Conference PaperPDF Available


In this paper, we present {GraRep}, a novel model for learning vertex representations of weighted graphs. This model learns low dimensional vectors to represent vertices appearing in a graph and, unlike existing work, integrates global structural information of the graph into the learning process. We also formally analyze the connections between our work and several previous research efforts, including the DeepWalk model of Perozzi et al. as well as the skip-gram model with negative sampling of Mikolov et al. We conduct experiments on a language network, a social network as well as a citation network and show that our learned global representations can be effectively used as features in tasks such as clustering, classification and visualization. Empirical results demonstrate that our representation significantly outperforms other state-of-the-art methods in such tasks.
GraRep: Learning Graph Representations with
Global Structural Information
Shaosheng Cao
Xidian University
Wei Lu
Singapore University of
Technology and Design
Qiongkai Xu
IBM Research - China
Australian National University
In this paper, we present GraRep, a novel model for learn-
ing vertex representations of weighted graphs. This model
learns low dimensional vectors to represent vertices appear-
ing in a graph and, unlike existing work, integrates global
structural information of the graph into the learning process.
We also formally analyze the connections between our work
and several previous research efforts, including the Deep-
Walk model of Perozzi et al. [20] as well as the skip-gram
model with negative sampling of Mikolov et al. [18]
We conduct experiments on a language network, a social
network as well as a citation network and show that our
learned global representations can be effectively used as fea-
tures in tasks such as clustering, classification and visualiza-
tion. Empirical results demonstrate that our representation
significantly outperforms other state-of-the-art methods in
such tasks.
Categories and Subject Descriptors
I.2.6 [Artificial Intelligence]: Learning; H.2.8 [Database
Management]: Database Applications - Data Mining
General Terms
Algorithms, Experimentation
Graph Representation, Matrix Factorization, Feature Learn-
ing, Dimension Reduction
In many real-world problems, information is often orga-
nized using graphs. For example, in social network research,
classification of users into meaningful social groups based on
social graphs can lead to many useful practical applications
such as user search, targeted advertising and recommenda-
tions. Therefore, it is essential to accurately learn useful
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full cita-
tion on the first page. Copyrights for components of this work owned by others than
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re-
publish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from
CIKM’15, October 19–23, 2015, Melbourne, Australia.
2015 ACM. ISBN 978-1-4503-3794-6/15/10 ...$15.00.
information from the graphs. One strategy is to learn the
graph representations of a graph: each vertex of the graph
is represented with a low-dimensional vector in which mean-
ingful semantic, relational and structural information con-
veyed by the graph can be accurately captured.
Recently, there has been a surge of interest in learning
graph representations from data. For example, DeepWalk
[20], one recent model, transforms a graph structure into a
sample collection of linear sequences consisting of vertices
using uniform sampling (which is also called truncated ran-
dom walk). The skip-gram model [18], originally designed
for learning word representations from linear sequences, can
also be used to learn the representations of vertices from such
samples. Although this method is empirically effective, it is
not well understood what is the exact loss function defined
over the graph involved in their learning process.
In this work, we first present an explicit loss function of
the skip-gram model defined over the graph. We show that
essentially we can use the skip-gram model to capture the
k-step (k= 1,2,3, . . . ) relationship between each vertex and
its k-step neighbors in the graph with different values of k.
One limitation of the skip-gram model is it projects all such
k-step relational information into a common subspace. We
argue that such a simple treatment can lead to potential
issues. The above limitation is overcome in our proposed
model through the preservation of different k-step relational
information in distinct subspaces.
Another recently proposed work is LINE [25], which has
a loss function to capture both 1-step and 2-step local re-
lational information. To capture certain complex relations
in such local information, they also learn non-linear trans-
formations from such data. While their model can not be
easily extended to capture k-step (with k > 2) relational
information for learning their graph representation, one im-
portant strategy used to enhance the effectiveness of their
model is to consider higher-order neighbors for vertices with
small degrees. This strategy implicitly captures certain k-
step information into their model to some extent. We believe
k-step relational information between different vertices, with
different values of k, reveals the useful global structural in-
formation associated with the graph, and it is essential to
explicitly take full advantage of this when learning a good
graph representation.
In this paper, we propose GraRep, a novel model for learn-
ing graph representations for knowledge management. The
model captures the different k-step relational information
with different values of kamongst vertices from the graph
directly by manipulating different global transition matrices
defined over the graph, without involving slow and complex
sampling processes. Unlike existing work, our model defines
different loss functions for capturing the different k-step lo-
cal relational information (i.e., a different k). We optimize
each model with matrix factorization techniques, and con-
struct the global representations for each vertex by combin-
ing different representations learned from different models.
Such learned global representations can be used as features
for further processing.
We give a formal treatment of this model, showing the
connections between our model and several previous mod-
els. We also demonstrate the empirical effectiveness of the
learned representations in solving several real-world prob-
lems. Specifically, we conducted experiments on a language
network clustering task, a social network multi-label clas-
sification task, as well as a citation network visualization
task. In all such tasks, GraRep outperforms other graph
representation methods, and is trivially parallelizable.
Our contributions are as follows:
We introduce a novel model to learn latent represen-
tations of vertices on graphs, which can capture global
structural information associated with the graph.
We provide from a probabilistic prospective an un-
derstanding of the uniform sampling method used in
DeepWalk for learning graph representations, which
translates a graph structure into linear sequences. Fur-
thermore, we explicitly define their loss function over
graphs and extend it to support weighted graphs.
We formally analyze the deficiency associated with the
skip-gram model with negative sampling. Our model
defines a more accurate loss function that allows non-
linear combinations of different local relational infor-
mation to be integrated.
The organization of this paper is described below. Sec-
tion 2 discusses related work. Section 3 proposes our loss
function and states the optimization method using matrix
factorization. Section 4 presents the overall algorithm. Sec-
tion 5 gives a mathematical explanation to elucidate the
rationality of the proposed work and shows its connection
to previous work. Section 6 discusses the evaluation data
and introduces baseline algorithms. Section 7 presents ex-
periments as well as analysis on the parameter sensitivity.
Finally, we conclude in Section 8.
2.1 Linear Sequence Representation Methods
Natural language corpora, consisting of streams of words,
can be regarded as special graph structures, that is, linear
chains. Currently, there are two mainstream methods for
learning word representations: neural embedding methods
and matrix factorization based approaches.
Neural embedding methods employ a fixed slide window
capturing context words of current word. Models like skip-
gram [18] are proposed, which provide an efficient approach
to learning word representations. While these methods may
yield good performances on some tasks, they can poorly cap-
ture useful information since they use separate local context
windows, instead of global co-occurrence counts [19]. On
the other hand, the family of matrix factorization methods
can utilize global statistics [5]. Previous work include La-
tent Semantic Analysis (LSA) [15], which decomposes term-
document matrix and yields latent semantic representations.
Lund et al. [17] put forward Hyperspace Analogue to Lan-
guage (HAL), factorizing a word-word co-occurrence counts
matrix to generate word representations. Levy et al. [4]
presented matrix factorization over shifted positive Point-
wise Mutual Information (PMI) matrix for learning word
representations and showed that the Skip-Gram model with
Negative Sampling (SGNS) can be regarded as a model that
implicitly such a matrix [16].
