[Solomonov Seminar] 247. Solomonov seminar

[Marko Grobelnik]

V cetrtek, 20. septembra bo ob 13h v Oranzni predavalnici IJS (drugo 
nadstropje glavne stavbe) 247. Solomonov seminar. Posnetki preteklih 
seminarjev so na http://videolectures.net/solomon/

Tokrat bomo gostili prof. Hiroshi Motodo iz Japonske, ki bo predaval na 
temo glasovanja in formiranja mnenja v socialnih omrezjih.

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Hiroshi Motoda, AFOSR/AOARD and Osaka University

        Opinion Formation by Voter Models in Social Networks

Large scale social networking applications have made it possible for
news, ideas, opinions and rumors to spread easily, which affects and
changes our daily life style substantially. Massive data are
constantly being produced and are made available to us, enabling the
study of the spread of influence in social networks. Much of the work
has treated information as one entity and nodes in the network are
either active (influenced) or inactive (uninfluenced), i.e., there are
only two states. In this work, we address a different type of
information diffusion, which is ``opinion formation'', i.e., spread of
opinions. This requires a model that handles multiple states. Since
each opinion (what is said) has its own value and an opinion with a
higher value propagates more easily/rapidly, we first extend the basic
voter model to be able to handle multiple opinions, and incorporate
the value for each opinion. We call this model the value-weighted
voter (VwV) model. We learn the weight from a limited number of
opinion propagation data and predict the future share. We further
added a new component to the VwV model reflecting the fact that there
are always people that do not agree with the majority,
i.e. anti-majoritarians.  The model is called the value-weighted
mixture voter (VwMV) model which combines the VwV and the anti-voter
models both with multiple opinions. We also learn the weight and the
anti-majoritarian tendency from the data. Learning the
anti-majoritarian tendency is much more difficult than learning the
weight, but we show that both are learnable from the data. We carry
out the mean field analysis to VwMV model to gain an insight into the
average behavior of opinion share and find some interesting features.
Finally, we address the problem of detecting the change in opinion
share caused by an unknown external situation change under the VwV
model with multiple opinions in a retrospective setting. This is the
double loop learning problem and the brute force approach is
infeasible. We show that the use of the first order derivative of the
log likelihood results in much faster solution.