<h3>Identifying Coordinated Accounts on Social Media through Hidden Influence and Group Behaviours</h3> <!-- .slide: data-transition="zoom" data-background="black"--> <img src="https://i.imgur.com/AeQovol.png" width="300"> > Karishma Sharma, Yizhou Zhang, Emilio Ferrara, Yan Liu <h5> presented by <a href="https://twitter.com/alorozco53">Albert M Orozco Camacho</a> </h5> --- <!-- .slide: data-transition="zoom" data-background="red"--> ## Preliminaries ---- <!-- .slide: data-transition="fade" data-background="white"-->  ---- <p class="fragment fade-up">Disinformation is promoted using coordinated groups of accounts</p> <p class="fragment fade-up">They all boost visibility of fake content</p> <p class="fragment fade-up">Since 2016, there is strong evidence of influence of such coordinated campaigns to influence Elections</p> ----  ---- Proposal: <p class="fragment fade-up"><b>AMDN-HAGE</b> (Attentive Mixture Density Network with Hidden Account Group Estimation)</p> ---- <!-- .slide: data-transition="fade" data-background="white"--> - _Unsupervised generative_ model - jointly models account activities and hidden account groups - Based on Neural Temporal Point Processes (**NTPP**) and Gaussian Mixture Models (**GMM**) - model the _distribution of future activities_ conditioned on past activities of all accounts with temporal differences - A bilevel optimization algorithm is proposed - uses both _stochastic gradient descent_ and _expectation-maximization_ --- <!-- .slide: data-transition="zoom" data-background="yellow"--> ## Coordination Detection Method ---- ### Selected Approach ("Assumptions") - To model **strong hidden influence** (or latent) between coordinated accounts' activities - To captiure **highly concerted activites** as such behaviour should be anoumalous w.r.t. other normal accounts with less organized activity patterns ---- <!-- .slide: data-transition="zoom" data-background="white"-->  ---- ### Task Definition - Assume only _observed_ activity traces are provided  > _sequence of posts from account $u$ at time $t$ > can be thought as an information cascade from the network ---- ### Temporal Point Processes - TPP is a stochastic process that represents a sequence of discrete events in continuous time - It can model the conditional density of future events on past using  ---- - $\lambda$ is the instantaneous rate of events or expected change at any time (based on Hawkes processes)  ---- ### But for this paper... - the authors propose encoding the history of activities using a neural network - i.e., a _context vector representation_ - in the _decoding_ stpe, the model will predict the distribution over the next event time; - **it is also possible to predict the type of event (account type)** <!-- .slide: data-transition="zoom" data-background="red"--> ---- <!-- .slide: data-transition="fade" data-background="white"--> ### Joint Learning  - the first term is the conditional density of the time and account type of the next event in the activity sequence - the last term is the log-likelihood of jointly learned account embeddings under a **GMM** ----  --- <!-- .slide: data-transition="fade" data-background="blue"--> ## Experiments ---- <!-- .slide: data-transition="fade" data-background="cyan"--> ### IRA Twitter Dataset ----  ---- <!-- .slide: data-transition="fade" data-background="brown"-->  ---- <!-- .slide: data-transition="fade" data-background="white"-->  ---- <!-- .slide: data-transition="zoom" data-background="green"--> ### Application to COVID-19 ---- <img src="https://i.imgur.com/hGzpeto.png" width="500">