Home · Viewer · How It Works · Comparison · Developer Guide · WebGPU · WebGL · GitHub

Layout Comparison

BlitZoom vs ForceAtlas2, UMAP, and t-SNE on 8 graph datasets.
Full report with methodology and all configurations.

BlitZoom positions nodes by property similarity, using topology as a secondary signal. This comparison measures both topology preservation (do connected nodes end up nearby?) and property-similarity preservation (do nodes with similar attributes end up nearby?) across 8 datasets against ForceAtlas2, UMAP, and t-SNE.

The results: on datasets where property structure diverges from graph connectivity, BlitZoom leads on property grouping by 2-6×. When adjacency correlates with properties, topology-based methods capture property structure incidentally. BlitZoom's auto-tune produces competitive layouts without manual configuration, at millisecond speed.

Which method should I use?

Method	Best when	Tradeoff
BlitZoom	Property similarity matters more than connectivity. Interactive exploration, zoom levels, millisecond speed.	Weaker on sparse chain topologies.
ForceAtlas2	Topology IS the signal. Social networks, connectivity analysis, chain-like graphs.	Slower (minutes). No property awareness.
UMAP	Dimensionality reduction with interpretable embeddings. High-dimensional feature data.	Seconds, not milliseconds. Seed-dependent.
t-SNE	Revealing fine-grained local clusters. Best local topology preservation.	Requires perplexity tuning. Slow.

What we measured

EdgeLen: How close are connected nodes? Normalized Euclidean distance between edge endpoints. Lower = tighter topology.
TopoNbrP: Do a node's graph neighbors end up as its layout neighbors? Jaccard overlap at k=10. Higher = better topology fidelity.
PropNbrP: Do a node's most property-similar nodes end up as its layout neighbors? Jaccard overlap at k=10. Higher = better property grouping. Measured using per-node token sets (the same tokens BlitZoom's MinHash operates on).
Silhouette: Do known communities form visible clusters? Score vs ground-truth labels, -1 to +1. Only available for Email-EU (42 departments).

We use k=10 neighbors as a balance between local structure fidelity and global layout quality.

Not all methods run on all datasets. UMAP requires computing pairwise Jaccard on dense n×n adjacency matrices. At 4K+ nodes this takes too long, so UMAP is only included for smaller datasets (Epstein, Pokemon, Email-EU, BZ Source). ForceAtlas2 and t-SNE run on all datasets but take 3–87 seconds vs BlitZoom's 1–5 milliseconds. Where a method is absent from a table, it was too slow to include.

The short version

On topology metrics, ForceAtlas2 and t-SNE lead. On property-similarity, the answer depends on whether properties correlate with graph connectivity. When they don't (MITRE, Synth Packages), BlitZoom with property strengths leads by 2.6-6x. When they do (BlitZoom Source, where call edges track file structure), topology-based methods capture property structure incidentally and score higher. BlitZoom's auto-tune (zero-config) beats or matches baselines on most property datasets.

Dataset	Nodes	Props	Best BlitZoom	PropNbrP	Auto-tune	Takeaway
Epstein	514	edge types	α=0 wt	1.44x	1.42x	Autotune discovers edgetype; matches hand-tuned
Pokemon	959	multi	autotune	1.13x	1.13x	Autotune beats hand-tuned and FA2
MITRE	4,736	rich	α=0 wt	2.42x	0.91x	Properties ≠ topology; hand-tuned leads
Synth Pkg	1,868	rich	α=0 wt	5.86x	2.62x	Properties ≠ topology; BlitZoom leads
BZ Source	917	rich	α=0.5 wt	0.84x	0.56x	Properties ≈ topology; topology-based methods lead
Email-EU	1,005	none	α=1.0	0.80x	—	Edge-only; topology comparable
Facebook	4,039	none	α=1.0	0.70x	—	Dense; smoothing works well
Power Grid	4,941	none	α=0.75	0.58x	—	Sparse chains; smoothing limited

PropNbrP ratios are BlitZoom / ForceAtlas2. Auto-tune column shows autotune / FA2. Higher = better. Edge-only datasets have no autotune (no property signal).

