PROJECT TITLE :

Unsupervised Spectral Feature Selection with Dynamic Hyper-graph Learning

ABSTRACT:

In order to produce interpretable and discriminative results from unsupervised spectral feature selection (USFS) methods, an embedding of a Laplacian regularizer within the framework of sparse feature selection was necessary. This was done in order to maintain the local similarity of the training samples. In order to accomplish this, USFS methods will typically construct the Laplacian matrix by employing either a general-graph or a hyper-graph on the data that was initially collected. In most cases, a general graph is able to measure the relationship between two samples, whereas a hypergraph is able to measure the relationship between at least two samples. The general graph is obviously a special case of the hypergraph, but the hypergraph may be able to capture a more intricate structure of the samples than the general graph. The construction of the Laplacian matrix, on the other hand, was treated as a step distinct from the procedure of feature selection in older USFS methodologies. In addition, there is typically some noise in the original data. Each of them makes it more difficult to output feature selection models that can be relied on. Within the context of sparse feature selection, this paper presents a novel method for the selection of features that is accomplished through the dynamic construction of a hyper-graph-based Laplacian matrix. Our proposed method outperformed the state-of-the-art methods in terms of both the clustering and segmentation tasks, as shown by experimental results based on real datasets.