Technical Summary: GLIDE – Graph-guided Leap Inference for Diffusion Estimation of Spatio-Temporal Point Processes

1. Problem Formulation

Spatio-Temporal Point Processes (STPPs) model asynchronous events in continuous time and space, aiming to estimate the conditional distribution of the next event given a historical sequence. Existing deep learning approaches face a fundamental trade-off:

• Deterministic models (e.g., recurrent or attention-based neural point processes) often exhibit "mean-seeking" behavior, collapsing predictions toward spatial centroids and failing to capture multi-modal future distributions.
• Standard Diffusion Models offer flexible distribution modeling but struggle in sparse spatio-temporal domains. Reverse sampling from pure Gaussian noise is computationally expensive (requiring many denoising steps) and lacks structural constraints, often leading to probability mass dispersion into invalid spatial regions before converging.

The core challenge is to design a generative framework that preserves the stochastic expressiveness of diffusion models while improving sampling efficiency and spatial localization in sparse, continuous domains.

2. Methodology

The authors propose GLIDE (Graph-guided Leap Inference for Diffusion Estimation), a unified framework integrating graph-structured historical encoding with conditional diffusion. The architecture consists of four primary components:

2.1 Multi-Scale Dynamic Graph Construction

Historical events are organized into a directed graph where nodes represent events and edges connect events within specific temporal windows ($K_{scales}$).

• Edge Features: Edges encode normalized time differences ($\Delta t$), Euclidean spatial distances ($\Delta s$), and window scale identifiers. These are modulated by decay functions inspired by self-exciting processes (e.g., Hawkes processes) to reflect temporal and spatial proximity.
• Node Features: Combine normalized event attributes with multi-scale local statistics (e.g., mean/variance of inter-event times, spatial drift vectors).

2.2 Dual-Stream Spatio-Temporal Encoder

To address the distinct nature of temporal evolution and spatial topology, GLIDE employs two separate encoding streams before fusing them:

• Temporal Stream: Processes event sequences using a Transformer with Rotary Positional Embeddings (RoPE) to capture relative order and long-range dependencies.
• Spatial Stream: Utilizes a Graph Attention Network (GATv2) operating on the constructed graph. It incorporates Fourier lifting of coordinates to expand the representational space and modulates attention coefficients using edge attributes to weigh neighbor influence based on spatiotemporal scales.
• Fusion: A gated interaction layer merges the two streams into a context matrix $C$, allowing the model to dynamically weigh temporal vs. spatial evidence.

2.3 Dual-Branch Diffusion Transformer (DiT)

The reverse diffusion process is parameterized by a conditional DiT.

• Dual Branches: Instead of a single shared path, the model uses parallel branches for temporal ($\Delta t$) and spatial ($s$) denoising. This preserves distinct inductive biases, as temporal prediction relies on progression while spatial prediction is often multi-modal.
• Interaction: Cross-branch interaction is delayed until deeper layers to prevent early feature over-sharing, allowing specialized denoising patterns to emerge before explicit spatio-temporal coupling.
• Conditioning: Context $C$ and diffusion step $k$ modulate the DiT blocks via Adaptive Layer Normalization.

2.4 Prior-Guided Leap Inference

To address the inefficiency of starting from pure noise, GLIDE introduces a "Leap" strategy:

1. Deterministic Anchor: A lightweight Joint Mean Predictor (JMP) generates a coarse deterministic estimate $\hat{x}_0$ from the context $C$.
2. Leap Initialization: The reverse process does not start at step $K$ (pure noise) but at an intermediate step $\tau$. The initial state $x_\tau$ is constructed by mixing the anchor $\hat{x}_0$ with Gaussian noise: $x_\tau = \sqrt{\bar{\alpha}_\tau}\hat{x}_0 + \sqrt{1-\bar{\alpha}_\tau}\epsilon$.
3. Elastic Guidance: During the reverse steps $t \in [\tau, 1]$, an elastic guidance rule gradually blends the standard denoised output with the anchor, decaying linearly as $t \to 0$. This transitions the process from coarse localization to fine stochastic refinement.

3. Key Contributions

The paper outlines three main contributions:

1. Graph-Guided Diffusion Framework: GLIDE integrates multi-scale dynamic correlation graphs with a dual-stream diffusion architecture to model complex spatio-temporal dependencies, addressing the lack of structural constraints in standard diffusion.
2. Prior-Informed Leap Inference: A novel sampling strategy that anchors the diffusion trajectory with a deterministic prior. This substantially reduces inference cost by bypassing the unguided early search phase while preserving the ability to explore multiple local modes.
3. Empirical Performance: The framework achieves state-of-the-art performance in distribution fitting (Negative Log-Likelihood) and next-event prediction across multiple real-world datasets, with significant gains in spatial accuracy and inference speed (up to $3.1\times$ faster than standard diffusion baselines).

4. Experimental Results

Experiments were conducted on four real-world datasets: Earthquake, COVID-19, Citibike, and Crime.

• Distribution Fitting (NLL): GLIDE outperformed baselines (including DSTPP, SMASH, and DeepSTPP) in spatial NLL across all datasets. The improvement was most pronounced on the sparse Earthquake dataset (Spatial NLL dropped from 0.452 to -0.067), indicating better concentration of probability mass in valid regions. Temporal NLL remained competitive.
• Next-Event Prediction: GLIDE achieved the lowest spatial error (Euclidean distance) on Earthquake and COVID-19 and matched the best results on Citibike. Temporal RMSE remained competitive, demonstrating that spatial gains were not achieved at the expense of temporal accuracy.
• Efficiency: The Leap inference strategy reduced inference time significantly. On GLIDE, average generation time dropped from 0.275 ms (without Leap) to 0.109 ms (with Leap), a $2.7\times$ speedup, with minimal parameter overhead (0.1685M additional parameters).
• Ablation Studies:
  • Leap Initialization: Removing the leap strategy resulted in slower convergence and higher spatial error, confirming that the prior anchor effectively skips the uninformative early diffusion phase.
  • Graph Guidance: Removing the graph module degraded performance most significantly on datasets with strong local propagation dynamics (COVID-19, Crime), validating the utility of graph-structured priors.
  • Cross-Stream Interaction: Removing the interaction between temporal and spatial branches hurt spatial learning, suggesting temporal phase information is crucial for spatial localization.

5. Significance and Claims

The paper claims that GLIDE successfully bridges the gap between the expressive power of diffusion models and the efficiency required for practical STPP applications.

• Spatial Fidelity: By leveraging graph-structured history and a deterministic anchor, GLIDE mitigates the "probability dispersion" issue common in standard diffusion, leading to more accurate spatial predictions in sparse domains.
• Efficiency: The leap inference mechanism transforms generation from a costly global search into a local refinement process, making diffusion-based STPPs viable for real-time or resource-constrained applications.
• Generality: The framework is applicable to diverse spatio-temporal dynamics, from epidemic spread and earthquake forecasting to urban mobility.

The authors note limitations: the current formulation handles continuous coordinates but does not explicitly model event marks or categorical types, and the graph component is most effective when local interactions are intrinsic to the underlying process. Future work is suggested to extend GLIDE to marked STPPs and dynamically evolving graph structures.