رسائل دكتوراة
English
2023
Developed Model Based on Text Analysis and Geo-Tweets for Detecting the Spread of Covid-19
أحمد سليم عباس جاسم, اقبال عبد الباقي محمد مهدي
كلية تكنولوجيا المعلومات-جامعة بابل · العراق
Twitter is a mine of gold; it provides incredibly valuable information for every working sector, mainly the health sector. Its reputation grew faster as the virus of COVID-19 spread universally and quickly. Previously, the geographical information provided by Twitter was limited. However, during the disease emergence, Twitter datasets were released with a broader scope, mainly provided with geographical information. This information is needed for detecting the spread of the virus and to identify infected areas.
For this dissertation, the demand for a geospatial dataset is crucial. Many datasets have been launched for the public; however, they are unsuitable for the purpose of this research because of the lack of some important information. The collected dataset used several controlled keys (logical statements) to reach the desired tweets, which enabled the authors to find the exact infected cases (self-reported users) among millions of tweets posted every hour regarding this crisis, and to analyze their exact location.
These tweets were split into four groups according to the type of geographical information provided with the tweet. Each type was exploited and examined to figure out the spatial location for each individual. These were taken from the place dictionary, the location field, the matched location and place information, and finally the text tweet content to be inferred. The authors built a small global gazetteer to investigate the spatial entities and to inspect the exact location for the infected user firstly, and his/her social network relations secondly.
This allowed identifying spatially close followers and friends of the infected user, determining those who live in the same area, then retrieving their tweets within a specified period of time and extracting their spatial information. Those in the same proxy environment were isolated by applying a clustering method—the k-means algorithm—to the longitude and latitude coordinate points for all of them. Those who gained the same cluster label were isolated as one of the proxy relations.
Their tweets were collected to be investigated about contracting the disease by applying a supervised classification method using machine learning. Before this step, the collected dataset was manually labeled according to specific criteria into three classes. The multilayer perceptron algorithm (MLP) was applied to the text tweets to classify the three types and identify the infected follower. This neural model achieved an accuracy of 0.88.
The positive cases were collected and their distribution was shown on the map of New York City as a case study, providing the number of infectious cases within that area. This research presents a complete spatial dataset of infected cases collected over one year, along with a labeled dataset of three classes to show the state of COVID-19 infectious disease. The classification process was able to identify self-reported infected users, excluding spam and declarative tweets.