On the Influence Propagation of Web Videos

1. ON THE INFLUENCE
PROPAGATION
OF WEB VIDEOS
ABIDHA.V.P
ROLL NO:01
MCSSE 108
08/12/2014

2. OBJECTIVES
INTRODUCTION
RELATED WORK
EXISTING SYSTEM
PROPOSED SYSTEM
TECHNICAL CONTRIBUTIONS
ADVANTAGES
APPLICATIONS
CONCLUSION
REFERENCES

3. INTRODUCTION
Web videos nowadays influence society like never
before in history.
It is utterly important to identify how the
propagation took place,
i.e., to determine if a popular video on a video
sharing website actually originated from that
website, or it is merely a projection of influence
from somewhere else of the cyberspace

4. Cont…

5. The problem to solve is extremely important to the video
sharing site owner
 video-sharing sites have developed schemes
to encourage users to upload content.
 Advertisers often exercise multiple partnership models
for online videos.
 Cyber cell sometimes need to know the origin of some video

6. RELATED WORK
 Fist we survey the recent studies in event discovery and other mining
tasks in social networks
 Twitter - sensor network for event detection ,with semantic analysis
of tweets establishes a probabilistic model which derives the
probability of an emerging event from the occurrence reading of
related tweets.
 Bayesian filters, it determines the location of the event. It is reported
effective for earthquakes, typhoons etc
 A statistical model PET is used to track events in social networks.

7. Cont…
 A model called Social Pixel is proposed for aggregating social
interests and detecting emerging events.
 A framework Social Transfer is to analyze trending topics on the
Internet for socialized video recommendation uses similar cross-domain
analysis to predict the popularity .
 Later leverages both Page Rank and video near-duplicate detection
techniques for the same task.
 A novel technique by jointly considering local regression
and global aligned when learning for a metric that could handle cross
media retrieval with multi-modalities.

8. EXISTING SYSTEM
Understanding Video Interactions using rank
 weighted Laplacian matrix (affinity matrix)
Anti-social behavior such as self-promotion and
other types of content pollution identification
The model is closely related to manifold learning
and local learning

9. PROPOSED SYSTEM
The proposed method inspired by previous but there
are two key differences.
Previous , which can not be applied to out-of-sample
data.
proposed method is not only inductive, but also very
efficient .
it is possible to use it to handle large-scale data.
our method simultaneously reduces the noise in the
data by dimension reduction , which is not considered
in previous.

10. TECHNICAL CONTRIBUTIONS
 To model an online video’s propagation and influence
in the cross-community cyberspace, A Unified
Virtual Community Space (UVCS)that captures
the propagation history of an online video.
 We propose an advanced learning method called
Noise-reductive Local-and-Global Learning (NLGL)

11. UNIFIED VIRTUAL COMMUNITY SPACE (UVCS)
which contains the web pages relevant to the video.
 Each page has a time stamp, indicating the publishing time
or modification time of the page
 Each page receives a set of inbound links , the link
relations among pages inside the UVCS
 Each page’s rank in the UVCS is known.
 Complementary information collected from other sources.

12. UVCS FEATURE REPRESENTATION
x = [ nt, tr, r, ir, ni, ipr, nip, L1 ◦W1,L2 ◦W2,NT,NB,NN, wi ].
 All the related pages are published with nt and tr.
 [r, ir, ni, ipr, nip,NT,NB,NN, wi] The second part contains
the search engine’s rank of the video page as well as
the related references in various communities of the UVCS.
[L1 ◦W1, L2 ◦W2] The rest of the feature vector helps
to evaluate both the propagation and the influence of the
online video.

13. UVCS FEATURE REPRESENTATION TABLE

14. NLGL LEARNING OBJECTIVES:
Noise-reductive Local-and-Global Learning
– THE METHOD SHOULD BE ABLE TO REDUCE NOISE.
 The UVCS feature is a combination of multiple semantic
components, the significance of each component is not specified
in the raw feature .Fields of the UVCS feature may be missing for
some feature vectors due to the diversified nature of web pages.
 The method should be able to learn a classification/ranking model
with a relatively limited number of learning data, and then apply
the obtained model on out-of-learning-set data. Consequently, it
should be able to preserve the structure of the learning set.

