ATWL Workflow Library Explorer (prototype)

sequential (forward) step, summed across the workflows shown above genuine loop repetition, summed across the workflows shown above

Rebuilt live, in the browser, from the same per-workflow parsed data used everywhere else in this tool (i.e. reflecting every regex/parsing fix made along the way — there is no separate, potentially-stale precomputed file behind this view). Reflects the same filter currently applied to the table below (search, topic, or Faceted Search) — narrow that filter to aggregate over just a subset. The minimum edge count is a percentage of however many workflows are currently shown (total or filtered), not a fixed number, so the threshold stays meaningful as the library grows; lower it to see rarer transitions.

outside any loop, summed across the workflows shown above inside a loop body, summed across the workflows shown above

The complementary, artifact-centred view: nodes are the 8 ATWL artifact categories (ordered left-to-right by an empirically-derived "abstraction pyramid" — entities in, knowledge out — see the per-workflow Artifact categories view for details), edges are input→output category transitions. Also reflects the current table filter, computed live the same way as the process map above.

Workflow	Diagram	Year	Venue	Nodes	Edges	Loops	Depth	Ifs	CFC	Human%	Machine%	Topic

Design note — "read from the web, or from our parsed representation?" This tool fetches every workflow directly from this folder at load time — manifest.json lists the case-study pages, and each page's own embedded ATWL source is parsed live in the browser (see stubWithStructure()/parseATWLBasic() below), rather than reading a frozen, previously-parsed dataset. Every graph measure, encoding, facet, and diagram is computed live from that same parsed source for every workflow in the library, so a future fix to the parser or the mining algorithms propagates everywhere automatically instead of silently drifting from what's actually on the page. The trade-off is the one this design accepted from the start: this file is no longer usable offline from a bare file:// URL (it needs to be served over http(s) alongside manifest.json and the case-study pages) — reliability and reproducibility come from the live parse having been checked, during development, against a hand-built reference set covering the original corpus, not from freezing a snapshot of it.

1 Vocabulary & encoding selection

Choose which token families go into the document–term matrix. Built live from this tool's own parsed data (no separate precomputed file). Histograms show each family's per-workflow token-count distribution across the whole library.

Minimum document frequency (drop terms appearing in fewer than N workflows)2

Maximum document frequency (drop near-universal, non-discriminative terms)38

—

2 NMF ensemble parameters

TF–IDF weighting raw term frequency Ensemble size: random initialisations

3 Topic interpretation

The full term×topic weight matrix across every topic at once, followed by one card per topic with a suggested interpretation and a sortable table of every vocabulary term that occurs at least once among its dominant workflows (weight and frequency side by side, so you can compare them directly rather than needing a separate note about how they diverge).

4 Pattern analysis & validation

Topics in term-weight space (PCA)

Workflows in topic-weight space (PCA)

Colour points by:

Each point is one workflow, positioned by projecting its topic-weight vector (W row, K=? dimensions) onto its top 2 principal components. Coefficients above are in topic space, not term space — compare with the term-space PCA on the left.

1 Feature selection

Built from the same calibrated facets used in Faceted Search (not Topic Modeling's token vocabulary — PCA needs comparable-scale binary/continuous features, not TF-IDF term frequencies).

—

2 Variance explained

3 Scatter (PC1 × PC2) and parallel coordinates (all 10 PCs)

Drag a rectangle on the scatter, or a vertical range on any parallel-coordinates axis, to select workflows. Multiple brushes combine (AND — narrows further, same convention as Faceted Search).

Colour points by:

4 Component loadings

Top features by |loading| per component — heuristic labels only, not authored interpretation (see the methods note on why: a confident narrative can make an unreliable analysis look trustworthy).

1 Feature selection

Same calibrated facet groups and z-scoring as the PCA Filter tab (kept as an independent selection here, since this view asks a different question — pairwise structural similarity rather than a global projection). Similarity between two workflows is the cosine similarity of their z-scored feature vectors over whichever groups are checked below.

—

ATWL Workflow Library Explorer — prototype v1

1 Vocabulary & encoding selection

2 NMF ensemble parameters

3 Topic interpretation

4 Pattern analysis & validation

Topics in term-weight space (PCA)

Workflows in topic-weight space (PCA)

1 Feature selection

2 Variance explained

3 Scatter (PC1 × PC2) and parallel coordinates (all 10 PCs)

4 Component loadings

1 Feature selection

2 Similarity network and reordered matrix

Overview — corpus-wide process map (all workflows, aggregated)

Overview — artifact category flow (all workflows, aggregated)

📊 Corpus insights (per-group stats, discriminating power, co-occurrence — computed live from this tool's own facet data)

1 Vocabulary & encoding selection

2 NMF ensemble parameters

3 Topic interpretation

4 Pattern analysis & validation

Topics in term-weight space (PCA)

Workflows in topic-weight space (PCA)

1 Feature selection

2 Variance explained

3 Scatter (PC1 × PC2) and parallel coordinates (all 10 PCs)

4 Component loadings

1 Feature selection

2 Similarity network and reordered matrix