Improving the Projection of Global Structures in Data through Spanning Trees

The connection of edges in a graph generates a structure that is independent of a coordinate system. This visual metaphor allows creating a more flexible representation of data than a two-dimensional scatterplot. In this work, we present STAD (Spanning Trees as Approximation of Data), a dimensionality reduction method to approximate the high-dimensional structure into a graph with or without formulating prior hypotheses. STAD generates an abstract representation of high-dimensional data by giving each data point a location in a graph which preserves the distances in the original high-dimensional space. The STAD graph is built upon the Minimum Spanning Tree (MST) to which new edges are added until the correlation between the distances from the graph and the original dataset is maximized. Additionally, STAD supports the inclusion of additional functions to focus the exploration and allow the analysis of data from new perspectives, emphasizing traits in data which otherwise would remain hidden. We demonstrate the effectiveness of our method by applying it to two real-world datasets: traffic density in Barcelona and temporal measurements of air quality in Castile and Le'on in Spain.

💡 Research Summary

The paper introduces STAD (Spanning Trees as Approximation of Data), a novel dimensionality‑reduction technique that represents high‑dimensional data as a graph whose structure is independent of any coordinate system. Starting from the full pairwise distance matrix Dₓ of an n‑point dataset, the method builds a complete weighted graph Gₓ and extracts its Minimum Spanning Tree (MST). The MST is then transformed into a unit‑distance graph U₀ where every edge has length 1, so that the distance between two vertices is measured by the length of the shortest path (i.e., the number of edges) rather than by the original Euclidean weight.

All remaining edges of Gₓ are sorted by their original weight and added one by one to U₀, producing a sequence of graphs U₁, U₂, …, Uₙ. After each addition, the shortest‑path distance matrix Dᵤᵢ of the current graph is computed and its Pearson correlation with the original distance matrix Dₓ is evaluated. Because correlation is invariant to scaling, a value close to 1 indicates that the graph distances faithfully reproduce the original geometry. The algorithm automatically selects the iteration i* that yields the maximal correlation, thereby determining the optimal number of extra edges without any user‑defined parameters.

Key innovations of STAD include:

1. Parameter‑free edge selection – the optimal graph density is discovered by maximizing correlation, eliminating the need for heuristic thresholds (e.g., ε‑neighborhoods in Isomap or perplexity in t‑SNE).
2. Preservation of global distances – unlike t‑SNE, UMAP, or LargeVis, which prioritize local neighborhoods, STAD explicitly optimizes the agreement of all pairwise distances, offering a more faithful view of the dataset’s overall shape.
3. Point‑level resolution – each original observation remains a distinct vertex, avoiding the aggregation typical of many Topological Data Analysis (TDA) methods such as Mapper or Reeb graphs.
4. Extensibility through filters – users can inject domain‑specific functions that modify edge addition or weighting (e.g., time‑based filters, density‑based emphasis), enabling focused exploratory analyses.

The authors compare STAD with classic multidimensional scaling (MDS), Sammon mapping, Isomap, and modern non‑linear embeddings, highlighting that STAD’s graph‑based representation provides a flexible visual metaphor: relationships are encoded by connectivity rather than by absolute coordinates.

Two real‑world case studies demonstrate the method’s practicality. In the Barcelona traffic‑density dataset, STAD reveals a high‑density subgraph corresponding to major roads and congested zones, while preserving the continuity of traffic flow as path lengths. In the Castile‑León air‑quality time series, seasonal and diurnal patterns emerge as distinct clusters and edge‑weight variations, and anomalous pollution events appear as abrupt changes in graph connectivity. Compared with t‑SNE and UMAP visualizations, STAD simultaneously captures global trends and local density variations, offering a more comprehensive overview.

Computationally, constructing the full distance matrix requires O(n²) time and memory, which can be prohibitive for very large n. However, the subsequent MST computation, edge sorting, and incremental updates run in O(n log n) time, making the approach feasible for moderate‑size datasets. The authors suggest possible extensions such as approximate nearest‑neighbor distances, sampling strategies, or parallel MST algorithms to scale to larger data.

In discussion, the paper acknowledges the trade‑off between exact distance preservation and graph sparsity, noting that the optimal correlation often plateaus after a relatively small number of added edges, reflecting a persistence‑like behavior. The flexibility of adding filters is praised as a way to tailor the graph to specific analytical questions, but the authors caution that inappropriate filters may distort the underlying geometry.

The conclusion positions STAD as a bridge between traditional distance‑preserving embeddings and topological summarization techniques. It provides a high‑resolution, coordinate‑free visualization that retains global structure while allowing domain‑specific emphasis through filters. Future work will explore interactive implementations, dynamic filter design, and integration with visual analytics platforms to support large‑scale, exploratory data analysis.