关键词: 主题演化, 深度动态聚类, 表示学习, 主题挖掘, 对比学习

Abstract: To address challenges related to topic ambiguity and alignment in existing deep document clustering methods when processing dynamic textual data with evolving themes, a Deep Evolutionary Topic Clustering Model (DETCM) was proposed. Temporal thematic evolution was captured by integrating historical topic patterns with features from the current time slice, allowing event theme trajectories to be uncovered and dynamic document-topic representations to be constructed. To reduce topic ambiguity during temporal transitions, a topic fusion learning module based on a hybrid encoder was designed, in which topic information from preceding time slices was utilized to enhance topic discrimination and feature extraction. Furthermore, a cross-temporal topic contrastive inheritance module was introduced to achieve topic alignment and consistency across different time slices. Historical thematic knowledge was effectively transferred and incorporated into current cluster formation through this mechanism. Experimental results based on the real-world arXiv evolving textual document dataset demonstrate that, compared with the Deep Evolutionary Document Clustering model with Instance-level Mutual Attention Enhancement (DEDC-IMAE), the proposed DETCM model achieves an average improvement of 3.08% in Normalized Mutual Information (NMI) across all temporal checkpoints. This confirms the superior capability of DETCM in tracking thematic evolution under dynamic scenarios, enabling more accurate capture of temporal variations in document topics and leading to enhanced clustering performance.

Key words: topic evolution, deep dynamic clustering, representation learning, topic mining, contrastive learning