=Paper=
{{Paper
|id=Vol-1548/087-Jakawat
|storemode=property
|title=OLAP Cube-based Graph Approach for Bibliographic Data
|pdfUrl=https://ceur-ws.org/Vol-1548/087-Jakawat.pdf
|volume=Vol-1548
|authors=Wararat Jakawat,Cécile Favre,Sabine Loudcher
|dblpUrl=https://dblp.org/rec/conf/sofsem/JakawatFL16
}}
==OLAP Cube-based Graph Approach for Bibliographic Data==
https://ceur-ws.org/Vol-1548/087-Jakawat.pdf
OLAP Cube-based Graph Approach for
Bibliographic Data
Wararat Jakawat, Cécile Favre, and Sabine Loudcher
Université de Lyon (ERIC LYON 2), France
{wararat.jakawat, cecile.favre, sabine.loudcher}@univ-lyon2.fr
Abstract. There is a growing number of different research fields that
are concerned with bibliographic data analysis. In many cases, data of
interest can be described as heterogeneous information networks. To ex-
plore knowledge from that networks in a multidimensinal way, OLAP
(Online Analytical Processing) analysis helps users to access data from
different points of views. The ability of OLAP for analyzing classical
data is clear. However it must be adapted to provide networked data by
considering both nodes and edges. In order to take into account linked
data in OLAP on networks, we propose a conceptual graph model to
represent bibliographic networks. Then we propose graphs enriched by
cubes. Each node and edge of the considered network are described by
a cube. It allows the user to quickly analyze the information summa-
rized into cubes. Our proposal also solves the slowly changing dimension
problem in OLAP analysis. To illustrate our approach, we integrate three
bibliographic databases and a computation process is defined according
to user’s needs. Then our implementation shows results on a real data.
1 Introduction
Information networks are ubiquitous due to the popular use of Web, blogs and
various kinds of online databases. Networks can be homogeneous or heteroge-
neous networks. Homogeneous networks contain a single object type and a single
edge type such as friends network, authors network and movies network. Hetero-
geneous networks are composed of multiple node and edge types. For example,
an author-paper network is an heterogeneous network with two types of nodes
(authors and papers) and three types of edges (”written” between authors and
papers, ”co-author relationship” between authors and the last one relates papers
written by the same author). A network can also be a multidimensional network
with multiple node attributes and edge attributes [8]. We take the example of
bibliographic data. In many fields, bibliographic databases store a collection of
information such as the publication title, authors, year, etc. To analyze biblio-
graphic data, there are many types of analysis (Statistics, Data Mining, Graph
Theory, OLAP analysis, etc.) to achieve different objectives in bibliometrics (re-
lationship studying, ranking, community mining, etc.). Among these different
types of analysis, we are more interested by OLAP analysis (Online Analytical
88 W. Jakawat, C. Favre, and S. Loudcher
Processing). OLAP provides the flexibility for navigating into data, for summa-
rizing data at different granularity levels and from different points of view.
Traditional OLAP did a great job on structured data, but OLAP faces chal-
lenges in processing networked data and it is called Graph OLAP. In several
recent approaches in Graph OLAP, a cube is created for a graph to provide
multidimensional and multilevel views [9]. In these approaches, we regret that
the interactions among objects are still hidden and that the slowly changing di-
mension problem is not taken into account [1]. For example, an author, Y. Sun,
published a paper when he was at Northeastern University. Then he published
another paper when he was at University of Illinois. There are two publications
of Y. Sun, one for each university. But from the authors network, if the user does
an OLAP operation like a Roll-Up in order to see the institutions network, these
two papers will be counted for both universities, and it is an incorrect answer.
In this case, networked data is non-summarizable: a higher level network cannot
be computed solely from the lower level network without accessing raw data.
So OLAP must be adapted to provide networked data by considering both data
objects and the interactions among objects. In order to take into account linked
data in the OLAP analysis of networks, we previously introduced a framework
to be able to analyze various networks built from bibliographic data [15]. In
the present paper, we extend the previous framework. Our contributions can be
summarized as follows:
– We use the properties of graph theory and we present a conceptual graph
model for bibliographic networks. Its content comes from multiple biblio-
graphic databases in a way that allows us to build several different networks
such as co-authorships, institutions of author, etc. The conceptual model is
mapped easily to support a variety of use cases.
