Research On Feature Reduction Method Based On Data Field And Cloud Model

With the development of data acquisition equipment and software, people can get more and more data, and the dimension of the data is higher and higher. Most of the pattern recognition algorithms in the face of high dimensional data, the time complexity of computing is increased exponentially, but the recognition effect instead. It has become a hot research direction that how to reduce the high dimension data and extract the feature of the recognition degree to improve the accuracy of pattern recognition. Feature reduction is aims at the original feature evaluation or transform formed a new feature space, in order to reduce the effect of redundant features for pattern recognition. The existing feature reduction algorithms are less in the study of the uncertainty and randomness of the feature itself as well as the intrinsic distribution of multi-scale features. It is necessary to actively propose new technologies to better study the data implied information, describe the internal distribution of the data, and extract more can reflect the nature of the data characteristics.In view of this, this paper introduces the cognitive physics method, through the distribution of the data itself to measure the importance degree of characteristics. Trying to extract data features from the multi scale perspective and reduce the number of features. The proposed algorithm is applied to solve some practical problems. The concept of data field and cloud model in cognitive physics is used to explore the internal feature relations from different perspectivesab, extract the relevant spatial features, and reveal the information of the data. The data field is used to explore the importance of the characteristics and the correlation between different features. The data object is influenced by other objects, which makes the feature description uncertain. Cloud model is an important model for uncertainty, which can realize the transformation between quantitative data and qualitative concepts. Using cloud model to construct three levels of image feature extraction methods:point, content and concept. It enriches the application range of cognitive physics, and analyzes the quality, distance and influence factors of the data field in a comprehensive analysis. The key technologies such as adaptive evolution mechanism of the dynamic data field and uncertainty of the cloud model are deeply studied. The main research works are as follows:(1) Sigma factor ? of potential function is calculated by density estimation, but the density estimation time complexity is too high. At the same time, it must traverse the entire data set to compute the distance between distance within class and distance between classes. For some classifiers, more attention should be paid to the data points near the sample, rather than the entire data set. For example, samples in the vicinity of the hyperplane are more important than other samples in SVM classifier. By introducing the idea of maximal distance, a k-local maximum margin feature extraction algorithm (KLMM) is proposed. The algorithm represents the anisotropic characteristics through influence factor a adaptive selection. Combining this strategy with the maximum margin criterion, maximum interval structure learning is performed in a generalized data field. KLMM has achieved good classification results on the general data sets, which provides a new reference method for measuring the distance between data points.(2) The existing feature importance measurement algorithms are less in the physical distribution of data and the anisotropy of the data characteristics. To this end, a feature importance measurement method based on potential entropy (FRGDF) is proposed. The method considers the distribution of the data itself when measuring the importance of the feature. The data field is extended to generalized multidimensional data field, the data is mapped into a high dimensional space which has better separability. Related experiments and analysis on the general data sets show that the FRGDF algorithm can effectively eliminate the unimportant or noise characteristics and improve the performance of the classification algorithm. A high classification accuracy is achieved by the combination of the selected feature subset and a variety of classifiers. It shows that the FRGDF algorithm is independent of the specific classifier.(3) Under the framework of the data field, the sample quality m is an important parameter to calculate the point potential value. The initial quality of all data points in the FRGDF is 1. In fact, there is a relationship between the quality of the sample and the sample density around it. At the same time, the important feature extraction is not only to consider the distribution of the feature itself, but also to consider the correlation with other features and categories. A feature subset selection algorithm based on mutual information (DFMIFS) is proposed. When the data is projected into the data field, the data is divided into the grids. The correlations between the candidate feature and selected feature are calculated by the introduction of mutual information theory. Only those weakly dependent features are added to the feature subset. Experiments show that DFMIFS can further improve or maintain the accuracy of the classifier by fewer features than FRGDF.(4) The existing dimension reduction algorithms are lack of uniform description of the features uncertainty in different scales, and the researches on the transformation between quantitative feature points and qualitative concepts are not enough. To this end, the concept of cloud model in cognitive physics is introduced to study the feature uncertainty in different scales. A feature extraction algorithm based on the combined cloud model (FECCM) is proposed. Taking image edge feature extraction as an example, the uncertainties of feature in micro and macro are analyzed. On the micro level, the cloud model and canny operator are combined to extract the pixel level digital features of the image. The traditional canny operator needs to manually set the dual threshold, FECCM can automatically select the threshold. Taking the face age recognition in FG-NET database as an example, the effectiveness of FECCM is further indicated. Macroscopically, the feature detection problem is transformed into the process of knowledge concept extraction which can effectively utilize the detected image. The distributions of different scale samples are obtained, and the corresponding common concept clouds are extracted. Crack extraction results in the elimination of noise points and the edge sharpness are excellent performance.To sum up, this paper shows that the potential value of the data field indicates the distance between sample distribution and the mutual influence of samples. In order to further express the fuzziness and randomness of feature in the space distribution, the cloud model is introduced to represent the data objects in different scales, so that the extracted features are more universal. The cognitive physics method and the specific feature reduction strategy are combined together, which makes the research of the cognitive physics method more deeply, and puts forward a new reference model for the feature reduction.