Hawkes process as a model of social interactions: a view on video dynamics

TL;DR

This paper uses computer simulations to evaluate the Hawkes process as a model for YouTube video view dynamics, confirming some predicted behaviors but highlighting limitations in its universality due to narrower decay exponent ranges.

Contribution

It provides a simulation-based assessment of the Hawkes process for modeling social interaction dynamics, specifically in online video viewership.

Findings

Simulated Hawkes process timeseries match predicted activity burst categories.

Hawkes process shows narrower decay exponent spread than actual YouTube data.

Limits to the universality of Hawkes-based analysis are identified.

Abstract

We study by computer simulation the "Hawkes process" that was proposed in a recent paper by Crane and Sornette (Proc. Nat. Acad. Sci. USA 105, 15649 (2008)) as a plausible model for the dynamics of YouTube video viewing numbers. We test the claims made there that robust identification is possible for classes of dynamic response following activity bursts. Our simulated timeseries for the Hawkes process indeed fall into the different categories predicted by Crane and Sornette. However the Hawkes process gives a much narrower spread of decay exponents than the YouTube data, suggesting limits to the universality of the Hawkes-based analysis.