EventBox: Temporal Visual Analytics

• EventBox is a visual encoding method that jointly represents temporal attributes, like duration and start time, along with categorical and numerical data within event groups.
• It integrates interactive transformations, outlier detection, and multivariate statistical summaries to enable detailed and flexible event sequence exploration.
• Embedded in the Sequen-C system, EventBox supports subgroup comparisons, pattern identification, and visual analytics for complex event data across diverse domains.

EventBox is a visual encoding and data representation for analyzing groups of event occurrences in temporal event sequences while jointly showing time-related attributes and other multivariate attributes. Introduced within the Sequen-C visual analytics system, it was designed to address a recurring limitation in event sequence analysis: many prior systems emphasize sequence order, pathways, or common patterns, but do not explicitly support the combined exploration of temporal attributes such as duration or start time together with categorical and numerical event or sequence attributes. EventBox therefore combines interactive transformation, multivariate visual encoding, and automatically generated statistical analyses in a single analysis workflow (Montana et al., 19 Jul 2025).

1. Conceptual basis and analytical scope

EventBox was proposed for the analysis of event sequence data across domains in which effective analysis and exploration are needed to facilitate decision-making. Its stated target is the joint analysis of temporal and multivariate attributes within groups of events, rather than sequence order alone. The analytical questions it is meant to support include how duration and start time relate within a specific event type, whether subgroups differ in temporal behavior, where frequent patterns and outliers occur, how categorical factors interact with temporal attributes, and which visually apparent attribute interactions are also statistically supported (Montana et al., 19 Jul 2025).

The design requirements reported for EventBox are fourfold. RE1 Main patterns addresses identification of the most frequent attribute values. RE2 Outliers addresses identification of infrequent values. RE3 Frequency distribution of multivariate attributes addresses joint study of up to five attributes. RE4 Statistical measures adds quantitative summaries and significance support. Within Sequen-C, these goals are embedded in a broader task structure that includes pattern and outlier detection at both sequence and event level, targeted focus on subsets, interactive transformation, flexible multivariate analysis, and quantitative summaries (Montana et al., 19 Jul 2025).

The paper presents EventBox as a response to two shortcomings in prior work. First, temporal information in sequence overviews was often reduced to order alone or to average durations, without exposing distributions and outliers. Second, multivariate attributes were often shown through averages, single-attribute encodings, or separate linked views, making it difficult to inspect interactions among multiple attributes within the same event group. This suggests that EventBox should be understood less as a standalone glyph than as a structured answer to a specific visual analytics problem: simultaneous inspection of temporal distributions, individual event instances, subgroup structure, and attribute interactions within comparable event groups (Montana et al., 19 Jul 2025).

2. Formal representation and visual encoding

The formal definition given for an EventBox is

$E = \{e^k, (p_h, p_v, s_h, s_v, b) \}$

where $k = 1, \dots, N$, and $N$ is the number of time-stamped event occurrences in the group. Each $e^k$ is an event occurrence. The tuple $(p_h, p_v, s_h, s_v, b)$ specifies the attributes visualized in that EventBox: $p_h$ and $p_v$ are two primary attributes, $s_h$ and $s_v$ are two secondary attributes, and $b$ is one breakdown attribute. The subscripts $k = 1, \dots, N$0 and $k = 1, \dots, N$1 indicate horizontal and vertical axes. Each event occurrence has a multivariate attribute vector

$k = 1, \dots, N$2

where attributes may be temporal $k = 1, \dots, N$3, categorical $k = 1, \dots, N$4, or numerical $k = 1, \dots, N$5. Users interactively choose which attributes play the primary, secondary, and breakdown roles, and the implementation supports simultaneous exploration of up to five attributes (Montana et al., 19 Jul 2025).

In the healthcare examples reported in the paper, the default primary attributes are typically $k = 1, \dots, N$6 as duration and $k = 1, \dots, N$7 as start time, but other assignments are possible, including categorical variables. All events in a given EventBox share the same event type, possibly after aggregation or substitution, and share a subset of attribute types represented by $k = 1, \dots, N$8 (Montana et al., 19 Jul 2025).

The visual encoding combines ideas from box plots, scatter plots, and histograms. It uses four principal visual marks: container area, quartile lines, data points, and histograms. A rectangular container area contains all $k = 1, \dots, N$9 events in the group. Hue encodes the event type. Height is proportional to the number of events $N$0. Width is proportional to the maximum value of the primary horizontal attribute $N$1 across events. This gives an immediate size cue for group frequency and horizontal extent for the main quantitative measure (Montana et al., 19 Jul 2025).

