We develop a graph structure for feature vectors in machine learning, which we denote as a feature network (FN); this is different from sample-based networks, in which nodes simply represent samples. FNs reveal the underlying relationship among feature vector components and re-represent features as functions on a network. Our study focuses on using FN structures to extract underlying information and thus improve machine learning performance. Upon the representation of feature vectors as such functions, so-called graph signal processing, or graph functional analytic techniques can be implemented, consisting of analytic operations including differentiation and integration of feature vectors. Our motivation originated from a study using infrared spectroscopy data, where domain experts prefer using the second derivative information rather than the original data; this is an illustration of the potential power of understanding the underlying feature structure.
We begin by developing a classification method based on the premise that is assuming data from different classes (e.g., different cancer subtypes) will have distinct underlying graph structures, for graphs consisting of genes as nodes and gene covariances as edges. That is, a feature vector from one class will tend to be "smooth" on the related FN, and "fluctuate" in the other FNs. This method, using an entirely new set of features from standard ones, on its own proves to somewhat outperform SVM and KNN in classifying cancer subtypes in infrared spectroscopy data and gene expression data. We are effectively also projecting high-dimensional data into a low dimensional representation of graph smoothness, providing a unique way of data visualization.
Additionally, FNs represent new ways of thinking about data. With a graph structure for feature vectors, graphical functional analysis can be used to extract various types of information not apparent in the original feature vectors. Specifically, operations such as calculus, Fourier transforms, and convolutions can be performed on the graph vertex domain. We introduce a family of calculus-like operators in reproducing kernel Hilbert spaces for feature vector regularization to deal with two types of data deficiency, which we designate as noise and blurring. Such operations are generalized from widely used ones in computer vision. The derivative operations on feature vectors provide additional information by amplifying differences between highly correlated features. Integrating feature vectors smooths and denoises them. Applications show that those denoising and deblurring operators can improve classification algorithms.
The feature network with deep learning can be naturally extended to graph convolutional networks. We proposed a deep multiscale clustering structure with small learning complexity on general graph distance structures. This framework substantially reduces the number of parameters, and it allows the introduction of general machine learning algorithms such as SVM to feed-forward in this deep structure.
| Identifer | oai:union.ndltd.org:bu.edu/oai:open.bu.edu:2144/42062 |
| Date | 17 February 2021 |
| Creators | Mu, Xinying |
| Contributors | Kon, Mark A. |
| Source Sets | Boston University |
| Language | en_US |
| Detected Language | English |
| Type | Thesis/Dissertation |
Page generated in 0.0015 seconds