FC-GAGA: Fully Connected Gated Graph Architecture for Spatio-Temporal Traffic Forecasting (2007.15531v2)

Abstract: Forecasting of multivariate time-series is an important problem that has applications in traffic management, cellular network configuration, and quantitative finance. A special case of the problem arises when there is a graph available that captures the relationships between the time-series. In this paper we propose a novel learning architecture that achieves performance competitive with or better than the best existing algorithms, without requiring knowledge of the graph. The key element of our proposed architecture is the learnable fully connected hard graph gating mechanism that enables the use of the state-of-the-art and highly computationally efficient fully connected time-series forecasting architecture in traffic forecasting applications. Experimental results for two public traffic network datasets illustrate the value of our approach, and ablation studies confirm the importance of each element of the architecture. The code is available here: https://github.com/boreshkinai/fc-gaga.

FC-GAGA: Fully Connected Gated Graph Architecture for Spatio-Temporal Traffic Forecasting

The paper introduces a novel model called FC-GAGA (Fully Connected Gated Graph Architecture) designed for spatio-temporal traffic forecasting. This model addresses complex interactions inherent in multivariate time-series (TS) data, which is commonly applicable in domains like traffic management and other networked system analyses. The architecture is built on a fully connected time-series forecasting base combined with innovative graph and time gating mechanisms, and it does not necessitate prior knowledge of the specific graph structure underlying the data relationships.

Model Architecture

FC-GAGA leverages a unique graph gating mechanism that differentiates it from existing TS models. The graph gate allows the network to selectively incorporate information from entity interactions across a fully connected architecture, leading to enhanced predictive performance. Notably, FC-GAGA integrates:

1. Graph Gate Block: This element uses node embeddings to compute weights representing the strength of interactions between nodes, gating the input data to form a selective representation of graph edges.
2. Time Gate Block: This component adjusts for temporal features, such as the time of day, which may influence the TS data, using it to modulate the impact of node embeddings.
3. Fully Connected Time-Series Blocks: This core part processes gated inputs using residual blocks to produce forecasts. Each layer in the architecture can learn its distinct graph structure, creating flexibility to adapt to non-Markovian diffusion processes across layers.

Unlike traditional approaches that rely heavily on pre-defined graphs, FC-GAGA aims to learn graph structures inherently from data alone, marrying flexibility with computational efficiency. Its designers implemented this architecture using the N-BEATS framework for fully connected TS forecasting, augmenting with graph gates to connect multivariate inputs effectively.

Empirical Validation

The paper presents extensive empirical results using two well-known traffic datasets, METR-LA and PEMS-BAY. These datasets provide real-world traffic sensor data for validating model performance:

• Performance Metrics: FC-GAGA's forecasting accuracy is assessed using MAE (Mean Absolute Error), MAPE (Mean Absolute Percentage Error), and RMSE (Root Mean Squared Error). The results demonstrate that FC-GAGA excels compared to state-of-the-art forecasts, especially when evaluated against baseline models and contemporary graph-based algorithms.
• Computational Efficiency: The model requires substantially fewer computational resources, both in terms of memory and processing time, compared to competing methods like DCRNN and Graph WaveNet. This efficiency is crucial for scalability in large-scale applications.
• Ablation Studies: These studies reveal the significance of each component, particularly the necessity of the graph gate for achieving strong predictive performance. The model performs best when stacking multiple layers, demonstrating an advantage over single-layer approaches due to its layered learning framework.

Implications and Future Directions

FC-GAGA's framework provides significant insight into optimizing TS models for scenarios involving unspecified or variable relational graph data. This capability is valuable in fields like cellular networks, commerce, and climate modeling, expanding the use of inherently learned graph architectures. Future research may explore extending the architecture to other application domains or integrating more sophisticated graph learning techniques to improve interpretability and accuracy further.

Moreover, the demonstrated computational efficiency presents opportunities for deploying similar technologies in scenarios constrained by hardware resources, potentially democratizing access to high-performance forecasting algorithms. The ability to train effective models without explicitly defined graphs represents a step forward in developing more autonomous and adaptive machine learning systems.