– In order to adapt OLAP to multidimensional networks by considering both
nodes and edges, we propose graphs enriched by cubes. Each node or edge is
weighted by an OLAP cube. It allows the user to quickly analyze information
that has been summarized into cubes and by viewing the graph. It supports
Graph OLAP operations such as informational and topological operations
and it solves the slowly changing dimension problem.
– We evaluate our proposal on real data set and we provide examples to show
how using our tool to analyze data.
The remainder of this paper is organized as follows. Section 2 briefly reviews the
related work. Section 3 presents our proposal of graphs with cubes. In Section 4,
we present the implementation and we show results on real data set. Section 5
concludes this paper and gives future directions.
2 Related Work
In a previous paper, we have already surveyed research work that combines
OLAP and informational networks [9]. In this paper, we focus only on OLAP
and bibliographic networks.
OLAP Cube-based Graph Approach ... 89
There are different types of design in order to analyze bibliographic data.
The first one is a model based on the entity-relationship diagram [3]. The se-
cond model is based on the classical multidimensional model used in data ware-
houses [2, 5]. These models use relational databases to store bibliographic data.
So, they deal poorly with edges and making complex queries is not easy be-
cause several join operations must be added to answer users’ queries. Moreover,
traditional OLAP analysis can not be done on a graph.
The concept of Graph OLAP was first proposed by J. Han’s team [5, 6, 8].
Chen et al. presented basic definitions of OLAP on information networks and
they introduced a framework for Graph OLAP [5]. They provided a cube of graph
where each cell stores a network instead of a numeric value. Two kinds of OLAP
dimensions were defined (informational and topological dimensions) with two
kinds of OLAP operations to navigate on the dimensions. The first operation is
the informational OLAP. For example, venue and time in author-paper network
are two informational attributes. They are used as information dimensions. For
instance, they allow to build a network of authors for the ICEIS Conference
for all years and another one for the data mining field in 2010. The second
operation is the topological OLAP. For example, the network of authors can be
generalized by merging all authors of a same institution as one node and building
a new graph at the institution level. In this more generalized network, an edge
between Stanford and the university of Lyon will aggregate all collaborations
occurred between Stanford’s authors and the authors of the university of Lyon.
However, Chen et al. did not mention how to design model for heterogeneous
networks. Hence, Yin et al. answered this problem by defining a concept of entity
dimensions to support two dimensions of heterogeneous networks [10]. There are
two novel operations of entity dimensions named Rotate and Stretch, which are
able to mine edges between different nodes.
While Qu et al. focused on an efficient topological OLAP, they presented two
techniques (T-Distributiveness and T-Monotonicity) in order to achieve efficient
query processing and cube materialization [6]. Zhao et al. defined the concept of
multidimensional networks to abstract the real networks. They introduced a new
multidimensional model, called Graph Cube [8]. They worked with structure-
enriched aggregate networks and they proposed a new type of query, called
crossboid query in contrast with traditional queries named cuboid query.
The closest works to those of Han’s team are those of Tian et al. [7]. Tian
et al. proposed new operations for summarizing graphs. The first one, called
SNAP, can produce a summary graph by grouping homogeneous nodes. More-
over, users can control the different resolutions of summaries by a k-SNAP
operation.
According to Chen’s framework, only nodes are described by attributes.
However, in reality, edges are associated with attributes as well. For example,
co-authorship network contains authors as nodes and collaboration relationship
as edges. The relationships may be described by time or the papers they wrote
together. To solve this problem, Zhang et al. and Wang et al. proposed models
to deal with both node and edge attributes. Zhang et al. defined a new multidi-
90 W. Jakawat, C. Favre, and S. Loudcher
mensional network that contains node and edge attributes [11]. Node attributes
were defined as dimensions in a graph cube while edge attributes were defined
as dimensions in a data cube. While, Wang et al. proposed a new conceptual
model with an hyper graph [13]. Graph aggregation is performed on node and
edge attributes. The aggregated graph is a multigraph, where several edges can
be between two nodes. It allows users to see the different views.
In a different way, Kaya et al. developed three different networks (authors,
topics and venues) with a cube-based modeling method [14]. In these networks,
each node is represented by a data cube which is analyzed by OLAP operations.
To sum up the short related work about OLAP on bibliographic data, we
can add two remarks. The first one is about the slowly changing dimension
problem [1]. This problem happens when an object (a fact, a node, etc.) changes
its content over time and when this causes a change in the structure. For example,
as we said in introduction, the author, Y. Sun, published a paper when he was
at Northeastern University then he published another paper when he was at
university of Illinois. To the best of our knowledge, the existing approaches in
Graph OLAP are not complete with this problem.
