<h3>Comparing information diffusion mechanisms by matching on cascade size </h3> <!-- .slide: data-transition="zoom" data-background="White"--> _Jonas L Juul, Johan Ugander_ presented by [Albert M. Orozco Camacho](https://twitter.com/alorozco53) --- <!-- .slide: data-transition="zoom" data-background="blue"--> ## Introduction ---- ### Why **cascades**? - Help scholars compare structural properties of _spreading content_ - Literature indicated that spreading phenomena can be explained by cascade statistical properties (_depth_, _breadth_, _viral index_). - Authors analyze two important cascade datasets: _true-fake-news_ and _videos, pictures, news spread_ ---- ### Twitter Rumour Dataset [Vosughi et al., 2018](https://www.science.org/doi/10.1126/science.aap9559)   ---- ### Media Cascades [Goel et al., 2016](https://pubsonline.informs.org/doi/10.1287/mnsc.2015.2158)  ---- ### In this paper... - The authors argue that _cascade size_ is an important confounder to explain the spread of false and real news. - Hence, structural differences can be _almost explained_ by differences in size. - Propagation of false news can be addressed by only controlling person-to-person infectiousness. --- <!-- .slide: data-transition="zoom" data-background="pink"--> ## Online Information Diffusion Data ---- #### From the True-False-News Data... - The authors reproduce 5 statistical quantities: cascade size (number of retweets), depth, maximal breadth, _virality_ (average pairwise distance of nodes), and geometric mean of time it takes a cascade to be retweeted by a number of unique users. - Previous authors concluded that cascade of fact-checked false news are bigger, breader, more viral, and spread faster. - Yet these properties collapse when controlling cascade size! ---- <!-- .slide: data-transition="zoom" data-background="white"-->  ---- <!-- .slide: data-background="white"--> #### What happens with images, videos, news, and petitions? > Properties keep holding even when controlling _size_!  --- <!-- .slide: data-transition="zoom" data-background="yellow"--> ## Size-Controlled Cascade Comparisons ---- ### Matching Procedure to Make Comparisons - The authors create a corpus of _matched_ cascades according to their sizes - Essentially, for every true-news cascade, a random false one is sampled - Unmatched cascades are NOT included in the corpus - The same is done for the video, image, news, and petition dataset. ---- <!-- .slide: data-transition="zoom" data-background="white"-->  ---- <!-- .slide: data-transition="zoom" data-background="white"-->  --- <!-- .slide: data-transition="zoom" data-background="red"--> ## Size Collapse in Models of Diffusion ---- <!-- .slide: data-transition="fade" data-background="cyan"--> ### SIR and IC _"In models of diffusion, when can the structural differences between resulting cascades be reduced to differences in size?"_ ---- - The IC model is parameterized by an _infectiousness_ parameter $p_{ij}$ between neighbours. - This parameter is fixed and adjusted for the purposes of this paper - The SIR is parameterized by the ratio $R_0 = r_I / r_R$ of a node's succeptibility to its ability to recover. - The authors perform simulations of the IC and SIR models. ---- <!-- .slide: data-transition="fade" data-background="white"-->  ---- <!-- .slide: data-transition="fade" data-background="white"-->  ---- <!-- .slide: data-transition="fade" data-background="white"-->  --- <!-- .slide: data-transition="fade" data-background="green"--> ## Discussion and Limitations