Visual Relation Identification Using BoFT Labels in Social Media Feeds

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395 -0056
Volume: 04 Issue: 06 | June -2017 www.irjet.net p-ISSN: 2395-0072
© 2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 1647
Visual Relation Identification Using BoFT Labels in Social Media Feeds
Syari Sasi1, Resmipriya M. G2
1PG Scholar, Dept. of Computer Science and Engineering, Amal Jyothi College of Engineering, Kerala, India
2Assistant Professor, Dept. of Computer Science and Engineering, Amal Jyothi College of Engineering, Kerala, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Social media is an online platform that became an
inevitable part of individuals. Interest of online users can be
detected from social graph. Socialgraphinterruptstheprivacy
of online users so it is dubious. This paper adapt users interest
from user shared images. Users share images on social media,
it depicts the interest of users or criticism about a real time
issue or despatch of users. These user shared images are
reachable to other users in contact which provide an easier
and systematic alternativefordiscoveringconnectionbetween
users. The proposed methodology investigatestodiscoveruser
connection through a better alternativecalledbag-of-features
tagging (BoFT). User shared images also helps to establish
connection between users. By BoFT, the users interest are
recognized and thereby estimate tag recommendation for
users. This paper identifies the connectionbetween usersusing
the image sharing mechanism and determines the tag
recommendation for friendship. These findings are useful for
interest discovery which categorize users based on propensity
of topic and the virality of user shared images are predicted.
Predicting the virality of user shared images are computed
based on the occurrence of images in social network.
Key Words: BoFT labels, connection discovery, user
shared images, user recommendation, content virality,
interest discovery.
1. INTRODUCTION
Nowadays, owing to the behaviour of contemporary society,
direct communication between individuals are decreasing
day-by-day. The social network plays a vital role for
communicating individuals. Smart phones, social media and
internet became an indispensable part of humans. Through
social media, the interaction link between the internet users
are exaggerated at higher rates. Billions of users share
billions of images in social network. It symbolise the social
link between users. Social network enables users to chat,
share images, videos, blogs etc. The proposed method relies
on users shared images. User shared images on social media
is a fundamental data to identify the user connection.
Most of the social media gather information of user profile
using social graph. A better alternative to recognize the
interconnections between users are user shared images. By
taking the advantages of extreme resemblance of features in
shared images for grouping users with their interest and
thereby discovering user connection. For this, a computer
vision technique called BoFT is adopted. Examples of user
shared images are demonstrated in Fig. 1. Both user A and
user B shared images of aircraft anduserBanduserCshared
images of bike in common. These user shared imagesexhibit
visual interrelation in features. The visually extreme
resembling images are utilized to identify interconnection
between users.
Fig -1: Example of user shared images
In the proposed framework, user shared images are the key
of connection discovery. Recommendations are based on
matching user shared images with other users. A better tag
recommendation is possible by estimating user connection
through analyzing the user profile which reflect users
interest. The proposed methods are evaluated withover 200
images from Facebook and it is proven thatuserconnections
can be discovered and user/tag recommendations are made
efficiently. The main contributions of this paper includes: 1)
user shared images with BoFT methods helps to discover
user connections 2) establish a tag recommendation
approach based on user connection 3) an approach to
predict the content virality based on occurrence on image 4)
propose a method to recognize the interest of users. Most of
the algorithm focus on predicting the content virality based
on time taken for the image to become viral. But the
proposed algorithm focus on occurrence of images.
This paper is organized as follows: section 2 bring out
related works. Section 3 introduces proposed method for
similarity calculation and connection discovery, along with
user/tag recommendation and interest discovery followed
by virality prediction of user shared images. Section 4
familiarize with dataset and come out with experimental
results. Section 5 concludes the proposed system and future
works.

