Geometric Visibility Graphs for Time Series

31 concepts23 papers

Time-series analysis has traditionally relied on two domains: the time domain (analyzing sequential steps) and the frequency domain (uncovering underlying periodicities). However, when dealing with highly non-linear, chaotic, or noisy signals, these traditional perspectives can struggle to capture structural patterns and long-range dependencies.

A powerful alternative approach transforms sequential time series into complex networks using geometric visibility rules. By treating a time series as a landscape of peaks and valleys, this methodology maps temporal data points to graph nodes and establishes edges based on direct lines of sight. This geometric transformation allows researchers to leverage the rich toolkit of graph theory and network topology to analyze temporal dynamics.

The Core Intuition: Time Series as Landscapes

To understand how a visibility graph works, imagine a 1D time series plotted as a sequence of vertical bars, much like a bar chart. Now, imagine standing on top of one of these bars and looking across the landscape of the data. Which other bars can you see directly?

      * (Node A)
     / \
    /   * (Node B)
   /   / \
  *   *   * (Node C)
---------------------
 t1  t2  t3

If a straight line can be drawn between the peak of bar $A$ and the peak of bar $B$ without intersecting any intermediate bar, an edge is created between node $A$ and node $B$ in the graph.

There are two primary geometric criteria used to construct these networks:

  1. Natural Visibility Graphs (NVGs): An edge is formed if a straight line of sight connects two data points over the intervening peaks. As demonstrated by ten Bosch et al. in Analyzing Reaction Time and Error Sequences in Lexical Decision Experiments, this formulation naturally preserves the overarching temporal structure and clustering behavior of sequences.
  2. Horizontal Visibility Graphs (HVGs): A stricter variant where nodes are connected only if a perfectly horizontal line can link them without being obstructed by any intermediate value. According to the authors of Assessment of visibility graph similarity as a synchronization measure for chaotic, noisy and stochastic time series, this horizontal restriction simplifies network construction by focusing primarily on the relative ordering and local peaks of the signal, effectively filtering out small-scale stochastic fluctuations.

Why It Works: Invariance and Topological Shortcuts

Transforming a sequence into a network is more than a conceptual novelty; it fundamentally alters how a model interacts with the data's underlying structure.

Bypassing the Sequential Bottleneck

In standard recurrent neural networks (RNNs) or sequential architectures, information must step through every single time interval. This creates a sequential bottleneck where long-term dependencies are easily lost due to vanishing or exploding gradients.

Visibility graphs elegantly solve this problem by creating topological shortcuts. If two major events occur far apart in time but are prominent enough to "see" each other over intermediate fluctuations, the visibility algorithm creates a direct edge between them. As highlighted by Wu et al. in Price graphs: Utilizing the structural information of financial time series for stock prediction, these direct links allow models to capture long-range dependencies across the timeline, bypassing the traditional limitations of sequential architectures.

Inherent Structural Invariance

Traditional time-series analysis often requires meticulous preprocessing, such as alignment, normalization, or phase-shift correction. Visibility graphs offer a shortcut through this complexity because their connectivity is determined purely by relative geometric heights and orderings.

Because of this, the resulting network topology is naturally invariant to linear scaling and translation shifts. Li et al. leverage this property in Extracting Statistical Graph Features for Accurate and Efficient Time Series Classification to build highly efficient classification frameworks that run up to 100x faster than traditional shapelet-learning methods. Similarly, Pletnev et al. utilize this structural invariance in Graph Neural Networks for Model Recommendation using Time Series Data to map relational similarities between entirely different datasets for automated model selection.

Unlocking Non-Linear Dynamics

One of the most profound features of the visibility graph transformation is its ability to expose high-dimensional, non-linear dynamics without requiring explicit state-space embedding. The network's resulting topological properties—such as degree distribution, clustering coefficients, and average path length—serve as a direct mathematical proxy for the system's underlying complexity.

The authors of Assessment of visibility graph similarity as a synchronization measure for chaotic, noisy and stochastic time series demonstrate that this graph-domain mapping is intrinsically non-linear, allowing researchers to evaluate synchronization in chaotic or noisy systems directly from raw signals.

