Files in this item



application/pdfQIAN-DISSERTATION-2015.pdf (5MB)
(no description provided)PDF


Title:Unsupervised feature analysis for high dimensional big data
Author(s):Qian, Mingjie
Director of Research:Zhai, ChengXiang
Doctoral Committee Chair(s):Zhai, ChengXiang
Doctoral Committee Member(s):Han, Jiawei; Roth, Dan; Hong, Liangjie
Department / Program:Computer Science
Discipline:Computer Science
Degree Granting Institution:University of Illinois at Urbana-Champaign
Subject(s):Feature selection
unsupervised feature selection
multi-view learning
multi-view topic discovery
multi-view unsupervised feature selection
Abstract:In practice we often encounter the scenario that label information is unavailable due to either high cost of manual labeling or unwillingness of users to label. When label information is not available, traditional supervised learning can not be directly applied so we need to study unsupervised methods which could work well even without supervision. Feature analysis has been proven effective and important for many applications. Feature analysis is a broad research field, whose research topics includes but are not limited to feature selection, feature extraction, feature construction, and feature composition e.g., in topic discovery the learned topics can be viewed as compound features. In many real systems, it is often necessary and important to do feature analysis to determine which individual or compound features should be used for posterior learning tasks. The effectiveness of traditional feature analysis often relies on labels of the training data examples. However, in the era of big data, label information is often unavailable. In the unsupervised scenario, it is more challenging to do feature analysis. Two important research topics in unsupervised feature analysis are unsupervised feature selection and unsupervised feature composition, e.g., to discover topics as compound features. This would naturally create two lines for unsupervised feature analysis. Also, combined with single-view or multiple-view for the data, we would generate a table with four cells. Except for the single-view feature composition (or topic discovery) where there're already many work done e.g., PLSA, LDA, and NMF, the other three cells correspond to new research topics, and there is few work done yet. For single view unsupervised feature analysis, we propose two unsupervised feature selection methods. For multi-view unsupervised feature analysis, we focus on text-image web news data and propose a multi-view unsupervised feature selection method and a text-image topic model. Specifically, for single-view unsupervised feature selection, we propose a new method that is called Robust Unsupervised Feature Selection (RUFS), where pseudo cluster labels are learned via local learning regularized robust NMF and feature selection is performed simultaneously by robust joint $l_{2, 1}$-norm minimization. Outliers could be effectively handled and redundant or noisy features could be effectively reduced. We also design a (projected) limited-memory BFGS based linear time iterative algorithm to efficiently solve the optimization problem. We also study how the choice of norms for data fitting and feature selection terms affect the ultimate unsupervised feature selection performance. Specifically, we propose to use joint adaptive loss and $l_2/l_0$ minimization for data fitting and feature selection. We mathematically explain desirable properties of joint adaptive loss and $l_2/l_0$ minimization over recent unsupervised feature selection models. We solve the optimization problem with an efficient iterative algorithm whose computational complexity and memory cost are linear to both sample size and feature size. For multiple-view unsupervised feature selection, we propose a more effective approach for high dimensional text-image web news data. We propose to use raw text features in label learning to avoid information loss. We propose a new multi-view unsupervised feature selection method in which image local learning regularized orthogonal nonnegative matrix factorization is used to learn pseudo labels and simultaneously robust joint $l_{2,1}$-norm minimization is performed to select discriminative features. Cross-view consensus on pseudo labels can be obtained as much as possible. For multi-view topic discovery, we study how to systematically mine topics from high dimensional text-image web news data. The application problem is important because almost all news articles have one picture associated. Unlike traditional topic modeling which considers text alone, the new task aims to discover heterogeneous topics from web news of multiple data types. We propose to tackle the problem by a regularized nonnegative constrained $l_{2,1}$-norm minimization framework. We also present a new iterative algorithm to solve the optimization problem. The proposed single-view feature selection methods can be applied on almost all single-view data. The proposed multi-view methods are designed to process text-image web news data, but the idea can be naturally generalized to analyze any multi-view data. Practitioners could run the proposed methods to select features that will be used in posterior learning tasks. One can also run our multi-view topic model to analyze and visualize topics in text-image web news corpora to help interpret the data.
Issue Date:2015-07-17
Rights Information:Copyright 2015 Mingjie Qian
Date Available in IDEALS:2015-09-29
Date Deposited:August 201

This item appears in the following Collection(s)

Item Statistics