2.2 Graph Representation Approaches
There exist several classical approaches to learning low di-
mensional graph representations, such as multidimensional
scaling (MDS) [8], IsoMap [28], LLE [21], and Laplacian
Eigenmaps [3]. Recently, Tang et al. [27] presented meth-
ods for learning latent representational vectors of the graphs
which can then be applied to social network classification.
Ahmed et al. [1] proposed a graph factorization method,
which used stochastic gradient descent to optimize matrices
from large graphs. Perozzi et al. [20] presented an approach,
which transformed graph structure into several linear vertex
sequences by using a truncated random walk algorithm and
generated vertex representations by using skip-gram model.
This is considered as an equally weighted linear combina-
tion of k-step information. Tang et al. [25] later proposed a
large-scale information network embedding, which optimizes
a loss function where both 1-step and 2-step relational in-
formation can be captured in the learning process.
In this section, we define our task and present our loss
function for our task, which is then optimized with the ma-
trix factorization method.
3.1 Graphs and Their Representations
Definition 1. (Graph) A graph is defined as G= (V, E ).
V={v1, v2,...,vn}is the set of vertices with each Vindi-
cating one object while E={ei,j }is the set of edges with
each Eindicating the relationship between two vertices. A
path is a sequence of edges which connect a sequence of ver-
We first define adjacency matrix Sfor a graph. For an
unweighted graph, we have Si,j = 1 if and only if there exists
an edge from vito vj, and Si,j = 0 otherwise. For a weighted
graph, Si,j is a real number called the weight of the edge ei,j ,
which indicates the significance of edge. Although weights
can be negative, we only consider non-negative weights in
this paper. For notational convenience, we also use wand c
to denote vertices throughout this paper.
The following diagonal matrix Dis known as the degree
matrix for a graph with adjacency matrix S:
Dij =PpSip,if i=j
0,if i6=j
Assume we would like to capture the transitions from one
vertex to another, and assume in the adjacency matrix Si,j
is proportional to the transition probability from vito vj, we
can define the following (1-step) probability transition matrix
A1 A2
A1 A2
A1 A2
A1 A2
B1 B2 B3 B4
A1 A2
A1 A2
C1 C2 C3 C4
A1 A2
A1 A2
A1 A2
A1 A2
C3 C4
A1 A2
B1 B2
Figure 1: The importance of capturing different k-step information in the graph representations. Here we
give examples for k= 1,2,3,4.
where Ai,j is the probability of a transition from vito vertex
vjwithin one step. It can be observed that the Amatrix
can be regarded as a re-scaled Smatrix whose rows are
Definition 2. (Graph Representations with Global Struc-
tural Information) Given a graph G, the task of Learning
Graph Representations with Global Structural Information
aims to learn a global representation matrix WR|Vdfor
the complete graph, whose i-th row Wiis a d-dimensional
vector representing the vertex viin the graph Gwhere the
global structural information of the graph can be captured
in such vectors.
In this paper, global structural information serves two
functions: 1) the capture of long distance relationship be-
tween two different vertices and 2) the consideration of dis-
tinct connections in terms of different transitional steps.
This would be further illustrated later.
As we have discussed earlier, we believe the k-step (with
varying k) relational information from the graph needs to
be captured when constructing such global graph representa-
tions. To validate this point, Figure 1 gives some illustrative
examples showing the importance of k-step (for k=1,2,3,4)
relational information that needs to be captured between
two vertices A1and A2. In the figure, a thick line indicates
a strong relation between two vertices, while a thin line in-
dicates a weaker relation. Here, (a) and (e) show the impor-
tance of capturing the simple 1-step information between
the two vertices which are directly connected to each other,
where one has a stronger relation and the other has a weaker
relation. In (b) and (f), 2-step information is shown, where
in (b) both vertices share many common neighbors, and in
(f) only one neighbor is shared between them. Clearly, 2-
step information is important in capturing how strong the
connection between the two vertices is – the more common
neighbors they share, the stronger the relation between them
is. In (c) and (g), the importance of 3-step information is
illustrated. Specifically, in (g), despite the strong relation
between A1and B, the relation between A1and A2can be
weakened due to the two weaker edges connecting Band C,
as well as Cand A2. In contrast, in (c), the relation between
A1and A2remains strong because of the large number of
common neighbors between Band A2which strengthened
their relation. Clearly such 3-step information is essential to
be captured when learning a good graph representation with
global structural information. Similarly, the 4-step informa-
tion can also be crucial in revealing the global structural
properties of the graph, as illustrated in (d) and (h). Here,
in (d), the relation between A1and A2is clearly strong,
while in (h) the two vertices are unrelated since there does
not exist a path from one vertex to the other. In the absence
of 4-step relational information, such important distinctions
can not be properly captured.
C1 C2 C3 C4
C2 C3
Figure 2: The importance of maintaining different
k-step information separately in the graph represen-
We also additionally argue that it is essential to treat
different k-step information differently when learning graph
representations. We present one simple graph in (a) of Fig-
ure 2. Let us focus on learning the representation for the
vertex Ain the graph. We can see that Areceives two types
of information when learning its representation: 1-step in-
formation from B, as well as 2-step information from all C
vertices. We note that if we do not distinguish these two dif-
ferent types of information, we can construct an alternative
graph as shown in (b) of Figure 2, where Areceives exactly
the same information as (a), but has a completely different
In this paper, we propose a novel framework for learning
accurate graph representations, integrating various k-step
information which together captures the global structural
information associated with the graph.
3.2 Loss Function On Graph
We discuss our loss function used for learning the graph
representations with global structural information in this
section. Assume we are now given a graph with a collec-
tion of vertices and edges. Consider a vertex wtogether
with another vertex c. In order for us to learn global repre-
sentations to capture their relations, we need to understand
how strongly these two vertices are connected to each other.
Let us begin with a few questions. Does there exist a path
from wto c? If so, if we randomly sample a path starting
with w, how likely is it for us to reach c(possibly within
a fixed number of steps)? To answer such questions, we
first use the term pk(c|w) to denote the probability for a
transition from wto cin exactly ksteps. We already know
what are the 1-step transition probabilities, to compute the
k-step transition probabilities we introduce the following k-
step probability transition matrix
Ak=A· · · A
| {z }
We can observe that Ak
i,j exactly refers to the transition
probability from vertex ito vertex jwhere the transition
consists of exactly kstep(s). This directly leads to:
pk(c|w) = Ak
where Ak
w,c is the element from w-th row and c-th column
of the matrix Ak.
Let us consider a particular kfirst. Given a graph G,
consider the collection of all paths consisting of ksteps that
can be sampled from the graph which start with wand end
with c(here we call w“current vertex”, and c“context ver-
tex”). Our objective aims to maximize: 1) the probability
that these pairs come from the graph, and 2) the probability
that all other pairs do not come from the graph.
Motivated by the skip-gram model by Mikolov et al. [18],
we employ noise contrastive estimation (NCE), which is pro-
posed by Gutmann et al. [11], to define our objective func-
tion. Following a similar discussion presented in [16], we first
introduce our k-step loss function defined over the complete
graph as follows:
Lk(w) = X
pk(c|w) log σ(~w ·~c)!+λEc0pk(V)[log σ(~w·~
Here pk(c|w) describes the k-step relationship between w
and c(the k-step transition probability from wto c), σ(·)
is sigmoid function defined as σ(x) = (1 + ex)1,λis a
hyper-parameter indicating the number of negative samples,
and pk(V) is the distribution over the vertices in the graph.
