduced to BACO for exploiting the commonalities and/or complementarities between different (yet possibly related) optimization tasks in a single multitasking environment. Then one task corresponding to one subpopulation, different subpopulation cooperate with each other, useful information transfer from one task to another. This cannot only enhance the solution quality, but also speed up the convergence trend. Then Coevolutionary Binary Ant Colony Optimization (CBACO) is proposed. By combing CBACO with Fractal Dimension, a novel method is proposed to solve multiple feature selection problems at the same time. The experimental results show that useful information transfer in the multitasking environment can speed up the convergence and enhance the solution quality.(4) Haze has brought great harm to human daily life, so it is very important and meaningful to forecast the haze day. For the orginial haze datas include redundant features, this leading to resource waste and inaccurate estimatimation. Therefore,feature reduction and nosiy data elimination is a tremendous need. Each feature in haze data has two states, be selected as key feature or not. Hence, kicking out the noisy features for haze data is essentially for feature selection problem. The aforementioned CBACO and Fractal Dimension have good performance in feature selection multitasking environment. Therefore, CBACO and FD are involved to deal with mulitiple haze data feature selection problem. Then the compact data and SVM are used to forecast the haze of the two cities, Beijing and shanghai. This provides an important reference for the effective prediction of haze. Numercial experiments reveal that our method has higher forecasting accuracy and provide a good tool for environmental improvement.