Research On Feature And Instance Selection Method For High Dimensional And Multi-scenario Data

With the advent of the information age,all kinds of data have shown explosive growth,and unlabeled multi-view data can be seen everywhere.However,troubles also follow when using complex data sets for research.For example,the huge amount of data will lead to high computing costs and degrade model performance;as the data information becomes more and more abundant,noise points and outliers also increase;data loss due to uncontrollable factors such as equipment failures occurs from time to time.At this time,how to achieve data dimensionality reduction for different complex scenarios becomes extremely important.For multi-view datasets with missing cases,this paper proposes an adaptive graph learning-guided incomplete multi-view unsupervised feature selection method.In order to alleviate the problem of missing data,we use the intrinsic information of the sample to learn the latent structure of the feature space,and use it as an important prior knowledge for effective adaptive recovery of incomplete data,and embed this filling process into the feature selection process make them mutually reinforcing.Considering the trade-off between the redundancy and discriminability of the selected features,we improved the traditional orthogonal constraints and constructed a generalized regression model to complete the feature subspace learning.Based on the consistency and difference between views,we propose a multi-view manifold structure adaptive construction method,so that the potential information and structural relationship between data can be fully learned and utilized to ensure that the selected features are more representative.Considering the influence of outliers,this paper also proposes a multi-view data structure-based instance and feature co-selection model to achieve double dimensionality reduction of the dataset.In order to select representative instances and features,according to the consistency and difference of data structures under different views,we use the reconstruction matrix of common structure and the reconstruction matrix of view deviation structure to complete the description of the potential structure of instances between views.At the same time,this process is extended to the feature space to achieve simultaneous selection,and the influence of the outlier degree of sample points on reconstruction is explored by learning the graph structure model.In addition,considering the diversity of the selected sample subsets,we limit the reconstruction contribution of similar samples based on the graph structure model to ensure that the selected subsets are more representative of the overall data situation,and define a new multi-view instance score calculation method.