The second remark is about the visualization of a multidimensional and mul-
tilevel view over graphs. For example, a cube, with a venue dimension and
time dimension, can contain a cell for (ICDE, 2008) and another one for
(DOLAP, 2008) cell. In the first Graph OLAP approaches, in each cell there is
a graph showing collaborations between authors for this venue and this year. Be-
tween two authors, we can see the collaborations only according to the venue and
the year, we don’t see a global view of the collaborations. Furthermore, Wang
et al. proposed a graph with multiple edges. However, their approach needs to
summarize a set of graphs with multiple edges and it is a complex task. In con-
trast, Zhang et al. used a single graph as input rather than a set of graphs. Kaya
et al. presented three networks where each node is represented by a cube.
Thanks to the related work, we can say that we want to :
– introduce a conceptual model for bibliographic networks based on graph
theory and not on the entity-relationship diagram.
– take into account the structure of the network in order to do topological
OLAP operations and not only classical or informational OLAP operations.
– deal with heterogenous networks and not only homogeneous networks.
– consider both node and edge attributes.
– have a global view of the network with multidimensional information.
– take into account the slowly changing dimension problem.
To extend OLAP on information networks, this paper presents graphs enriched
by cubes. The global idea is that each node or each edge is couple with a cube
according to user’s requirements. This graph model supports OLAP operations
for analysis.
OLAP Cube-based Graph Approach ... 91
3 Graphs Enriched by Cubes
In this section, we introduce graphs enriched by cubes. We first give an overview
of the overall process, then we explain the approach to create graphs with cubes.
Afterwards, we extend the OLAP operations to graphs enriched by cubes.
3.1 The Process
Figure 1 illustrates all of the components of our architecture in three steps.
User
NAVIGATION
OLAP Analysis Interface Properties: author’s name
Properties: year, title,
abstract
Meta Data
User’s requirements Results Navigation
cited by
write
author paper
1 2 3 Properties: order
institutions, country
GRAPHS WITH CUBES
publish in
PROCESSING
Build
Graph Computing nt
ain
co
keyword
venue
Graphs with Cubes
PRE-PROCESS
Properties: venue’s name
Integrate XML ETL Properties: keyword,
Conceptual year, research area
Graph category
Database Graph Model
Bibliographic XML Files
Databases
Fig. 2. The conceptual
graph model
Fig. 1. The Architecture
PRE-PROCESS. We first access bibliographic databases to extract data into
XML files. After integration, ETL process is used to extract and load data into
a graph database. In order to build an heterogeneous multidimensional network
G = (V, E, AV , AE ) where V is the set of multiple nodes, E the set of multi-
ple edge types, AV and AE respectively the set of attributes describing nodes
and edges, we introduce a conceptual graph model (Figure 2). The conceptual
graph model contains four types of nodes (author, paper, venue, keyword) and
four types of edges among these nodes. Each node and edge are described by
attributes. More details of nodes and edges are presented in Figure 2.
In reality, bibliographic data may have two problems. First, an entity con-
cerns many different values in the same property. For example, author named
Bin Yang works at Aalborg university and Fudan university in the same time.
Secondly, a property value is changing over time such as a change of institution.
For instance, Yzhou Sun published a paper in 2009 when he was at university
of Illinois (Urbaba-Champaign), whereas his other publications were published
for Northeasten university. In order to keep this information correctly, we design
institution as an edge property between author and paper. It is useful to track
changes over time.
92 W. Jakawat, C. Favre, and S. Loudcher
GRAPHS with CUBES PROCESSING. A graph enriched by cubes may be
used easily to perform OLAP operations on a network and it provides mul-
tiple network views at different levels of granularity. It takes a single graph
rather than a set of graphs. With user’s requirements, the first graph enriched
by cubes is built. For example, the user chooses as fact the co-authorship. Co-
authorship is a network where nodes are authors and an edge between two of
them indicates they coauthored papers. Formally, we use the concept of path as-
sociated with co-authorship network. There are different paths of co-authorship
from the conceptual model such as author − write − paper or author − write −
paper − publish − venue. Each path gives the different dimensions. For exam-
ple, dimensions as keywords, the year and the venue of papers are taken from
author − write − paper − publish − venue path. Therefore, dimensions can be
derived from node and edge attributes. Then, a first graph enriched by cubes
like in Figure 3 (d) is built. The network is the co-authorship network enriched
by a cube for each edge in order to count the number of papers written by two
authors according to keywords, years and venues. It has no sense to build a cube
for each node (author) because the fact is the co-authorship. While the fact is
the scientific production, the network is the authors network and cubes for nodes
and cubes for edges are created. It has a sense to count the scientific production
of an author or between two authors. In order to view the constructed network
from different perspectives, dimensions of cubes allow to perform multidimen-
sional analysis over networks. For enriched graph computing, we propose a new
algorithm (see section 3.2 for more details). Finally, graphs with cubes are sent
as the result to OLAP analysis interface.
