Multi-view non-negative matrix factorization by patch alignment
framework with view consistency
Weihua Ou
a,
n
, Shujian Yu
b
, Gai Li
c
, Jian Lu
d
, Kesheng Zhang
e
, Gang Xie
a
a
School of Big Data and Computer Science, Guizhou Normal University, Guiyang 550001, China
b
Department of Electrical and Computer Engineering, University of Florida, Gainesville, FL 32601, USA
c
Department of Electronics and Information Engineering, Shunde Polytechnic, Foshan 528300, China
d
School of Information Science and Engineering, Southeast University, Nanjing 210096, China
e
School of Information Engineering, Guizhou Institute of Technology, Guiyang 550003, China
article info
Article history:
Received 9 March 2015
Received in revised form
13 September 2015
Accepted 15 September 2015
Available online 9 April 2016
Keywords:
Multi-view non-negative matrix factoriza-
tion
Patch alignment framework
Geometric structure
Locally linear embedding
View consistency
abstract
Multi-view non-negative matrix factorization (NMF) has been developed to learn the latent repre-
sentation from multi-view non-negative data in recent years. To make the representation more mean-
ingful, previous works mainly exploit either the consensus information or the complementary infor-
mation from different views. However, the latent local geometric structure of each view is always
ignored. In this paper, we develop a novel multi-view NMF by patch alignment framework with view
consistency. Different from previous works, we take the local geometric structure of each view into
consideration, and penalize the disagreement of different views at the same time. More specifically, given
a data in each view, we construct a local patch utilizing locally linear embedding to preserve its local
geometrical structure, and obtain the global representation under the whole alignment strategy.
Meanwhile, for different views, we make the representations of views to approximate the latent
representation shared by different views via considering the view consistency. We adopt the
correntropy-induced metric to measure the reconstruction error and employ the half-quadratic techni-
que to solve the optimization problem. The experimental results demonstrate the proposed method can
achieve satisfactory performance compared with single-view methods and other existing multi-view
NMF methods.
& 2016 Elsevier B.V. All rights reserved.
1. Introduction
In real applications, the images or instances can be represented
by multi-view features, such as color, shape, texture and so on. For
example, a web page can be described by visual features from the
image and the text features surrounding the image. Different fea-
tures provide consistent and complementary information, and
integrating multi-view features has been shown to outperform
using only a single view feature [1,2]. A simple method to exploit
multi-view features is to concatenate all the features into a single
one, and then apply existing algorithms to this single feature.
Obviously, this approach ignores the differences of statistical prop-
erties between different views. By exploring the consistency and
complementary properties of different views, multi-view learning is
more effective and has better generalization capability than single-
view learning. The representative multi-view learning methods
include co-training, multiple kernel learning and subspace learning
[2].Co-training[3–5] trains the classifiers by maximizing the
mutual agreement on two distinct views of the unlabeled data.
Multiple kernel learning [6–8] combines different kernels to
improve the performance based on the fact that different kernels
correspond to different views. Subspace learning-based method
[9–11] aims to obtain a latent subspace based on the assumption
that different views are generated from this latent subspace. For
example, canonical correlation analysis (CCA) obtains the latent
subspace via maximizing the correlation between different views.
Though these approaches are successful in multi-view learning, the
multi-view features are often non-negative in real applications
[2,12].
Traditional non-negative matrix factorization (NMF) [13],
which has been widely used in non-negative features extraction
[14–16], can only deal with single-view data. Works on extending
NMF to multi-view data has attracted much attention. For exam-
ple, Liu et al. [17] proposed a joint non-negative matrix factor-
ization for multi-view data clustering based on the relationship
between NMF and probabilistic latent semantic analysis. By posing
a novel normalization strategy and regularization, this multi-view
NMF (MultiNMF) method uncovered the latent subspace shared by
Contents lists available at ScienceDirect
journal homepage: www.elsevier.com/locate/neucom
Neurocomputing
http://dx.doi.org/10.1016/j.neucom.2015.09.133
0925-2312/& 2016 Elsevier B.V. All rights reserved.
n
Corresponding author. Tel.: þ86 18198247234.
E-mail address: ouweihuahust@gmail.com (W. Ou).
Neurocomputing 204 (2016) 116–124