Leveraging complex network features improves vaccine stance classification

Abstract

The widespread use of social media allows unprecedented ways to monitor opinions and stances regarding critical public health issues globally. Advanced Natural Language processing algorithms are being used routinely to extract information and classify vaccination hesitancy or stance. However, communication on online social networks such as Twitter (now X) is carried by short messages, the meaning of which can be difficult to understand in the absence of context. Therefore, in this study we propose the use of complex-network features extracted from the social network to integrate and enhance text-based Deep Learning models. Leveraging a dataset of about 20 million Italian language posts (of which about 7000 were manually annotated), we showed how the integration of text and network features improves vaccine stance classification, especially for the most polarized classes. Additionally, network features overperformed text features in a dataset collected a year after model training, possibly indicating how the social network changes more slowly than the trending words or topics.

Fig. 1

Community distribution using Leiden algorithm fixing the 90% coverage, The first six communities cover the 90% of the first dataset, while the remaining 10% of the users is placed in the seventh community.

Fig. 2

Network layout using fa2 algorithm. Different colors denote the distinct Leiden communities represented in the visualization. a: first dataset, b: second dataset, N: number of users.

Fig. 3

F1 scores per class in (a) first test set and (b) in the second dataset, for the integrated classifier (fa2 + text), and classifiers based on fa2 and text separately.