As expected, the biggest urban centers have the highest concentration of outdoor advertisements, but the viewership distribution was not equal, showing the potential for better distribution of outdoor advertisements.
This challenge tackled a less discussed problem in cities: visual pollution. Many cities are flooded by countless outdoor advertising panels. It is known that visual aspects are crucial in the urban planning process since each modification can generate obstruction to urban elements. Besides the city aesthetics, it is also possible to argue that too much outdoor advertising can decrease well-being and increase social inequality. It is clear that reducing outdoor advertisements can create a positive impact on the city. The goal of this challenge is to do so with a minimum impact on the reduction of the audience.
Create a model that optimizes the number and location of the outdoor advertising positions to minimize the visual impact in urban environments. Better dimensioning and integrating outdoor advertising positions in cities should also be considered.
United Nations SDG
GOAL 11: Sustainable Cities and Communities
The following datasets were provided to the participants:
List of outdoor advertisements in the entire territory of mainland Portugal with the location, number of average views, maximum visibility, and height provided by PSE.
Using domain knowledge, it was noted that some panels were missing in the dataset. It was also recommended to include in the dataset the type of outdoor advertisement, advertisement cost, and the direction the advertisement panel is facing.
More concretely, on the dataset, it would also be essential to have the time frame to which the dataset is related (e.g., does it include the COVID period?) and how the different metrics are calculated (e.g., daily views and max visibility).
Due to the nature of this challenge, very different approaches were proposed by the competing teams. One proposal was to use a clustering algorithm (K-Means and K-ProtoTypes) to cluster the location of advertisement placement and to remove the advertisements with fewer views from the cluster.
Other teams decided on a metaheuristics approach. One approach was based on a local or neighborhood search that optimizes for the spread of advertisement panels while simultaneously maintaining the number of views high. Another approach was to train a regressor with a Gradient Boosting model to predict the number of views at a particular location, generate random geographical points, and for those points predict the number of views. Later on, optimize on a set of parameters the points to be selected (e.g., spread from concentration points or a number of views).
It was noted that if it were possible to predict the number of unique viewership for each geographical point (e.g., by using mobility data), it would be possible to create an optimization model which could be used for optimizing the distribution of outdoor panels while also minimizing the number of panels.