NAVIGATION. The OLAP interface manages both the user’s needs and interac-
tions, the input and the output of graphs with cubes during analysis. The OLAP
interface uses meta-data in order to know the relationships between facts, mea-
sures, dimensions, nodes, edges, etc. It helps users to specify the first enriched
graph to start OLAP analysis. Then the interface allows users to explore graphs
and cubes from different views with OLAP operations.
In the next two subsections, we give more details about the algorithm for
computing graphs enriched by cubes and about the OLAP operations.
3.2 How to Build Graphs Enriched by Cubes?
The graph enriched by cubes construction involves two algorithms: BUILD-
GRAPH for computing the aggregated graph and BUILDCUBES for constructing
cubes on nodes or edges.
BUILDGRAPH (Algorithm 1) starts with the user’s requirements with a fact
F , a measure M , a set of dimensions D. We first generate a set of paths, P ,
which depends on fact, measure and dimensions (line 1) from G. Subsequently,
the algorithm creates the structure of nodes Vf0 , which contains node names
according to the fact and a list of paths id. Then, we traverse the set of paths. For
each path, we create a new node Vf0 , if there is no such value (line 4-5). Otherwise,
we simply update a path id for the node Vf0 (line 7). After the loop, the algorithm
creates the structure of edges Ef0 . Each edge contains edge name coming from
OLAP Cube-based Graph Approach ... 93
two any nodes and a list of measure’s values. For each vf0 in Vf0 , we compare the
list of measure’s values with the adjacent vf0 by using intersection operator. If
the comparison result is not empty, we create a new edge e0f (vf0 , vf0 + 1).
After the creation of the aggregated graph G0 = (V 0 , E 0 , VP0 ) where V 0 is
a set of generalized nodes, E 0 is a set of generalized edges and VP0 is a set of
paths associating with nodes, cubes are computed by BUILDCUBES. Due to
the limited of space, the idea of BUILDCUBES is that if the fact needs cubes
on nodes, the algorithm scans though Vf0 . Otherwise, it scans though Ef0 . The
measure’s value is computed from each path. Its value puts into cell that belongs
to its dimensions. Each Ef0 , we can find a list of measure’s values. We don’t keep
a set of paths id in edges because more than one path have the same measure’s
value. For example, paper44 is got from the path 1 and the path 3 because it is
written by two authors.
Figure 3 illustrates both algorithms. BUILDGRAPH takes as input the user’s
parameters. As previously, the fact is the co-authorship, the measure is the
number of papers, the dimensions are the year, the venue and the keywords.
In order to obtain the first graph, a set of paths is generated like author −
write − paper − publish − venue. In our example, there are 13 paths (Figure
3a). The next step is to compute a set of nodes. We get a list of authors with
their paths (Figure 3b). Then, any two authors who wrote papers together, are
added to a list of edges (Figure 3c). The number of papers on edges is com-
puted by using intersection operators. For instance, J. Han published paper33,
paper47, paper44 and etc. Y. Sun published paper44, paper10, paper47 and etc.
A set of papers between them is computed by {paper33, paper47, paper44, ...}∩
{paper44, paper10, paper47, ...} = {paper44 , paper47, ... . Due to needing
edges cubes of co-authorship network, the output graph of co-authorship net-
work is built by selecting a set of nodes from edges like in Figure 3d. Authors
who only write papers alone are not in the network.
3.3 OLAP Operations
Operations on cubes (e.g., roll-up, drill-down, slice, etc.) are supported to ex-
plore different multidimensional views and allow interactive queries and ana-
lysis. As we said before, two different types of operations are introduced in
Graph OLAP [5]. The first one is an informational OLAP, and it uses informa-
tional attributes. This operation doesn’t change the structure of the network.
