Toward equitable, transparent, and collaborative human mobility computing for smart cities

Main text

Cities constitute vast and intricate systems in which diverse entities (e.g., people, vehicles, and roads) interact collaboratively and dynamically. Analyzing and understanding the core elements of people in complex urban mobility systems provides a crucial means for smart city applications. In the era of global digitization, the exponential surge in geolocation data linked to human travel has profoundly transformed our understanding of human travel behavior. Data science enables us to comprehensively capture human mobility characteristics¹ at the individual and population levels; these characteristics include regularity, diversity, and predictability. Empowered by data science, human mobility computing research has shaped a closed-loop scientific ecosystem involving data training models, model serving applications, and application feedback data (Figure 1).

Figure 1.

Framework for equitable, transparent, and collaborative human mobility computing

Here, we highlight three pivotal challenges, equitable, transparent, and collaborative human mobility computing, and delve into solutions to address them. Our emphasis on equity entails mitigating discrimination or bias throughout the data life cycle, from collection to analysis and modeling. Transparency requires that the processes and outputs of human mobility computing be easily understandable and interpretable. Collaboration underscores the need for mutual understanding and enhancement between users and machine algorithms in user-centered interactive travel systems.

Equitable data modeling is imperative

Today, vast amounts of GPS trajectory and social media data are used to analyze the travel patterns and scaling laws of individuals and populations. These insights have become instrumental in informing urban planning and business decision-making. Multi-source data facilitate the recreation of the spatiotemporal panorama of individuals' activity chains within and between cities. Ensuring that the insights gained from data science regarding human travel patterns are trustworthy and unbiased is critical. Unequitable data modeling can result in misleading decision-making instructions for managers, exacerbating unequal access to resources.

To this end, at the data collection level (Figure 1, box 1), ensuring the integrity and diversity of the data used for modeling is crucial but often tricky because of users' privacy concerns. High-quality data should accurately reflect the population diversity, encompass different socio-economic backgrounds and geographic regions, and exhibit fewer missing fields across different samples. Therefore, by leveraging technologies such as blockchain to enhance data anonymization, encrypted transmission, and access control, users can confidently share more comprehensive travel information for human mobility research. Under rigorous data review, introducing fairness measures at the individual and group levels² is indispensable for evaluating the representativeness of data distribution and the consistency of expert labeling. However, no unified standard exists for the definition of fairness, and we have a limited capability to identify and consider only a few sensitive attributes, such as income, race, and gender. In particular, the sensitive boundaries of some attributes (e.g., the division between the rich and poor) are difficult to clarify. This challenge can be ameliorated at the algorithmic level. On the one hand (Figure 1, box 2), data-driven methods can be used to identify which groups are underrepresented, and adaptive sampling strategies can be devised to adjust data sampling probabilities from unbalanced datasets, facilitating the self-balancing of samples for model training. On the other hand, the fairness measure can be incorporated to constrain and extend the model’s objective function, identify sensitive attributes containing unfairness characteristics, and design adversarial learning networks to achieve model self-correction (Figure 1, box 3). These solutions aim to produce the same model output for different sensitivity groups, assuming that all other variables remain consistent.

Transparent mobility computing is challenging

Transparent model inferences imply that managers can scrutinize the features employed by mobility models in their decisions, ensuring that the model outputs are reasonable and equitable. Unfortunately, a prevalent challenge lies in community scientists' perception of data-driven human mobility models as black boxes. Unlike physical models, data-driven models automatically extract travel features from large-scale datasets, yielding prediction results characterized by high-dimensionality and abstract features.

One of the most promising and straightforward approaches to promoting transparent mobility computing research involves the integration of explainable artificial intelligence technology³ (Figure 1, box 4). This allows the derivation of understandable rules that clearly explain the dependencies between the input variables and outputs. Meanwhile, universal scaling laws can be incorporated to construct a physics-informed mobility model, thereby enhancing its interpretability. Moreover, large language models (LLMs), such as ChatGPT, have emerged that demonstrate remarkable contextual awareness and generative reasoning capabilities. This has inspired us to leverage LLMs in inverse learning of the travel representation results of traditional data-driven models and explaining the inference reasons in an easy-to-understand manner. Furthermore, we propose integrating multi-modal travel data (e.g., text, sequence, images, and videos) using chain-of-thought technology.⁴ By conducting domain-adaptive fine-tuning on multi-modal LLMs⁴ (Figure 1, box 5), we can guide LLMs to output prediction results and explanations directly. Finally, struggling to understand the model itself may not be as beneficial as understanding humans. By employing neuroscience theory⁵ to delve into the underlying mechanisms of human travel choices¹ (Figure 1, box 6), we can build brain-like intelligent human mobility decision-making models that enhance transparency in travel modeling from a brain-inspired perspective.

Collaborative human-machine decision-making is lacking

With the advancement of technology, various human mobility models empowered by data science have been integrated into diverse software and platforms, enabling automatic interaction and collaboration with users for travel-related decisions. Consequently, research on data-driven human mobility computing needs to focus on understanding how humans interact with these tools, how automated recommendations or suggestions affect people’s decision-making processes, and how integrating human and machine (model) decisions can provide better solutions for travel choices.

Nonetheless, existing data-driven human-machine collaboration faces substantial disparities in human and machine expression and decision-making mechanisms.⁵ An immediate need is to combine human brain decision-making mechanisms and brain-like intelligence methods to develop representation methods catering to humans and machines. For example, based on diverse wearable physiological recorders, employing neuroscience methods such as electroencephalography experiments allows us to analyze human travel selection mechanisms (Figure 1, box 7). Furthermore, human-machine collaborative decision-making emphasizes the cooperative participation of humans and intelligent systems (machines) in joint management decisions. There is a drive to analyze the mechanism of human-machine collaborative decision-making and develop human-like user interfaces, interactive decision support, and information-sharing systems (Figure 1, box 8) to achieve dynamic human-machine collaboration in complex travel scenarios. Additionally, throughout the human-machine interaction process, the machine continually receives user feedback while encountering diverse users with different preferences and evolving scenarios. Therefore, integrating a dynamic continuous learning module (Figure 1, box 9) is vital to enhance the machine’s flexibility and empower it to intelligently adapt to humans. For example, the initial model of an intelligent travel-planning platform may learn only from a limited set of users' typical modes, routes, and destination selection behaviors. Nevertheless, during actual operations, user behavior and the travel environment may be influenced by various dynamic factors, such as traffic conditions and special events. In such scenarios, dynamic continuous learning allows the system to capture these changes and make the necessary adjustments.

Conclusion

Data-driven human mobility computing offers profound insights into travel behavior at previously unattainable scales and granularities. Nevertheless, computational models have many limitations regarding equity, transparency, and collaboration, and new insights must be introduced to propel methodological and conceptual advancements. Approaches such as equitable data modeling through adversarial learning, transparent mobility computing utilizing LLMs, and the construction of travel interaction systems based on brain-like intelligence and continuous learning are promising solutions for addressing these challenges. These solutions promote the effectiveness, inclusivity, and sustainability of human mobility systems in smart cities.

Published Online: July 2, 2024