2. RELATED WORKS
A relation identification system compares the system with
existing technologies. Connection between users in online
social network is identified through the interest of users in
sharing of images. Users interest are analyzed by accessing
social graph [14]. But social graph interferes with the
protection of online users thus it become improbable. Non-
user generated labels with differentcolor-basedandfeature-
based methods are used in connection discovery [7]. But
larger social network data to prove the effectiveness of the
feature based method. One of the conceivable method for
finding user connection is content based approach, in which
the visual appearance of an image is considered, inorder to
create a label for the image [23]. However, determining the
relationship between the appearance and the label is not a
trivial task because the same object can be visually different
among images. To conquer these drawbacks, BoFT isusedin
[4], [27]. BoFT is a computer vision approach for image
feature classification.
The virality prediction algorithm is based popularity of
images. But most of the algorithm [10] is based on time
constrain using basic reproduction number. But the
popularity of images in social media is on the basis of
number of views or sharing of images. In proposedapproach
the virality prediction is based on the number of occurrence
of image, which precisely shows the viral images.
Recommending users from billions of user shared content is
always a challenging task. The Probabilistic Prediction
Framework [2] are used to improve tag recommendation.
However, it is based on assumption of addictive
independence which is not acceptable. So it leads a poor
recommendation. One of the possible ways to make the user
recommendation is by the existing connections among
people [20]. User recommendation is also possible with
users interests [24] or user generated content [27] and
personalized information [2]. One ofthepossiblesolutionsis
to analyse the interests reflected in the content they have
shared. Thus user interests are identified from user shared
images for user recommendation.
The proposed technique is based on BoFT approach[4],[27]
for connection discovery using the following ways: 1)
screening usersharedimagesandmeasuringBoFTsimilarity
2) Discovering the user connection using BoFT similarity
calculation 3) recommending users with respect to the
discovered connection 4) occurrences of images are
obtained for measuring the virality of content 5) group of
users with special area of interest are identified from user
shared images. The connection discovery approach is useful
for user recommendation, virality prediction and interest
discovery.
3. PROPOSED METHODOLOGY
The proposed technique identifiestheinterrelationbetween
users, for this, it precisely measures the similarity between
user shared images in which they shared extreme
comparative images. The proposed technique comprises of
four phases. The first part distinguishes user connection
based on similarity calculation. The second part comes out
with proposal for user tag recommendation for friendship
based on user interest which is figured by utilizing the
similarity in user shared images. The third part presents the
virality in user shared images based on popularityofimages.
The fourth part presents interest discovery, which
demonstrates the interest of a group of users. Fig 2 shows
the system flow of proposed technique.
Fig -2: System flow of the proposed method.
Visual Relation identification in social media is based BoFT
[4], which is a computer vision approach and is used for
analyzing images. Converting local image descriptors of an
image into feature vector is termed as BoF. BoF can be
adapted to different methods such image classification,
object detection, image retrieval. An orderless collection of
visual features of an image constitute BoF, which is
computationally cheaper and theoretically straightforward.
The proposed technique focuses on identifying the
interrelation between internet users and subsequently
categorizing the users for tag recommendation. BoF is a
multistep approach and each step denotes a different
process. Fig 3 (a) is BoFT label annotation and Fig 3(b)isthe
similarity calculation of user shared images using BoFT
labels. The key steps of proposed techniques are:
• Feature Extraction
• Codebook Generation
• Coding and Pooling
• Clustering and BoFT Labeling
• Profile Learning
• BoFT Similarity Calculation