The term Ec0pk(V)[·] is the expectation when c0follows the
distribution pk(V), where c0is an instance obtained from
negative sampling. This term can be explicitly expressed as:
Ec0pk(V)[log σ(~w ·~
=pk(c)·log σ(~w ·~c) + X
pk(c0)·log σ(~w ·~
This leads to a local loss defined over a specific (w,c):
Lk(w, c) = pk(c|w)·log σ(~w ·~c) + λ·pk(c)·log σ(~w ·~c)
In this work, we set a maximal length for each path we
consider. In other words, we assume 1 kK. In fact,
when kis large enough, the transition probabilities converge
to certain fixed values. The distribution pk(c) can be com-
puted as follows:
pk(c) = X
q(w0)pk(c|w0) = 1
Note that here Nis the number of vertices in graph G,
and q(w0) is the probability of selecting w0as the first vertex
in the path, which we assume follow a uniform distribution,
i.e.,q(w0) = 1/N. This leads to:
Lk(w, c) = Ak
w,c ·log σ(~w ·~c) + λ
w0,c ·log σ(~w ·~c)
Following [16], we define e=~w ·~c, and setting Lk
∂e = 0.
This yields the following:
~w ·~c = log Ak
w0,c !log(β)
where β=λ/N.
This concludes that we essentially need to factorize the
matrix Yinto two matrices Wand C, where each row of W
and each row of Cconsists of a vector representation for the
vertex wand crespectively, and the entries of Yare:
i,j =Wk
j= log Ak
t,j !log(β)
Now we have defined our loss function and showed that
optimizing the proposed loss essentially involves a matrix
factorization problem.
3.3 Optimization with Matrix Factorization
Following the work of Levy et al. [16], to reduce noise,
we replace all negative entries in Ykwith 0. This gives us a
positive k-step log probabilistic matrix Xk, where
i,j = max(Yk
i,j ,0)
While various techniques for matrix factorization exist, in
this work we focus on the popular singular value decompo-
sition (SVD) method due to its simplicity. SVD has been
shown successful in several matrix factorization tasks [9, 14],
and is regarded as one of the important methods that can be
used for dimensionality reduction. It was also used in [16].
For the matrix Xk, SVD factorizes it as:
where Uand Vare orthonormal matrices and Σ is a diagonal
matrix consisting of an ordered list of singular values.
We can approximate the original matrix Xkwith Xk
where Σk
dis the matrix composed by the top dsingular val-
ues, and Uk
dand Vk
dare first dcolumns of Ukand Vk, re-
spectively (which are the first deigenvector of XXTand
This way, we can factorize our matrix Xkas:
2, Ck= (Σk
The resulting Wkgives representations of current vertices
as its column vectors, and Ckgives the representations of
context vertices as its column vectors [6, 5, 30]. The fi-
nal matrix Wkis returned from the algorithm as the low-d
representations of the vertices which capture k-step global
structural information in the graph.
In our algorithm we consider all k-step transitions with
all k= 1,2,...,K, where Kis a pre-selected constant. In
our algorithm, we integrate all such k-step information when
learning our graph representation, as to be discussed next.
Note that here we are essentially finding a projection from
the row space of Xkto the row space of Wkwith a lower
rank. Thus alternative approaches other than the popular
SVD can also be exploited. Examples include incremental
SVD [22], independent component analysis (ICA) [7, 13],
and deep neural networks [12]. Our focus in this work is on
the novel model for learning graph representations, so we
do not pursue any alternative methods. In fact, if alterna-
tive method such as sparse auto-encoder is used at this step,
it becomes relatively harder to justify whether the empiri-
cal effectiveness of our representations is due to our novel
model, or comes from any non-linearity introduced in this
dimensionality reduction step. To maintain the consistency
with Levy et al. [16], we only employed SVD in this work.
We detail our learning algorithm in this section. In gen-
eral, graph representations are extracted for other applica-
tions as features, such as classification and clustering. An
effective way to encode k-step representation in practice is to
concatenate the k-step representation as a global feature for
each vertex, since each different step representation reflects
different local information. Table 1 shows the overall algo-
rithm, and we explain the essential steps of our algorithm
Step 1. Get k-step transition probability matrix
Akfor each k= 1,2,...,K.
As shown in Section 3, given a graph G, we can calcu-
late the k-step transition probability matrix Akthrough the
product of inverse of degree matrix Dand adjacent matrix
S. For a weighted graph, Sis a real matrix, while for an
unweighted graph, Sis a binary matrix. Our algorithm is
applicable to both cases.
Step 2. Get each k-step representation
We get k-step log probability matrix Xk, then subtract
each entry by log (β), and replace the negative entries by
zeros. After that, we construct the representational vectors
as rows of Wk, where we introduce a solution to factorize the
positive log probability matrix Xkusing SVD. Finally, we
get all k-step representations for each vertex on the graph.
Step 3. Concatenate all k-step representations
We concatenate all k-step representations to form a global
representation, which can be used in other tasks as features.
GraRep aims to learn representations for graphs where
we optimize the loss function based on matrix factorization.
On the other hand, SGNS has been shown to be successful
in handling linear structures such as natural language sen-
tences. Is there any intrinsic relationship between them? In
this section, we provide a view of SGNS as a special case of
the GraRep model.
5.1 Explicit Loss of Skip-gram Model on Graph
SGNS aims at representing words in linear sequences, so
we need to translate a graph structure into linear structures.
Deepwalk reveals an effective way with uniform sampling
Table 1: Overall Algorithm
GraRep Algorithm
Adjacency matrix Son graph
Maximum transition step K
Log shifted factor β
Dimension of representation vector d
1. Get k-step transition probability matrix Ak
Compute A=D1S
Calculate A1, A2,...,AK, respectively
2. Get each k-step representations
For k= 1 to K
2.1 Get positive log probability matrix
calculate Γk
p,j ) respectively
calculate {Xk
i,j }
i,j = log Ak
assign negative entries of Xkto 0
2.2 Construct the representation vector Wk
[Uk,Σk,(Vk)T] = SV D(Xk)
End for
3. Concatenate all the k-step representations
W= [W1, W 2,...,WK]
Matrix of the graph representation W
(truncated random walk). This method first samples uni-
formly a random vertex from the graph, then walks ran-
domly to one of its neighbors and repeats this process. If the
length of the vertex sequence reaches a certain preset value,
then stop and start generating a new sequence. This pro-
cedure can be used to produce a large number of sequences
from the graph.
Essentially, for an unweighted graph, this strategy of uni-
form sampling works, while for a weighted graph, a proba-
bilistic sampling method based on the weights of the edges is
needed, which is not employed in DeepWalk. In this paper,
we propose an Enhanced SGNS (E-SGNS) method suitable
for weighted graphs. We also note that DeepWalk optimizes
an alternative loss function (hierarchical softmax) that is
different from negative sampling.
First, we consider total K-step loss Lon whole graph
L=f(L1, L2,...,LK)
where f(·) is a linear combination of its arguments defined
as follows:
f(ϕ1, ϕ2,· · · , ϕK) = ϕ1+ϕ2+· · · +ϕK
We focus on the loss of a specific pair (w,c) which are i-
th and j-th vertex in the graph. Similar to Section 3.2, we
assign partial derivative to 0, and get
i,j = log Mi,j
PtMt,j log(β)
where Mis transition probability matrix within Kstep(s),
and Mi,j refers to transition probability from vertex ito
vertex j.YESGNS
i,j is a factorized matrix for E-SGNS, and
M=A1+A2+· · · +AK
The difference between E-SGNS and GraRep model is on the
definition of f(·). E-SGNS can be considered as a linear com-
bination of K-step loss, and each loss has an equal weight.
