Background
User-generated geo-data from sources like social media, micro blogs, or mobile networks carry valuable information (spatial, temporal, semantic), which is able to support decision-making in humanitarian action. However, these data are methodologically challenging to analyse because of various factors of uncertainty, including location uncertainty, temporal uncertainty, semantic ambiguity, lacking structure, unknown spatial interrelationships, and unstructured text.For these reasons, no standardised set of methods to analyse social media data exists and established methods are mostly not capable of dealing with the complexity and the unstructured nature of social media data. Therefore, artificial intelligence (in particular machine learning) algorithms need to be researched to analyse social media posts, and to assess the quality of user-generated and related information products.
Methods
This project aims to investigate new cross-disciplinary machine learning methods to analyse social media, dealing with the data’s noisy characteristics. Therefore, this project covers the workflow from data gathering and filtering to analysing and classifying social media posts. In the next step, the analysed data will be visualised in an expressive manner. We design and validate our approach for the use case of disaster management and then transfer it to other use cases, i.e., epidemiology and the detection of refugee movement to support humanitarian action.
Results
As result, we expect multi-modal (spatial, temporal, semantic) machine learning algorithms for social media analysis to support decisions in humanitarian action, disaster management and epidemiology through information provision in near real-time. Additionally, a demonstrator of a decision-support dashboard for user-tailored result visualisation will be created.
Helen Ngonidzashe Serere, Clemens Havas, Andreas Petutschnig