For example, venue and time are two informational attributes with their respec-
tive hierarchies {year, decade, all} and {conf erence, area, all}. The second one,
topological OLAP, implies a new structure of the network; if we do a topologi-
cal Roll-Up, the network is generalized by some merging nodes. This operation
uses topological attributes. In the authors network, for instance, the hierarchy
{institution, country, all} associated with the node attribute author can be used
for merging authors from a same institution into a generalized node. In graphs
enriched by cubes, we can perform both informational and topological OLAP.
Informational OLAP operations are classically done, so we don’t give details
in the present. The most difficult problem we have to solve is how to support
94 W. Jakawat, C. Favre, and S. Loudcher
Algorithm 1 BUILDGRAPH
Input: An heterogeneous multidimensional network G = (V, E, AV , AE ), a fact F , a
measure M , a set of dimensions D
Output: A graph G0 = (V 0 , E 0 , VP0 ).
1: Generate a set of paths (P ) according to F , M and D
2: V 0 (value of node fact (Vf ), a set of paths id (pids)) = null
3: for each p ∈ P do
4: if p(Vf ) not in Node(Vf ) then
5: Create V 0 (Vf , p)
6: else
7: add p at Node(Vf )
8: end if
9: end for
10: E 0 (values of edge fact Ef , a list of measure0 s value) = null
11: for each i = 0 to V 0 .size do
12: for each j = i + 1 to V 0 .size-1 do
13: list1 = get the values of measure from node[i].pids
14: list2 = get the values of measure from node[j].pids
15: if list1 ∩ list2 != null then
16: Create E 0 (Vi and Vj , list1 ∩ list2)
17: end if
18: end for
19: end for
20: Build graph according to fact.
topological OLAP operations over networks. This problem is even more diffi-
cult if we take into account the slowly changing dimension over time. A higher
level of network cannot be computed from lower level of network without access-
ing raw data. Networked data is often non-summarizable. The idea of keeping
a set of paths into nodes in the previous algorithms allows us to solve this prob-
lem.
T opological roll − up. Figure 4b shows an example of a topological roll-up
of the co-authorship network to the institutions network. While all authors of
a same institution are merged as one node, edges are created when any two
institutions published papers together. In case of many institutions of an author
in the same time, the author is counted into all his institutions. After the roll-up,
in the more generalized network, new cubes have to be computed. In our example,
co-authorship network involves edge cubes, whereas institutions network needs
both cubes on nodes and edges. To build the institutions network, we use both
BUILDGRAPH and BUILDCUBES. Before computing a set of nodes (line 2 in
algorithm 1), we need to filter paths instead of generating a set of paths (line
1 in algorithm 1). We have to filter paths because all nodes of data set are
collected in V 0 , but some nodes may not be in co-authorship network (because
some papers are written by only one author). The step of path filtering is called
when the previous network needs cubes on edges. Then we compute a new set
of nodes from line 2 in algorithm 1. Refer to example in Figure 4b, nodes are
OLAP Cube-based Graph Approach ... 95
!"#$ %&'()*+,*"'-+!%!-*+!&./"0(+1-2&-$ #$"%&'(" H!I7IGJ"
!" #$"%&'()*+,-.)/'/.+00)/123,4&)566" =$">1?" H1?)*+,-.)/'/.+00)/123,4&)566" G" G"
0" =$">1?)*+,-.)/'/.+!@)/123,4&)566" K2L">.-"(M"?(N.4"*,-&"3,4-"(M"/'-&",N"
A" =$">1?)*+,-.)/'/.+08)/123,4&)9:5;"
B" C$"D'?)*+,-.)/'/.+<<)/123,4&)566" #$"%&'(I" H/'/.+00I"GJ"
=$">1?"
8" C$"D'?)*+,-.)/'/.+08)/123,4&)9:5;"
#$"%&'(I" H/'/.+08I"GJ" >,Q,3'+,-S" <"
E" C$"D'?)*+,-.)/'/.+00)/123,4&)566" C$"D'?" 5.S*(+N4" ;,?,?P"
P+'/&"
F" C$"D'?)*+,-.)/'/.+<<)/123,4&)566" =$">1?I" H/'/.+00I"/'/.+08I"GJ" 7@!7"
G" G" C$"D'?" =.'+" 7@!!"
7@!@" !"