• User Tag Recommendation
• Virality Prediction
• Interest Discovery
Fig -3: BoFT approach for connection discovery
3.1. Feature Extraction
The first key step of the proposed method is feature
extraction. Feature extraction is a process to obtain the
unique local features. SIFT based feature extraction method
[15] is used in the proposed framework.SIFT withmaximum
key points extraction is opted. SIFT includes a feature
detector and a feature descriptor. The extracted featuresare
compatible under different viewing angle and lightning
condition and is liberated of size and orientation. SIFT
includes four functions:
 Difference of Gaussian
 Keypoint localization
 Orientation
 Keypoint description
The primary phase of keypoint detection is to recognize the
locations and scales. In Gaussian, finding the keypoints scale
space and rotation of images. DifferenceofGaussianfunction
is for identifying images under different orientation and
angles. That is, points that are invariant to scale and
orientation. In keypoint localization, the selection of key
points is based on the parameters such as threshold, edge
threshold and remove boundary points. Keypoints are
selected based on invariant points. One or moreorientations
are assigned to each keypoint location based on gradient
information on the image and compute the best orientation
for each keypoint region. The Key point descriptor is
computed for the local image region about each keypoint
that is highly distinctive and invariant points.
3.2. Codebook generation
Codebook generation is for acquiring the image feature
categorization. It is a clustering process, so k-means
clustering is adopted, which is used to obtain a set of visual
words. K-means clustering is used to cluster bag-of-features
of user shared images. Each cluster accommodates with
different features of the images. Clustering is based on the
extreme resemblance of extracted features. Afterwards,
proper clustering is acquired based on theclusterparameter
inorder to create Bag-of-visual word. Then visual word is
obtained by taking the mean of each cluster.
3.3. Coding and Pooling
Feature coding is the pivotal component of image
classification. Coding is to map local features into a compact
representation. The closest visual word is obtained from
codebook. Features are trained in coding. The inputs of this
step are feature descriptors extracted from all training
images and the outputs are codewords. The codewords are
typically generated by clustering technique over feature
descriptors. Feature coding is the core component, which
links feature extraction and feature pooling. Each image is
represented by a feature vector in the feature pooling. The
output is a pooling vector for each image.
3.4. Clustering and BoFT Labeling
The feature vector obtained in pooling is used in clustering
process to group images that are visually similar. K-means
clustering method is adopted for grouping similar feature

vectors of image. In BoFT labeling, it assigns each cluster a
BoFT label and images with the same BoFT label arevisually
similar. User connections can be discovered through BoFT
labels. Each cluster contains a cluster id. The cluster id
represents labels of each cluster and determines which
cluster the images belong to. Consequently, identify the
category of user shared images.
3.5. Profile Learning
The key of connection discovery is profile learning. Unique
users are identified from the dataset and user category and
category count are calculated from the cluster. The number
of unique users who share images can beidentified. The user
category indicates in which cluster the user have images.
Based on user category similarity between users are
calculated. The category countiscalculatedtocheck whether
the user have images in every cluster or not.
3.6. BoFT Similarity Calculation
The images with extreme resemblance will have high BoFT
similarity. The similarity between the user shared images
helps to identify the connection between users. The number
of pairs and user pairs are calculated. The number of pairs
are calculated using,
Number of Pairs = (n ∗ (n − 1))/2; (1)
Owing to the homogeneous features of user shared images
included in user profiles, calculate similarity offeatures with
respect to the user category. Analyzing the user categoryisa
better alternative for similarity calculation. Based on the
number of occurrences of user category, similarity between
the images are calculated.
Sa,b = sum(count)/(length(a) + length(b)); (2)
Here Sa,b denotes similarity calculation, sum(count)
represents total number of common imagessharedbyusera
and user b, length(a) represents images sharedbyuseraand
length(b) represents images shared by user b.
Recommendations aremade basedonthesimilarity between
two users. Inorder to recommend users to other users, the
proposed technique accurately measures the similarity
between users. With respect to similarity calculation,
connection between theusersarediscovered withuserpairs.
User pair depicts that two users are related. The user pairs
are ascertained on the basis of visual appearance of images
and similarity between the users are calculated accordingly.
Thus related pairs are identified and user connections are
discovered.
3.7. User/Tag Recommendation
Tagging is a most important way for information
propagation. Tag recommendation is used to predict the
possible number of tags for a special content. Co-occurrence
of tag for user shared images are computed using similarity
between images. User shared images and frequent
occurrence of tags depicts personalized interest of user
based on a special topic. With respect to the similarity
calculation, interconnection between users are identified
and user/tag recommendations [2] are made. Interest
matching users are a sensible way to recommend users for
friendship through tagging. Unique users are searched out
and in addition, checking each user shared image against
another inorder to find the matching ones. Based on that,
recommendations are made.
3.8. Virality Prediction
The virality prediction is more challengingtask of predicting
whether an image is viral or not by looking at its content.
The images in social network are augmented by virtueofthe
shared content. Data spread more rapidly on the extensive
world system. Viral data spread through interpersonal
organizations. The estimation of virality can be done by
assessing the popularity. Individuals who shared, voted or
saw content are characterized as infected node on the
framework. Most of the algorithm concentrate on the time
taken by the content to become viral. The proposed
technique focus on predicting theimagetoendupnoticeably
popular in light of the number of occurrences of image.
Consequently the occurrence of viral image is predicted
based on a threshold value. Checking whether the images
keep running ahead than threshold value, then the images
are opted as viral. The basic figure in spreading of viral
image is social cascade. A biggercascadesizeevaluatehigher
virality [10] of the image.
3.9. Interest Discovery
The interest discovery [6] in onlinecommunicationplatform
gain much importance because it helps to connect users
based on their interest. Identifying the interest of user is a
challenging task. Usersina grouprequireslike-mindedusers
and collect information about their specific area of interest.
Tags on images helps to gather users with similar interest.
Users in social network assign tags to theimagethatthey are
interested in. In addition to user/tag recommendation,their
exhibit a list of users having common interest. Discover the
interest of each communities with specific interest. On the
basis of users interest, the interest discovery approachfinds
interest of each group of users who share similar images.
4. EXPERIMENTAL RESULTS
This chapter describes the datasetandresultoftheproposed
technique. The results show how the experiments are
conducted and connections are discovered. The proposed
framework is a four phase process, where the first stage is
for connection discovery of users from user shared images,
second stage indicates user/tagrecommendation[2]system.

