Boosting Fine-Grained Urban Flow Inference via Lightweight Architecture and Focalized Optimization

Yuanshao Zhu, Xiangyu Zhao, Zijian Zhang, Xuetao Wei, James Jianqiao Yu · 2025

Read Boosting Fine-Grained Urban Flow Inference via Lightweight Architecture and Focalized Optimization by Yuanshao Zhu, Xiangyu Zhao, Zijian Zhang, Xuetao Wei, James Jianqiao Yu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-07098.txt.

Introduction

Yuanshao Zhu and four co-authors infer fine-grained city traffic flows with a model small enough to deploy. Estimating flows at high spatial resolution supports planning and intelligent transport, but leading models are heavily over-parameterized, and standard loss functions handle the strongly skewed distribution of urban flows poorly. Their answer pairs a lightweight design with a redesigned loss.

The design, PLGF, progressively fuses local detail with global context, capturing both street-level variation and citywide dependencies. DualFocal Loss supervises in two spaces and focuses adaptively on hard-to-predict regions. Across four real-world cases the method set state-of-the-art accuracy while shrinking model size by up to ninety-seven percent against high-performing rivals. Under equal parameter budgets it improved accuracy by over ten percent, and code is available.

Catalog record

CREATORYuanshao Zhu, Xiangyu Zhao, Zijian Zhang, Xuetao Wei, James Jianqiao Yu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE47 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-07098

