Statistical testing of proxy-based influence maxim

Időpont: 
2021. 01. 07. 14:15
Hely: 
Az előadás online formában lesz megtartva.
Előadó: 
Sziklai Balázs ( Budapesti Corvinus Egyetem, Matematikai és Statisztikai Modellezés Intézet / Operációkutatás és Aktuáriustudományok Tanszék)

                                                                                MEGHÍVÓ

                                          Szeretettel meghívjuk Sziklai Balázs előadására

                                    az Adatelemzés és Optimalizálás szeminárium keretében

                                         2021. január 07-én csütörtökön 14:15 órai kezdettel

                                               Az előadás online formában lesz megtartva.

                                                                     https://meet.google.com/umt-wxju-oma

 

Előadó: Sziklai Balázs ( Budapesti Corvinus Egyetem, Matematikai és Statisztikai Modellezés Intézet / Operációkutatás és Aktuáriustudományok Tanszék)

Előadás címe:  Statistical testing of proxy-based influence maximisation algorithms

Absztrakt:

In the Influence Maximisation Problem we seek to identify the most prominent agents in a network, whose activation would result in the largest influence spread. Proxy-based Influence maximisation algorithms are network centralities that quantify the potential of the agents. These algorithms are typically ranked by how their top choices perform in a diffusion simulation. This comparison method does not necessarily reflect how these proxies are used in real-life applications. Viral marketing campaigns may consider complex individual features limiting the possible choices of agents. We may have to choose our spreaders from a small set that might not contain any highly ranked agents at all.  There is no guarantee that a proxy that is better at predicting the performance of the most popular agents will be equally successful for an arbitrary group of prospective individuals. In this paper, we provide a systemic test for comparing proxies based on node sampling and on a novel statistical method, the Sum of Ranking Differences. Using a Linear Threshold diffusion model on an online social network, we evaluate seven measures. We demonstrate that statistical assessment of these proxies, when low-ranked individuals are present as well, is remarkably different from what we obtain by examining the performance of their top choices exclusively. The results highlight, that the standard simulation test must be accompanied by statistical tests, for which Sum of Ranking Differences provides a widely applicable framework.