This capability is particularly valuable in biomedical signal processing:

Advanced Variations and Extensions

As the field has matured, several specialized variations of the visibility algorithm have emerged to handle specific data challenges.

Preserving the Arrow of Time: Directed Graphs

Standard visibility graphs are undirected, meaning that if point $A$ can see point $B$, the connection is symmetrical. However, time has a strict direction. To capture temporal flow and system irreversibility, Fan et al. utilize the Directed Horizontal Visibility Graph (DHVG) algorithm in Instability of Financial Time Series Revealed by Irreversibility Analysis. By assigning directions to edges based on the chronological order of the nodes, this method preserves temporal asymmetry, allowing for the detection of subtle market instabilities and anomalies.

Handling Noise: Penetrable and Weighted Graphs

In real-world physiological signals, a single random noise spike can completely block the line of sight between two otherwise highly correlated regions. To mitigate this vulnerability, Liu et al. developed the "Weighted Limited Penetrable Visibility Graph" in Cardiopulmonary Analysis of Sleep Apnea Based on Weighted Limited Penetrable Visibility Graph. This variation allows lines of sight to "penetrate" through a limited number of intermediate obstacles, ensuring that long-range dependencies are not severed by localized noise. Furthermore, it assigns weights to the resulting edges based on horizontal distance and vertical height differences, ensuring that the network topology retains the precise geometric intensity of the original signal.

Scaling to Deep Learning: Vector Visibility

While visibility graphs originated in statistical physics, they are increasingly integrated into modern deep learning architectures. In VISPool: Enhancing Transformer Encoders with Vector Visibility Graph Neural Networks, Alikasifoglu et al. extend the concept to multivariate token sequences. By treating word embeddings as a multivariate time series, they construct "Vector Visibility Graphs" (VVGs) on the fly based on the relative magnitudes of embedding vectors. This allows a Graph Neural Network to act as a dynamic, highly parameter-efficient pooling mechanism for Transformer encoders.

Broad Applications

Because visibility graphs convert abstract temporal sequences into tangible structural networks, they have been successfully deployed across a remarkably diverse array of domains:

  • Environmental Science: Hu et al. use visibility graphs to map carbon dioxide emissions across Chinese provinces during COVID-19 lockdowns, uncovering hidden regional dynamics that traditional statistical models missed.
  • Epidemiology: Zhu et al. transform daily growth rates of disease outbreaks into complex networks, using network connectivity and clustering coefficients to provide a real-time indicator of epidemic control.
  • Clinical Diagnostics: Ahmadi et al. apply horizontal visibility algorithms to short-term, seizure-free EEG data to extract discriminative clustering coefficients, successfully differentiating between epilepsy and psychogenic non-epileptic seizures.
  • Information Fusion: Liu et al. combine visibility graphs with evidence theory in A Data-Driven Dynamic Data Fusion Method Based on Visibility Graph and Evidence Theory to automatically determine how much to trust historical data based purely on its structural network topology.
  • Multi-Modal Physiology: Dimitri et al. model simultaneous spikes in intracranial pressure and heart rate as a multiplex visibility network, discovering that the topological "cross-talk" between these distinct physiological layers serves as a powerful digital biomarker for patient survival.

By translating the sequential passage of time into the static geometry of a network, visibility graphs provide a unique lens that unifies time-series analysis, non-linear dynamics, and graph machine learning.

Key Foundational Work

Ictal time-irreversible intracranial EEG signals as markers of the epileptogenic zone (2016) — contributed 1 concept elements to this direction.
Unified functional network and nonlinear time series analysis for complex systems science: The pyunicorn package (2015) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 21 papers
  2. Version 1 · 2026-10-04 · 24 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_visibility_graph_transformation_5114d671,
  title = {visibility graph transformation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/visibility-graph-transformation/rdr_5114d6716312438f82fc3c82530656ab},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_visibility_graph_transformation_5114d671_v2,
  title = {visibility graph transformation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/visibility-graph-transformation/rdr_5114d6716312438f82fc3c82530656ab/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake