Rbfaco And Its Application To Dimensionality Reduction Of High-dimensional Medical Data

Because of the informatization of the medical and health care,and the development of new approaches for data collection and storage,the type and scale of medical data are both in explosive growth.These medical data,which are often high-dimensional and containing irrelevant or redundant features,have sparse available information and a low density of information.What is worse,these characteristics would heavily affect the performance of data mining algorithms on these data sets and might even lead to a disaster called dimension curse.Feature selection,which is always used to reduce the dimension of the data set and improve the density of information,is a type of non-deterministic polynomial time problem.This problem makes the search space and the computing cost of feature selection increase exponentially when its number of features in data sets increases,leading to the difficulty to generate an effective solution in a high-dimensional data set.Nowadays,although ant colony optimization algorithm,which is a famous algorithm of swarm intelligent,is widely used in salving feature selection problem,it has some shortcomings.The standard fully connected ant colony optimization algorithm must give the stopping criteria in advance,the binary one loses the diversity of the generated solutions due to its fixed feature sequence,and the one based on a mixed graph has larger search space and computing cost.Therefore,in order to improve the performance of this algorithm and reduce its time complexity and the cost of computation,the study of ant colony optimization algorithm and its representation is still of great significance.The main work of this paper are as follows:(1)The representation ways of ant colony optimization algorithm are analyzed in this thesis,and the limitation of existing ant colony optimization algorithm based on the standard graph,binary graph and the mixed graph are pointed out.A new solution to solve feature selection problem using ant colony optimization algorithm is designed.(2)A new presentation,which is called random binary fully-connected ant colony optimization algorithm,is proposed in this thesis.In this representation,when the ant is observing the accessible feature nodes,all the sub-nodes of features are hidden.After the next feature to visit is confirmed,the sub-nodes of corresponding feature are visible to the ant.(3)The random binary fully-connected ant colony optimization algorithm is proposed,and the core elements of this algorithm are designed in detail.In this algorithm,based on the newly proposed accessible probability and the rule of minimum redundancy-maximum relevances,a novel two-step feature node selection mechanism is proposed.The time complexity and computing cost of this algorithm in generating solutions are analyzed.(4)The performance of the existing ant colony optimization algorithm based on the mixed graph and the random binary fully-connected ant colony optimization algorithm proposed in this thesis in solving feature selection problem were compared.To verify the validity of the proposed random binary fully-connected ant colony optimization algorithm,experiments were carried out on six medical data sets with gradient changing number of features,and the experimental results were analyzed.