Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

Pavithra.R Sugantha Priyadharshini.

P
Department of Computer Science and Engineering Department of Computer Science and Engineering
Sri Ramakrishna Engineering College, Sri Ramakrishna Engineering College,
Coimbatore, TamilNadu, India Coimbatore, TamilNadu, India
pavithra.r@srec.ac.in sugantha.ravee@srec.ac.in

Abstract Graph matching (GM) is matching a collection of Graph matching problems has two important limitations. It is
edges that has no common vertices in a graph. It is an essential NP-hard, also hard to estimate. To deal with the above
problem in computer vision, biosciences, information technology, mentioned drawbacks, the factorized graph matching approach
distributed control and facility allocation and has an important is used (FGM). It reduces the huge pair wise similar matrices
role in solving correspondence problems. Here, optimal graph to smaller matrices which determine the intermediate structure
matching problem for real time DNA datasets are considered. of every graph as well as pair wise similarity among the edges.
The Weighted Graph Matching and Maximum Weighted Graph The factorization has many benefits like: Affinity matrix that
Matching algorithms are used for deciding the optimal matching
is costly (in terms of space and time) need not be computed.
between two weighted graphs. Our work uses an analytic
approach to optimize matching problem. The images are
Also factorization methodology uses path following algorithm
converted to adjacency matrices and by computing the Eigen which gives improved optimization approaches and matching
decompositions of the adjacency matrices in case of undirected performance.
graphs or Hermitian matrices in case of directed graphs, a Our proposed approach finds an approximate solution by
matching nearer to the best solution can be obtained once the firstly reducing a large affinity matrix into smaller ones by
graphs are satisfactorily nearer to one another. using factorization approach and later uses weighted graph
matching algorithm for optimizing the problem.
Keywords Graph Matching; Factorized Graph Matching; The weighted graph matching algorithm, if given two
Path Following Algorithm; Maximum Weighted Graph Matching weighted graphs, will search an optimal placement of nodes in
such a manner that the edge weight difference is minimal
I. INTRODUCTION between them. Graph that has weighted edges and same
Graph matching is an extensively researched area in number of nodes is taken into consideration and approaches
different topics of computer vision[14] like ranking in large analytically by using the adjacency matrix for undirected
graph matching.
graphs, finding similarity of nodes in graphs, large scale
complex graphs and telephone problems and clustering of
graphs. Graph theory is the study of graphs that computes the II. RELATED WORK
relationship between edges and vertices. It is used to find This section deals with the various approaches related to
shortest path in road or a network. A matching graph is a sub graph edit distance, similarity search and optimizing the graph
graph of a graph in which edges are not adjacent to each other. structure in graph matching.
Basically, there should not be any common vertex between Feng Zhou et al. deals with the problem of QAP by using
any two edges. That is, each vertex in matching M has degree factorized graph matching (FGM) method. FGM factorizes the
one. [12]