Our GraRep model does not make such a strong assump-
tion, but allows their (potentially non-linear) relationship
to be learned from data in practice. Intuitively, different
k-step transition probabilities should have different weights,
and linear combination of these may not achieve desirable
results for heterogeneous network data.
5.2 Intrinsic Relation Between Sampling and
Transition Probabilities
In our approach, we used transition probabilities to mea-
sure relationship between vertices. Is this reasonable? In
this subsection, we articulate the intrinsic relation between
sampling and transition probabilities.
Among the sequences generated by random walk, we as-
sume vertex woccurs a total of T=γ·Ktimes:
#(w) = γ·K
where γis a variable related to #(w). We regard vertex w
as the current vertex, then the expected number of times
that we see c1as its direct neighbor (1-step away from w)
#(w, c1) = γ·K·p1(c|w)
This holds for both uniform sampling for unweighted graphs
or our proposed probabilistic sampling for weighted graphs.
Further, we analyze the expected number of times of co-
occurrence for wand cof context window size 2,
#(w, c2) = γ·KX
p(c2|c0)·p(c0|w) = γ·K·p2(c|w)
where c0can be any vertex bridging wand c2. That is, c0is
shared neighbor between wand c2. Similarly, we can derive
the equations for k= 3,4,· · · , K:
#(w, c3) = γ·K·p3(c|w)
#(w, cK) = γ·K·pK(c|w)
then we add them up and divide both sides by K, leading
#(w, c) = γ·
where #(w, c) is the expected co-occurrence count between
wand cwithin Kstep(s).
According to definition of Mw,c, we can get
#(w, c) = γ·Mw,c
Now, we can also compute the expected number of times
we see cas the context vertex, #(c):
#(c) = γ·X
where we consider the transitions from all possible vertices
to c.
Finally, we plug these expected counts into the equation
i,j , and we arrive at:
w,c = log #(w , c)· |D|
where Dis the collection of all observed pairs in sequences.
Here |D|=γK N where Nis the total number of vertices
in the graph. This matrix YESGNS becomes exactly the
same as that of SGNS as described in [16].
This shows SGNS is essentially a special version of our
GraRep model that deals with linear sequences which can
be sampled from graphs. Our approach has several advan-
tages over the slow and expensive sampling process, which
typically involves several parameters to tune, such as maxi-
mum length of linear sequence, sampling frequency for each
vertex, and so on.
In this section, we assess the effectiveness of our GraRep
model through experiments. We conduct experiments on
several real-word datasets for several different tasks, and
make comparisons with baseline algorithms.
6.1 Datasets and Tasks
In order to demonstrate the performance of GraRep, we
conducted the experiments across three different types of
graphs – a social network, a language network and a cita-
tion network, which include both weighted and unweighted
graphs. We conducted experiments across three different
types of tasks, including clustering, classification, and vi-
sualization. As we mentioned in earlier sections, this work
focuses on proposing a novel framework for learning good
representation of graph with global structural information,
and aims to validate the effectiveness of our proposed model.
Thus we do not employ alternative more efficient matrix
factorization methods apart from SVD, and focused on the
following three real-world datasets in this section.
1. 20-Newsgroup1is a language network, which has ap-
proximately 20,000 newsgroup documents and is partitioned
by 20 different groups. In this network, each document is
represented by a vector with tf-idf scores of each word, and
the cosine similarity is used to calculate the similarity be-
tween two documents. Based on these similarity scores of
each pair of documents, a language network is built. Fol-
lowing [29], in order to show the robustness of our model,
we also construct the following 3 graphs built from 3, 6 and
9 different newsgroups respectively (note that NG refers to
“Newsgroups”):, and talk.politics.guns;
6-NG: alt.atheism, comp.sys.mac.hardware,,, soc.religion.christian and talk.religion.misc;
9-NG: talk.politics.mideast, talk.politics.misc,
windows.misc, sci.crypt,,, sci.electronics,, and
Besides randomly sampling 200 documents from a topic
as described in [29], we also conduct experiment on all doc-
uments as comparison. The topic label on each document is
considered to be true.
This language network is a fully connected and weighted
graph, and we will demonstrate the results of clustering,
using graph representation as features.˜jason/20Newsgroups/
2. Blogcatalog2is a social network, where each vertex
indicates one blogger author, and each edge corresponds to
the relationship between authors. 39 different types of topic
categories are presented by authors as labels.
Blogcatalog is an unweighted graphs, and we test the per-
formance of the learned representations on the multi-label
classification task, where we classify each author vertex into
a set of labels. The graph representations generated from
our model and each baseline algorithm are considered as
3. DBLP Network3is a citation network. We extract
author citation network from DBLP, where each vertex in-
dicates one author and the number of references from one
author to the other is recorded by the weight of edge be-
tween these two authors. Following [26], we totally select 6
different popular conferences and assign them into 3 groups,
where WWW and KDD are grouped as data mining, NIPS
and ICML as machine learning, and CVPR and ICCV as
computer vision. We visualize the learned representations
from all systems using a visualisation tool t-SNE [31], which
provides both qualitative and quantitative results for the
learned representations. We give more details in Section
To summarize, in this paper we conduct experiments on
both weighted and unweighted graphs, and both sparse and
dense graphs, where three different types of learning tasks
are carried out. More details of the graph we used are shown
in Table 2.
6.2 Baseline Algorithms
We use the following methods of graph representation as
baseline algorithms.
1. LINE [25]. LINE is a recently proposed method for
learning graph representations on large-scale informa-
tion networks. LINE defines a loss function based on
1-step and 2-step relational information between ver-
tices. One strategy to improve the performance of
vertices with small degrees is to make graph denser
by expanding their neighbors. LINE will get the best
performance, if concatenating the representation of 1-
step and 2-step relational information and tuning the
threshold of maximum number of vertices.
2. DeepWalk [20]. DeepWalk is a method that learns
the representation of social networks. The original
model only works for unweighted graph. For each ver-
tex, truncated random walk is used to translate graph
structure into linear sequences. The skip-gram model
with hierarchical softmax is used as the loss function.
3. E-SGNS. Skip-gram is an efficient model that learns
the representation of each word in large corpus [18].
For this enhanced version, we first utilize uniform sam-
pling for unweighted graph and probabilistic sampling
proportional to weight of edges for weighted graph, to
generate linear vertex sequences, and then introduce
SGNS to optimize. This method can be regarded as a
special case of our model, where different representa-
tional vector of each k-step information is averaged.
4. Spectral Clustering [23]. Spectral clustering is a
reasonable baseline algorithm, which aims at minimiz-
ing Normalized Cut (NCut). Like our method, Spec-
tral clustering also factorize a matrix, but it focuses on
a different matrix of the graphs – the Laplacian Matrix.
Essentially, the difference between spectral clustering
and E-SGNS lies on their different loss function.