The third stage is virality prediction [10] and fourth stage
shows interest discovery [6].
4.1. Dataset
Facebook is a social networking site which is utilized for
connection discovery. Facebook permits users to upload
images and videos, recordings, exchange messages, create
groups and share both internal and external information.
Social networks rely vigorously on user-generated content,
which spreads rapidly across social networks. Due to the
privacy concerns, the vast majority of the social networking
sites like Flicker, Friendster and Twitter does not provide
their dataset. So images are gathered from Facebook with
image id and user id. Missing information in social media is
caused by numerous reasons like no access authorization on
data, imperfect knowledge acquisition processes etc. The
dataset is created in Microsoft Excel based on the data
collected from Facebook.
The dataset comprised of nodes and image id where nodes
represents users, image id represent images shared by
nodes. The dataset of Facebook isresizedintoa standard size
250x250. 200 users shared images of different categories
such as aircraft, turtle, lamp, buildings, camera, flower,
mountain, ship, watchesare collectedfortheimplementation
of the proposed approach. 139 users shared 200 images
based on their interest. Some of the users shared more than
one category of images. Through this, we can also identify
the interest of users. Particular images areattracted bysome
of the users, so many individuals shared the same images
and thus the image becomes viral. The virality of the images
can be predicted from the Facebook dataset.
4.2. Connection Discovery
The first key step is feature extraction, SIFT [15] based
feature extraction is used. SIFT with maximum key point
extraction is opted. The extracted features are invariant to
scale and orientation. Keypoint descriptor is computed for
the local image region about each keypoint that is highly
distinctive and invariant points. After SIFT extraction, 128
descriptors together with scales are obtained.
The images should be in proper cluster, in order to get the
accurate result in the clustering process. K-means clustering
is used to cluster descriptor vectors of BoF with respect to
keypoints. Different values are assigned to the cluster
parameter, but the proper cluster is obtainedwhenthevalue
of the number of clusters is assigned with 8. The Fig 4 and
Fig 5 shows the precision and recall rate for k-means
clustering based on the categories of image. The precision
and recall rate are computed as the following,
Precision = Tp /( Tp + Fp) (3)
Recall = Tp /( Tp + Fn) (4)
Fig -4: Precision rate of clustering
Fig -5: Recall rate of clustering
where Tp is the true positive (correctly clustered image),
while Fp and Fn are false positive and false negative
respectively. Fp is the case that the clustered image that is
predicted as correct. Fn is not properly clustered image.The
precision rate, p, measures the percentage of properly
clustered images upto a limit while recall rate, r, measures
the percentage of properly clustered images from the whole
dataset. Inorder to evaluate the precision rate, different
dataset are used to check the accuracy of clustering. The
higher value of p and r gives better result of clustering. The
users and category of users in identified from users profile.
Table 1 shows the category count of 12 user and Table 2
shows the user category of 12 users. The user category of
user 3, user 11 and user 12 having image in only one cluster
and user 5 having images in more number of cluster. The
category count of user 3 have images in only one cluster and
most of the users have images in more than one cluster. User

