Recurrent Coupled Topic Modeling over Sequential Documents

The abundant sequential documents such as online archival, social media, and news feeds are streamingly updated, where each chunk of is incorporated with smoothly evolving yet dependent topics. Such digital texts have attracted extensive research on dynamic topic modeling to infer hidden topics their temporal dependencies. However, most the existing approaches focus single-topic-thread evolution ignore fact that a current may be coupled multiple relevant prior In addition, these also incur intractable inference problem when inferring latent parameters, resulting in high computational cost performance degradation. this work, we assume evolves from all corresponding coupling weights, forming multi-topic-thread . Our method models dependencies between thoroughly encodes complex multi-couplings across time steps. To conquer challenge, new solution set novel data augmentation techniques proposed, which successfully discomposes A fully conjugate model thus obtained guarantee effectiveness efficiency technique. Gibbs sampler backward–forward filter algorithm efficiently learns time-evolving parameters closed-form. Indian Buffet Process compound distribution exploited automatically overall number customize sparse proportions for document without bias. proposed evaluated both synthetic real-world datasets against competitive baselines, demonstrating its superiority over baselines terms low per-word perplexity, coherent topics, better prediction.