6.3 Parameter Settings
As suggested in [25], for LINE, we set the mini-batch size
of stochastic gradient descent (SGD) as 1, learning rate of
starting value as 0.025, the number of negative samples as
5, and the total number of samples as 10 billion. We also
concatenate both 1-step and 2-step relational information
to form the representations and employ the reconstruction
strategy for vertices with small degrees to achieve the opti-
mal performance. As mentioned in [20], for DeepWalk and
E-SGNS, we set window size as 10, walk length as 40, walks
per vertex as 80. According to [25], LINE yielded better
results when the learned graph representations are L2 nor-
malized, while DeepWalk and E-SGNS can achieve optimal
performance without normalization. For GraRep, we found
the L2 normalization yielded better results. We reported
all the results based on these findings for each system ac-
cordingly. For a fair comparison, the dimension dof repre-
sentations is set as 128 for Blogcatalog network and DBLP
network as used in [25] and is set as 64 for 20-NewsGroup
network as used in [29]. For GraRep, we set β=1
Nand max-
imum matrix transition step K=6 for Blogcatalog network
and DBLP network and K=3 for 20-NewsGroup network.
To demonstrate the advantage of our model which captures
global information of the graph, we also conducted exper-
iments under various parameter settings for each baseline
systems, which are discussed in detail in the next section.
6.4 Experimental Results
In this section, we present empirical justifications that our
GraRep model can integrate different k-step local relational
information into a global graph representation for different
types of graphs, which can then be effectively used for dif-
ferent tasks. We make the source code of GraRep available
6.4.1 20-Newsgroup Network
We first conduct an experiment on a language network
through a clustering task by employing the learned repre-
sentations in a k-means algorithm [2].
To assess the quality of the results, we report the aver-
aged Normalized Mutual Information (NMI) score [24] over
10 different runs for each system. To understand the effect
of different dimensionality din the end results, we also show
the results when dimension dis set to 192 for DeepWalk,
E-SGNS and Spectral Clustering. For LINE, we employ the
reconstruction strategy proposed by their work by adding
neighbors of neighbors as additional neighbors to improve
performance. We set k-max=0, 200, 500, 1000 for experi-
ments, where k-max is a parameter that is used to control
how the higher-order neighbors are added to each vertex in
the graph.
As shown in Table 3, the highest results are highlighted
in bold for each column. We can see that GraRep con-
sistently outperforms other baseline methods for this task.
For DeepWalk, E-SGNS and Spectral Clustering, increasing
Table 2: Statistics of the real-world graphs
Language Network Social Network Citation Network
Name 20-NewsGroup 20-NewsGroup Blogcatalog DBLP
(200 samples) (all data) (author citation)
Type weighted weighted unweighted weighted
#(V) 600, 1200 and 1800 1,720, 3,224 and 5,141 10,312 7,314
#(E) Fully connected Fully connected 333,983 72,927
Avg. degree 64.78 19.94
#Labels 3, 6 and 9 3, 6 and 9 39 3
Task Clustering Clustering Classification Visualization
Table 3: Results on 20-NewsGroup
200 samples all data
Algorithm 3NG(200) 6NG(200) 9NG(200) 3NG(all) 6NG(all) 9NG(all)
GraRep 81.12 67.53 59.43 81.44 71.54 60.38
LINE (k-max=0) 80.36 64.88 51.58 80.58 68.35 52.30
LINE (k-max=200) 78.69 66.06 54.14 80.68 68.83 53.53
DeepWalk 65.58 63.66 48.86 65.67 68.38 49.19
DeepWalk (192dim) 60.89 59.89 47.16 59.93 65.68 48.61
E-SGNS 69.98 65.06 48.47 69.04 67.65 50.59
E-SGNS (192dim) 63.55 64.85 48.65 66.64 66.57 49.78
Spectral Clustering 49.04 51.02 46.92 62.41 59.32 51.91
Spectral Clustering (192dim) 28.44 27.80 36.05 44.47 36.98 47.36
Spectral Clustering (16dim) 69.91 60.54 47.39 78.12 68.78 57.87
Table 4: Results on Blogcatalog
Metric Algorithm 10% 20% 30% 40% 50% 60% 70% 80% 90%
GraRep 38.24 40.31 41.34 41.87 42.60 43.02 43.43 43.55 44.24
LINE 37.19 39.82 40.88 41.47 42.19 42.72 43.15 43.36 43.88
DeepWalk 35.93 38.38 39.50 40.39 40.79 41.28 41.60 41.93 42.17
E-SGNS 35.71 38.34 39.64 40.39 41.23 41.66 42.01 42.16 42.25
Spectral Clustering 37.16 39.45 40.22 40.87 41.27 41.50 41.48 41.62 42.12
GraRep 23.20 25.55 26.69 27.53 28.35 28.78 29.67 29.96 30.93
LINE 19.63 23.04 24.52 25.70 26.65 27.26 27.94 28.68 29.38
DeepWalk 21.02 23.81 25.39 26.27 26.85 27.36 27.67 27.96 28.41
E-SGNS 21.01 24.09 25.61 26.59 27.64 28.08 28.33 28.34 29.26
Spectral Clustering 19.26 22.24 23.51 24.33 24.83 25.19 25.36 25.52 26.21
(a) Spectral Clustering (b) DeepWalk (c) E-SGNS (d) LINE (e) GraRep
Figure 3: Visualization of author citation network. Each point indicates one author. Green: Data Mining,
magenta: computer vision and blue: Machine Learning.
the dimension dof representations does not appear to be
effective in improving the performance. We believe this is
because a higher dimension does not provide different com-
plementary information to the representations. For LINE,
the reconstruction strategy does help, since it can capture
additional structural information of the graph beyond 1-step
and 2-step local information.
It is worth mentioning that GraRep and LINE can achieve
good performance with a small graph. We believe this is be-
cause both approaches can capture rich local relational in-
formation even when the graph is small. Besides, for tasks
with more labels, such as 9NG, GraRep and LINE can pro-
vide better performances than other methods, with GraRep
providing the best. An interesting finding is that Spectral
Clustering arrives at its best performance when setting di-
mension dto 16. However, for all other algorithms, their
best results are obtained when dis set to 64. We show these
results in Figure 5 when we assess the sensitivity of param-
Due to limited space, we do not report the detailed results
with d=128 and d=256 for all baseline methods, as well as
k-max=500, 1000 and without reconstruction strategy for
LINE, since these results are no better than the results pre-
sented in Table 3. In Section 6.5, we will further discuss the
issue of parameter sensitivity.
6.4.2 Blogcatalog Network
In this experiment, we focus on a supervised task on a so-
cial network. We evaluate the effectiveness of different graph
representations through a multi-label classification task by
regarding the learned representations as features.
Following [20, 27], we use the LibLinear package [10] to
train one-vs-rest logistic regression classifiers, and we run
this process for 10 rounds and report the averaged Micro-
F1 and Macro-F1 measures. For each round, we randomly
sample 10% to 90% of the vertices and use these samples
for training, and use the remaining vertices for evaluation.
As suggested in [25], we set k-max as 0, 200, 500 and 1000,
respectively, and we report the best performance with k-
Table 4 reports the results of GraRep and each baseline al-
gorithm. The highest performance is highlighted in bold for
each column which corresponds to one particular train/test
split. Overall, GraRep significantly outperforms other meth-
ods, especially under the setting where only 10% of the ver-
tices are used for training. This indicates different types
of rich local structural information learned by the GraRep
model can be used to complement each other to capture
the global structural properties of the graph, which serves
Table 5: Final KL divergence for the DBLP dataset
Algorithm GraRep LINE DeepWalk E-SGNS
KL divergence 1.0070 1.0816 1.1115 1.1009
as a distinctive advantage over existing baseline approaches,
especially when the data is relatively scarce.
