Forty-two clusterings later, the elbow is still a matter of opinion.
DBC · densityPBC · partitionTM · topicsCD · community detectionclick the counter if you feel the urge to re-run everything
The paper
SmartIterator: Visual Analytics Workflows for Supervising Unsupervised Data Grouping
Gennady Andrienko and Natalia Andrienko — Fraunhofer IAIS, Sankt Augustin, Germany · City, University of London, UK
Currently under review in IEEE TVCG
The paper is currently under review in IEEE Transactions on Visualization and Computer Graphics (IEEE TVCG), 2026.
A preprint is available on arXiv; this site collects all material referenced in the appendices.
@misc{andrienko2026smartiterator,
title = {SmartIterator: Visual Analytics Workflows for Supervising
Unsupervised Data Grouping},
author = {Andrienko, Gennady and Andrienko, Natalia},
year = {2026},
note = {Under review in IEEE Transactions on Visualization and
Computer Graphics},
eprint = {2605.28219},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2605.28219}
}
Supplementary
Supplementary Appendices and Notebooks
Each appendix provides full-resolution figures and detailed analysis steps for one demonstration
from the paper. The corresponding Google Colab notebooks implement key parts of the computational pipeline
(parameter sweeps, quality metrics, archetype detection, embedding computation, word cloud generation) in
reproducible Python code.
Note: The notebooks implement the computational engine described in the paper (the section on computational components).
The IteraScope visual display is a separate Java application within the V-Analytics environment; the notebooks
produce the data files that IS consumes, and include standalone visualizations for exploration without IS.
Appendix A — Density-Based Clustering (VAST Challenge 2011)
Simulated microblog messages from the fictional city of Vastopolis; the analysis clusters the 59,761 outbreak-related messages. A contamination
event caused spatially concentrated health complaints that vary substantially in density—from tight hospital
clusters to elongated, diffuse wind- and water-dispersed patterns. The workflow sweeps DBSCAN parameters
(ε and min_samples) across three progressive rounds, using map and space-time cube views for domain
validation against the known event geography.
Highlights: Multi-round progressive refinement; transition-class propagation to isolate
borderline members; temporal layering revealed in space-time cube; validation against ground truth.
Appendix B — Partition-Based Clustering (EU NUTS-3 Population)
Demographic indicators (gender and age structure) for ~1,500 European NUTS-3 regions from Eurostat. The
workflow sweeps the number of K-means clusters (K = 3…50), compares candidate configurations through violin
plots and transition-class propagation, validates geographic coherence on choropleth maps, and verifies
robustness through a 30-seed stability sweep.
Highlights: Violin-based confidence comparison between K = 20 and K = 24; transition
decomposition revealing peripheral losses; seed-stability verification; parallel coordinates for demographic
profiling.
Appendix C — Topic Modeling (IEEE VIS Papers, NMF)
~3,800 IEEE VIS papers (1990–2024) with titles and abstracts. The workflow begins with data cleaning
(removal of encoding artifacts), sweeps the number of NMF topics (5–25), diagnoses a “wastebasket” topic at
ntopics = 16 through Sankey transition analysis, investigates the progressive differentiation of
the “time” topic, and validates the final 15-topic solution through 34-year temporal prevalence charts.
Highlights: data cleaning (a preliminary step) via artifact detection; wastebasket diagnosis through
multi-parent Sankey pattern; term-level transition word clouds; temporal prevalence alignment with known
community evolution (VAST track founding, growth of evaluation culture).
Note: The notebooks do not implement the full IteraScope interactive display (Sankey rendering, violin
plots, hover/click interactions, domain-view linking). These are part of the Java-based V-Analytics
environment. The notebooks produce the input data files for IS and provide standalone exploratory
visualizations.
For researchers
The Framework
SmartIterator (SI) is a visual analytics approach that treats the full sequence of
unsupervised grouping results across a parameter sweep as a first-class analytical object. Rather than seeking a
single “optimal” parameter configuration, SI guides the analyst through systematic exploration of how data
structure emerges, persists, and transforms across configurations—supporting cumulative understanding of data
structure along the way.
The approach is operationalized through IteraScope (IS), a coordinated visual display that combines:
A quality-metrics chart with semantic color encoding (warm = higher-is-better, cool = lower-is-better)
A 1D group embedding with Sankey-style transition flows showing splits, merges, and member movements
Violin plots of assignment confidence and outlier scores (globally normalized for cross-group comparison)
A 2D group embedding with HDBSCAN-detected recurrent patterns and interactive threshold control
Word clouds and term-level transition tooltips (for text corpora)
Three method-specific workflows are provided—for density-based clustering, partition-based clustering, and NMF
topic modeling—each following a common six-step structure:
Step
What happens
What you get
1 · Configure
Define the parameter ranges to be explored for the chosen method.
