Skip to main content



Can Cascades be Predicted?

Source: “Can Cascades be Predicted?” by Justin Cheng, Lada A. Adamic, P. Alex Dow, Jon Kleinberg, Jure Leskovec

Cascades are a fascinating facet of social networks that have many characteristics and properties to study. One interesting question is how well can cascades be predicted? The paper, “Can Cascades be Predicted?,” explores this exact topic. The paper performs experiments on Facebook network data, specifically cascades based on resharing to tackle this question. Before performing experiments, they identify several factors that may contribute to the growth and spread of cascades. These factors are content features, original poster/resharer features, structural features of the cascade, and temporal features. Most feature sets are fairly self explanatory, for example, structural features describe the network structure and temporal features describe properties related to the speed of the cascade. Utilizing these feature, they apply many machine learning models predict whether a cascade will reach the median cascade size given the first 5 reshares of the cascade. However, they report scores on the simple and easily comparable logistic regression model, which is able to achieve a respectable 80% accuracy on their dataset. They find that even though all feature sets performed better than random predictions, the temporal features were quite predictive, outperforming all others.

In addition, to predicting cascade size they decided to try to predict the resulting structure of the cascade. Again they observe 5 reshares of a cascade and attempt to predict where the cascade structure falls in relation to the median. To quantify this they utilize a metric called the Wiener index, which is the average distance between all pairs of nodes in a cascade, an interpenetrate structural measure. Thus they attempt to predict if a cascade has a Wiener index above or below the median. Here they are able to achieve a decent 72.5% accuracy. Clearly one will expect structural features to be the most important, much more so than temporal features, but it turns out that they are surprisingly almost equally useful for predicting cascade structure. Nonetheless, independently structural features are the most accurate. For predicting both size and structure of cascade, when they conduct experiments using more than 5 reshares, they reasonably observe that cascades become more predictable.

Overall, this paper shows that they fairly accurate predictions on both cascade size and structure given a fairly small amount of observed reshares are achievable on Facebook data. However, it is important to note that the experiments performed here were done solely on Facebook data, so these results cannot be expected to remain constant for other social networks.

Comments

Leave a Reply

Blogging Calendar

December 2019
M T W T F S S
 1
2345678
9101112131415
16171819202122
23242526272829
3031  

Archives