6.4.3 DBLP Network
In this experiment, we focus on visualizing the learned rep-
resentations by examining a real citation network – DBLP.
We feed the learned graph representations into the standard
t-SNE tool [31] to lay out the graph, where the authors from
the same research area (one of data mining, machine learn-
ing or computer vision, see Section 6.1) share the same color.
The graphs are visualized on a 2-dimensional space and the
Kullback-Leibler divergence is reported [31], which captures
the errors between the input pairwise similarities and their
projections in the 2-dimensional map as displayed (a lower
KL divergence score indicates a better performance).
Under the help of t-SNE toolkit, we set the same param-
eter configuration and import representations generated by
each algorithm. From Figure 3, we can see that the lay-
out using Spectral Clustering is not very informative, since
vertices of different colors are mixed with each other. For
DeepWalk and E-SGNS, results look much better, as most
vertices of the same color appear to form groups. However,
vertices still do not appear in clearly separable regions with
clear boundaries. For LINE and GraRep, the boundaries
of each group become much clearer, with vertices of differ-
ent colors appearing in clearly distinguishable regions. The
results for GraRep appear to be better, with clearer bound-
aries for each regions as compared to LINE.
Table 5 reports the KL divergence at the end of iterations.
Under the same parameter setting, a lower KL divergence
indicates a better graph representation. From this result, we
also can see that GraRep yields better representations than
all other baseline methods.
6.5 Parameter Sensitivity
We discuss the parameter sensitivity in this section. Specif-
ically, we assess the how the different choices of the maximal
k-step size K, as well as the dimension dfor our represen-
tations can affect our results.
Labeled Nodes
1% 2% 3% 4% 5% 6% 7% 8% 9%
(a) Blogcatalog, Micro-F1
Labeled Nodes
1% 2% 3% 4% 5% 6% 7% 8% 9%
(b) Blogcatalog, Macro-F1
Figure 4: Performance over steps, K
Figure 4 shows the Micro-F1 and Macro-F1 scores over
different choices of Kon the Blogcatalog data. We can ob-
serve that the setting K=2 has a significant improvement
over the setting K=1, and K=3 further outperforms K=2.
This confirms that different k-step can learn complementary
local information. In addition, as mentioned in Section 3,
when kis large enough, learned k-step relational informa-
tion becomes weak and shifts towards a steady distribution.
Empirically we can observe that the performance of K=7
is no better than K=6. In this figure, for readability and
clarity we only present the results of K=1,2,3,6 and 7. We
found the performance of K= 4 is slightly better than K=3,
while the results for K=5 is comparable to K=4.
32 64 128 256
(a) 3NG, NMI
32 64 128 256
(b) 9NG, NMI
Figure 5: Performance over dimensions, d
Figure 5 shows the NMI scores of each algorithm over dif-
ferent settings of the dimension don 3NG and 9NG data.
We can observe that GraRep consistently outperforms other
baseline algorithms which learn representations with the same
dimension. This set of experiments serves as an additional
supplement to Table 3. Interestingly, all algorithms can ob-
tain the optimal performance with d= 64. As we increase d
from 64 to larger values, it appears that the performances of
all the algorithms start to drop. Nevertheless, our GraRep
algorithm is still superior to the baseline systems across dif-
ferent dvalues.
128 128*2 128*3 128*4 128*5 128*6 128*7
Cost of time/seconds
(a) Blogcatalog, running time
Nodes of graph
600 1200 1800 3224 5141
Cost of time/seconds
(b) 20NG, running time
Figure 6: Running time as a function of a) dimension
and b) size of graph.
Figure 6 shows the running time over different dimensions
as well as over different graph sizes. In Figure 6(a), we set
Kfrom 1 to 7 where each complementary feature vector has
a dimension of 128. This set of experiments is conducted
on the Blogcatalog dataset which contains around 10,000
vertices. It can be seen that the running time increases ap-
proximately linearly as the dimension increases. In Figure
6(b), we analyze running time with respect to graph sizes.
This set of experiment is conducted on different size of the
20-NewsGroup dataset. The result shows a significant in-
crease of running time as graph size increases. The reason
resulting in the large increase in running time is mainly due
to high time complexity involved in the computation of the
power of a matrix and the SVD procedure.
In this paper, GraRep, a novel model for learning better
graph representations is proposed. Our model, with our k-
step loss functions defined on graphs which integrate rich
local structural information associated with the graph, cap-
tures the global structural properties of the graph. We also
provide mathematical derivations justifying the model and
establish the connections to previous research efforts. Em-
pirically, the learned representations can be effectively used
as features in other learning problems such as clustering
and classification. This model comes with one limitation:
the expensive computation of the power of a matrix and
SVD involved in the learning process. Future work would
include the investigation of efficient and online methods to
approximate matrix algebraic manipulations, as well as in-
vestigation of alternative methods by employing deep ar-
chitectures for learning low-dimensional representations in
place of SVD.
The authors would like to thank the anonymous reviewers
for their helpful comments. The authors would also like to
thank Fei Tian, Qingbiao Miao and Jie Liu for their sug-
gestions or help with this work. This paper was done when
the first author was an intern at SUTD. This work was sup-
ported by SUTD grant ISTD 2013 064 and Temasek Lab
project IGDST1403013.
[1] A. Ahmed, N. Shervashidze, S. Narayanamurthy,
V. Josifovski, and A. J. Smola. Distributed large-scale
natural graph factorization. In WWW, pages 37–48.
International World Wide Web Conferences Steering
Committee, 2013.
[2] D. Arthur and S. Vassilvitskii. k-means++: The
advantages of careful seeding. In SODA, pages
1027–1035. Society for Industrial and Applied
Mathematics, 2007.
[3] M. Belkin and P. Niyogi. Laplacian eigenmaps and
spectral techniques for embedding and clustering. In
NIPS, volume 14, pages 585–591, 2001.
[4] J. A. Bullinaria and J. P. Levy. Extracting semantic
representations from word co-occurrence statistics: A
computational study. BRM, 39(3):510–526, 2007.
[5] J. A. Bullinaria and J. P. Levy. Extracting semantic
representations from word co-occurrence statistics:
stop-lists, stemming, and svd. BRM, 44(3):890–907,
[6] J. Caron. Experiments with lsa scoring: Optimal rank
and basis. In CIR, pages 157–169, 2001.
[7] P. Comon. Independent component analysis, a new
concept? Signal processing, 36(3):287–314, 1994.
[8] T. F. Cox and M. A. Cox. Multidimensional scaling.
CRC Press, 2000.
[9] C. Eckart and G. Young. The approximation of one
matrix by another of lower rank. Psychometrika,
1(3):211–218, 1936.
[10] R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang,
and C.-J. Lin. Liblinear: A library for large linear
classification. JMLR, 9:1871–1874, 2008.
[11] M. U. Gutmann and A. Hyv¨
arinen. Noise-contrastive
estimation of unnormalized statistical models, with
applications to natural image statistics. JMLR,
13(1):307–361, 2012.
[12] G. E. Hinton and R. R. Salakhutdinov. Reducing the
dimensionality of data with neural networks. Science,
313(5786):504–507, 2006.
[13] C. Jutten and J. Herault. Blind separation of sources,
part i: An adaptive algorithm based on neuromimetic
architecture. Signal processing, 24(1):1–10, 1991.
[14] V. Klema and A. J. Laub. The singular value
decomposition: Its computation and some
applications. Automatic Control, 25(2):164–176, 1980.