The distribution of $N$2 is encoded along the horizontal axis using the five standard boxplot statistics: minimum, $N$3, $N$4, $N$5, and maximum. These are shown as vertical line marks inside the box. Alternating bands between quartile lines are shaded using different saturation levels. White indicates no events, and light grey indicates outlier regions. Outliers are computed using Tukey’s rule exactly as

$N$6

with

$N$7

so points outside this interval are treated as outliers (Montana et al., 19 Jul 2025).

Each individual event occurrence $N$8 is also plotted as a point. The x-position encodes $N$9, and the y-position encodes $e^k$0. The vertical axis is scaled top-to-bottom over the range of $e^k$1. Points can additionally encode the breakdown attribute $e^k$2 by color hue, or use transparency to create a density-like heatmap effect. Histograms summarize distributions of the primary attributes when no secondary attributes are selected. Once secondary attributes are selected, histogram bars become stacked bar charts, with bar height encoding frequency and color encoding secondary attribute value. A breakdown operation can split one EventBox into one EventBox per distinct value of $e^k$3, such as separate EventBoxes for Monday, Tuesday, and other weekdays (Montana et al., 19 Jul 2025).

This arrangement gives EventBox its hybrid character. It is simultaneously a distribution summary, a scatterplot of individual occurrences, and a basis for subgroup comparison. A plausible implication is that its analytical utility depends on this layered reading: quartiles summarize, points localize, histograms contextualize, and breakdown partitions support comparative analysis (Montana et al., 19 Jul 2025).

3. Integration into Sequen-C and user-driven transformations

EventBox is not presented as a standalone plot. It is integrated into Sequen-C, described as a multilevel visual analytics system for temporal event sequences. Sequen-C includes coordinated panels for general settings and controls, events, clusters, EventBox, unique sequences, individual sequences, attribute analysis, and filters. Selections propagate across all views. EventBox appears both inside overview panels and as a dedicated detailed analysis panel (Montana et al., 19 Jul 2025).

In the Events panel, users choose which event types should use EventBox encoding. In Clusters, Unique Sequences, and Individual Sequences, EventBoxes can summarize selected event types within clusters. In the dedicated EventBox panel, users set primary, secondary, and breakdown attributes, show or hide outliers, points, and heatmap views, merge EventBoxes, and generate statistical reports. In the Attribute Analysis panel, selected sequence groups can be compared to the full dataset using stacked bars. In the Filters panel, interactions become structured queries, and users can define compound queries such as $e^k$4 (Montana et al., 19 Jul 2025).

A major part of the system contribution is a set of user-driven transformations that define meaningful event groups for EventBox analysis. Substitution allows users to replace existing event types with a new user-defined type, reducing semantic complexity by mapping multiple fine-grained events to a broader concept. Aggregation merges consecutive occurrences into one aggregated event, with attributes merged appropriately. The purpose is to reduce clutter, shorten sequences, and enable group-level EventBox analysis over broader process stages or even entire sequences. One reported example aggregated all events in 9,003 sequences into a single event type so that one EventBox could analyze whole-sequence duration by weekday (Montana et al., 19 Jul 2025).

Alignment adds interactive manual alignment to a system that previously had automatic alignment. Users define event types of interest as hard or soft: hard events are aligned first, and soft events are aligned between pairs of hard-aligned events. Sequences are padded and gaps inserted to achieve uniform alignment while preserving event order. The paper explicitly notes that the inserted gaps are not real time intervals; they are only for alignment and comparison. Sorting uses a selected event index to create suffix sub-sequences from that point to sequence end, then compares those sub-sequences to determine sequence order. The stated purpose is improved visual comparability, especially after alignment (Montana et al., 19 Jul 2025).

The intended workflow is reported as an ordered sequence: start with an overview of clusters or sequences; identify sequence groups or event types of interest; apply substitutions and aggregations; align around important hard or soft events; sort to improve local comparability; select an event type for EventBox encoding; choose primary, secondary, and breakdown attributes; inspect quartiles, outliers, distributions, subgroup breakdowns, and heatmaps; use lasso selection or bar selection to isolate subsets; generate the statistical report; and propagate findings through coordinated views and queries. This suggests that EventBox is tightly bound to Sequen-C’s transformation pipeline rather than functioning as a static summary graphic (Montana et al., 19 Jul 2025).

4. Statistical support and empirical case studies

EventBox is complemented by an automatically generated statistical report driven by the user-selected attributes. The report is explicitly intended to reduce bias from purely visual exploration and to increase confidence. It includes three kinds of summaries: continuous summaries and mean comparison tests at different user-defined granularities, contingency tables for associations among selected categorical attributes, and ANOVA tables relating a user-selected continuous attribute to one or more categorical attributes to support model building and variable selection. The paper states that the tests rely on distributional assumptions justified by the central limit theorem due to large group sizes. No explicit effect size, correction method for multiple testing, or p-value adjustment procedure is provided in the reported text (Montana et al., 19 Jul 2025).