Dataset by dataset

Epstein: edge-type properties

514 nodes (people, organizations), 494 edges with rich edge types (ABUSED, ASSOCIATED_WITH, BUSINESS_RIVAL, etc.). Properties: group + edge types.

Layout	Time	EdgeLen	TopoNbrP	PropNbrP	Note
BlitZoom α=0	1ms	0.510	0.005	0.035	No property strengths
BlitZoom α=0 wt	1ms	0.456	0.004	0.118	group:5, edgetype:8 — best PropNbrP
BlitZoom autotune	1ms	0.277	0.013	0.116	group:8, edgetype:3, α=0.5 — discovers edgetype
ForceAtlas2	3s	0.045	0.099	0.082	Topology-driven
UMAP	4s	0.376	0.004	0.082	Similar to FA2
t-SNE	0.4s	0.428	0.006	0.090	Third-highest PropNbrP

Hand-tuned edgetype:8 scores 44% higher than ForceAtlas2 on PropNbrP (0.118 vs 0.082). The dual-pass autotune at 0.116 nearly matches — it discovers edgetype:3 through its α=0.5 pass, where edge-type structure becomes visible via topology smoothing. This is the key insight of dual-pass: some properties only show their value when combined with moderate topology.

Pokemon: multi-property, property-dominant

959 Pokemon with 4+ property groups (type1, type2, generation, rarity, stats). Edges exist but properties dominate the similarity structure.

Layout	Time	EdgeLen	TopoNbrP	PropNbrP	Note
BlitZoom α=0	1ms	0.500	0.004	0.010	No property strengths
BlitZoom α=0 wt	1ms	0.476	0.013	0.019	Hand-tuned: type1:8, type2:4, gen:3, rarity:2
BlitZoom autotune	1ms	0.373	0.009	0.027	generation:8, α=0.15 — beats hand-tuned
ForceAtlas2	17s	0.024	0.063	0.023	Strong topology
UMAP	5s	0.073	0.012	0.022	Similar to hand-tuned
t-SNE	2s	0.182	0.011	0.017	Lowest baseline PropNbrP

Autotune outperforms the hand-tuned preset by 41% (0.027 vs 0.019) and beats ForceAtlas2 by 18% (0.027 vs 0.023). It correctly identified that generation alone with a touch of topology (α=0.15) is more discriminative than the multi-property hand-pick. Types spread across generations; generation groups tightly.

Email-EU: a dense communication network

1,005 researchers at a European institution, 16.7K emails, 42 departments as ground truth. Edge-only (no node properties beyond auto-generated tokens).

Layout	Time	EdgeLen	TopoNbrP	PropNbrP	Silhouette	Note
BlitZoom α=0	1ms	0.466	0.006	0.007	-0.47	Near-random layout
BlitZoom α=1.0	1ms	0.221	0.056	0.007	-0.29	Best BlitZoom for topology
ForceAtlas2	19s	0.009	0.064	0.009	-0.38	Shortest edges
UMAP	5s	0.182	0.107	0.010	+0.01	Only positive silhouette
t-SNE	1s	0.161	0.108	0.009	-0.11	Highest TopoNbrP

PropNbrP is uniformly low (0.007-0.010) across all methods. Without real node properties, auto-generated tokens provide little differentiation. UMAP and t-SNE recover department structure best on topology metrics.

Facebook: dense ego networks

4,039 users, 88K friendship edges. Dense community structure.

Layout	Time	EdgeLen	TopoNbrP	PropNbrP	Note
BlitZoom α=1.0	3ms	0.063	0.110	0.003	74% of FA2's TopoNbrP
ForceAtlas2	65s	0.011	0.151	0.004	Shortest edges
t-SNE	5s	0.072	0.176	0.003	Highest TopoNbrP

Dense ego-network structure responds well to topology smoothing. BlitZoom at α=1.0 reaches 74% of ForceAtlas2's TopoNbrP. PropNbrP is uniformly low (edge-only dataset).

