A Computational Framework for Cross-Temporal Sentiment and Topic Analysis of Reddit Data

This paper develops and validates a modular computational framework for longitudinal sentiment and thematic analysis of unstructured social media data. While computational social science increasingly relies on automated tools, a gap remains in validated pipelines that rigorously integrate temporal dynamics with reliable sentiment metrics for noisy text. The proposed technical contribution is a three-stage pipeline comprising: (1) a comparative validation protocol establishing the superiority of lexicon-based rules (VADER) over un-tuned transformers (BERT) for specific informal domains; (2) a hybrid hierarchical topic modeling approach; and (3) a robust regression architecture cross-validated with LightGBM machine learning models. The framework's engineering utility is demonstrated through a 15-year dataset (N = 2267) of cross-cultural Reddit discourse. Results confirm the pipeline's ability to reliably detect event-driven signal shifts and isolate feature importance. This study offers engineers and data scientists a validated, replicable solution for high-noise public opinion mining and crisis monitoring applications.