15. THE NLGL ALGORITHM
Utilizing The Un-annotated Sample Data
 Obtain a training dataset with n total videos and n‘ manually
annotated videos (n’< n), as X = [x1, . . . , xn] ^T ∈ R ^n×d,
// where xi is the d-dimensional UVCS feature vector for the
i th video in the dataset.
 X^l = [x1, . . . , xn ]^T. We use Y^l = [y1, . . . , yn ]^T ∈ R^n×c to
denote the class lables or ranks for the manually annotated videos X^l,
// where c is an integer greater or equal to one.
 Note that Y^l is a unified presentation of the class labels for the
propagation and the ranks for the influence ranking.
// c is the number of classes, and the label yi for the
ith labeled video in X^l

16. Cont…
• To predict the label of new videos, a straightforward method is to
minimize the empirical error for the following regression model:
………….(1)
• where Wo R^d×c is the classifier ∈ which is learnt from manually
annotated data and is able to predict the labels for the un-annotated
data.
• To leverage the unlabeled data for a better performance, it is
desirable to design an algorithm which is able to predict the labels
of the unlabeled videos by exploiting the data distribution of X.

17. CLASSES OF VIDEO PROPAGATION PATTERNS

18. Preserving Local Structure for the Annotated Data
o To predict the ranking score of each datum and then all the local
linear regression models are optimized globally.
o Predicted label of all the training data , sum up the loss functions
for each xi and minimize the total loss.
o Denoting the column vector with all ones, we derive the objective as:
equation is written as:
………….(2)
Despite the high accuracy
It is not able to predict the labels of the
videos outside the training set.

19. Dimension Reduction and the Complete Model
o If linear transformation, which transforms the video data X in feature
space to a more compact and accurate representation Z ∈ R^n×d
// where d’ ≤ d is the reduced dimensionality.
o The transformation between X and Z can be formulated as:
Z = XQ
………….(3)
// where Q ∈ R^d’×d is the transformation matrix.
o If Q is properly trained, the classification based on Z would result in
better performance
o we integrate training data label prediction, dimension reduction and
classifier inference into a joint framework and propose to minimize
the final objective function of NLGL:

20. where λ1 and λ2 are parameters.
………….(4)
o The first term aims at learning an optimal dimension reduction
transformation and a classifier simultaneously.
o The second and third terms predict the labels of the unlabeled
training data.
o Besides, ||XQW − F||F ^2 can be regarded as a regularizer of the
local learning algorithm, which considers the global information
to the predicted labels F.

21. Iterative Optimization
1. Fix F and optimize Q and W. By setting the derivatives of
Equation (4) with respect to W to be 0,
………….(5)
Substituting W in Equation (4) by Equation (5), we arrive at:
………….(6)
2. Q and W are fix to optimize F. By setting the derivatives
in Equation (14)
The two steps are iteratively performed until convergence.
………….(7)

23. ADVANTAGES
 With this model we are able to predict the propagation
pattern and influence ranking for any new online video.
 Extensive experiments are conducted on popular online
videos to verify the effectiveness of the proposed
framework.
 Some observations on propagation and owners’
behaviors are also provided.

24. APPLICATIONS
 In future research, there is great potential for us to
extend this research and enable it with other interesting
capabilities.
 The most likely problem would be the identification
of the actual origin of the video. Also investigating
possibilities of establishing inter-sharing-site influence
model to analyze how video sharing site influence each
other.

25. CONCLUSION
 Determine how it got popular through it’s analyzing life cycle. Also
give a rough estimation of its general influence on the Internet.
 First we have made pioneer effort to model a video’s propagation and
influence in cyberspace by a unified presentation, namely UVCS.
 NLGL exploits the benefits of local learning, manifold structure and
dimension reduction .
 Successfully model the propagation classification and influence
estimation tasks into a classification problem and a ranking problem
under the same learning framework.

26. REFERENCES
1. J. Liu et al., “Presenting diverse location views with real-time near-duplicate
photo elimination,” in Proc. IEEE 29th ICDE, Brisbane, QLD,
Australia, 2013, pp. 505–516.
2. S. D. Roy, T. Mei, W . Zeng , and S. Li, “Towards cross domain
learning for social video popularity prediction,” IEEE Trans.
Multimedia, vol. 15, no. 6, pp. 1255–1267, Oct. 2013.
3.S. D. Roy, T. Mei, W . Zeng , and S. Li, “ Social transfer : Cross
domain transfer learning from social streams for media applications,” in
Proc. ACM Multimedia, New York, NY, USA, 2012, pp. 649–658.
4. Z. Wang et al. “Propagation-based social-aware replication for
social video contents,” in Proc. ACM Multimedia, New York, NY, USA,
2012, pp. 29–38.

27. QUESTIONS