[MLNS | NetSci] A Generative/ Discriminative Approach to De-construct Cascading Events
2 likes365 views
The document presents a generative/discriminative approach for analyzing cascading events in social networks, focusing on how information spreads through interactions on platforms like Twitter and Reddit. It discusses the probabilistic generation of cascades, inferring user participation and timing, and combines these with genetic algorithms to improve predictions. The methodology aims to capture dependencies between user behavior, timing, and cascade structure for enhanced predictive accuracy in understanding social influence and information diffusion.
[MLNS | NetSci] A Generative/ Discriminative Approach to De-construct Cascading Events
- 2. Diffusion in Social Networks • News, opinions, rumors, fads, urban legends, … • Virus, disease propagation • Change in social priorities: smoking, recycling • Saturation news coverage: topic diffusion among bloggers • Internet-energized political campaigns • …. “Viral Marketing” -- Patterns of Influence in a Recommendation Network 2
- 3. Information Cascade • A diffusion process initiates activations through an underlying network • Activations leave a trace – cascade • E.g., re-tweet chain, online conversation 3
- 5. Cascading Events • who did what to whom when where • E.g., @Sameera mentions @Anda in the Retweet “RT: TheHackerNews: Biggest…” at 2019:03:27 08:23 5
- 6. Collective Behavior in Cascading Events • “Information do no spread in isolation, independent of all other information currently diffusion in the network” Myers et al. • Competing or Cooperative cascades • E.g., Reddit conversations on Bitcoin scaling debate Introducing Bitcoin Cash(BCH) that splits Bitcoin’s (BTC) original blockchain via a hard fork — a consequence of popular Bitcoin scaling debate “We need off-chain scaling and on-chain scaling.Our stupid politics is hurting Bitcoin because we're separating the two necessary parts of the overall solution.” Discussion on the advantages of Bitcoin Cash (BCH) on scalability — “Bitcoin Cash (BCH) totally fixes the quadratic scaling of sighash operations bug” ..“I miss those days man and got tired of these scaling topics.”.. “Don't fall for the big block argument while we are so close to have the scaling solution.” “..but he made a mistake in the scaling debate” 6
- 7. Probabilistic Cascade Generation • Branching Model • E.g., person transmits the disease to each people she meets independently with a probability p, an infected person meets k (new) people while she is contagious • Used to accurately model cascade subtree structure, Cheng et al. 7
- 8. Probabilistic Generation of a Cascade Pool • Cascade pool consists of a set of cascades • Each cascade originated from a given set of initial seeds (e.g., tweet, or a post in Reddit), • Apply branching model for each seed separately. • Based on empirically- benchmarked distributions; • E.g., conditional degree distribution by the level • Branching model generates only the cascade structure. 8
- 9. Cascading Events: Inferring Users • How do we map users to the generated cascade trees? • Overlay the cascade structure on top of an underlying diffusion network • Given a user in an underlying network, we take a random weighted sample of user’s incoming edges • Underlying Networks: • Reddit, shared subreddit network weighted by direct interactions • Twitter, follower network where edges are weighted by the number of previous direct interactions ua ub uc ud ue uf 0.45 0.35 0.2 1 0.5 0.5 1 post comment 1 comment 2 comment 3
- 10. Cascading Events: Inferring Time • How do we estimate the time of adoption (timestamp of the comment or retweet)? • Given a size of the cascade, we randomly sample a sequence of long propagation delays in a previous cascade of a similar size 10 post comment 1 comment 2 comment 3 post comment 1 comment 2 comment 3 2017/08/01 13:00 + 00:35 + 00:40 + 00:55 2018/02/01 23:00 + 00:35 + 00:40 + 00:55 Previously seen cascade Simulated cascade
- 12. Generative/ Discriminative Approach 12 How to capture the dependence between user, timing and structure in a cascade? Task is to accurately predict a pool of cascades with user and timing information • Generate N pools of cascades using probabilistic models • Reconstruct a pool of cascades using genetic algorithm • We develop a fitness function based on the output of two trained LSTM models • The models assess how realistic to observe the generated cascade in two spatio-temporal properties (i.e., branching and propagation delay)?
- 13. Discriminative Models 13 Two Prediction Objectives: • Predict the branch and leaf nodes • critical factor to accurately predict the cascade structure • Predict the early and late adopters • critical factor to identify the temporal position of a node in the cascade • E.g., assume node 1 appeared in timestep x, and node 2 and 3 appeared in timestep x+1 and x+2 respectively: • node 2 is a branch, and an earlier adopter (propagation delay = 1) than node 3 (propagation delay = 2)
- 15. Re-construction of a Cascade Pool • How realistic is the generated cascade with the user and timing information? • Terminology: • “Gene” – individual cascade • “Individual” – a pool of cascades • “Population” – a set of cascade pools
- 16. Fitness Score 16 How realistic is the generated cascade with the attached user and timing information? seed Nx Ny Nz 1 0 0 1 1 0 Generated Cascade Branch LSTM Delay LSTM 0 0 1 0 1 1 T10 T1 T2 Ny T1 Nz T2 Nx T10 Branch vector Delay vector
- 17. Fitness Score 17 Genetic Test • The objective is to come up with an individual (a pool of cascades) which outperforms any existing pool in the initial search space. • We feed each generated cascade into two machine- learning models: • Predict a binary vector that represents branch/leaf using the branch discriminator model, • Predict a binary vector that represents the early/late adopters using the delay discriminator model. • We consider these two binary vectors as the inferred ground truth to assess the generated cascade. • Report AUC comparing the inferred ground truth with the simulated cascade output.
- 18. Performance (prediction) • Two classification tasks • Predict branch or leaf node • Predict early or late adopter 18 Training Data Test Data time 2015 January 2016 December 2017 August
- 19. Performance (simulation) • Simulation Task • Given a set of initial seeds (e.g., tweet, or a post in Reddit), predict the full cascade structure with user and timing information • E.g., who did what to whom when where 19 Training Data Test Data 2015 January 2017 July 2017 August Simulation