Discovering Relation Based on User-Shared Image in Social Media by Using Big Data is the title for Term Paper Assignment in Artificial Intelligence class, Yuan Ze University, Taiwan
Discovering Relation Based on User-Shared Image in Social Media by Using Big Data
1. Discovering Relation Based on
User-Shared Image in Social
Media by Using Big Data
Radhiyatammardhiyyah (1056037)
CSE - YZU
2. Abstract
◆ Billions of user-shared images are generated by individuals in
many social networks. User social graphs are only accessible to
exclusive parties, these user-shared images are proved to be an
easier and effective alternative to discover user connections.
◆ Data : 360000 user shared images from two social networks,
Skyrock and 163 Weibo (3 million follower/followee
relationships)
◆ A multimedia big data system is proposed as an alternative to
user-generated tags and social graphs for follower/followee
4. ◆Social graphs (SGs), representing online friendships among
users, are fundamental data for many applications, such as
recommendation, virality prediction and marketing in social
media.
◆Problem : some users limit the information due to privacy
concerns and accessing the SGs is getting more difficult and
costly.
◆Solution : Use annotated tag (user tagging) with each shared
image to discover user connections.
5. The contributions of this paper includes:
1. Intensive measurements and characterizations of user shared images from
two new dataset
2. Methods using bag-of-features tagging (BoFT) are proposed as a
recommendation system to discover user connections and recommend
follower/followee relationships by their shared images;
3. extensive verifications of the proposed formulation, methods and system
are provided with the data sets from two social networks and two
practical use cases to prove the effectiveness of using user shared images
through BoFT.
7. A. BoF-Based Tagging
(a) annotation with BoFT labels,
(b) user resemblance calculation
based on BoFT labels
8. 1. Feature Extraction: is a
process to obtain the
unique local features in
step 1 of Fig. 3(a).
2. Codebook Generation:
the step 2 of fig.3(a) is a
clustering process to
obtain a set of visual
words, a representative
and distinct set of
unique visual features.
4. Clustering and BoFT Labeling:
(step 4 of Fig. 3(a))
when two images contain cars in
the countryside, the feature
vectors of the two images are
similar in terms of the number of
occurrences of each unique
visual word.
→ the two images will be
assigned the same BoFT label to
indicate that they are visually
similar.
3. Feature Coding and
Pooling: (tep 3 of Fig. 3(a))
represented by the closest
visual word. Each image is
represented by a feature
vector in the feature
pooling.
9. B. Connection Discovery with BoFT labels
1.BoFT Labels and User Profile : key
connection discovery
2.User Profile and User BoFT Similarity :
user who share highly similar images will
have a high BoFT similarity.
10. 3. MEASUREMENT ON USER-SHARED IMAGES
Characteristic
s of User-
Shared
images
Dataset
BoFT
Similarity
11. A.Dataset
176,547 images uploaded by 722 users on Skyrock, and 187,491 images
uploaded by 493 users on 163 Weibo are experimented in this work.
12. B. Characteristics of User-Shared images
The characteristics of user shared images are investigated and modeled as
exponential distributions based on the analysis of 3 million
follower/followee relationships from two social networks; Skyrock and 163
Weibo, for which similar observations are found.
13. C. BoFT Similarity
two types of user pairs:
1. related pairs, which are the pairs of users that are follower/followee,
2. Non-related pairs, which are the pairs in which a follower/followee relationship
does not exist between the two users.
Related pairs and non-related pairs can be considered as two classes, and the class of
each pair, C, can be defined as
where C=1 is the class in which the pair is a related pair, and C=0 is
15. A. Image Collection → Skyrock and 163 Weibo
B. Connection Discovery Using BoFT → a computer vision approach to
give a label to user generated images, which is not affected by the
language, culture or other characteristics of the user. It only affected
depend on the image’s visual appearance.
C. Follower/Followee Recommendation → The probability that two user
are a related pair, or C = 1, given the BoFT similarity of user i and j,
P(C=1|Si,j) can be calculated by (1) based on Li and Lj. Also, by
applying Bayes’ theorem, follower/followee should be created
depend on P(C=1|Si,j) which is from the highest to the lowest.
17. A.SETUP
3 connection discovery methods are implemented for comparison
1. FoF. It is an achievable upper bound when difficult and limited access
SGs are available. The recommendation is from the similarity of the SGs.
2. using user annotated tags (UserT), and the recommendation is based
on the connection discovered based on the similarity among the user
annotated tags on shared images between two users.
3. a random method (Rand), in which follower/followee relationships are
recommended randomly.
18. A.RESULTS
Fig. 13. Performance of follower/followee recommendations with connection discovery,
in precision p(upper part) and recall r (lower part): (a) and (b) are the p from 5 to 10, for
Skyrock and 163 Weibo, respectively, (c) and (d) shows the r from 5 to 10 for skyrock
and 163 weibo, respectively.
19. 6. CONCLUSION
1. Follower/followee recommendation using discovered
connections by user shared images is possible, and the
recommendation is 60% better than UserT and achieves 25% of
the performance of FoF, a method used when limited access
SGs are available.
2. Create a potential long term impact and contribution to
scientific research and commercial application
20. References
[1] M. Cheung, J. She, and Z. Jie, “Connection discovery using big data of
user-shared images in social media,” Multimedia, IEEE Transactions on, vol.
17, no. 9, pp. 1417–1428, 2015.
[2] Suganya. M , U. priyadharshini , Akil, D. Kanagalakshmi, R. Raja prabha
“Associating User Shared Images on Social Media” International Journal of
Advanced Research Trends in Engineering and Technology (IJARTET), Vol. 3,
Special Issue 16, March 2016
[3] Li, Xiaopeng, Cheung, Ming, She, James (2016), “Connection Discovery
using Shared Images by Gaussian Relational Topic” Model IEEE International
Conference on Big Data 2016