Zoomable Multi-Level Tree (ZMLT) Overview

• ZMLT is a hierarchical, multi-resolution tree structure that encodes data at varying abstraction levels for smooth semantic zoom.
• The approach uses tailored algorithms such as agglomerative clustering, incremental construction, and Steiner tree extraction to maintain interactivity and scalability.
• ZMLT implementations, like DendroMap, HETree, and Graph ZMLT, provide domain-specific layouts and statistics, enhancing visualization and analytical insight across diverse data types.

A Zoomable Multi-Level Tree (ZMLT) is a hierarchical, rooted, ordered structure designed to support interactive, scalable visualization and exploration of large, complex datasets. ZMLTs deliver semantic zoom functionality by encoding data at multiple levels of abstraction, such that users can smoothly traverse from coarse overviews to fine local detail. ZMLTs have been instantiated for diverse data modalities including multivariate tables, large-scale image corpora, and massive graphs. Three signature implementations—DendroMap for embedding-based image collections, HETree for generic numeric and temporal data, and De Luca et al.’s ZMLT for graph visualization—illustrate the unified structural principles and the domain-specific extensions characteristic of the approach (Bertucci et al., 2022, Bikakis et al., 2015, Luca et al., 2019).

1. Mathematical Formulation and Hierarchical Structure

ZMLTs generalize hierarchical aggregation to a tree $T=(V,E)$ with the following core formal properties, with parametrizations that vary by application:

• In DendroMap (Bertucci et al., 2022), given embeddings $X=\{x_1, ..., x_n\}\subseteq \mathbb{R}^d$, $T$ is a full binary dendrogram whose leaves $\ell_i$ correspond to individual data points and whose internal nodes represent the union of clusters formed during agglomerative clustering. Tree depth directly encodes abstraction level.
• In HETree (Bikakis et al., 2015), for sorted data $S[1\,..\,N]$ and user-specified degree $d$ and leaf count $\ell$, the tree is constructed as a rooted $d$-ary tree of height $h=\lceil \log_d \ell\rceil$ where all leaves reside at the same depth. Each internal node $n$ is associated with the interval $[I^-(n), I^+(n)]$ (for numeric data) covering its descendants. Specializations for content-based and range-based splits are supported.
• In the graph-visualization context (Luca et al., 2019), level-$i$ tree $T_i=(V_i, E_i)$ is a (multi-level) node-weighted Steiner tree—a subgraph of the original $G=(V,E)$ spanning the $k_i$ most "important" nodes with minimum-completion cost (where node importance is user-defined or grounded in objective metrics). Nested property holds: $T_1\subset T_2\subset\cdots\subset T_n=G$.

This hierarchical structure ensures that at each abstraction level, the ZMLT encodes a coherent, disjoint partition of the data, supporting both efficient data summarization and information-preserving navigation.

2. Construction Algorithms and Adaptations

The tree construction pipeline depends on data type and user objectives:

• DendroMap: Implements bottom-up agglomerative clustering with Ward’s linkage and Euclidean metric, initialized as singletons for each image. Pairs of clusters $(A,B)$ are merged minimizing Ward’s criterion:

$$d_{\text{Ward}}(A,B) = \frac{|A|\cdot|B|}{|A|+|B|}\|\mu_A-\mu_B\|_2^2,$$

yielding a full binary tree. The selection of $k$ top-level clusters for zoom levels is performed breadth-first, maintaining the user’s requested granularity (Bertucci et al., 2022).

• HETree: Supports all-at-once and incremental construction strategies (ICO). For range-based ZMLT-R, sorted data is partitioned into $\ell$ equi-width or equi-content leaves, then recursively grouped into internal nodes of arity $d$. ICO enables on-the-fly node expansion during user zoom, prefetching necessary subtrees and minimizing in-memory and computational cost (Bikakis et al., 2015).
• Graph ZMLT: Relies on node-weight filtration and polynomial-time approximation algorithms for node-weighted Steiner tree extraction at multiple resolutions. Each level is explicitly constructed, with monotonic level nesting, enforcing the property that no node or edge is dropped as zoom proceeds deeper (Luca et al., 2019).

All algorithms annotate tree nodes with summary statistics or cluster sizes, enabling efficient layout and interaction.

3. Layout, Visualization, and Zoom Interactions

Domain-specific layout methodologies realize the spatial and semantic arrangement of tree nodes:

• Treemap Layout (DendroMap): Utilizes slice-dice treemap partitioning recursively. Given a rectangle $R_v$ for node $v$, children receive subrectangles proportionally to their cluster size and divided horizontally or vertically as dictated by aspect ratio. Up to $k$ clusters are shown at overview; clicking a cluster zooms to view its descendants, with animated transitions interpolated for smoothness. Only visible nodes are drawn, maintaining $O(k)$ rendering complexity (Bertucci et al., 2022).
• Multi-Level Numeric Data (HETree): Uses parent/child positional relationships to support table, treemap, and statistical visualizations. ICO prefetches siblings and parents during zoom-in (drill-down) and zoom-out (roll-up), so response times remain sub-interactive even for $N\sim10^5$ (Bikakis et al., 2015).
• Graph ZMLT: Employs a force-directed optimization (ImPred) combining spring attraction, node-node repulsion, node-edge repulsion, and label-label repulsion. Edge lengths are scaled by level, and newly revealed subtrees are inserted into angular gaps around persistent nodes, preserving planarity and label non-overlap. Per-level layouts are precomputed offline and switched interactively in the front end (OpenLayers/GeoJSON), achieving seamless semantic zoom (Luca et al., 2019).