Power Grid: sparse chains

4,941 substations, 6.6K transmission lines. Average degree 2.7, diameter ~46.

Layout	Time	EdgeLen	TopoNbrP	PropNbrP	Note
BlitZoom α=0.75	4ms	0.271	0.003	0.002	Best BlitZoom; α=1.0 is worse
ForceAtlas2	61s	0.006	0.195	0.002	Traces chains via global forces
t-SNE	10s	0.171	0.041	0.002	Limited by sparse adjacency

ForceAtlas2 dominates. Its global optimization traces long chains (diameter ~46) that 5-pass local smoothing cannot reach. α=0.75 outperforms α=1.0 because pure topology with few passes oversmooths hubs while leaving chains unresolved.

MITRE ATT&CK: the property-similarity test

4,736 nodes (techniques, tactics, software, mitigations) with rich properties: platforms, kill chain phases, aliases. This is the dataset that tests BlitZoom's core claim.

Layout	Time	EdgeLen	TopoNbrP	PropNbrP	Note
BlitZoom α=0	5ms	0.461	0.001	0.007	No property strengths
BlitZoom α=0 wt	4ms	0.534	0.002	0.034	Property strengths: best PropNbrP
BlitZoom autotune	4ms	0.486	0.002	0.013	killchain:8, α=0.5 — near FA2
ForceAtlas2	87s	0.181	0.005	0.014	Shorter edges; lower PropNbrP
t-SNE	12s	0.298	0.004	0.026	Second-highest PropNbrP

BlitZoom with property strengths (group=5, platforms=6, killchain=4) scores 2.4x higher than ForceAtlas2 on PropNbrP (0.034 vs 0.014). Autotune finds killchain:8 at α=0.5, scoring 0.013 — competitive with FA2 but below hand-tuned. The multi-group interaction (group + platforms + killchain) that makes the hand-pick work is not discoverable from coordinate descent alone. Without property strengths, BlitZoom's PropNbrP drops to 0.007.

Synth Packages: designed group structure

1,868 synthetic packages, 4K co-reference edges. Properties: group, downloads, license, version, depcount. Edges are co-reference links, not property-based.

Layout	Time	EdgeLen	TopoNbrP	PropNbrP	Note
BlitZoom α=0	1ms	0.468	0.003	0.010	No property strengths
BlitZoom α=0 wt	2ms	0.387	0.005	0.050	Property strengths: best PropNbrP (6x FA2)
BlitZoom autotune	1ms	0.535	0.002	0.022	downloads:8 — 2.6x above FA2
ForceAtlas2	54s	0.028	0.018	0.009	Shortest edges; low PropNbrP
t-SNE	19s	0.222	0.001	0.013	Low on both metrics

BlitZoom with strengths scores 5.9x higher than ForceAtlas2 on PropNbrP (0.050 vs 0.009). Autotune at 0.022 still beats FA2 by 2.6x. Graph connectivity and property similarity diverge: edges are co-reference links, not property-based. The autotune found that downloads alone is the most discriminative property for this dataset.

BlitZoom Source: when topology tracks properties

916 nodes (functions, methods, classes, imports) from this project's source code, 2,096 call edges. Properties: kind, file, lines, bytes, age, edge types.

