KDD_FULL_PRESENTATION.pdf

Paper: https://arxiv.org/abs/2006.07812
Full Research Presentation: https://www.youtube.com/watch?v=MWoNQvclpoo

Transcript

1. Deep Exogenous and Endogenous Influence Combination for Social Chatter Intensity Prediction Subhabrata Dutta 1, Sarah Masud 2, Soumen Chakrabarti 3, Tanmoy Chakraborty 2 1 Jadavpur University, India; 2 IIIT-Delhi, India; 3 IIT Bombay, India ACM SIGKDD 2020 August 23-27 Virtual Conference
2. Introduction
3. • Chatter intensity Research Motivation - how much people will talk, given a content
4. • Chatter intensity Research Motivation - how much people will talk, given a content - need to detect as early as possible - may help containing spread of malicious contents
5. • Chatter intensity • Exogenous influence Research Motivation Season arrivals Discovery Detection Fig.1 Submissions and comments containing two different keywords on Reddit
6. • Chatter intensity • External influence • Endogenous influence Research Motivation Fig.1 Submissions and comments containing ‘Black Mirror’ on different subreddits
7. Research Motivation • Chatter intensity • External/Internal influence • Existing prediction frameworks
8. Research Motivation • Chatter intensity • External influence • Existing prediction frameworks Data Available Publicly • Discussion threads • Subreddit mapping • Who follows whom
9. Unified Chatter Model • News, submissions and comments come in bulk and continuously, so Online prediction • Zero-shot or few-shot prediction • No social network is visible, rely on exogenous and endogenous signals
10. Unified Chatter Model • News, submissions and comments come in bulk and continuously, so Online prediction • Zero-shot or few-shot prediction • No social network is visible, rely on exogenous and endogenous signals
11. Unified Chatter Model • News, submissions and comments come in bulk and continuously, so Online prediction • Zero-shot or few-shot prediction • No social network is visible, rely on exogenous and endogenous signals
12. Unified Chatter Model • News, submissions and comments come in bulk and continuously, so Online prediction • Zero-shot or few-shot prediction • No social network is visible, rely on exogenous and endogenous signals All in one! We propose ChatterNet!
13. Proposed Framework
14. ChatterNet
15. ChatterNet
16. ChatterNet
17. ChatterNet
18. Experiments & Evaluations
19. Dataset OCT 1, 2019 OCT 31, 2019 NOV 30, 2019 Submission: 751,866 Comments: 2,604,839 Submission: 1,334,341 Comments: 4,264,177 Articles: 1,851,022 Sources: 4,757 Articles: 2,010,985 Sources: 5054 43 subreddits Training data Testing data
20. ChatterNet: Main results + and ++ refers to zero-shot and 1hr early observations, respectively
21. ChatterNet: Ablation variants Kendall-τ for ablation variants of ChatterNet with zero-shot and observation
22. ChatterNet: Subreddit-wise Mean Absolute Percentage error with ablation on different subreddits
23. ChatterNet: Absolute error vs. size Prediction error increases for larger discussions
24. Details of our framework at: github.com/LCS2-IIITD/ChatterNet For more interesting research follow us at: @lcs2iiitd ACM SIGKDD 2020 August 23-27 Virtual Conference