A sweep plan
2 · Iterate
Run the method once per configuration, keeping all intermediate results.
A family of groupings
3 · Evaluate
Compute internal quality measures per run and compare them across the sweep.
Metric profiles
4 · Compare
Track how group memberships change between neighbouring configurations.
Transition structure
5 · Interpret
Inspect content and context in domain-linked views; identify recurring structures and discard single-setting artefacts.
Defensible groupings
6 · Decide
Accept or refine: adopt the candidate configuration(s), or narrow the ranges and return to step 1.
A decision — or the next round
Step 1: Metric Overview & Sweep Configuration ↓ Step 2: Iteration over the Parameter Range ↓ Step 3: Quality-Metric Assessment ↓ Step 4: Transition Assessment (Sankey flows) & Confidence Evaluation (violin plots) ↓ Step 5: Content & Context Inspection (domain views) ↓ Step 6: Decide — accept or refine ⤻ back to Step 1
Key insight: Each step yields evidence about the data structure. In our demonstrations, studying the
transitions between configurations—which groups persist, which split, what vocabulary migrates, where boundaries
lie—surfaced observations beyond the final parameter choice. Whether this improves analysts’ understanding or
decisions has yet to be tested in a user study; the demonstrations illustrate how the workflows can be used.
Supported scenarios
Scenario
Method family
Key parameters
Demonstration dataset
DBC
Density-based clustering (DBSCAN)
ε, min_samples
VAST Challenge 2011 microblogs
PBC
Partition-based clustering (K-means)
number of clusters K, random seed
EU NUTS-3 population structure
TM
Topic modeling (NMF)
number of topics, vocabulary pruning
IEEE VIS papers 1990–2024
CD
Community detection
resolution γ, random seed
Additional worked example (demo only)
The paper discusses the three scenarios DBC, PBC and TM in detail; community detection is
included here as an additional worked example in the demonstrator.
Try it · learn it · teach it
Demonstrator and Educational Materials
Everything below is openly accessible and needs no account. The demonstrator runs entirely in the
browser; the student package is self-contained and runs locally or in Colab; the teaching material consists of
the fully executed notebook and separate instructor notes.
Why numbers differ between the implementations. The Java IteraScope, the Colab notebooks A–C, the browser demonstrator and the student
notebook are separate implementations. They use different libraries, random seeds, preprocessing and, in places, different definitions (for example how the
distance between two locations is computed). Results are therefore comparable in kind but not identical, and they are not expected to match exactly.
Online demo
Interactive demonstrator
Explore precomputed parameter sweeps in IteraScope: quality-metric profiles, Sankey transitions, confidence
violins, group embeddings and the linked domain views (maps, space-time cubes, profiles, calendars, word clouds) — for all example datasets,
including the cherry blossom data used in the student notebook. Explore… runs your own sweeps in the browser, and you can add your own data.
Density-based clustering uses the real (great-circle) distance in km. The caption of the demonstrator shows its release time, so you can check that you are not looking at a cached copy.
A guided practical in three parts: density-based clustering of ~9,700 geolocated cherry blossom photos
(real distance in km, space-time cubes, quality indicators),
K-means on weekly regional photo counts, and a closing exercise in auditing over-confident claims about a
single clustering run. Predict-then-reveal checkpoints and interactive sliders throughout.
Contains the notebook, both datasets, the base map, requirements.txt and a README
with local and Colab setup instructions.
The complete student notebook with every cell executed — all figures, tables and printed results in place —
so you can preview the expected outcome of each stage, project it in class, or recover from a failing
environment without running anything.
The instructor notes add learning objectives, suggested timing per stage, the findings students should
arrive at, typical misconceptions, discussion prompts and answers to the exercises.
Students used the notebook and could try the web demonstrator. Twenty of them answered a voluntary anonymous questionnaire. The sample is small and self-selected, and the summary reports counts only.
The following recommendations generalize from our experience applying SmartIterator across three
method families and diverse datasets. They are intended as actionable guidance for analysts beginning their own
parameter-supervision workflows.
Begin with a wide parameter sweep at coarse resolution (e.g., K = 3…50 with step 1, or ε = 0.05…1.0 with step
0.05) to identify the region of interest. Examine the metrics chart for elbows, peaks, and plateaus; note where
complete iterations cluster. Then narrow the range and reduce the step size for a second round. Two to three
rounds typically suffice—each informed by the findings of the previous one. This avoids wasting computation on
irrelevant regions and gives early orientation before committing to detailed inspection.
2Let Sankey Bands, Not Metrics Alone, Define Stability
A silhouette peak or coherence plateau narrows the candidate range, but the definitive stability evidence comes
from transition flows. A “good” configuration is one where the groups at the selected parameter persist with
minimal change at neighboring values—visible as wide, horizontal Sankey bands connecting the same vertical
positions across consecutive iterations. Specifically:
Persistent wide bands: The group is stable—it neither splits nor absorbs neighbors.
