The workflow supports visual summarization and stage analysis of event sequence data. Raw heterogeneous event sequences (varying in length, event types, and order) are first preprocessed into a uniform tensor representation through filtering low-importance events via TF-IDF, aligning sequences to a common temporal origin, folding events into fixed stages, and modeling the result as a three-way tensor (Entity × Stage × Event).
The tensor is then decomposed to extract a set of latent threads—orthogonal clusters of similar event sequences—each characterized by event distributions, stage-wise occurrence probabilities, and entity affinities. Threads are grouped into latent stage categories at each stage using a similarity-based layout and clustering algorithm with an adjustable distance threshold.
The analyst interactively explores the thread visualization—a line-map display showing thread evolution, branching, and merging across stages—together with coordinated context views (event flow, entity profiles, ranked event lists, thread comparison treemaps). The clustering threshold and number of threads are iteratively adjusted until the stage categories reveal meaningful, interpretable evolution patterns. Once satisfactory structure is found, the analyst identifies evolution patterns (branches, merges, divergences) and formulates domain-specific insights by interpreting what the discovered structures represent in terms of the underlying events and entities.