Policy-aware cross-city learning: A framework for sustainable urban governance

Main text

Incorporating knowledge of policy transfer into urban governance frameworks fosters cross-city learning and facilitates a transition from prediction-based to policy-guided decision-making. This approach combines data with policy insights, expanding the scope of cross-city learning and fostering collaborative governance. Unlike traditional qualitative policy transfer studies, the envisioned policy-aware framework computationally translates policy contexts into quantitative representations for direct integration into machine learning, enabling automated strategy adaptation rather than relying solely on human interpretation.

The need for policy awareness in cross-city learning

Cross-city learning holds significant potential for advancing smart cities and fostering sustainable urban development, promising to make global urbanization processes more effective and sustainable.¹ However, its practical applications remain limited, requiring further exploration of effective implementation strategies.

Cross-city learning is fundamentally data driven in rationale and methodology. Conceptually, this data-driven approach is necessitated by the complexity of urban systems, inter-city heterogeneity, and the need for evidence-based governance at scale. Methodologically, it employs computational techniques, statistical analysis, and machine learning to extract transferable knowledge and quantify the effectiveness of strategies. These mechanisms enable cross-city learning to translate urban data into transferable knowledge, providing empirical foundations for the adaptation and implementation of urban policies.

Transfer learning has demonstrated remarkable success in addressing data sparsity and domain inconsistency across diverse fields, including computer vision, natural language processing, medical diagnostics, and autonomous driving.² Cross-city learning seeks to transfer knowledge from source cities—prediction models, decision strategies, and data distributions—to target cities for prediction, detection, and deployment tasks.³ Despite its promise, existing research predominantly focuses on data features and algorithmic performance, neglecting deeper contextual factors like policy regulations, social culture, and public acceptance. Consequently, relying solely on prediction accuracy or error metrics leads to real-world challenges: execution difficulties, social rejection, and inadequate adaptability.⁴

To address these challenges, we advocate for a policy-aware framework for cross-city learning. This framework accounts for variations in data distribution between cities and incorporates key contextual factors, such as policy environments, legal frameworks, social feedback, and cultural differences, into the learning models. By integrating these elements, the framework aims to generate actionable and contextually relevant strategic recommendations tailored to the specific needs of target cities.

A framework for policy-aware cross-city learning

Transitioning from purely data-driven approaches to policy-sensitive frameworks requires incorporating quantitative representations of policy and social environments into cross-city learning models—“policy-aware embedding.”⁵ We present a conceptual perspective on policy-aware cross-city learning and propose an integrative framework (Figure 1). By embedding quantitative representations and dynamically integrating policy and social factors, this framework ensures that strategy outputs extend beyond data pattern recognition to incorporate local regulations, implementation conditions, and socio-cultural acceptance.

Figure 1.

Policy-aware cross-city learning framework

Four steps with policy-aware integration. (1) S1: incorporates and quantifies policy/societal data alongside traditional urban data to capture critical local context. (2) S2: aligns data and quantified policy features between cities, integrating contextual nuances for coherence. (3) S3: embeds policy features into model training, optimizing for both technical performance and policy compliance. (4) S4: utilizes feedback on policy outcomes and social responses from the target city for iterative model adaptation and refinement.

In S1 of Figure 1 (data acquisition and preprocessing), we extract traditional features—infrastructure, environmental, and transportation data from source and target city datasets. To complement these, advanced techniques like text mining, knowledge graph construction, and survey analysis derive computable policy and social features from diverse sources, including policy documents, legal provisions, public opinion feedback, and socio-economic statistics. This process converts regulatory constraints, economic incentives, social acceptability, and cultural preferences into numerical indicators, forming a comprehensive “policy feature representation.” To address policy and social features heterogeneity across cities, we propose employing a series of standardization techniques. For instance, policy stringency can be quantified by analyzing keyword frequency, regulation mandates, and penalty severity in policy documents, creating a standardized index. Similarly, social acceptance becomes comparable to numerical indicators through statistical analysis of survey data. Furthermore, embedding techniques (e.g., Word2Vec or BERT) can convert policy texts into vector representations, enabling semantic alignment across cities. These methods enable the effective integration of heterogeneous data, providing consistent inputs for subsequent stages.

In S2 of Figure 1 (feature alignment), conventional feature alignment is extended by introducing “policy-aware alignment.” This step incorporates the target city’s local policies and social parameters, aligning with relevant data features. The model transcends data distribution similarities by integrating these elements, accounting for the target city’s legal frameworks, societal values, and implementation conditions. This enhanced alignment improves the adaptability and feasibility of the generated strategies.

