Mapping Hidden Social Networks in Marginalized Communities Using Machine Learning

Abstract

This paper takes a hybrid experimental methodology that integrates qualitative and quantitative methods in the investigation of the possible potential of machine learning in order to uncover implicit social networks in disadvantaged populations.  Supervised and unsupervised learning models were combined with digital trace examination, focus groups (participatory), and ethnographic interviews to develop and test latent relational linkages.  The results depict that there is significant disparity in the dynamics of trust, structural centrality, and community involvement where discrete clusters are observed all over the network.  Although model analysis showed the ensemble and neural network classifiers to be the most effective, with F1-scores of more than 0.90, quantitative data, including communication frequency, semantic similarity, co-location information, and others, turned out to be potent predictors of hidden relationships.  High levels of correspondence with relationships inferred were achieved through contextual dependability through community member validation.  These findings were also reinforced by the visualizations that revealed the connection in modularity in cluster memberships, variance in trust and reciprocity distributions and patterns of communication and tie strength.  Ethical considerations such as anonymization and differential privacy ensured protection of the participants and analytical rigour was ensured.  The results of the study posit that mapping concealed societal systems through machine learning and participatory validation offers an effective, ethically sound model that has implications on formulating inclusive policy, promoting resilience, and empowering the underexemplified communities.