The first healthcare case study concerns ANCU, an outpatient antenatal clinic dataset covering three months, with 73,279 recorded events and 9,623 pregnant women. Event types include consultations, ultrasound scans, and blood tests. The stated question was to understand patient waiting times for consultation, especially for patients undergoing an ultrasound scan. The analyst selected 8 clusters, focused on cluster C1, aligned the subsequence NYA, IS, WC, IC, and COM, used WC and IC as soft events to ensure they fall between IS and COM, sorted by WC, and built an EventBox for WC using duration and start time as primary attributes and day of week as breakdown. The reported insights were that Fridays had a higher proportion of waiting-time outliers; Mondays and Fridays mainly had morning consultations with longer waits; Tuesdays to Thursdays had both morning and afternoon sessions with shorter waits; and adding ClinicCode as a secondary attribute on the vertical histogram showed distinct clinic allocation by time slot, including a Wednesday split between Clinic_21 in the morning and Clinic_58 in the afternoon (Montana et al., 19 Jul 2025).

The second case study concerns CUREd ambulance service calls in Yorkshire and Humber. The dataset spans three months, includes 25,243 calls, 21,805 patients, and 34 attributes. The stated question was how call pathways relate to patient characteristics and what factors influence call duration. Fifteen clusters were selected, and clusters C1 and C6 were compared because they had similar pathways; attribute analysis showed that C6 predominantly involved children. After alignment on event CAL, EventBox used duration and start time as default primary attributes. Reported insights include a very long duration outlier in C1 of over 5 hours; outlier proportions of 10.58% for C1 and 8.48% for C6; concentration of long-duration outliers among non-urgent transfer requests between healthcare facilities and ED; more outliers in the afternoon for both clusters; and quartile-band evidence that ambulance arrival occurred within 40 minutes for 75% of calls in C1 and within 30 minutes for 75% of calls in C6. A second EventBox heatmap, with duration as $e^k$5, urgency as $e^k$6, and reported symptom as secondary vertical attribute, showed that most Red calls were dispatched quickly within 10 minutes, that Amber calls had more variable durations, and that among frequent symptoms the C1 Red group was associated with Chest Pain while the C6 Red group was associated with Convulsions Fitting. The accompanying ANOVA on call duration with day of week, urgency, symptom, and cluster number reported a significant three-way interaction among cluster, day of week, and urgency. Among 264 main effects and interactions, the most statistically significant increases in call time were associated with Psychiatric Suicide Attempt, GreenT, and GreenF (Montana et al., 19 Jul 2025).

The third case study uses a MIMIC-IV-derived dataset from admissions, diagnoses_icd, patients, prescriptions, services, and transfers. After filtering patients with at least two of hypertension, chronic kidney disease, and diabetes, the resulting dataset contained 1,776 patients. The initial cluster contained 102 event types and sequences up to 153 events long. The analyst substituted and aggregated all drugs into DRU, substituted and aggregated transfers into TRA, reduced event types from 102 to 7, aligned and sorted sequences, and obtained a final overview clarifying the proportion of transfers and how transfers related to medication changes before discharge (DIS). In this case, the emphasis is less on one particular EventBox insight than on how the transformation pipeline makes EventBox and overview analysis feasible in highly complex event sequences (Montana et al., 19 Jul 2025).

5. Evaluation, strengths, and limitations

The reported evaluation involved 21 participants: 3 domain experts and 18 novice data analysts. The evaluated system was Sequen-C with EventBox rather than EventBox in isolation. The novice analyst study involved participants aged 22–40, with 9 female and 9 male engineers who had data analysis experience but no prior exposure to Sequen-C or EventBox. The study included training, performance tasks, and an ICE-T questionnaire with free-text feedback. Screens and mouse clicks were recorded to verify timings and usage patterns, and no time limit was imposed (Montana et al., 19 Jul 2025).

The performance-task component used a 15-question multiple-choice questionnaire. Q1–Q7 did not require EventBox, Q8–Q9 could be answered without EventBox but with much greater effort, and Q10–Q15 required EventBox. The reported average accuracy was 90.37% $e^k$7, and the average response time per question was 95 seconds, with a range from 27 seconds for Q7 to 221 seconds for Q9. Q3 was the hardest question, with 8 of 18 correct, corresponding to 44.44%. Q9, despite requiring multiple steps and panel interactions, was answered correctly by all participants (Montana et al., 19 Jul 2025).