In S3 of Figure 1 (model refinement and generalization), a unified representation of fused data and policy features undergoes further training to achieve robust “policy embedding.” Technically, policy embedding integrates into model training through several approaches. One straightforward method involves incorporating quantified policy and social features as additional inputs alongside traditional data features, making it suitable for modeling the explicit impacts of policy. Another approach employs multi-task learning, where one task optimizes predictive performance while another ensures policy compliance, using shared representation layers to balance objectives, which is particularly useful when weighing multiple goals. Alternatively, policy constraints are incorporated as regularization terms in the loss function, directly guiding the model toward policy adaptability during training, which is effective in contexts with stringent policy requirements. The choice among these methods depends on the specific cross-city learning task, data characteristics, and policy context.

An adaptive feedback cycle for continuous learning

Policy transfer is not instantaneous but a continuous, evolving, and adaptive process. Traditional cross-city learning methods often employ a “one-way” transfer model, delivering strategy recommendations from offline training. However, dynamic urban environments and variable social attitudes necessitate real-time monitoring and adaptive feedback to ensure the successful implementation of policies.

To address these challenges, we propose an adaptive feedback cycle. This cycle anchors in S4 of Figure 1 (deployment and continuous learning), where strategy recommendations are piloted within a limited scope in target cities. During this phase, policy outcome data are collected from various sources, including sensor networks, enforcement statistics, and public opinion surveys. These data capture key metrics, including environmental indicators, public satisfaction levels, enforcement costs, and compliance rates.

Operationally, reintegrating collected feedback into the learning model involves several technical pathways. Policy outcome data, such as environmental indicators or public satisfaction levels, can directly update policy and social feature representations in S1, serve as new objectives or constraints guiding model refinement in S3, or directly inform the tuning of deployment parameters in S4. The update cycle frequency is adaptable—rapid feedback loops are suitable for dynamic data, such as social media sentiment. In contrast, longer cycles (e.g., quarterly or annually) are ideal for slower institutional changes or cumulative environmental impacts.

The feedback collected during this process is reintegrated into the cross-city learning model, facilitating updates of policy, societal, and data feature representations. This iterative refinement enables the generation of strategies that are increasingly tailored to the unique conditions of the target cities. The framework facilitates deeper knowledge transfer grounded in practical adaptation by capturing and encoding lessons from real-world implementation outcomes (policy effectiveness and social responses) into the updated model. While the internal model refinement operates as a feedback loop, the system actively engages the target cities’ open environment to gather crucial feedback data. Transcending static, one-off transfer models fosters co-evolutionary paradigms where models and policy environments adapt through real-time feedback and continuous optimization.

Challenges and outlook

Integrating policy features into transfer learning models presents substantial challenges. These challenges stem from the inherent heterogeneity of policy structures across cities, cultural differences, and the difficulty of obtaining high-quality social and regulatory data. Moreover, the lack of standardized methodologies for quantifying and representing policy constraints and regulatory enforcement disparities complicates the creation of universal and scalable frameworks.

From a governance perspective, implementing policy-aware cross-city learning faces several critical challenges that require targeted solutions. Institutional fragmentation within cities often results in siloed data and decision-making processes, complicating the integration of comprehensive policies. Power imbalances between source and target cities can lead to inappropriate policy impositions rather than contextually sensitive adaptations. Additionally, temporal mismatches between rapid technological change and slower policy evolution create friction in real-time adaptation. To address these challenges, we propose establishing cross-departmental coordination mechanisms within city administrations, developing ethical guidelines that emphasize mutual benefit and context sensitivity, creating intermediary institutions as knowledge brokers between technical and policy domains, implementing regulatory sandboxes for controlled experimentation of transferred strategies, and engaging citizens through participatory governance structures that ensure public input shapes adaptation processes.

Additionally, we advocate establishing interdisciplinary collaboration mechanisms. Future research should prioritize technology-policy convergence through partnerships among academia, industry, and government entities. Such collaboration is essential for advancing cross-city learning models from theoretical constructs to practical applications, supporting the goals of smart cities and sustainable urban development.

Funding and acknowledgments

This paper was funded by the NSFC (nos. 72401174 and 52220105001); the Independent Research Project of the State Key Laboratory of Intelligent Green Vehicle and Mobility, Tsinghua University (no. ZZ-GG-20250403); and Tsinghua University (State Key Laboratory of Intelligent Green Vehicle and Mobility)-Hangzhou Airport Economic Demonstration Zone Joint Research Center for Integrated Transportation.

Declaration of interests

The authors declare no competing interests.

Published Online: June 10, 2025