Abstract
Predicting urban traffic patterns is a high-dimensional spatio-temporal challenge with significant implications for city planning and navigation. This paper presents a novel "Sedenion U-Net" architecture developed for the Traffic4cast 2020 challenge. By leveraging 16-dimensional hypercomplex numbers (sedenions), the model efficiently captures complex correlations between static city layouts and dynamic traffic modalities. The proposed approach achieves competitive performance with significantly fewer parameters than standard real-valued networks, demonstrating the potential of hypercomplex deep learning for multi-modal sensor data.
Introduction
Urban traffic flow prediction—encompassing both speed and volume—is a classic big-data problem. The Traffic4cast 2020 challenge required participants to predict these variables across three major cities (Berlin, Istanbul, and Moscow) using high-resolution map data. The primary difficulty lies in multimodality: traffic data consists of correlated directional speeds, volumes, and incident reports, which must be fused with static geographic information like city layout.
Standard neural networks often struggle to model these inter-channel relationships without becoming excessively large. This research addresses the challenge by introducing a parameter-efficient architecture based on sedenion algebra, which mathematically enforces the learning of relationships between different data components.
Methodology: The Sedenion U-Net
The core innovation of this work is the replacement of standard real-valued convolutions with Sedenion convolutions (SConv) within a U-Net framework. Sedenions are 16-dimensional hypercomplex numbers. In an SConv layer, weights are shared across these 16 dimensions according to specific algebraic rules. This weight sharing allows the model to maintain high expressive power while drastically reducing the total parameter count to approximately 628,592—a lean profile for high-resolution spatio-temporal tasks.
Input Mapping and Early Fusion
Unlike traditional models that process static and dynamic data through separate branches (late fusion), the Sedenion U-Net employs early fusion. Data is mapped directly into the 16 sedenion components as shown in the figure below. One component represents static city data, twelve represent the preceding hour of traffic (12 frames at 5-minute intervals), and three are reserved as zeros for future use.
Architecture Overview
The network follows a U-Net structure, an architecture widely successful in image segmentation. The encoder path progressively reduces spatial resolution using encodeBlock (ResNet-based) layers, while the decoder path recovers resolution using decodeBlock layers. Skip connections are utilized to preserve fine-grained spatial details from the encoder to the decoder.
Experimental Setup and Results
The study utilizes the Traffic4cast 2020 dataset, covering 362 days of data for three major cities. Each data point includes 9 channels (8 for directional speed/volume and 1 for incidents) at 5-minute intervals. The task is to use 12 past frames (1 hour) to predict 6 future time steps (5, 10, 15, 30, 45, and 60 minutes ahead).
The Sedenion U-Net demonstrated consistent performance across all test cities, achieving a validation Mean Squared Error (MSE) of $1.33893 \times 10^{-3}$ and a test MSE of $1.30845 \times 10^{-3}$.
Key Insights
- Algebraic Correlation: By using sedenion multiplication, the network is mathematically forced to consider the relationships between the 16 components as a single entity rather than independent channels.
- Parameter Efficiency: The model's low parameter count makes it highly suitable for deployment on edge devices or systems with limited memory.
- Static-Dynamic Synergy: Treating static city maps as the "real" part of the sedenion and dynamic traffic as the "imaginary" parts provided a superior mechanism for handling city-specific variations compared to traditional methods.
Conclusion and Future Work
Hypercomplex networks offer a mathematically grounded approach to high-dimensional sensor fusion. While hardware constraints limited the number of experimental iterations possible during the challenge, the results prove that sedenion-based architectures are effective tools for complex spatio-temporal prediction. Future work will involve populating the currently unused sedenion dimensions with additional data sources, such as weather or calendar events, to further enhance prediction accuracy.
Summary updated