A Network Science Approach to Granular Time Series Segmentation

Time series segmentation (TSS) is one of the time series (TS) analysis techniques, that has received considerably less attention compared to other TS related tasks. In recent years, deep learning architectures have been introduced for TSS, however their reliance on sliding windows limits segmentation granularity due to fixed window sizes and strides. To overcome these challenges, we propose a new more granular TSS approach that utilizes the Weighted Dual Perspective Visbility Graph (WDPVG) TS into a graph and combines it with a Graph Attention Network (GAT). By transforming TS into graphs, we are able to capture different structural aspects of the data that would otherwise remain hidden. By utilizing the representation learning capabilities of Graph Neural Networks, our method is able to effectively identify meaningful segments within the TS. To better understand the potential of our approach, we also experimented with different TS-to-graph transformations and compared their performance. Our contributions include: a) formulating the TSS as a node classification problem on graphs; b) conducting an extensive analysis of various TS-to-graph transformations applied to TSS using benchmark datasets from the TSSB repository; c) providing the first detailed study on utilizing GNNs for analyzing graph representations of TS in the context of TSS; d) demonstrating the effectiveness of our method, which achieves an average F1 score of 0.97 across 59 diverse TSS benchmark datasets; e) outperforming the seq2point baseline method by 0.05 in terms of F1 score; and f) reducing the required training data compared to the baseline methods.

💡 Research Summary

The paper tackles the problem of Time‑Series Segmentation (TSS) from a novel graph‑centric perspective. Traditional deep‑learning approaches to TSS rely on sliding windows with fixed lengths and strides, which inherently limit the granularity of the segmentation and require large amounts of labeled data. To overcome these constraints, the authors propose to (i) transform a univariate time series into a graph, and (ii) treat the segmentation task as a node‑level classification problem solved with a Graph Attention Network (GAT).

Graph construction.
Seven graph‑generation schemes are examined, ranging from classic Visibility Graphs (NVG, HVG) to weighted variants, Transition Networks (quantile‑based, ordinal‑partition, phase‑space), and proximity‑based k‑NN graphs. The centerpiece of the study is the Weighted Dual‑Perspective Visibility Graph (WDPVG). WDPVG combines the standard Natural Visibility Graph with its reflected counterpart (obtained by mirroring the series on the time axis) and assigns edge weights based on Euclidean distance, slope, or temporal gaps. This dual‑perspective design captures both peaks and troughs and handles uneven sampling, which are common shortcomings of plain NVG.

Problem formulation.
Given a series (S = {s_1,\dots,s_N}), a transformation (T) yields a graph (G = (V,E)) where each node corresponds to a time point. The segmentation objective becomes: for each node (v_i) predict a label (y_i \in {0,\dots,C-1}) indicating the segment class. This is expressed as a function (F_G(G;\theta)) learned by a neural model.

Model architecture.
The authors employ a multi‑head Graph Attention Network. Each GAT layer computes attention coefficients (\alpha_{ij}) for neighboring nodes, aggregates their hidden states, and updates node embeddings. Stacking several such layers enables the model to capture both short‑range and long‑range temporal dependencies without the need for explicit windowing. The final layer applies a soft‑max to produce per‑node class probabilities, and training uses cross‑entropy loss with L2 regularization.

Experimental setup.
The evaluation uses the TSSB benchmark repository, comprising 59 diverse univariate time‑series datasets (different lengths, domains, and noise levels). Baselines include the seq2point model, U‑Time CNN, ClaSP, and auto‑encoder based change‑point detectors. Performance is measured primarily by the F1‑score, as well as accuracy and robustness to reduced training data.

Results.
The WDPVG + GAT pipeline achieves an average F1‑score of 0.97, outperforming the seq2point baseline by 0.05 points. It also demonstrates remarkable data efficiency: when the training set is reduced to 10 % of its original size, the F1‑score drops only to ~0.94, whereas baselines suffer larger degradations. Among the seven graph constructions, plain NVG and HVG perform noticeably worse, especially on series with negative values or irregular sampling, confirming the importance of weighting and dual‑perspective mechanisms.

Analysis of strengths and limitations.
The superior performance stems from two factors: (1) WDPVG encodes rich structural cues (visibility, temporal distance, slope) that directly reflect the dynamics of the series, and (2) GAT’s attention mechanism selectively emphasizes the most informative neighbors, enabling precise detection of subtle transition points. However, the graph‑building step for visibility‑based methods has quadratic time complexity (O(N²)), which may become prohibitive for very long series. The authors suggest possible mitigations such as edge‑pruning thresholds or sampling‑based sparsification.

Future directions.
The study opens several avenues: extending the approach to multivariate time series (e.g., constructing multiplex graphs or inter‑channel edges), developing incremental graph updates for streaming data, and exploring alternative GNN architectures (e.g., GraphSAGE, GIN) that may further reduce computational overhead.

Conclusion.
By converting a time series into a weighted dual‑perspective visibility graph and applying a graph attention network for node classification, the paper delivers a highly granular, accurate, and data‑efficient TSS solution that surpasses existing sliding‑window deep‑learning baselines. This work demonstrates the practical viability of graph‑based representations for time‑series analysis and suggests a promising new paradigm for future research.