Traditional time series analysis treats data as a sequence of values ordered in time, often relying on convolutional kernels or recurrent units to extract patterns. However, a growing body of research suggests that we can gain deeper insights by changing our perspective: treating a time series not as a 1D array, but as a landscape where data points "see" one another. This transformation, known as a Visibility Graph (VG), maps temporal dynamics into a complex network, allowing researchers to apply the powerful toolkit of graph theory to sequential data.
The Geometry of Sight
The core intuition behind visibility graphs is simple: imagine the time series as a series of skyscrapers of varying heights. If you stand on top of one building, which other buildings can you see?
In a Natural Visibility Graph (NVG), two nodes (time points) are connected if a straight line can be drawn between their peaks without being obstructed by any intermediate data point. A common variation is the Horizontal Visibility Graph (HVG), where connections are only made if a horizontal line can link two points without hitting an intermediate value.
As Daoyuan Li, Jessica Lin 0001, and colleagues demonstrate in Extracting Statistical Graph Features for Accurate and Efficient Time Series Classification, this mapping is inherently invariant to vertical scaling and horizontal translation. Because the graph depends on the relative heights and "line of sight" rather than absolute values, it provides a robust structural representation that is naturally shielded from certain types of noise and signal drift.
Bypassing the Limits of Recurrence
One of the most significant advantages of this approach is how it handles long-term dependencies. In standard Recurrent Neural Networks (RNNs), information must pass through every intermediate step, often leading to the vanishing gradient problem.
By contrast, visibility graphs create "topological shortcuts." As Junran Wu, Jichang Zhao, and their team show in Price graphs: Utilizing the structural information of financial time series for stock prediction, two distant but prominent events in a financial time series can be directly connected in the graph domain if they have a clear line of sight. This allows models to capture long-range structural dependencies that are often lost in Euclidean sequence modeling.
Capturing Non-Linear Dynamics
Visibility graphs do more than just simplify long-range connections; they act as a bridge to the world of non-linear dynamics. Kamyar Zeinalipour and Marco Gori argue in Graph Neural Networks for Topological Feature Extraction in ECG Classification that this mapping captures geometric and topological features—such as the "shape" of a heartbeat—that standard convolutional kernels overlook.
Researchers have extended the basic visibility criteria to handle specific signal challenges:
- Limited Penetrability: Chengyu Liu and Jianqing Li utilize "Limited Penetrable" graphs to allow lines of sight to pass through a small number of obstructions, making the network more resilient to high-frequency noise in sleep apnea data.
- Curvature-based Visibility: Qi Xuan and Jinchao Zhou introduced the Circular Limited Penetrable Visibility Graph (CLPVG), which uses circular arcs instead of straight lines to better capture the non-linear "bend" of complex signals like epilepsy detections.
From Brain Waves to Language
The utility of visibility graphs spans diverse domains, particularly where the "cross-talk" between different signals is vital. Giovanna Maria Dimitri, Pietro Lio, and Adam Young have pioneered the use of multiplex visibility graphs to study traumatic brain injury. By mapping heart rate and intracranial pressure into two layers of a single network, they identified "interlayer mutual information" as a digital biomarker for physiological resilience.
The concept has even moved into Natural Language Processing. In VISPool: Enhancing Transformer Encoders with Vector Visibility Graph Neural Networks, Tuna Alikasifoglu and Aykut Koç treat word embeddings as a multivariate time series. By building visibility graphs between tokens, they created a dynamic pooling mechanism that captures relational structures more efficiently than standard linear layers.
Generative Frontiers
The most recent evolution of this field involves moving from analysis to synthesis. Xuyang Li and Lele Long introduced TSGDiff, a framework that generates synthetic time series by performing diffusion in a latent graph space. By ensuring the graph structure of the generated data matches the original, the model preserves the underlying structural integrity and seasonal links that traditional sequence-based generators often fail to replicate.
By transforming the temporal into the topological, visibility graphs provide a mathematically rigorous way to "see" the hidden architecture of data, turning the noisy fluctuations of a signal into a clear map of its underlying system.