shared images appeared in different cluster indicates the
interest of users in different image category.
Table -1: Category count of 12 users
Users Category count
1 1 0 0 2
2 2 0 0 2
3 1
4 1 1
5 0 2 3 1
6 0 2 1
7 0 2 0 1
8 0 1 0 1
9 2
10 1 0 1
11 0 0 1
12 0 0 1
The number of pairs and user pairs are computed inorder to
identify similarity between users. The user category is used
to measure the similarity between users.
Number of Pairs = (n ∗ (n − 1))/2 (5)
Sa,b = sum(count)/(length(a) + length(b)) (6)
User connection is calculated on the basis of similarity
calculation. Related pair exhibit user connections alongwith
its range of similarity. Table 3 demonstrates the discovered
connection between user pairs. The total number of
connections obtained is 9529. Out of 9529 user connections,
10 are displayed in the table. Some of the related pairs have
higher similarity in visual featuresandsomeappearswith an
average similarity range. The related pair with higher
similarity range is occurred between 7701-7702. The user
pairs with similarity range 0 is considered as non-related
pairs. The non-related pairs does not have connection
between users.
Table -2: User category of 12 users
Users User category
1 4 4 1
2 4 1 4 1
3 1
4 1 1
5 3 3 4 3 2 2
6 3 2 3
7 2 2 4
8 4 2
9 1 1
10 1 3
11 3
12 3
Table -3: Discovered connection between user pairs
SL No. User pairs Discovered
connection
1 7701-7702 1
2 7701-7703 0.5000
3 7701-7704 0.6000
4 7701-7705 0.3333
5 7701-7706 0
6 7701-7707 0.5000
7 7701-7708 0.6000
8 7701-7709 0.6000
9 7701-7710 0.4000
10 7701-7711 0
Fig -6: Example of connection discovery
4.3. User/Tag Recommendation
User/tagrecommendationdependsonconnectiondiscovery.
The similarity calculation gives a precise information about
interrelation between users. Comparative interest shared
users having high similarity. Based on their interest,
recommendations are made. Users are sorted in an order
with respect to similarity calculation.

List out the users having common category of images. The
unique users are then filtered out from the list and check for
the matching pairs. The user idofinterestmatchingusers are
stored during each iteration process. The Fig 7 and Fig 8
shows the precision and recall rate for a list of
recommendation based on the categories of image. The
precision and recall rate are computed using equation 3 and
4. The higher value of p and r gives better recommendation
based on the image category. The Fig 9 demonstrates the
recommendation of list of users based on the discovered
connection between users. Most of the users are
recommended to the images appeared in category 3. The
category 2 is considered as least recommended category.
Fig -7: Precision rate of recommendation list
Fig -8: Recall rate of recommendation list
Fig -9: List of recommended users based on discovered
connection
4.4. Virality Prediction
Spreadability and propagativity are the two reason for
content virality. Virality is a phenomenon strictly connected
to the character of the content being spread, insteadofusers
who spread it. The concept of spreadabilityspeaksthatmore
rapidly the content flow across the social media. Most of the
algorithm in predicting the content virality focusonthetime
taken by the social cascade to become viral. But the
proposed viralityprediction methodisbasedoncounting the
number of occurrences of the image to become viral. The
content virality is measured based on a threshold value. For
checking the virality of images, unique user shared images
are taken from the database and checked the number of
users who shared same image. Consequently, a threshold
value is set and checking whether the occurrence of images
exceeds the given threshold value. The threshold valueisset
as 20. If the number of occurrence of images isgobeyond20,
then the image is treated as a viral. The Fig 10 shows viral
images from user shared images. The result shows
occurrence of images are accurately calculated without any
prediction error. Large datasetfromFacebook canbeusedto
obtain a more comprehensive evaluation of the approach.