[15] T. K. Landauer, P. W. Foltz, and D. Laham. An
introduction to latent semantic analysis. Discourse
processes, 25(2-3):259–284, 1998.
[16] O. Levy and Y. Goldberg. Neural word embedding as
implicit matrix factorization. In NIPS, pages
2177–2185, 2014.
[17] K. Lund and C. Burgess. Producing high-dimensional
semantic spaces from lexical co-occurrence. BRMIC,
28(2):203–208, 1996.
[18] T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and
J. Dean. Distributed representations of words and
phrases and their compositionality. In NIPS, pages
3111–3119, 2013.
[19] J. Pennington, R. Socher, and C. D. Manning. Glove:
Global vectors for word representation. EMNLP, 12,
[20] B. Perozzi, R. Al-Rfou, and S. Skiena. Deepwalk:
Online learning of social representations. In SIGKDD,
pages 701–710. ACM, 2014.
[21] S. T. Roweis and L. K. Saul. Nonlinear dimensionality
reduction by locally linear embedding. Science,
290(5500):2323–2326, 2000.
[22] B. Sarwar, G. Karypis, J. Konstan, and J. Riedl.
Incremental singular value decomposition algorithms
for highly scalable recommender systems. In ICIS,
pages 27–28. Citeseer, 2002.
[23] J. Shi and J. Malik. Normalized cuts and image
segmentation. PAMI, 22(8):888–905, 2000.
[24] A. Strehl, J. Ghosh, and R. Mooney. Impact of
similarity measures on web-page clustering. In
Workshop on Artificial Intelligence for Web Search
(AAAI 2000), pages 58–64, 2000.
[25] J. Tang, M. Qu, M. Wang, M. Zhang, J. Yan, and
Q. Mei. Line: Large-scale information network
embedding. In WWW. ACM, 2015.
[26] J. Tang, J. Zhang, L. Yao, J. Li, L. Zhang, and Z. Su.
Arnetminer: extraction and mining of academic social
networks. In SIGKDD, pages 990–998. ACM, 2008.
[27] L. Tang and H. Liu. Relational learning via latent
social dimensions. In SIGKDD, pages 817–826. ACM,
[28] J. B. Tenenbaum, V. De Silva, and J. C. Langford. A
global geometric framework for nonlinear
dimensionality reduction. Science,
290(5500):2319–2323, 2000.
[29] F. Tian, B. Gao, Q. Cui, E. Chen, and T.-Y. Liu.
Learning deep representations for graph clustering. In
AAAI, 2014.
[30] P. D. Turney. Domain and function: A dual-space
model of semantic relations and compositions. JAIR,
pages 533–585, 2012.
[31] L. Van der Maaten and G. Hinton. Visualizing data
using t-sne. JMLR, 9(2579-2605):85, 2008.
... By minimizing the KL-divergence of the two distributions and the empirical distributions, respectively, embedding representations of vertices that can maintain first-and second-order closenesses can be obtained. Considering that LINE preserves only firstorder and second-order affinities, GraRep (Cao et al., 2015) shows that k-step (k>2) proximity should also be captured when generating global representations of the vertices. Given an adjacency matrix , the elements of the -step probability transition matrix represent the In this definition, , is the distribution over the vertices of the RAG and is the peak obtained with negative sampling. ...
Full-text available
The image classification problem is a process that many machine learning methods are trying to solve. Graphs, which are combinatorial mathematical structures, are frequently used in machine learning problems. In this study, a method using machine learning based embeddings of weighted regional graphs for image classification problem is proposed.
In recent years, lots of recommendation methods based on heterogeneous information network has been proposed. However, traditional heterogeneous information network mainly relies on the similarity of meta-path to recommend, which cannot fully mine the user’s potential preference features. We propose a heterogeneous information network model using self-attention mechanism for recommendation (HINS). In our method, local neighbor information is modeled for users and items respectively via collaborative attention mechanism. Moreover, by optimizing the relationship representation between learning nodes of multi label classification problem based on meta-path, self-attention mechanism is added to the relationship representation to make the model personalized. Finally, the two parts are unified into one model for top-N recommendation. We do evaluation experiments on four different datasets, and our model achieves the best results.
Network embedding is a technique that aims at inferring the low-dimensional representations of nodes in a semantic space. In this article, we study the problem of inferring the low-dimensional representations of both nodes and attributes for attributed networks in the same semantic space such that the affinity between a node and an attribute can be effectively measured. Intuitively, this problem can be addressed by simply utilizing existing variational auto-encoder (VAE) based network embedding algorithms. However, the variational posterior distribution in previous VAE based network embedding algorithms is often assumed and restricted to be a mean-field Gaussian distribution or other simple distribution families, which results in poor inference of the embeddings. To alleviate the above defect, we propose a novel VAE-based co-embedding method for attributed network, F-CAN, where posterior distributions are flexible, complex, and scalable distributions constructed through the normalizing flow. We evaluate our proposed models on a number of network tasks with several benchmark datasets. Experimental results demonstrate that there are clear improvements in the qualities of embeddings generated by our model to the state-of-the-art attributed network embedding methods.
Full-text available
Data analysis is the foundation of Internet of Things (IoT) based applications, and clustering is an effective technology of data analysis. Clustering ensemble integrates multiple base clustering results to obtain a consensus result and thus improves the clustering performance in stability and robustness. However, it is difficult for existing clustering ensemble algorithms to achieve a satisfying ensemble result, when the base clustering results are unreliable. Concerning this problem, we develop a new clustering ensemble model in this paper, which has several advantages compared with traditional algorithms: (i) structure information about the data is effectively extracted from the base clusterings; (ii) data characteristics and structure information are integrated in an elegant fashion, in the production of the consensus clustering result; and (iii) our model has the generative ability that makes the model achieve outstanding performance when training samples are insufficient. In our model, the structural information is extracted by explicating the coupling relationships between base clusterings and between samples in clustering members. Then, data characteristics and structure information are combined in a generative graph representation learning framework. And the objectives of representation learning and consensus clustering are integrated into a unified optimization model, in which the prior distribution of the data is approximated by a Gaussian mixture model (GMM). Extensive experiments are conducted on multiple IoT datasets; the results prove that our model not only performs better than the conventional clustering ensemble algorithms but also outperforms the state-of-the-art deep clustering methods.
Unsupervised network embedding learns low-dimensional vector representations of nodes based on the network structure. However, typical graphs only contain positive edges. Hence, most network embedding models take sampled negative edges as input to avoid trivial solutions. However, conventional negative sampling methods follow fixed distributions to generate negative samples. Such a static sampling strategy could dramatically hurt the performance of node embeddings in downstream tasks. Researchers have proposed several adaptive sampling strategies to improve the quality of negative samples. However, existing methods do not sufficiently use the network structure to explore high-quality samples, resulting in complex and inefficient negative sampling. This paper develops a novel dynamic negative sampling scheme, which maintains a candidate sample population for each node. We update the populations dynamically and retain high-quality samples in each iteration. We develop a novel network embedding algorithm based on the proposed model, which selects high-quality negative samples adaptively from the population. We conduct extensive experiments to evaluate its effectiveness based on several benchmark datasets. The experimental results verify the superiority of the proposed methods over state-of-the-art approaches.