G" G" 7@@F" 7" 7"
566"
9:5;"
U6VW"
>:R:X"
T.?1."
3&.-$)4$567%2$%2#$86$9&2$
K'L">.-"(M"/'-&4" KOL">.-"(M".NP.4"*,-&"3,4-"(M"Q.'41+."" KNL"R+'/&4".?+,O&.N"2S"O12.4"
Fig. 3. Computation of a co-authorship network
+>BC"6D@$ =>?>6"5>9>"%@$
!79:;95$ <$
567&8$
'()'$
'())$
'()($ )$ '$
'((*$ '$ '$
+,,$
-.+/$
0,12$
3.4.!$
A>9%>$
+>BC"6D@$/;9;95$ '$
F>?C"6:$
!"##$%&$ '()<$
E>76$ '()'$
'((*$
)$
'$
+,,$
0,12$
A>9%>$
G7H$-"I7%?8"6@8;&$9>?C"6:$G9%JK>6$"L$&7&>6@$M$'H$ GKH$.9@N?%N"9$9>?C"6:$G9%JK>6$"L$&7&>6@$M$'H$
Fig. 4. Roll up from the co-authorship network to the institution network
grouped into institutions. For example, university of Illinois contains path6 and
path7 because J. Han and P.S. Yu belong to this university.
Slice. Traditional slice operation selects one particular dimension from
a given cube and provides a new sub-cube. In our context, slice operation can not
be like the classical one, it should be adapted to graphs. The slice operation se-
lects a part of the graph and provides a new sub-graph. For example, if a whole
co-authorship network is too big to be comprehensive, the user can focus on
a smaller subgraph more interesting to analyze information clearly.
96 W. Jakawat, C. Favre, and S. Loudcher
4 Experiments
In this section, we present how we have implemented our solution and we give
some examples of analysis.
4.1 Implementation
The implementation has been done as follows:
– We get data from three bibliographic databases. First we use the well known
database DBLP. But in order to complete our conceptual graph model, we
access on ACM and Microsoft Research databases for taking keywords, in-
stitutions and research areas. In theses three sources, we keep only three re-
search areas (data mining, databases and information retrieval) and we pick
only a few representative conferences for the three areas (PODS, EDBT,
KDD, DOLAP, ASONAM, SIGIR and CIKM). At the end, we build a data
set which contains 4,727 papers and 8,238 authors since 2009.
– The ETL process is used to fulfill data into the model. After cleaning, data
is loaded into an unified structure with a graph model.
– A new type of NoSQL databases, called graph databases, is used to im-
plement our conceptual graph model. We choose Neo4j1 because it is an
open-source software, it supports the properties of our graph model.
– Finally, an OLAP interface analysis is developed in Java and tested on
a Mac OS X version 10.9.2 with Intel core i5 2.4 GHz and 8 GB of Ram. For
graph visualization, we use the GraphStream2 library because it is a library
to model and analyze the dynamic of graphs and it is an open source library.
4.2 Example of Analysis
We first use the example of the co-authorship network introduced before Figure 5
shows the co-authorship network in three areas since 2009. Each edge of this
network has a cube. In order to reduce the graph, we filter edges in order to
keep only edges with a number of papers over than 10.
For example, look at the edge between Iadh Ounis and Craig Macdonald;
these authors published 29 papers together. We consider these papers like a cube
with two dimensions. It could be interesting to have two ways of visualization.
The first way is to focus on time, having the count of papers per year. Each
year has the count of papers by venues. The second way is to focus on the venue,
having the count of papers per venue’s name. Each venue has the count of papers
by year.
Now, if we want to roll up the co-authorship network of the conference KDD
between 2009 and 2013 to the institutions network (for the same conference and
years) with a topological OLAP operation, we obtain Figure 6. The institutions
are filtered with a number of papers over than 10. For example, the university of
Illinois at Urbana - Champaign published 31 papers in the KDD conference from
1
http://neo4j.com/
2
http://graphstream-project.org
OLAP Cube-based Graph Approach ... 97
!"
!"
Fig. 5. The co-authorship network (on three areas and all years) with a number of
papers over than 10
!"
#"
$"
Fig. 6. The institution network for the KDD conference (2009-2013) with a number of
papers over than 10
2009 to 2013. Look at the big number 1 in Figure 6, it means that the university
of Illinois at Urbana - Champaign has one collaboration with the New York State
Museum and one with the New York State department of Environment in 2010
by publishing in the KDD conference. The big number 2 shows the number
of papers written by several authors but all belonging to the same university
(Illinois at Urbana - Champaign).