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Boosting Fine-Grained Urban Flow Inference via Lightweight Architecture and Focalized Optimization Yuanshao Zhu, Xiangyu Zhao, Zijian Zhang, Xuetao Wei, James Jianqiao Yu Abstract Fine-grained urban flow inference is crucial for urban planning and intelligent transportation systems, enabling precise traffic management and resource allocation. However, the practical deployment of existing methods is hindered by two key challenges: the prohibitive computational cost of over-parameterized models and the suboptimal performance of conventional loss functions on the highly skewed distribution of urban flows. To address these challenges, we propose a unified solution that synergizes architectural efficiency with adaptive optimization. Specifically, we first introduce PLGF , a lightweight yet powerful architecture that employs a Progressive Local-Global Fusion strategy to effectively capture both fine-grained details and global contextual dependencies. Second, we propose DualFocal Loss, a novel function that integrates dual-space supervision with a difficulty-aware focusing mechanism, enabling the model to adaptively concentrate on hard-to-predict regions. Extensive experiments on 4 real-world scenarios validate the effectiveness and scalability of our method. Notably, while achieving state-of-the-art performance, PLGF reduces the model size by up to 97% compared to current high-performing methods. Furthermore, under comparable parameter budgets, our model yields an accuracy improvement of over 10% against strong baselines. The implementation is included in the https://github.com/Yasoz/PLGF. 1 Introduction Fine-grained urban flow data provides the foundational insights for modern intelligent transportation systems and smart city infrastructures, enabling applications like precise traffic management and responsive urban planning ( 17 ) . However, acquiring such high spatial resolution data directly through dense sensor deployment is often infeasible due to prohibitive long-term costs for equipment, operation, and maintenance ( 18 ) . To address this, Fine-Grained Urban Flow Inference (FUFI), which infers fine-grained flow maps from available coarse-grained observations, has emerged as a pressing and cost-effective solution for developing sustainable and economically viable smart cities ( 12 ; 16 ) . To accurately infer fine-grained urban flow, early FUFI methods draw from the super-resolution techniques of computer vision, where urban flow maps can be viewed as an image. Foundational works like UrbanFM ( 6 ) and UrbanPy ( 10 ) established the core paradigm, introducing deep learning frameworks with crucial spatial constraints and pyramid architectures to handle the mapping between coarse and fine-grained flows. Follow-up efforts are mostly focused on refining architectural designs for better city-wide traffic profile capturing ( 24 ) . More recently, the field has shifted from pursuing pure accuracy to addressing practical deployment challenges, such as catastrophic forgetting in dynamic systems ( 19 ) , robustness to noisy data ( 22 ) , and cross-city knowledge transfer ( 20 ) . Figure 1: The proposed PLGF addresses efficiency barriers and optimization bottleneck issues with a lightweight architecture design and focalized optimization. Despite these significant advancements, the very pursuit of higher accuracy has given rise to two critical limitations that hinder practical deployment, as shown in Figure 1 . First, an efficiency barrier has emerged. The push for higher performance has led to increasingly complex and over-parameterized architectures to capture intricate spatio-temporal dependencies ( 2 ) . The scale of model parameters has increased from 1 million to 100 million. This trend towards “model bloat” incurs extreme computational costs for both training and inference, posing significant challenges for real-world applications. Second, a persistent optimization bottleneck remains. Existing models are typically optimized with conventional regression losses, such as Mean Squared Error, which treat all prediction errors equally. By failing to account for the highly skewed, non-uniform nature of urban flow data, this generic approach leaves the optimization process unfocused, limiting the ultimate accuracy and robustness potential of the model, especially when dealing with high-variance samples ( 6 ) . To address these fundamental limitations, we propose a unified framework that rethinks both architectural design and optimization strategy. First, we introduce PLGF ( P rogressive L ocal- G lobal F usion), a lightweight yet powerful architecture designed for parameter efficiency. Specifically, PLGF adopts an efficient progressive local-global fusion framework and a context-aware integration mechanism. This design enables it to capture complex multi-scale spatial dependencies with remarkable parameter efficiency, while ensuring that the entire process is subject to modulation by specific spatio-temporal contexts. In addition, we propose the DualFocal Loss, a novel loss function for focalized optimization. By integrating dual-scale (linear and logarithmic) supervision with a difficulty-aware focusing mechanism, our loss allows the model to adaptively concentrate on hard-to-predict, low-flow regions and high-variance samples that are often ignored by traditional methods. The main contributions of this paper are as follows: • We propose PLGF, a novel, lightweight, and parameter-efficient architecture. It effectively captures multi-scale spatial dependencies based on a progressive local-global fusion and context-aware strategy, thus significantly reducing the computational cost. • We introduce DualFocal Loss, a universal and plug-and-play loss function with a difficulty-aware focusing mechanism. It can be flexibly applied to various FUFI models, improving both accuracy and robustness. • Extensive experiments on 4 real-world scenarios demonstrate that PLGF achieves over 10% performance improvement under the same parameter budget and reduces parameter count by approximately 97% while maintaining state-of-the-art performance. 2 Related Work Fine-grained urban flow inference seeks to recover high-resolution urban flow maps from coarse-grained observations. The foundational paradigm was established by UrbanFM ( 6 ) , which introduced a distributional upsampling module enforcing strict spatial constraints, ensuring that the sum of fine-grained flows matches their corresponding coarse-grained region. Subsequent works advanced FUFI from several perspectives. UrbanPy ( 10 ) proposed a cascading pyramid architecture for progressive upsampling, improving scalability to large-scale urban maps. DeepLGR ( 7 ) further enhanced spatial modeling by introducing a dual-path network that jointly learns global and local features. Alternative approaches, such as FODE ( 24 ) and UrbanODE ( 23 ) , leveraged neural ordinary differential equations to capture continuous urban dynamics and address numerical instability. More recently, research has shifted to practical deployment challenges in dynamic, real-world environments. CUFAR ( 19 ) addressed catastrophic forgetting in dynamic environments via an adaptive knowledge replay strategy, while UNO ( 2 ) proposed data-free incremental learning using neural operators for privacy-preserving, scale-invariant modeling. Robustness to data quality has also been explored: multi-task frameworks have been developed for simultaneous missing data imputation and inference ( 5 ) , and denoising strategies have been integrated to enhance resilience to noisy sensor inputs ( 22 ) . To tackle data scarcity in target cities, cross-city transfer learning methods have been introduced to leverage knowledge from data-rich source cities ( 21 ; 20 ) . Despite these advancements, most existing approaches remain limited by large model sizes and high computational costs. Furthermore, optimization strategies tailored to the highly skewed and non-uniform nature of urban flow data are still lacking. In this work, we address these gaps by proposing a lightweight architecture and a focused optimization strategy, aiming to improve both efficiency and adaptability for real-world FUFI applications. 3 Preliminary This section formally defines the key concepts and the problem of FUFI, providing the foundation for our methodology. Urban Flow Map. Following standard practice, a city or a region of interest is partitioned into a uniform I\times J grid map based on longitude and latitude. At a given time interval, the traffic flow across the whole area is represented by an urban

Read the complete text