IVESA supports domain experts in the interactive visual analysis of time-stamped event sequences (TSEQs) — temporal data consisting only of event timestamps without associated values. The workflow enables the exploration of temporal occurrence patterns across potentially thousands of sequences with up to one million events, supported by metrics, feature-based clustering, filtering, motif detection, and metadata contextualization.
The process begins with preprocessing, where long raw event streams are segmented into manageable sub-sequences (e.g., by year, month, or day) according to a domain-appropriate criterion, and aligned using a global temporal reference to enable meaningful comparison. A comprehensive set of 22 metrics is then computed, yielding 8 temporal (event-based) features (velocity, acceleration, density, entropy, etc.) and 14 static (sequence-based) features (mean gap, variance, regularity, skewness, etc.).
The system generates an initial overview consisting of coordinated views: a Sequence Overview displaying all events as vertical line marks on a shared time axis with sequences stacked vertically, a Metadata View cross-tabulating event distributions by two user-selected criteria (metadata attributes, features, or clusters) as a heatmap, and a Summary View providing entry points to data simplification tools.
The expert enters an iterative exploration loop. They inspect the overview to identify temporal patterns — seasonal trends, periodic behaviours, gaps, dense phases, or anomalies. They sort/rank sequences by selected features or metadata to reveal ordering-based patterns, select interesting sequences for detailed comparison in the Details View (which shows individual events, temporal features as line charts, and motif highlights), and cross-tabulate event distributions in the Metadata View to relate patterns to contextual attributes (e.g., weekday, device type, location). After each exploration cycle, the expert assesses whether their understanding is sufficient or whether data simplification would help.
When simplification is needed, the expert can apply one or more data reduction strategies: (a) filtering by metadata attributes or feature ranges to reduce the number of sequences, (b) feature-based clustering using interactive hierarchical clustering with user-selected and normalised features, producing a steerable dendrogram with linked colour-coded groups, or (c) motif detection where the expert defines a query-by-example event pattern from the Details View, configures similarity parameters (Euclidean or DTW distance, index-based or time-based kernel, threshold), and the system identifies and substitutes all matching occurrences across the dataset. After simplification, views are updated and exploration continues.
The cycle of exploration and simplification repeats — with the expert progressively building understanding through sorting, comparing, relating, clustering, and motif substitution — until they have identified and contextualized all patterns of interest. Finally, the expert synthesises knowledge: confirmed expectations, discovered anomalies, contextual explanations for temporal patterns, and actionable conclusions.