Gramian Angular Difference Field (GADF)

• Gramian Angular Difference Field is a representation technique that maps time series data into 2D images using the sine of angular differences, capturing temporal dynamics.
• GADF facilitates multimodal fusion by integrating CNN-extracted spatial features with raw temporal signals, leading to enhanced performance in tasks like ECG classification.
• GADF images highlight phase transitions and local interactions, offering improved interpretability and robustness for applications in biomedical and activity recognition.

The Gramian Angular Difference Field (GADF) is a two-dimensional representation technique that encodes the temporal dynamics of a time series into an image, enabling the use of convolutional neural networks and other visual feature extractors for time-series data. While the Gramian Angular Field (GAF) encompasses two related families—Gramian Angular Summation Field (GASF) and Gramian Angular Difference Field (GADF)—the latter specifically encodes the difference of angles between time points, facilitating the capture of relative temporal transitions with rotationally invariant properties. GADF transformations have become integral in multimodal architectures for biomedical signal analysis, human activity recognition, and other domains requiring robust temporal pattern classification (Qin et al., 2024).

1. Mathematical Formulation of GADF

The GADF transformation proceeds through three primary stages: scaling, angular encoding, and Gramian matrix construction. Given a normalized time series segment $S_i = \{ x_j \}_{j=1}^{w}$, standardization maps the sequence onto $[-1,1]$: $$\tilde{x}_j = \frac{(x_j - \min S_i)\cdot(1 - (-1))}{\max S_i - \min S_i} + (-1)$$ Each $\tilde{x}_j$ is then mapped to angular coordinates: $\phi_j = \arccos(\tilde{x}_j)$ where $\phi_j\in[0,\pi]$. For GADF, the Gramian matrix $A$ is computed by

$$GADF_{j,k} = \sin(\phi_j - \phi_k), \quad j, k = 1, \ldots, w$$

whereas GASF uses the sum. This representation yields a $w \times w$ image encoding the sine of angular differences between all time point pairs, capturing phase relationships and temporal dynamics.

2. Role in Multimodal and Multibranch Architectures

GADF is commonly employed alongside raw time series and other transforms to enrich feature sets for multimodal fusion. For example, GAF-FusionNet utilizes a branch dedicated to extracting spatial features from GAF images (often GASF, occasionally GADF) and fuses them with temporal features processed by recurrent or temporal CNN encoders. The transformation enables integration with standard “2D” computer vision backbones such as ResNet, increasing feature diversity and enhancing classification robustness (Qin et al., 2024). The dual-branch paradigm allows models to exploit both direct time-domain patterns and higher-level textures or symmetries implicit in the GADF image.

3. Visualization Properties and Interpretability

GADF images are characterized by their symmetric, often periodic, visual structure. Each pixel $(j,k)$ encodes the sine of the phase difference between two time points, highlighting transitions, reversals, and rhythmicity. Unlike the GASF, which encodes cumulative phase information, the GADF emphasizes differences, making it particularly sensitive to change points and local interactions. This property is advantageous when distinguishing between classes whose main discriminants are transient or differential in nature, as in arrhythmia or gesture recognition.

A plausible implication is that GADF-based feature maps could offer improved interpretability regarding phase reversal, periodicity, or abrupt temporal changes, especially when coupled with attention-based visualization methods (e.g. Grad-CAM applied to GADF-derived CNN activations as suggested in future work) (Qin et al., 2024).

4. Integration and Fusion Strategies

GADF-derived feature maps can be fused with other modalities through several strategies:

• Concatenation: Direct merging of CNN-spatial features from GADF images with features from the raw or preprocessed time series.
• Attention mechanisms: Application of dual-layer cross-channel split-attention, where both intra-modality and cross-modality self-attention modules operate on the fused feature space, as in GAF-FusionNet. The attended features are passed through an MLP classifier, enabling the network to learn complex interdependencies between temporal and spatial domains (Qin et al., 2024).
• Residual and gating techniques: Incorporating elementwise gates or residual paths ensures stable integration and improved gradient flow in deep multimodal settings.

5. Empirical Performance and Ablation Results

Integration of GADF (along with or replacing GASF) in multimodal architectures yields substantial performance gains in benchmark classification tasks. In GAF-FusionNet, the joint use of time-series and GAF image domains, with advanced attention-driven fusion, resulted in classification accuracies of 94.5% (ECG200), 96.9% (ECG5000), and 99.6% (MIT-BIH) (Qin et al., 2024). Ablation studies demonstrate that omitting the attention layers or relying on a single modality (either raw time series or GAF-based image branch) leads to accuracy drops of 1.5–2.6%, underlining the complementary role of GADF-type representations. These findings suggest that GADF-based encoding contributes crucial discriminative power, especially when harnessed via cross-domain attention and careful fusion.

6. Limitations and Prospective Research

No publications to date report large-scale deployment of GADF in real-world clinical or non-laboratory contexts. Limitations include increased computational cost when integrating image-based and sequential representations, lack of explicit interpretability for the domain expert (since the meaning of structures in GADF images is not immediately transparent), and sensitivity to preprocessing steps such as window length or normalization regime. Future work is suggested in extending the fusion paradigm to additional physiological modalities, improving interpretability through advanced visualization, and enabling dynamic windowing for online monitoring (Qin et al., 2024).

7. Connections to Related Tensorial and Visual Representations

While GADF is related to other time-series-to-image encoding techniques (such as recurrence plots or Markov transition fields), a key distinction is the rotationally invariant and phase-differential nature of the mapping. GADF and related Gramian fields thus provide a mathematically grounded, lossless mapping that is well-suited to CNN-based analysis, feature fusion, and transfer learning scenarios. Combined with multibranch networks and attention mechanisms, GADF facilitates robust and scalable multimodal sequence classification (Qin et al., 2024).