Reproduce “Predicting Opinion Dynamics via Sociologically-Informed Neural Networks” on OpenResearch

About this paper

Maya Okawa, Tomoharu Iwata

Researchers at NTT Human Informatics Labs and NTT Communication Science Labs developed Sociologically-Informed Neural Networks (SINN) to predict opinion dynamics in social networks by integrating deep learning with established sociological models. This approach, which adapts Physics-Informed Neural Networks to social science, achieves higher accuracy and interpretability than purely data-driven or theoretical methods on various datasets.

Key insights

• The concept of Physics-Informed Neural Networks (PINNs) can be successfully adapted to social science, enabling the integration of sociological theories with data-driven deep learning for complex social phenomena.
• A hybrid modeling approach, combining theoretical sociological knowledge (via ODEs) with deep learning, yields more accurate and interpretable predictions of opinion dynamics than either purely theoretical or purely data-driven methods.
• Mechanisms like Gumbel-Softmax reparameterization and pre-trained language models effectively address challenges of stochasticity and rich textual side information within a differentiable social science model.

Summary by alphaXiv.