A band splitting into two: Potential meaningful sub-structure emerging at finer granularity. Inspect both children in domain views.
Multiple thin bands converging into one new group: Possible wastebasket—a group collecting heterogeneous members from multiple sources without a unifying identity.
A band thinning progressively: The group is dissolving under stricter parameters—its members are borderline.
If a group splits at the very next parameter step, it may be an artifact of the current granularity rather than
genuine structure.
3Use Domain Views to Break Metric Ties
When multiple configurations score similarly on quality metrics, propagate each to domain-linked views and
compare. Domain coherence frequently discriminates between statistically equivalent alternatives:
Spatial data: Do clusters form geographically contiguous regions, or are they scattered? Contiguity is a powerful external validator.
Temporal data: Do groups exhibit temporal coherence—emerging at specific times, persisting over defined periods?
Text data: Do word clouds reveal interpretable themes, or are top terms generic and overlapping?
Multivariate data: Do parallel-coordinates profiles show clearly distinct patterns, or do cluster profiles overlap substantially?
The domain view often reveals the “right” answer instantly when metrics are ambiguous.
4Inspect What Changes, Not Only What Persists
Transition-class propagation is most informative when applied to understand losses: which members
leave a group at the next parameter step, and where do they go? The spatial, temporal, or semantic distribution
of lost members reveals the nature of the transition:
Losses at cluster edges (spatially coherent): The finer configuration is trimming genuine boundary members—the core structure is preserved.
Losses scattered randomly: The configuration change may be over-fitting to noise.
Losses forming a coherent sub-group: A genuine sub-structure is emerging—consider whether the finer granularity captures meaningful differentiation.
Losses with distinctive vocabulary (topic modeling): A sub-theme is separating out—examine whether it merits its own topic.
This “what changes” perspective often generates more domain insight than examining static group composition at
a single configuration.
5Treat Archetype Completeness as Necessary, Not Sufficient
A “complete” iteration (containing representatives of all HDBSCAN-detected archetypes) captures the full
diversity of recurrent patterns—making it a strong candidate. However:
Completeness alone does not guarantee interpretability or confidence. Always verify through the transition, confidence and domain-inspection steps.
An iteration missing one archetype may still be preferable if domain context explains the absence (e.g., a noise archetype capturing only preprocessing artifacts, or a rare pattern that only appears under extreme parameter settings).
If no iteration is complete, consider whether the archetype threshold is set too aggressively. Lower the min_cluster_size to allow more archetypes—some of which may represent genuine but infrequent patterns.
Use the HDBSCAN threshold sweep chart to find a stability plateau where small changes do not alter the archetype count.
6Verify Seed Robustness for Partition-Based Methods
K-means and other partition-based methods depend on random initialization. Once a promising K is identified, run 20–30 seeds at that K and look at the score and at the groups:
Silhouette spread: a large spread across seeds suggests instability, but a small spread does not prove that the groups are stable.
Agreement between seeds: in the Sankey view, bands that run horizontally without fanning out mean that the same items stay together; bands that fan out mean that individual items change groups between seeds. Quantify the agreement if you can (for example with the adjusted Rand index between the partitions of different seeds).
2D embedding: tight archetype clusters show that the group profiles recur, even when single items change groups.
If one seed produces a structurally different result (visible as crossed Sankey bands), investigate whether it represents a legitimate alternative interpretation or an inferior local minimum.
Example (EU population data, K = 20, 30 seeds): the silhouette varied only between 0.18 and 0.19, and the group profiles recur, yet partitions of different seeds agreed only moderately (median adjusted Rand index 0.53). The result is stable in quality and in group profiles, but not in the membership of single regions.
For topic modeling (NMF), seed sensitivity is typically lower due to the non-negative constraint, but verification remains good practice.
7Document Rejections Explicitly
When an alternative configuration is rejected, record:
Which step provided the discriminating evidence (metrics? transitions? violins? domain views?).
What specific observation led to rejection (boundary ambiguity? wastebasket pattern? spatial fragmentation?).
What the rejection reveals about data structure (e.g., “K=24 rejected because the additional clusters shave off geographic periphery without introducing new demographic archetypes—confirming that Europe's demographic typology is adequately captured at K=20”).
Documented rejections serve two purposes: they strengthen confidence in the final selection (by demonstrating
that alternatives were considered and found wanting), and they contribute to cumulative understanding (by
revealing what the data does not support).
8When in Doubt, Keep Multiple Candidates
The SI workflow does not require converging to a single “winner.” If two configurations offer complementary
views of the data—e.g., a coarse typology (K=7) and a fine-grained one (K=20)—both may be worth retaining as
analytical lenses for different purposes. The knowledge gained from understanding how the coarse groups
differentiate into fine-grained ones is itself a valuable analytical product.