Furthermore, in the interface, the user can slice to consider only a sub-graph.
98 W. Jakawat, C. Favre, and S. Loudcher
There are several groups of authors in co-authorship network. Suppose that the
user needs to consider only the interest group; with a slice operation, the user
can select the sub co-authorship network. Finally, a roll-up operation is done on
this sub-graph.
5 Conclusion
In this paper, we wanted to enhance decision support on networks by combining
OLAP on networked data. So we first presented a conceptual graph model to
support OLAP on bibliographic networks. In order to consider both nodes, edges
and multidimensional information for the analysis, we proposed the graphs en-
riched by cubes. The graphs enriched by cubes perform multidimensional views of
a heterogeneous graph rather than a set of graphs. Cubes are provided for nodes
or edges according to the user’s requirements (fact, measure, dimensions, etc.) In
order to compute graphs enriched by cubes, we proposed the algorithms which,
in addition, solve the slowly changing dimension problem in OLAP analysis.
Then we adapted the OLAP operations to graphs enriched by cubes. We showed
an implementation with real data sets from three bibliographic databases. We
focus our approach on bibliographic data in this paper, and it can be applied to
other use cases as well.
References
1. Waqas, A., Zimányi, E., Wrembel, R.: Temporal Data Warehouses: Logical Models
and Querying. In: EDA’15, vol. RNTI-B-11, pp. 33–48 (2015)
2. Ferrara, A., Salini, S.: Ten challenges in modeling bibliographic data for bibliomet-
ric analysis. In: Scientometrics, vol. 3, pp. 765–785 (2012)
3. Mallig, N.: A relational database for bibliographic. Journal of Informatics,
pp. 564-580 (2010)
4. Trifonova, T. G.: Warehousing and OLAP Analysis of Bibliographic Data. Intelli-
gent Information Management, vol. 3, pp. 109–197 (2011)
5. Chen, C., Yan, X., Zhu, F., Han, J., Yu, P.S.: Graph OLAP: Towards online ana-
lytical processing on graphs. In: ICDM’08, pp. 103–112 (2008)
6. Qu, Q., Zhu, F., Yan, X., Han, J., Yu, P.S., Li, H.: Efficient Topological OLAP
on Information Networks. In: DASFAA’11, Part I. LNCS, vol. 6587, pp. 389–403
(2011)
7. Tian, Y., Hankins, R.A., Patel, L.M.: Efficient Aggregation for Graph Summariza-
tion. SIGMOD Conference, pp. 567–580 (2008)
8. Zhao, P., Li, X., Xin, D., Han, J.: Graph cube: on warehousing and OLAP multi-
dimensional networks. In: SIGMOD’11, pp. 853–864 (2011)
9. Loudcher, S., Jakawat, W., Morales, E.P.S., Favre, C.: Combining OLAP and in-
formation networks for bibliographic data analysis: a survey. Scientometrics, vol.
103(2), pp. 471-487 (2015)
10. Yin, M., Wu, B., Aeng, Z.: HMGraph OLAP: a Novel Framework for Multi-
dimensional Heterogeneous Network Analysis. In: DOLAP’12, pp. 137–144 (2012)
11. Zhang, J., Hong, X., Peng, Z., Li, Q.: Nestedcube: Towards Online Analytical Pro-
cessing on Information-Enhanced Multidimensional Network. WAIM Workshops,
pp. 128–139 (2012)
OLAP Cube-based Graph Approach ... 99
12. Beheshti, S.M.R., Benatallah, B., Motahari-Nezhad, H.R.: A Framework and
a Language for On-Line Analytical Processing on Graphs. In: WISE’12,
pp. 213-227 (2012)
13. Wang, Z., Fan, Q., Wang, H., Tan, K.L., Tan, Agrawal, D., El Abbadi, A.: Pagrol:
Parallel graph olap over large-scale attributed graphs. ICDE’14, pp. 496-507 (2014)
14. Kaya, M., Alhajj, R.: Development of multidimensional academic information net-
works with a novel data cube based modeling methord. Information Sciences,
pp. 211–224 (2014)
15. Jakawat, W., Favre, C., Loudcher, S.: OLAP on Information Networks: A New
Framework for Dealing with Bibliographic Data. SoBI’13 in conjunction with
ADBIS, pp. 361-370 (2013)