The common principle is the synchronization of zoom level with tree hierarchy depth, affording users continuous, information-preserving transitions.

4. Statistical Aggregation and Metadata Propagation

ZMLT nodes capture aggregate statistics over their covered data, enabling fast summary queries and on-demand analytics:

• HETree Aggregates: Each node $n$ maintains count $N(n)$, sum $\Sigma(n)$, mean $\mu(n)$, variance $\sigma^2(n)$, minimum, and maximum, computed recursively from children. Efficient propagation rules (including for adaptive hierarchy reshaping) ensure minimal recomputation during zoom or dynamic degree/leaf modifications (Bikakis et al., 2015).
• DendroMap: Nodes are annotated with cluster sizes $N_v=|C_v|$. For display, systematic sampling strategies select representative images from clusters exceeding grid capacity, ensuring diversity in the visualization (Bertucci et al., 2022).
• Graph ZMLT: No synthetic meta-nodes are introduced; labels and vertices correspond to actual nodes, with font size and presence determined by level. Compactness (label area to canvas area ratio), planarity, label overlap, and stress (deviation of embedded from graph-theoretic distances) are tracked as formal quality metrics (Luca et al., 2019).

5. Complexity, Performance, and Scalability

Computational performance is governed by clustering, layout, and interactive rendering phases:

• Clustering: Ward’s method in DendroMap yields $O(n^2)$ complexity for $n$ images, taking $<1$ min for $n=10,000$ (CIFAR-10) on commodity hardware (Bertucci et al., 2022).
• Tree Construction: HETree achieves $O(|D|\log|D| + d^2 \ell/(d-1))$ offline construction, but in practice uses $O(N_\text{zoom} \cdot (d^2+\max_\text{leaf}))$ with ICO since only visible/recently visited subtrees are materialized (Bikakis et al., 2015).
• Graph Layout: Polynomial-time for Steiner tree extraction, $O(|V|^3)$ worst-case, but empirically manageable for sparse graphs at the 5K–10K node scale (Luca et al., 2019).
• Interactive Rendering: All implementations decouple offline construction from real-time navigation. At view-switching, time and memory cost is proportional to the number of visible nodes or clusters, independent of total tree size, enabling interactive performance even for very large datasets.

6. Comparison of ZMLT Variants and Properties

Key ZMLT properties, as established across these systems, are summarized:

Property	DendroMap (Bertucci et al., 2022)	HETree (Bikakis et al., 2015)	Graph ZMLT (Luca et al., 2019)
Branching Factor	Strictly binary	User-param d-ary	Varies per Steiner tree
Adaptivity	Dynamic k, static tree	Dynamic ($d,\ell$)	Precomputed thresholds
Zoom↔Hierarchy Coupling	Tree depth	Level, leaf count	Sequence of trees
Aggregation	Cluster size	Rich numeric stats	None (graph vertices)
Layout	Treemap (slice-dice)	Table/treemap/graph	Planar force-directed
Representative Nodes/Edges	Yes	Yes	Only real graph nodes
Label/Aspect Optimization	Yes, minimal skew	N/A	Overlap-free, planar

All approaches ensure seamless semantic zoom: a property set comprising monotonicity (nodes/edges persist across zoom), planarity (for graphs), overlap-free labeling, and compactness of drawing area.

7. Evaluations, Applications, and Empirical Analysis

Empirical measurements confirm the effectiveness and scalability of ZMLT-based methods:

• DendroMap: Outperforms gridified t-SNE for user-comprehensibility on image grouping/search tasks. Supports discovery of dataset structure, identification of subgroup outliers, and integration with ML model accuracy overlays. Zooming remains responsive across zoom levels $k=4$–20, clusters with $10$–$100$ images each (Bertucci et al., 2022).
• HETree/SynopsViz: Demonstrates interactive exploration, rich per-group analytics (mean, variance), and adaptive redesign of hierarchy (on-demand changes to arity or overview coarseness), even on hundreds of thousands of records (Bikakis et al., 2015).
• Graph ZMLT: On “GRAM Topics” graph with $5947$ nodes, stress/compactness metrics substantially outperform direct circular layouts: e.g., at deepest level, stress is reduced from $5.1\times10^6$ (Direct) to $2.3\times10^6$ (ZMLT), and compactness from $1740$ to $287$, with tight tracking of decreasing edge lengths by zoom level (Luca et al., 2019).

Applications extend to machine learning dataset inspection, time-series and attribute-driven exploration, and large-scale knowledge graph visualization.

---

ZMLTs provide a unifying framework for interactive, information-preserving, multi-resolution exploration across heterogeneous data types, validated by a spectrum of instantiations with formal guarantees on visual consistency, scalability, and analytic expressiveness (Bertucci et al., 2022, Bikakis et al., 2015, Luca et al., 2019).