Fig -10: Viral images
4.5. Interest Discovery
In online social network, the recommendation list will
categorize the user shared imagesthatdepictstheinterestof
the user. The user shared images are more concise and
closely reflect the interest of user. The interest of users are
identified from recommendation system. Users with similar
interests are discovered through similarity calculation in
connection discovery. Based on the occurrence of different
category in image sharing we can identify the interest of
users.
In recommendation system, users with special area of
interest are identified and based on thatinterestofthegroup
of users are identified. The resultsshowthatbyanalyzingthe
recommendation list, users interest is captured and is
accurately measured. The interest discovery approach is
highly focused on list of recommended users based on
discovered connection. It is proved by experiments that the
proposed approach is very effective to discover common
interest of users in online social networks with the
information on the online connections among users. The
table 4 shows the interest of users based on 8 category.
Based on user recommendation, interest of users are
identified.
Table -4: Interest discovered for a group of users
SL
No.
User
id
User
recommendation
Cluster
id
User
interest
1 7701 7702 7703 7704
7708 7715 7717 ...
5 Flower
2 7702 7703 7704 7708
7715 7717 7723 ...
1 Bus
3 7704 7706 7710 7712
7713 7716 7718 ...
6 Flight
4 7705 7707 7708 7709
7711 7712 7718 ...
2 Ship
5 7708 7713 7718 7720
7723 7725 7734 ...
3 Watch
6 7711 7719 7720 7722
7723 7784 7785 ...
8 Mountain
7 7714 7715 7716 7717
7718 7719 7720 ...
7 Turtle
8 7721 7722 7729 7733
7757 7761 7763 ...
4 Camera
5. CONCLUSION
The proposed connection discovery method and the system
for user/tag recommendation, prediction of virality and
interest discovery approach are based on user shared
images in Facebook. The proposed method is a novel BoF
based Tagging approach to make better recommendations.
The matching users interests are discovered from images
uploaded. BoF based approach can help discover hidden
users connections from the shared images on a social
network for a better recommendation. The proposed
technique helps to classify image according to their visual
features. Those visual features represent the user interest
and hence recommendation. As a result of discovered
connection, the proposed approach can recommend users
based on the interest of user.
Visual similarity of users shared images helps to identify
connection between users and friendship recommendations
can be made based on the similarity. Also, identified the
interest of users and content virality. Interest identification
is a system to discover common interest topics which helps
to connect more people with common interests and
encourage people to contributeandshare morecontentsand
based on that friendship recommendationscanbe made.The
virality
of the images are predicted by counting the number of
occurrence of image.
REFERENCES
[1] V. Leroy, B. B. Cambazoglu, and F. Bonchi,“Cold start link
prediction,” in Proc. 16th ACM SIGKDD Int. Conf. Knowl.
Discovery Data Mining, pp. 393402, 2010.
[2] A. Rae, B. Sigurbjrnsson, and R. van Zwol,“Improving tag
recommendation using social networks,” in Proc. Int.
Conf. Adaptivity, pp. 9299, 2010.
[3] I. Guy, N. Zwerdling, I. Ronen, D. Carmel, and E. Uziel,“
Social media recommendationbasedonpeopleandtags,”
in Proc. 33rd Int. ACM SIGIR Conf. Res. Develop. Inf.
Retrieval, pp. 194201, 2010.
[4] M. Cheung and J. She,“Bag-of-features tagging approach
for a better recommendation with social big data,” in