Layout	Time	EdgeLen	TopoNbrP	PropNbrP	Note
BlitZoom α=0 wt	1ms	0.433	0.012	0.236	kind:8, group:3, rank quant
BlitZoom α=0.5 wt	1ms	0.308	0.017	0.244	Topology helps slightly
BlitZoom autotune	1ms	0.297	0.013	0.163	kind:3, edgetype:8 — finds edgetype
ForceAtlas2	11s	0.018	0.110	0.290	High PropNbrP via adjacency
UMAP	7s	0.105	0.056	0.303	Highest PropNbrP
t-SNE	3s	0.208	0.035	0.249	High PropNbrP via adjacency

Topology-based methods score higher on PropNbrP (0.25-0.30) because call-graph edges track file/kind similarity. BlitZoom hand-tuned at 0.244 reaches 84% of FA2. Autotune at 0.163 found edgetype:8 as a useful dimension but the hand-tuned multi-group combination (kind + group) scores better. On this dataset, topology correlation means all methods score well on PropNbrP.

What each method is good at

Aspect	ForceAtlas2	UMAP / t-SNE	BlitZoom
Edge length	Best (optimizes for this)	Moderate	Improves with α
Topology preservation	Strong; global forces	Best overall (t-SNE)	Comparable on dense graphs
Property grouping	Incidental; depends on adjacency	Moderate (via adjacency)	Best when props ≠ topology
Sparse / chain graphs	Strong (global forces)	Limited	Limited (local smoothing)
Speed	Minutes (O(n log n)/iter)	Seconds	Milliseconds (O(n))
Hierarchical zoom	No	No	14 levels from 4 bytes/node
Determinism	Seed-dependent	Seed-dependent	Fully deterministic

Auto-tune: zero-config performance

BlitZoom's auto-tune uses a dual-pass optimizer: it searches the strength parameter space at both α=0 (property-only) and α=0.5 (moderate topology), then picks whichever produces the better layout. This discovers two kinds of useful properties: those that cluster independently (generation, group) and those that only show value when topology connects related nodes (edgetype, platforms). Bearing auto-tune then rotates each group to maximize 2D spread via closed-form trace maximization.

Dataset	Hand-tuned	Autotune	Best baseline	Autotune vs baseline
Epstein	0.118	0.116	0.090 (t-SNE)	1.28× — discovers edgetype via dual-pass
Pokemon	0.019	0.025	0.022 (FA2)	1.13× — beats all
MITRE	0.034	0.013	0.026 (t-SNE)	0.49× — multi-group interaction not discoverable
Synth Pkg	0.050	0.022	0.013 (t-SNE)	1.76× — beats all
BZ Source	0.244	0.163	0.303 (UMAP)	0.54× — topology-dominant dataset

Autotune beats all baselines on 3 of 5 property datasets (Epstein, Pokemon, Synth Packages) while being 1,000-70,000× faster. Epstein is the dual-pass showcase: the α=0.5 pass discovers edgetype:3 (invisible at α=0), reaching 98% of the hand-tuned PropNbrP. MITRE remains the weak spot — the optimal three-group combination (group + platforms + killchain) isn't discoverable from coordinate descent alone. Domain experts can still do 2-3× better on datasets where the optimal property mix requires multi-way interactions.

Rank vs Gaussian quantization

BlitZoom supports two quantization modes. Rank quantization sorts nodes by position and assigns grid cells uniformly. Gaussian quantization uses fixed CDF boundaries, preserving density structure: tight clusters stay tight, sparse regions stay spread out.

Caveats

PropNbrP uses the same similarity BlitZoom optimizes. Token-set Jaccard is both the ground-truth similarity and the signal BlitZoom's MinHash approximates. This is partially circular. A fully independent property-similarity metric (e.g., domain-expert labels) would be stronger evidence.

Edge-only datasets show no property differentiation. Three of eight datasets have no real node properties. PropNbrP is uniformly low across all methods on these graphs.

When adjacency correlates with properties, topology-based methods win PropNbrP too. On BlitZoom Source, ForceAtlas2 and UMAP outscore BlitZoom on PropNbrP because call-graph edges track file/kind similarity. BlitZoom's advantage is specific to datasets where property structure differs from graph connectivity.

Hierarchical zoom is not measured. BlitZoom derives 14 aggregation levels from 4 stored bytes per node. No other method produces a zoom hierarchy. This capability is not captured by any metric above.