Leveraging Variational Information Bottleneck for Fine-Grained Urban Traffic Flow Inference

Accurately inferring fine-grained traffic flow from coarse-grained observations can significantly reduce traffic monitoring costs while enhancing urban traffic analysis and decision-making. Existing methods typically rely on direct deep-network fitting, which struggles to effectively handle noise and redundancy in multimodal inputs. To address this issue, we propose a novel multimodal fine-grained traffic flow inference framework based on the Variational Information Bottleneck (VIBFI). By imposing information-theoretic constraints, VIBFI explicitly compresses redundant content while preserving prediction-relevant factors, enabling the learning of more robust spatial distributions of fine-grained traffic flow. Specifically, we first construct multimodal representations encompassing road structure, external factors, and traffic flow; then model the spatial influence of temporal and meteorological conditions through discrete embeddings and grid-aligned mappings; and finally capture local flow patterns via residual convolutions. Importantly, we introduce the Variational Information Bottleneck (VIB) into each modality as an information regularizer to suppress modality-specific but prediction-irrelevant redundancy, thereby strengthening the most critical representation factors for fine-grained inference. After VIB compression, the fused multimodal features are processed through short-range convolutions to encode local dynamics and a Transformer to capture global long-range dependencies, producing structurally consistent and detail-preserving fine-grained flow maps. Extensive experiments on multiple real-world datasets demonstrate that VIBFI consistently outperforms state-of-the-art methods across diverse urban scenarios.