Dynamic Dual Sparse Topic Model: Integrating Temporal Dynamics and Sparsity with Spike and Slab Priors into Topic Model

Topic modeling is a statistical technique that identifies underlying abstract 'topics' in a corpus of texts. Notable among current topic modeling methodologies are the Dynamic Topic Model (DTM), which captures the temporal dynamics of topics, and the Dynamic Sparse Model (DSM), which emphasizes the sparsity in the distributions of topics and words. However, no existing model appropriately integrates both the temporal progression and the sparsity within these distributions. Therefore, in this study, we propose the Dynamic Dual Sparse Topic Model (DDSTM), which employs the Spike and Slab Prior as its prior distribution, thereby extending the capabilities of both DTM and DSM. The DDSTM is distinct in its ability to simultaneously manage the temporal dynamics and sparsity in topic distributions, significantly enhancing the model's interpretability. Theoretically, we prove that the DDSTM maintains the same level of sparsity as the existing DSM models. In the empirical analysis, we use real-world text data to examine the characteristics of the DDSTM compared to existing topic models.