Prediction of Community Behavior in News Social Media using Deep Learning
1. Introduction
Related Work
Conceptual Approach
Timeplan
References
Prediction of Community Behavior in News Social Media
using Deep Learning
Svetlana Pavlitskaya
308415
12.05.2016
Supervisors: Prof. Dr. Matthias Jarke
Chair of Information Systems
RWTH Aachen University
Prof. Gerhard Lakemeyer, Ph.D.
Knowledge-Based Systems Group
RWTH Aachen University
Advisor: Zinayida Petrushyna, Ph.D.
Chair of Information Systems and Database Technologies
RWTH Aachen University
Licensed under CC BY-SA 2.5 1
2. Introduction
Related Work
Conceptual Approach
Timeplan
References
Introduction
Problem: conflicts during discussions of controversial political issues
Solution: track news discussions to foresee and avoid aggressive behavior and
attacks through adequate moderation
Goal: learn deep neural network to predict reaction on news
1 user communities in news discussions in social media
2 community transformation over time
3 discussed topics
4 attitude towards topics within community
Licensed under CC BY-SA 2.5 2
3. Introduction
Related Work
Conceptual Approach
Timeplan
References
Community Detection and Evolution
Deep Learning
Community Detection and Evolution
Community detection using propinquity dynamics [ZWWZ09]
compute propinquity for each possible edge
iteratively update current graph topology to make it more consistent with propinquity
perform depth-first search over the resulting graph to find communities
Community evolution: an event-based framework[APU09]
network states at consecutive timestamps (snapshots)
critical events for communities and for nodes
Licensed under CC BY-SA 2.5 3
6. Introduction
Related Work
Conceptual Approach
Timeplan
References
Architecture
Community Behavior Prediction Model
Experimental Data Collection
all news in 2015 with keyword ”refugee(s)” from Facebook pages of CNN,
BBCWorldNews and Euronews
average number of comments for one news post: > 1K
=> expected number of comments for all news from CNN, BBCWorldNews and
Euronews in 2015: > 23M!
strategy: limit the number of news posts using keywords, retrieve only limited
number of comments per post
Licensed under CC BY-SA 2.5 6
9. Introduction
Related Work
Conceptual Approach
Timeplan
References
Architecture
Community Behavior Prediction Model
Dataset Extension using AlchemyAPI
for each comment: sentiment score between -1 and +1
for each news: relevancy-ranked list of concepts
for each community and news: common sentiment or distribution of sentiments
Licensed under CC BY-SA 2.5 9
10. Introduction
Related Work
Conceptual Approach
Timeplan
References
Architecture
Community Behavior Prediction Model
Recurrent Neural Network
given a sequence of feature vectors describing community behavior during several
snapshots, predict sentiment score at the next snapshot
cross-validation (variant for time series using sliding window)
Licensed under CC BY-SA 2.5 10
12. Introduction
Related Work
Conceptual Approach
Timeplan
References
References
Sitaram Asur, Srinivasan Parthasarathy, and Duygu Ucar.
An event-based framework for characterizing the evolutionary behavior of
interaction graphs.
ACM Transactions on Knowledge Discovery from Data (TKDD), 3(4):16, 2009.
Yuzhou Zhang, Jianyong Wang, Yi Wang, and Lizhu Zhou.
Parallel community detection on large networks with propinquity dynamics.
In Proceedings of the 15th ACM SIGKDD international conference on Knowledge
discovery and data mining, pages 997–1006. ACM, 2009.
Licensed under CC BY-SA 2.5 12