Network embedding is an efficient and effective approach for network analysis. Current network embedding approaches focus on learning representations based on the proximity similarity principle, which are effective in capturing strong assortative structural relations but fail to model complex relations in many real-world networks such as disassortative structure. To this end, we introduce a novel approach – structure-characteristic-aware network embedding (SCANE) model. SCANE learns multi-view structure characteristics and optimizes them by non-negative matrix factorization to faithfully encode diverse complex graph structures. Furthermore, to adaptively aggregate structure information according to the characteristics of a given network and mitigate unnecessary workload caused by Grid Search, SCANE exploits the differential evolution to joint constrain its network embedding process and raise efficiency. From the experimental results of SCANE compared with seven state-of-the-art related network embedding models on eleven real-world benchmark datasets, the effectiveness and superior graph representation ability of the developed SCANE approach is observed.
Community embeddings are useful in node classification since they allow nodes to aggregate relevant information regarding the network structure. Modularity maximization-based algorithms are the most common approach to detect communities in networks. Moreover, spectral theory plays an important role in community detection by providing a matrix representation of the network structure. However, the literature is still scarce on spectral and modularity maximization-based community embedding methods. Besides, the node features of attributed graphs are usually not considered for producing the community embeddings. This paper introduces a community embedding algorithm based on spectral theory, called SpecRp, that has an overlapping modularity maximization-based step also herein proposed. SpecRp is a community detection method that considers node attributes and vertex proximity to obtain community embeddings. Computational experiments showed that SpecRp outperformed the literature in most of the tested benchmark datasets for the node classification task. Moreover, we observed that to detect disjoint communities, SpecRp and the reference literature algorithms presented a conflicting behavior concerning performance measures. While the reference methods achieved better results for modularity, SpecRp performed better concerning the Normalized Mutual Information to the ground-truth partitions. On detecting overlapping communities, SpecRp was considerably faster than the state-of-the-art algorithms, despite presenting worse results in most of the datasets.
A network embedding approach reduces the analysis complexity of large biological networks by converting them to low-dimensional vector representations (features/embeddings). These lower-dimensional vectors can then be used in machine learning prediction tasks with a wide range of applications in computational biology and bioinformatics. Several network embedding approaches have been proposed with different methods of generating vector representations. These network embedding approaches can be quite diverse in terms of data representation and implementation. Moreover , most were not originally developed for biological networks. Therefore comparing and assessing the performance of these diverse models in practice, in biological contexts, can be challenging. To facilitate such comparisons, we have developed the BioNE framework for integration of different embedding methods in prediction tasks. Using this framework one can easily assess, for instance, whether combined vector representations from multiple embedding methods offer complementary information with regards to the network features and thus better performance on prediction tasks. In this paper , we present the BioNE software suite for embedding integration, which applies network embedding methods following standardised network preparation steps, and integrates the vector representations achieved by these methods using three different techniques. BioNE enables selection of prediction models, oversampling methods, feature selection methods, cross-validation type and cross-validation parameters. Availability and implementation: BioNE pipeline and detailed explanation of implementation is freely available on GitHub, at
User recommendation aims at recommending users with potential interests in the social network. Previous works have mainly focused on the undirected social networks with symmetric relationship such as friendship, whereas recent advances have been made on the asymmetric relationship such as the following and followed by relationship. Among the few existing direction-aware user recommendation methods, the random walk strategy has been widely adopted to extract the asymmetric proximity between users. However, according to our analysis on real-world directed social networks, we argue that the asymmetric proximity captured by existing random walk based methods are insufficient due to the inbalance in-degree and out-degree of nodes. To tackle this challenge, we propose InfoWalk, a novel informative walk strategy to efficiently capture the asymmetric proximity solely based on random walks. By transferring the direction information into the weights of each step, InfoWalk is able to overcome the limitation of edges while simultaneously maintain both the direction and proximity. Based on the asymmetric proximity captured by InfoWalk, we further propose the qualitative (DNE-L) and quantitative (DNE-T) directed network embedding methods, capable of preserving the two properties in the embedding space. Extensive experiments conducted on six real-world benchmark datasets demonstrate the superiority of the proposed DNE model over several state-of-the-art approaches in various tasks.
Full-text available
Recently deep learning has been successfully adopted in many applications such as speech recognition and image classification. In this work, we explore the possibility of employing deep learning in graph clustering. We propose a simple method, which first learns a nonlinear embedding of the original graph by stacked au- Toencoder, and then runs it-means algorithm on the embedding to obtain clustering result. We show that this simple method has solid theoretical foundation, due to the similarity between autoencoder and spectral clustering in terms of what they actually optimize. Then, we demonstrate that the proposed method is more efficient and flexible than spectral clustering. First, the computational complexity of autoencoder is much lower than spectral clustering: The former can be linear to the number of nodes in a sparse graph while the latter is super quadratic due to eigenvalue decomposition. Second, when additional sparsity constraint is imposed, we can simply employ the sparse autoencoder developed in the literature of deep learning; however, it is non- straightforward to implement a sparse spectral method. The experimental results on various graph datasets show that the proposed method significantly outperforms conventional spectral clustering, which clearly indicates the effectiveness of deep learning in graph clustering. Copyright © 2014, Association for the Advancement of Artificial Intelligence ( All rights reserved.
The independent component analysis (ICA) of a random vector consists of searching for a linear transformation that minimizes the statistical dependence between its components. In order to define suitable search criteria, the expansion of mutual information is utilized as a function of cumulants of increasing orders. An efficient algorithm is proposed, which allows the computation of the ICA of a data matrix within a polynomial time. The concept of ICA may actually be seen as an extension of the principal component analysis (PCA), which can only impose independence up to the second order and, consequently, defines directions that are orthogonal. Potential applications of ICA include data analysis and compression, Bayesian detection, localization of sources, and blind identification and deconvolution.
We analyze skip-gram with negative-sampling (SGNS), a word embedding method introduced by Mikolov et al., and show that it is implicitly factorizing a word-context matrix, whose cells are the pointwise mutual information (PMI) of the respective word and context pairs, shifted by a global constant. We find that another embedding method, NCE, is implicitly factorizing a similar matrix, where each cell is the (shifted) log conditional probability of a word given its context. We show that using a sparse Shifted Positive PMI word-context matrix to represent words improves results on two word similarity tasks and one of two analogy tasks. When dense low-dimensional vectors are preferred, exact factorization with SVD can achieve solutions that are at least as good as SGNS's solutions for word similarity tasks. On analogy questions SGNS remains superior to SVD. We conjecture that this stems from the weighted nature of SGNS's factorization.
Scientists working with large volumes of high-dimensional data, such as global climate patterns, stellar spectra, or human gene distributions, regularly confront the problem of dimensionality reduction: finding meaningful low-dimensional structures hidden in their high-dimensional observations. The human brain confronts the same problem in everyday perception, extracting from its high-dimensional sensory inputs—30,000 auditory nerve fibers or 106 optic nerve fibers—a manageably small number of perceptually relevant features. Here we describe an approach to solving dimensionality reduction problems that uses easily measured local metric information to learn the underlying global geometry of a data set. Unlike classical techniques such as principal component analysis (PCA) and multidimensional scaling (MDS), our approach is capable of discovering the nonlinear degrees of freedom that underlie complex natural observations, such as human handwriting or images of a face under different viewing conditions. In contrast to previous algorithms for nonlinear dimensionality reduction, ours efficiently computes a globally optimal solution, and, for an important class of data manifolds, is guaranteed to converge asymptotically to the true structure.