AbstractTextual data analysis is critical for monitoring changing themes over time. To overcome challenges posed by data richness, graph theory emerges as a tool for investigating word-topic associations. We present an approach to clustering co-occurrence word networks that prioritises network similarity quantification over time. Addressing theoretical and network geometrical constraints, a statistical framework for manifold data analysis facilitates the grouping of semantic networks, partitioning the observed time frame into periods, and identifying dominant topics in each period via tensor decomposition. The analysis of Brexit-related tweets demonstrates the efficacy of modern methods for identifying social media patterns on public discourse.