| In the era of big data information,complex high-dimensional data is gradually increasing,so it is an inevitable choice to effectively extract the key features of data.With the increase of data dimension,traditional feature learning methods are difficult to capture the latent dependence between data dimension information in the feature extraction of static spatial data and time-series data.Therefore,this paper proposes two feature learning methods to solve the above problems,which can effectively realize the feature extraction of complex high-dimensional data.The main contributions of this paper are as follows:(1)Aiming at the problem that it is difficult to effectively obtain the spatial structure of data caused by information dispersion,a feature learning method is proposed based on the spatial information depth aggregation.The proposed method utilizes the Cauchy-Schwarz regular term to capture the spatial dependence of data in latent space and preserve the spatial structure of data manifold.At the same time,the method proposes a spatial information aggregation strategy to realize the weighted aggregation of data features and extract the global structure information of data.By capturing the deep spatial dependence and extracting the global features of data,this method solves the problems of local capture of spatial features and disentanglement of features to improve the effect of feature learning.(2)Aiming at the problem that it is difficult to obtain the latent feature correlations between time-series data and high estimation uncertainty caused by information dispersion,a feature learning method is proposed based on the temporal information depth aggregation.The proposed method utilizes a temporal information aggregation strategy to extract the temporal feature between data dimensions.The aggregation strategy captures the temporal dependence between observation data in continuous time and accurately models the inherent structure of time-series data;At the same time,the method designs a neural Kalman filter strategy to integrate the uncertainty estimation into the temporal depth modeling to update the latent state of the data in real-time and reduce the uncertainty of the estimation.This method solves the problems of significant deviation of data state estimation and minor feature receptive field by modeling target temporal dependence and target state uncertainty.This method realizes feature extraction and state prediction of time-series data,and improves the ability of feature learning of time-series data.(3)The experimental verification and data analysis of the two methods are carried out on datasets.The experimental results show that the two methods have better performance of obtaining latent features.The method based on spatial information depth aggregation proves the feature learning ability of static spatial data,and the method based on temporal information depth aggregation reflects the effectiveness and superiority of feature extraction of time-series data.In conclusion,the feature learning methods for high-dimensional complex data proposed in this paper can learn latent features of data and improve the effectiveness of feature learning to provide a solid foundation for a wide range of downstream tasks. |
