Introduction
Temporal Interactive Networks (TINs) model evolving systems where entities interact over time, such as e-commerce transactions or social media engagements. Traditional approaches often treat future interaction prediction as a coarse-grained multi-sequence problem, focusing on individual user histories while neglecting the underlying network topology. While integrating Graph Convolutional Networks (GCNs) with Temporal Point Processes (TPPs) can address this, standard GCNs are computationally intensive and often impractical for large-scale applications. This paper introduces the Deep Graph Neural Point Process (DGNPP), a model designed to capture both network topology and temporal dynamics with high efficiency.
Methodology
The DGNPP model operates on snapshots of the interaction graph and consists of two primary modules that decouple static and dynamic feature learning.
Node Aggregation Layer (NAL)
The NAL generates static embeddings to capture the network's topological structure. To maintain efficiency, the authors omit the heavy linear transformations and non-linear activations typical of GCNs. Instead, the NAL employs a simplified weighted sum of neighboring node features, normalized by degree, to extract essential topological information without significant computational overhead.
Self-Attentive Layer (SAL)
The SAL generates dynamic embeddings using a self-attention mechanism to capture long-range dependencies in historical interaction sequences. It includes a "Temporal Shift Layer" that models continuous time passage, allowing embeddings to evolve smoothly between discrete events.
Intensity Function
The model combines static and dynamic embeddings into an intensity function inspired by the Multivariate Hawkes Process. The static embeddings represent the baseline intensity, while dynamic embeddings represent the influence of historical events. The system is optimized end-to-end using Maximum Likelihood Estimation (MLE) to predict both the next interaction and its precise timing.
Experiments
The model was evaluated on three benchmark datasets: Reddit (user-subreddit interactions), Wikipedia (page edits), and Last.FM (music listening behaviors). These datasets represent varying scales and interaction densities.
Results and Discussion
DGNPP demonstrates significant improvements over state-of-the-art baselines in three key areas:
- Predictive Accuracy: On the Last.FM dataset, DGNPP outperformed competitors by 16.0% in Recall@10 and 8.1% in Mean Reciprocal Rank (MRR). It maintained top-tier performance across all datasets.
- Temporal Precision: The model significantly reduced time prediction error, achieving a Root Mean Square Error (RMSE) reduction of 30.4% on Last.FM, 28.8% on Wikipedia, and 13.3% on Reddit compared to the DSPP baseline.
- Computational Efficiency: By simplifying the Node Aggregation Layer, DGNPP reduced training runtime per epoch by up to 25.1% (on Reddit) compared to DSPP, proving that complex GCN layers are not always necessary for effective topological feature extraction.
Ablation studies confirmed that both the NAL and SAL are critical, as removing either module led to a measurable decline in performance (e.g., a 5.7% MRR drop on Reddit without the SAL).
Conclusion
DGNPP provides a scalable and accurate framework for learning in Temporal Interactive Networks. By decoupling static topological features from dynamic temporal sequences and simplifying the graph aggregation process, the model achieves superior predictive performance and computational efficiency. Future work may explore the model's robustness in cold-start scenarios or its application to richer node attribute sets.
Summary updated