Extending Theory-Guided Data Science to Consider Social Science Domains

Abstract:

Data-driven analyses are often an epitome of modern data science assessments. However, the prevalence of black box model implementations in data science for scientific applications hinder subsequent investigations due to lack of explainability. Theory-guided data science (TGDS) aims to address these issues by explicitly integrating data science with physical science domain knowledge, the latter guides everything from parameter selection and initialization to model design. This study extends TGDS to consider theoretical integrations of social science domains for data science applications. Specifically, our analysis focuses on the training of a neural network as a basis for coupling data about wildfire events with insights about journalism and established social science theories to improve understanding of natural disaster coverage in media. Preliminary results indicate that such integration of data-driven and theory-based knowledge can address issues of low representative datasets and model explainability. These advancements demonstrate significant potential for data sciences to serve as an integrative platform for diverse knowledge domains towards improving our understanding of complex societal behaviors of geoscience-relevant problems.