An unsupervised feature extraction and fusion framework for multi-source data based on copula theory

Abstract

With the development of big data technology, people are increasingly facing the challenge of dealing with massive amounts of multi-source or multi-sensor data. Therefore, it becomes crucial to extract valuable information from such data. Information fusion techniques provide effective solutions for handling multi-source data and can be categorized into three levels: data-level fusion, feature-level fusion, and decision-level fusion. Feature-level fusion combines features from multiple sources to create a consolidated feature, enhancing information richness. This paper proposes an unsupervised feature extraction and fusion method for multi-source data that utilizes the R-Vine copula, denoted as CF. The method starts by performing kernel density estimation to extract each data source's marginal density and distribution. Next, the maximum spanning tree is employed to select a vine structure for each attribute, and the corresponding copulas are chosen using maximum likelihood estimation and the AIC criterion. The joint probability density of each attribute across all information sources can be obtained by utilizing the relevant vine structure and copulas, serving as the final fusion feature. Finally, the proposed method is evaluated on eighteen simulated datasets and six real datasets. The results indicate that compared to several state-of-the-art fusion methods, the CF method can significantly enhance the classification accuracy of popular classifiers such as KNN, SVM, and Logistic Regression.