1
reduces the huge pair wise similar matrices to smaller matrices This paper proposed an indexing approach using the Q- Gram
which determine the intermediate structure of every graph as
well as pair wise similarity among the edges. FGM method has the graph into small grams based on the Q-Gram idea for
low computation cost in pair wise affinity matrix and leads to candidate filtering. Also Guoren Wang et al. used many
better optimization strategy as mentioned already. constructions of inverted index methods as well as query
Because of its combinatorial nature in GM, it is not easy to processing via online. By this way many graphs required to
predict the actual solution. Hence it is converted to inexact continue into exact matching can be decreased to the maximum.
graph matching that effort for computing a best possible Other sections of our paper are arranged as: Section
matching for the original problem. Tianshu Yu et al. used an III briefs existing system. Section IV is the proposed system.
approach for the relaxation of the rank into convex form, the Section V is the investigational results and Section gives the
proposed approach makes the graph matching to a standard inference for this paper.
Semi- Definite Programming, which is easy to solve. III. EXISTING ALGORITHM
As a detailed representation, graph can define objects with
significant facts. Ruonan Zhang et al. recommended the K- A. Path following algorithm
Nearest-Neighbor-Pooling Matching (KNNPM) method,
This algorithm was used to resolve the matching problems
which incorporates the feature pooling method into the graph
in undirected graphs and produced effective results in graph
matching problem with lower solving complexity.
matching. The path following algorithm can also be extended
Xinbo Gao et al. briefed about graph edit distance to work
by implementing it in directed graphs by utilizing concave
out on the similarity issue among pair wise graphs in inexact
relaxation. Both synthetic and real time data sets can be used
graph matching and is also applied in pattern analysis as well
for implementation.
as recognition.
According to Ayser Armiti et al. inexact graph matching is
B. Pseudo code for Path Following Algorithm
one of major problems for geometric graphs in two
Initialize the inputs Affinity matrices with edge pair namely
dimensional spaces. They try to systematize the GM problem
Kp,Kq and Pair of graphs with vertices g1, g2
in an extreme probability evaluation framework. Then, to
Initialize output (K) variable as zero
calculate match between two graphs approximately, the
Set X to be a doubly stochastic matrix
expectation maximization method is incorporated. The
Factorize L=UV^T with singular valued decomposition
proposed system calculates the density function of a graph
:1 and start path-following
based on its sub structures to improve the system scalability.
Junchi Yan et al. concern the problems of matching
Update the matrix with X-X0
between multiple graphs stating to matching. They used multi-
Improve the factorized value by convex optimization and
graph matching methodology to integrate the dual aspects by
modified Frank-wolfe algorithm
improving the affinity score, in intervening time steadily
else if optimized value < X Jg(X)
infusing the consistency.
Optimize using Frank-wolfe method
Most of the graph data available in the graph database are a
Update X-X0(weighted matching occurs in graphs)
lot noisy and incomplete, increasing its importance to retrieve
End
graphs from the graph database. Weiguo Zheng et al says that
The pair of graphs, g1 and g2 is considered and the dual
for retrieving graphs by the edit-distance based similarity
affinity matrices, namely Kp R n1×n2 and Kq R m1×m2
search problem from the graph database. Since searching
are utilized to determine harmony of every node. The edge
similarity between graphs is complex, the method recovers
pair is also computed. When two graphs and the associated
graphs which are alike to a known given query graph. The U -
tree based pruning method is used for similarity search and affinity matrices are given, the best correspondence X amongst
retrieval of graphs from the graph database. The proposed nodes is the totality of the node as well as edge compatibility
can be increased by means of Jgm(X)
technique is better in relation with pruning power as well as
query response time says Weiguo Zheng .
Jgm(X) = tr(KpTX)+tr(K qT(G 1T*G2.H 1T*H2)/Y) (1)
Silvio Cesare et al. used graph matching for malware
detection. Using the distance metrics of the graphs, a similarity Where,
search is made to detect the malwares. The distance metric is
X is the input data
utilized to perform pre-filtering and it is effective although it is
a less computationally efficient metric.
Kq is the edge pair of matrix
Similarly, graph indices can be used to perform graph g1 is the vertex1
similarity search. Guoren Wang et al. mainly concerns for g2 is the vertex2
similarity search using index structure over a large set of The major disadvantage in path following algorithm is that
graphs. concave and convex optimizations are approximated and the

2
inputs are selected in random manner so, the results are not retrieved from factorized matrix A of size (350*350). The
accurate. below graph shows factorized graph from matrix A.

IV. PROPOSED ALGORITHM

A. Weighted Matching algorithm


Each node in the graph contains the information about the
nearer nodes as well as the edge weights. It can interact with
all the nearer nodes by exchaning messages. We will be able
to predict the node with heaviest weight by considering an
edge e = (u, v) where, u and v see the other edges and
compares the weight and consider the heaviest of all the edges
from the neighboring nodes. If e is the node with heaviest
weight than all the edges nearer to it means, then that node is
considered to be nearby heaviest edge. If the occurrence of Fig.5.1. Factorized Graph generated from matrix A
matching pair is found then all the neighbor nodes will get
notified and also stops looking for the edges that are incident. Similarly, factorized graph from matrix B is shown below.
All the nodes are traversed iteratively until all the matching
pairs are found.

