| Feature selection is a typical technique of data dimensionality reduction,which has been widely used in many practical problems such as image classification and disease diagnosis.In recent years,with the rapid development of edge computing and artificial intelligence,data describing the same learning task in real life may be stored in different institutions(called participants)in distributed mode,and all the participants cannot share these data due to the restriction of privacy protection.The feature selection algorithm for such scenarios is called federated feature selection algorithm for short.Since participants cannot share any sensitive information,the existing traditional feature selection algorithms for centralized data storage and the distributed feature selection algorithms for data sharing environment are no longer applicable.In view of this,this thesis studies the federated particle swarm feature selection algorithms by combining the global search ability of evolutionary optimization technology in two scenarios of vertical and horizontal data distribution.(1)Considering the scenario where the data is horizontally distributed among all participants and and participants don’t include imbalanced classes,a horizontal federated particle swarm feature selection method with trusted third party is proposed.Firstly,a trusted third party is introduced to process and integrate the feature subsets obtained from each participant,and the basic framework of federated evolutionary feature selection is established for the first time.Then,based on the framework,particle swarm optimization is used to search the optimal feature subset on each participant,and a federated evolutionary feature selection algorithm based on particle swarm optimization is proposed.By designing some new operators suitable for federated learning,such as the privacy-protected cooperative integration strategy and the integration result-guided swarm initialization strategy,the ability of particle swarm optimization to deal with federated feature selection is significantly improved.The proposed algorithm is applied to 15 test data sets,and compared with 7 typical filtering or wrapper integrated feature selection algorithms.Experimental results show that the proposed algorithm can significantly improve the classification performance of the feature subset selected by each participant under the premise of protecting the data privacy.(2)Considering the scenario where data is horizontally distributed among all participants and participants include imbalanced classes,a horizontal federated particle swarm feature selection method with joint sample filling mechanism is proposed.Firstly,by effectively fusing the sample distribution information of multiple participants,a multi-layer joint sample filling strategy with multi-party participation,namely sampling-rough selection-fine tuning strategy,is proposed to fill the imbalanced or empty classes on each participant while ensuring the data privacy.Then,a multi-participant federated evolutionary feature selection algorithm based on particle swarm optimization is proposed by periodically sharing the optimal feature subset obtained by the particle swarm optimization algorithm on each participant.The proposed algorithm is applied to 18 test data sets,and compared with 7 typical filtering or wrapper integrated feature selection algorithms.Experimental results show that the proposed algorithm can significantly improve the ability of all participants to deal with imbalanced data and significantly improve the classification accuracy of the selected feature subsets on the basis of sufficient to protect the privacy of data.(3)Considering the scenario where data is vertically distributed among all participants and only one participant contains label information,an embedded vertically federated feature selection method based on particle swarm optimization is proposed.Firstly,by integrating particle swarm feature selection technology into the Secure Boost framework,the basic framework of embedded vertical federated feature selection is established.Then,a feature importance integration ranking strategy based on XGBoost tree model is proposed to reduce the search space of subsequent particle swarm optimization algorithms.A particle hybrid encoding strategy is proposed to simultaneously optimize the hyper-parameters and the corresponding feature subsets of XGBoost tree model.In order to improve the quality of initial population,a particle swarm initialization strategy guided by feature importance degree is designed.The proposed algorithm is applied to 10 test datasets,and compared with 3 typical vertically federation learning algorithms.Experimental results show that the proposed algorithm can significantly improve the classification performance of feature subsets selected by each participant on the basis of fully protecting the data privacy of participants.(4)Furthermore,the proposed horizontal federated feature selection method is applied to multi-participant joint medical diagnosis problem,and a disease diagnosis model construction method with feature selection is given to verify the ability of the proposed method to solve practical problems.Compared with the single diagnosis model without feature selection,the joint diagnosis model without feature selection,and the single diagnosis model with feature selection,experimental results show that the proposed method can build a better disease diagnosis model and produce high-precision diagnosis results on the basis of fully protecting data privacy.The above research results successfully extend evolutionary optimization to the feature selection problem with multi-participation under privacy protection,improve the performance of the evolutionary feature selection method,enrich the theory and method of feature selection,expand the application scenario and scope of the feature selection method,and have important theoretical significance and practical value.This thesis has 26 pictures,26 tables and 150 references. |