Proc. 5th Int. Conf. Adv. Inf. Mining Manag., pp. 8388,
2014.
[5] E. Gilbert and K. Karahalios,“Predicting tie strength with
social media,” in Proc. SIGCHI Conf. Human Factors
Comput. Syst., pp. 211220, 2009.
[6] X. Li, L. Guo, and Y. E. Zhao,“Tag-based social interest
discovery,” in Proc. 17th Int. Conf. World Wide Web, pp.
675684, 2008.
[7] Ming Cheung, James Shey, Xiaopeng Li,“Non-user
Generated Annotation on User Shared Images for
Connection Discovery,” Conf., DOI:
10.1109/DSDIS.2015.113, Dec. 2015.
[8] F. Jurie and B. Triggs,“Creating efficient codebooks for
visual recognition,”in Proc. 10th IEEE Int. Conf. Comput.
Vis., vol. 1, pp. 604610, Oct. 2005.
[9] Z. Jie, M. Cheung, and J. She,“A cloud-assisted framework
for bag-offeatures tagging in social networks,” in Proc.
IEEE 4th Symp. Netw. Cloud Comput. Appl., to be
published in 2015.
[10] Ming Cheung, James She, Lei Cao,“Predicting Content
Virality in Social Cascade,” Article number 6682181, p.
970-975, 20-23 Aug. 2013.
[11] Indika Kahanda andJennifer Neville,“Using Transactional
Information to Predict Link Strength in Online Social
Networks,” in Proc. 3rd
International AAAI Conf. on Weblogs and Social media, pp.
7481, 2009.
[12] J. Leskovec, D. Huttenlocher, and J. Kleinberg,“Predicting
positive and negative links in online social networks,” in
Proc. 19th Int. Conf. World Wide Web, pp. 641650, 2010.
[13] B. Sigurbjrnsson and R. Van Zwol,“Flickr tag
recommendationbasedoncollectiveknowledge,”inProc.
the 17th Int. Conf.World WideWeb, pp. 327336, 2008.
[14] H. Kim, J. Jung, and A. El Saddik,“Associative face co-
occurrence networks for recommending friendsinsocial
networks,” in Proc. 2nd ACM SIGMM Workshop Social
Media, pp. 2732, 2010.
[15] D. G. Lowe,“Distinctive image features from scale-
invariant keypoints,” Int. J. Comput. Vis., vol. 60, pp.
91110, 2004.
[16] A. Shepitsen, J. Gemmell, B. Mobasher, and R.
Burke,“Personalized recommendation in social tagging
systems using hierarchical clustering,”inProc.ACMConf.
Recommender Syst., pp. 259266, 2008.
[17] Xiaopeng Li, Ming Cheungy, James Shez,“Connection
Discovery using Shared Images by Gaussian Relational
Topic Model,” Hong Kong University of Science and
Technology, arXiv:1612.03639v1 [cs.SI], Dec 2016.
[18] I. Konstas, V. Stathopoulos, and J. M. Jose,“On social
networks and collaborative recommendation,” in Proc.
32nd Int. ACM SIGIR Conf. Res. Develop. Inf.Retrieval,pp.
195202, 2009.
[19] X. Zhang, Z. Li, and W. Chao,“ Tagging image by merging
multiple features in a integrated manner,” J. Intell. Inf.
Syst., vol. 39, pp. 87107, 2012.
[20] Sara Mazhari, Seyed Mostafa Fakhrahmad, Hoda
Sadeghbeygi,“ A user- profile-based friendship
recommendation solution in social networks,” Journal of
Information Science, pp. 112, 2015.
[21] Jing Liu, Zechao Li, Jinhui Tang, Yu Jiang, and Hanqing
Lu,“Personalized Geo-Specific Tag Recommendation for
Photos on Social Websites,” VOL. 16, NO. 3, APRIL 2014.
[22] I. Guy et al.,“Personalized recommendation of social
software items based on social relations,” in Proc. 3rd
ACM Conf. Recommender Syst., pp. 5360, 2009.
[23] X. Zhang et al.,“Social image tagging using graph-based
reinforcement on multi-type interrelated objects,”Signal
Process., vol. 93, no.8, pp. 21782189, 2013.
[24] A. Shepitsen, J. Gemmell, B. Mobasher, and R.
Burke,“Personalized recommendation in social tagging
systems using hierarchical clustering,”inProc.2008ACM
Conf. Recommender Syst., pp.259266, 2008.
[25] X. Xie,“Potential friend recommendation in online social
network,” in Proc. IEEE Int. Conf. Cyber, Physical Social
Comput.,pp.831835, Dec. 2010.
[26] F. Tian, X. Shen, F. Shang, and K. Zhou,“Automatic image
tagging by multiple feature tag relevance learning,” in
Pattern Recognition Anonymous Springer, pp. 505-513,
2012.
[27] Ming Cheung, James She, and Zhanming Jie,“Connection
Discovery Using Big Data of User-Shared ImagesinSocial
Media,” IEEE Transactions on Multimedia, vol. 17, no. 9,
Sep. 2015.

Visual Relation Identification Using BoFT Labels in Social Media Feeds

More Related Content

What's hot (20)

Similar to Visual Relation Identification Using BoFT Labels in Social Media Feeds (20)

More from IRJET Journal (20)

Recently uploaded (20)

Visual Relation Identification Using BoFT Labels in Social Media Feeds