The ICE-T framework was used to assess value beyond task accuracy. ICE-T evaluates Insight, Confidence, Essence, and Time on a 1–7 scale, with scores greater than or equal to 5 interpreted as strengths. The reported overall visualization value was 5.82 $e^k$8. Component scores include Insight at 6.04, which was the highest, and Time at 5.7, which was the lowest. The highest-rated Essence heuristic was 6, and the lowest Confidence heuristic was 5.06. The highest-rated Essence heuristic was quoted as: “The visualization helps understand how variables relate in order to accomplish different analytic tasks.” The lowest Confidence heuristic indicated a need for more explicit communication of inconsistent, duplicate, missing, or invalid data (Montana et al., 19 Jul 2025).

The domain-expert sessions involved three healthcare delivery experts, each in a 90-minute session comprising a walkthrough of new features, expert-formulated questions answered using Sequen-C and EventBox by an analyst, and feedback using ICE-T categories, usability, and future directions. The reported findings were strong support for rapid identification of patterns, outliers, and bottlenecks; value of EventBox breakdowns; usefulness of coordinated views and propagated selections; statistical analyses that increased confidence; and good accessibility relative to SQL-heavy workflows. At the same time, the paper reports that advanced features require learning, though experts considered them worth it (Montana et al., 19 Jul 2025).

The strengths explicitly claimed for EventBox include combining summary statistics, individual points, subgroup comparison, and multivariate encoding in one view; supporting explicit temporal analysis, including duration and start time; preserving outlier visibility; strong integration with interactive transformations; coordinated views for iterative exploration; and automated statistics for confidence and interpretation. The reported limitations are scalability on very large or complex datasets, possible cognitive overload, a current heatmap strategy that could be improved, lack of analytic provenance that may hurt reproducibility, a steep learning curve especially for novice users, and a need for better data quality signaling. Future work includes analytic provenance capture, integration with LLMs for automated storytelling, interactive predictive modeling based on EventBox and statistical results, standalone evaluation of EventBox outside Sequen-C, and better heatmaps with automatic feature selection (Montana et al., 19 Jul 2025).

6. Position within adjacent event-centric systems

Within the literature represented here, EventBox occupies a distinct position. It is a visual encoding for grouped event occurrences in temporal event sequences, whereas several adjacent systems operate at different layers of the event-analysis stack. podio’s Frame, for example, is introduced as “a thread safe, generalized event data container” intended to aggregate all relevant data for one user-defined processing unit and provide a uniform, thread-safe interface. It is therefore an EventBox-style abstraction in the sense of a generalized container, but not a visual representation. Its emphasis is on ownership transfer, immutability after insertion, read-only concurrent access, type erasure, and backend-neutral collection materialization rather than on visual analysis (Declara et al., 2023).

Other systems are event-visualization environments, but with different targets. ELAINA for JUNO is a Unity-based detector and event visualization system for observing detector geometry, tuning reconstruction algorithms, and analyzing physics events, including reconstruction and Monte Carlo truth overlays and 2D projections. A plausible implication is that ELAINA addresses spatial detector-event inspection, whereas EventBox addresses temporal and multivariate attribute analysis over grouped event occurrences (Zhu et al., 2018). A similar distinction applies to the Phoenix-based CEPC event display, which integrates detector geometry, reconstructed objects, and Monte Carlo truth in a web-based 3D environment oriented toward detector design, software development, and physics analysis rather than multivariate event-sequence summarization (Zeng et al., 21 Sep 2025).

Outside HEP detector visualization, RESIN-EDITOR is an interactive event graph visualizer and editor for hierarchical event graphs extracted from multimedia and multi-document news clusters with guidance from human-curated event schemas. Its main features are hierarchical graph visualization, comprehensive source tracing, and interactive user editing. This suggests a different event-analysis paradigm: graph-structured, schema-guided complex-event understanding, as opposed to EventBox’s grouped occurrence analysis with quartiles, points, histograms, and automated statistical summaries (Nguyen et al., 2023). ECAT similarly addresses event-centric analysis, but as an event capture annotation tool for Kinect-based RGB-D recordings, emphasizing event segmentation, participant labeling, subevents, trajectories, and mapping to VoxML (Do et al., 2016).

ETA, the Extensible Time-tag Analyzer, is again distinct. It is a toolbox for analysis of time-tagged measurements in which event-stream analysis is expressed through state diagrams plus code snippets and then compiled just-in-time into optimized native code. Its focus is correlation extraction, histogramming, and efficient processing of time-tagged measurements rather than visual encoding of grouped event occurrences. This suggests that EventBox belongs to visual analytics, whereas ETA belongs to event-stream processing infrastructure (Lin et al., 2021).

Taken together, these comparisons situate EventBox as a specialized contribution to event-sequence visual analytics. It is neither a generalized event container, nor a detector-event display, nor an event-graph editor, nor an event-stream processing engine. Its distinctive contribution is the joint visual treatment of temporal attributes, multivariate attributes, outliers, subgroup breakdowns, and statistical summaries within transformed and aligned groups of event occurrences (Montana et al., 19 Jul 2025).