Authors: Fangcao Xu*, Penn State University
Topics: Political Geography, Temporal GIS, Quantitative Methods
Keywords: Networks of Diffusion, Spatiotemporal Pattern Discovery, Pairwise Specific Distance, Topic Modeling, Multi-Instance Multi-Label Learning
Session Type: Paper
Start / End Time: 1:20 PM / 3:00 PM
Room: Bayside B, Sheraton, 4th Floor
Presentation File: No File Uploaded
Networks play a fundamental role in the information diffusion process. Innovations often spread over the time and space through individual transmissions via our social networks, as well as through external media, including published news. However, the underlying social networks and paths through which the diffusions and propagations pass is difficult to trace. The recent availability of massive event data from the digital news media allows us to study the diffusion network in more detail.
Usually a good network model performance can only be achieved when a good distance metric is obtained. Some unobserved linkage information (e.g. the sequence or order of events in a specific pattern), however, has not been exploited in distance metric calculation. Moreover, a node (e.g. place name) in the network could be detected in multiple instances and associated with multiple event types. It greatly influences how the information is transmitted through the network because instances with different properties, even they are detected at the same node, may have different semantic meanings.
To tackle these challenges, the Multi-Instance Multi-Label learning (MIML) framework is proposed in this research to deal with the complicated problem of non-uniqueness. Rather than generate a global distance as in previous MIML models, I have developed a pairwise specific distance (PSD), which calculates different distances for different linkage pairs with the aim of providing a more detailed measure of how the information is transmitted through the linkages of spatialtemporal events reported in the news media.