TLDR: TGM is a new open-source library for machine learning on temporal graphs, which are networks that change over time. It’s the first to unify continuous-time and discrete-time approaches, offering native support for dynamic node features, time-granularity conversions, and various prediction tasks. TGM significantly outperforms existing libraries in efficiency, achieving up to 7.8x speedup in training and 175x speedup in graph discretization. Its modular ‘hook’ mechanism and research-oriented design simplify complex workflows and enable new research into dynamic graph property prediction and the impact of time granularity and batching strategies.
Machine learning on temporal graphs, which are networks that change and evolve over time, is a rapidly growing field with applications ranging from social networks to financial transactions. However, researchers have long faced significant challenges due to the lack of a unified, efficient, and flexible software infrastructure. Existing libraries are often specialized for particular model types, making it difficult to compare different approaches or develop new ones. This fragmentation has created a divide between methods that handle continuous-time dynamic graphs (CTDG) and those that use discrete-time dynamic graphs (DTDG), hindering progress.
Introducing TGM: A Unified Framework for Temporal Graph Machine Learning
To address these critical gaps, a team of researchers has introduced Temporal Graph Modelling (TGM), a new open-source library designed specifically for machine learning on temporal graphs. TGM stands out as the first library to unify both continuous-time and discrete-time approaches, providing a comprehensive platform for diverse models and research questions. It offers native support for dynamic node features, allowing for the modeling of events like social media posts, and includes built-in tools for converting between different time granularities. Furthermore, TGM natively handles various machine learning tasks, including link, node, and graph-level predictions.
Overcoming Current Limitations
Current temporal graph libraries often suffer from several limitations. Many are tailored to a single algorithm family, limiting their extensibility and support for the wide array of models emerging in the field. For instance, some libraries are optimized for temporal message passing but don’t support newer transformer-based methods. Crucially, most existing tools lack essential time conversion operations, which are vital for analyzing how temporal granularity impacts graph dynamics. They also fall short on usability features like profiling tools, test suites, and modular abstractions, making reproducible research and rapid prototyping challenging.
TGM tackles these issues head-on. By providing a unified framework, it eliminates the need for entirely different data pipelines for continuous- and discrete-time models. Core operations like temporal neighbor sampling and negative edge construction are standardized, fostering fair benchmarking and easier combination of ideas across different approaches.
Key Innovations and Efficiency
TGM introduces several groundbreaking features:
- First Unified Library: It supports both CTDG and DTDG methods, treating them as distinct views of the same underlying data. This includes implementations of eight methods from both literatures, including frontier models.
- Time as a First-Class Citizen: TGM’s API natively incorporates time granularity, offering built-in support for graph discretization (converting fine-grained events into coarser snapshots) and snapshot iteration.
- Efficiency: Empirical results show TGM achieves an average 7.8 times speedup in training across multiple models, datasets, and tasks compared to the widely used DyGLib. For graph discretization, it boasts an average 175 times speedup relative to existing implementations. This efficiency is partly due to its fully vectorized, PyTorch-native implementation of operations like recency sampling.
- Research-Oriented Design: Built for rapid prototyping, TGM emphasizes modularity and ease-of-use. Its novel “hook” mechanism standardizes temporal graph transformations, making complex workflows simpler to express.
The library’s architecture is divided into three layers: a data layer for efficient storage and slicing of temporal graphs, an execution layer built around a hook manager for transparently performing transformations, and an ML layer that materializes data on-device for model computation. This design ensures both efficiency and extensibility.
Unlocking New Research Possibilities
Beyond its efficiency, TGM opens doors to entirely new research avenues. Experiments demonstrate how it enables dynamic graph property prediction, allowing researchers to forecast the future evolution of graph properties. For example, it can predict whether a transaction network snapshot will grow, which is crucial for understanding network dynamics.
TGM also facilitates the study of how time granularity affects DTDG model performance. Researchers can easily adjust snapshot granularities (e.g., hourly, daily, weekly) with a single line of code, treating it as a hyperparameter. This has revealed that the choice of granularity significantly impacts model accuracy, with optimal performance varying by dataset and model.
Furthermore, TGM allows for the investigation of how different batching strategies (by fixed edges versus by time) influence CTDG model performance, highlighting the validation process itself as a critical, yet often overlooked, hyperparameter.
Also Read:
- A New Benchmark for Understanding Evolving Communities in Dynamic Networks
- HTMformer: Enhancing Time Series Forecasting Accuracy and Efficiency with Hybrid Embeddings
Conclusion
TGM represents a significant step forward for machine learning on temporal graphs. By providing a modular, efficient, and unified framework, it simplifies complex workflows, accelerates experimentation, and bridges the long-standing divide between continuous- and discrete-time methods. The researchers envision TGM as a foundational tool for a shared ecosystem where models, hooks, and analytics can be seamlessly composed and reused, ultimately accelerating the pace of temporal graph learning research. You can find the full research paper here: TGM: A Modular and Efficient Library for Machine Learning on Temporal Graphs.