B. Maximum Weighted Graph Matching


Maximum weighted graph matching has many advantages
over weighted graph matching. Here, the augmented path is
discovered to match the edges and a vertex in bipartite graphs
and results are accurate and yield the better optimal solution.
A maximum weighted matching is the summation of all
weighted edges that gets matched will hold the maximum
weight. It is also known as assignment problem. To solve
assignment problem, the graphs are pruned into smaller
matrices from large matrix such that the graphs are easily
matched. Fig.5.2. Factorized Graph generated from matrix B
Now, the Path following algorithm is used and the matched
C. Maximum Weighted Graph Matching pseudo code edges and nodes are shown from Factorized Graph A and
Initialize the input values Factorized Graph B with approximated result values which is
Introduce super source nodes to create the adjacency matrix rounded further. The below figure shows the graph obtained
Use ford Fulkerson algorithm to find the augmented path after applying path following algorithm.
Traverse from root node by choosing maximum weight
Send request to neighbor nodes

Set the maximum weight and compare the other maximum


weight to check the heaviest node
Match the maximum weighted edges and vertices
End

V. EXPERIMENT AND RESULT


Matrix Representation for Gene Sequences (DNA)
The adjacency matrix A (350*350) and adjacency matrix B
Fig.5.3. Graph generated after using Path following
(350*350) are taken from the dataset of gene sequences algorithm
(DNA) images through imperfect domain theory. The DNA
sequence points are removed to identify the restrictions among
exons and introns. The below figure depicts the result obtained from
A large pair wise affinity matrix is pruned into smaller Maximum weighted graph matching algorithm from
matrix through factorization. The actual structure of every Factorized Graph A and Factorized Graph B.
graph gets converted and the pair wise affinity among edges is
3
Fig.5.7. Image Matching using weighted graph matching algorithm.
Fig.5.4. Graph generated by using Maximum weighted graph matching
The structure of the DNA is plotted from Matrix C which The below figure depicts the accuracy comparison
is the intermediate result from factorized graph matching. The obtained by comparing the path following algorithm and
factorized matrix of size (50*50) is shown below. Maximum weighted graph matching algorithm.

Fig.5.5. Image plotted from Matrix C


Fig.5.8. Accuracy comparison
The structure of the DNA is plotted from Matrix D which
The below figure depicts the objective ratio obtained by
is the intermediate result from factorized graph matching. The
comparing the path following algorithm and Maximum
factorized matrix of size (50*50) is shown below.
weighted graph matching algorithm.

Fig.5.6. Image plotted from Matrix D Fig.5.9. Objective ratio


Matching between two DNA structures is by using The below figure depicts the edge affinity matrix obtained
maximum weighted graph matching. The below result by comparing the path following algorithm and Maximum
obtained shows the perfect matching between the two images weighted graph matching. The Maximum weighted graph
by using Maximum weighted graph matching algorithm. matching yields efficient result while comparing to path
following algorithm.
4
Analysis And Machine Intelligence, Vol.38,Issue 6, pp1-15,2016.
Weiguo Zheng, Lei Zou, Xiang Lian, Dong Wang, and Dongyan Zhao,
Transactions On Knowledge and Data Engineering, Vol. 27, No. 4, pp
978,2015.
-Based Malware Variant
EEE Transactions on Dependable and Secure
Computing, Vol.11 , Issue: 4, pp 307-317, 2014.

Indexing Large Sparse Graphs For Similarity


Transactions on Knowledge and Data Engineering, Vol
24, issue 3, pp 451, 2014.

Fig.5.10. Edge affinity matrix

-634, 2015.
IV.CONCLUSION
The WGMP has been used for approximate solution in
undirected graphs case. The Weighted Graph Matching
Algorithm method always results the improved matching
when the graphs are nearer to each other. The challenging
part of the algorithm is that it needs more number of
iterations in order to factorize to get the improved results.
Initially local optimal solution is achieved by using path
following algorithm. In the proposed system, the nearer
global optimal solution is achieved and yields exact
matching between graphs. It finds an augmented path
between many nodes to find the matching node.The
converged solution is guaranteed to be binary and further
discrete rounding step is not needed. Consequently, the
advanced methods can be used to enhance factorization.
Besides we try to extend factorization methodology to
additional higher-order graph matching problems.

References
[1] Feng Zhou,Fernando
Transactions On Pattern Analysis And Machine Intelligence, Vol. 24, no.
6, pp. 381 395, 2015.
[2] -Rank
f Computer
Vision (WACV),2014.
[3] -Nearest

Conference on Multimedia & Expo Workshops (ICMEW),2016.


[4] Survey Of Graph
-129 2010.
[5]
Similarity:A
26th International Conference on Scientific and Statistical Database
Management, Vol No. 27, pp 1 12,2014.
[6]
Edit Distance In Quadratic T
Computational Biology and Bioinformatics,Vol.17,Issue 2,2014.
[7] Junchi Yan, Minsu Cho, Hongyuan Zha, Xiaokang Yang, And Stephen
-Graph Matching Via Affinity Optimization With
5

You might also like