COMMUNIWAVE: A Machine Learning Model for Quantifying the Degree of Temporary Informal Behavior in Urban Communities

Urban planning often relies on top-down approaches that conflict with the spontaneous, bottom-up activities of residents—such as street vending or informal gatherings. Traditional methods for evaluating these behaviors, such as manual surveys or post-occupancy evaluations, are often inefficient and lack the precision needed for dynamic urban management. This study seeks to bridge this gap by developing a computational tool to quantify the Degree of Informal Behavior (DIB) in urban spaces.

The authors developed CommuniWave, a multi-stage machine learning pipeline. First, the Behavior Capture Net (BCN) uses the mmaction2 framework and a custom YOLOv10-based model (YLX) to detect specific informal actions in 10-second video clips. Second, the Behavior Eval Model (BEM) uses a random forest regression algorithm to score the intensity of these behaviors on a scale of 1 to 5. The model was trained on 180 video segments across three different road classes and validated against human ratings provided by 10 volunteers. The researchers also used SHAP values to interpret which specific behaviors most heavily influence the DIB scores in different urban environments.

CommuniWave successfully generates DIB fluctuation charts that allow managers to visualize how informal behavior changes throughout the day. The model identified that street vending is a primary driver of high DIB scores across various road types, while other behaviors like three-wheeled motorcycling or square dancing have context-specific impacts depending on the road class. Validation tests showed that the model's predictions align reasonably well with human assessments, offering a more efficient and continuous alternative to manual field surveys.

By providing a quantitative metric for informal behavior, CommuniWave enables urban designers and managers to move beyond anecdotal evidence. This allows for more refined, data-driven decision-making that can accommodate resident needs while maintaining territorial resilience. The model's open-interface design allows other researchers to retrain the system for different cultural contexts or specific community needs, potentially reducing the friction between formal urban planning and the lived reality of city residents.