Research On Swarm Intelligence Feature Selection Algorithm For Small Sample(Medical) Data

In recent years,with the vigorous development of machine learning technology in the medical industry,traditional medical treatment is gradually transforming into an intelligent model.It is a new direction of intelligent medical development how to dig out the information of contribution to medicine from massive electronic medical records so as to realize accurate diagnosis.In this direction,feature selection,as an efficient data processing technology,also shines.It can screen out the most effective feature subset from the initial feature subset according to the classification model to achieve the purpose of dimension reduction.Feature selection methods and classification algorithms are usually established on the premise of sufficient sample size,but in the actual medical field,some data are difficult to collect or the cost of collection is huge,and there is usually a potential relationship between the features of medical data sets,which will lead to the decline of the prediction ability of classification models.In addition,with the rapid increase of feature dimensions,a lot of irrelevant and redundant information will be generated,which may greatly reduce the performance of classification model.The conventional feature selection method is difficult to select the optimal feature subset,resulting in low accuracy of classification algorithm.How to use feature selection method to effectively analyze small sample medical data sets is a hot issue in current research,which has strong practical significance.To solve the above problems,this thesis proposes a population intelligent feature selection algorithm to assist clinical diagnosis,improve the accuracy of classification model,and provide help for the recognition and diagnosis of rare diseases in medicine and the identification of related biomarkers.The main work is as follows:Since traditional feature selection methods cannot completely remove redundant features when dealing with high-dimensional data,and swarm intelligence algorithms are susceptible to the interference of a large number of irrelevant features in the process of finding the optimal feature subset,which leads to the reduction of the search accuracy of the model.We combine machine learning algorithms with swarm intelligence algorithms and propose a Ridge regression and Q-learning strategy based Bee Swarm Optimization(RQBSO).The algorithm uses the Ridge regression feature selection algorithm to eliminate the redundant features and improve the efficiency of the bee swarm algorithm search.Q-learning strategy is also used in the search process to improve the adaptability of bee swarm search.Compared with the traditional feature selection method,RQBSO improves the effectiveness of feature selection and the classification accuracy of the model.Since the RQBSO feature selection algorithm does not fully consider the correlation between features when initializing feature subsets,this thesis addresses this problem by proposing a new RQBSO algorithm based on Weighted Near-set Aggregating and Incremental Feature Selection(Weighted Near-set Aggregating based on LASSO-Incremental Feature Selection-RQBSO,WNAL-IFS-RQBSO).This algorithm can fully take into account the correlation between features when initializing feature subsets,thus avoiding the RQBSO algorithm from falling into local optimal solutions.Compared with the RQBSO algorithm,WNAL-IFS-RQBSO further improves the effectiveness of feature selection and the classification accuracy of small sample data.In this thesis,the existing feature selection algorithm is improved for small sample datasets in the medical field to improve the overall prediction accuracy of the model while also achieving the feature selection of predictors for some specific diseases and biomarkers,mining effective information and providing help for the identification and diagnosis of specific diseases in medicine.The effectiveness and robustness of the algorithm are verified by conducting experiments on several medical small sample datasets,and the feasibility of the study is also demonstrated.