According to the European Environment Agency (EEA) report on environmental noise in Europe, health risks are posed when the population is exposed to increased noise. Some vulnerable groups have been specifically identified:
However, not only are humans affected by noise, but the report also shows that biodiversity negatively affects terrestrial and aquatic species. This indicates that there are several benefits to reducing noise pollution in cities. One source of noise pollution is the recreational nightlife noise, which comes from loud conversations during nighttime on the streets.
The city of Torino has been studying the noise levels of recreational nightlife noise in the San Salvario area, which is home to many bars and clubs. They have installed several sensors to measure the noise levels with records since 2016.
Predict future noise levels and, if possible, explain complaint trends that can be attributed to leisure noise levels.
GOAL 11: Sustainable Cities and Communities
The following datasets were provided to the participants:
The amount of data used by the teams also varied. Some teams decided to use fewer datasets to develop the solution. Other teams incorporated data points such as weather from OpenStreetMaps, holidays in Torino, academic calendar, COVID restrictions, football matches in Torino, and business locations. One team encoded the COVID measures into six different categories and did other feature extraction such as converting the night before a holiday, Friday, and Saturday into holiday days.
The teams had two criticisms towards the dataset: a larger dataset regarding the number of people at a specific location would have been helpful for a better estimation of how the number of people can influence the noise levels. The other criticism is regarding the police complaints: Many complaints have missing hours of complaint.
Other interesting data to include when modeling mentioned in the submitted solutions included socio-economic data, local events, business opening hours, the more precise location of complaints, a better context of the complaints (e.g., party next door, party on the street), and more sensors in the streets.
Most teams conducted an extensive EDA to understand the data by looking at daily averages and how the different variables were correlated. One team noted that it is vital to use logarithmic averages for calculating averages during a specific period. However, there were different modeling approaches for predicting noise levels: some used tree-based algorithms (such as XGBoost, Random Forest, and linear regression for baseline model), others used time series (SARIMAX and Moving Average) and Artificial Neural Network (ANN) such as CNN and LSTM (Long Short-Term Memory).
One team experimented with GAM (Generalised Additive Models) to associate sound intensity with the features. They have also developed a bivariate multimodal distribution from gaussian kernels. Another team focused on predicting outliers in data to detect unexpected changes in noise.
Regarding the feasibility study for predicting complaints, while some teams did a theoretical analysis, others developed models by applying tree-based algorithms. One team used a K-NN algorithm to obtain patterns characterizing demographics in order to derive those profiles with a higher propensity to complain.
During the initial analyses, it was found that the obvious assumption holds true: COVID (see Figure 1), weekends and holidays, weather, and football games influence the noise levels.
A surprising discovery was that most complaints happened during the week when the noise level was lower than on the weekend. It was also found that many of the complaints are logged in the morning during low noise levels.
While some teams attempted to predict complaints, the results varied. The general conclusion is that it is impossible to predict complaints based on noise levels, as the current data needs more quality.
Product
One team proposed to create an app where Torino’s residents could notify the police anonymously, and the city could collect data about complaints. Another team suggested creating an annoyance score to enrich the complaints dataset by combining the probability of noise exceeding the threshold level, causing annoyance, and causing a complaint.
The teams saw that these models could predict noise levels and apply preventive measures during more problematic situations.
One team recommended creating a traffic light system to inform the authorities about the upcoming noise level changes. Another noted that it is essential for the model to be explainable, as it can give a better intuition to the authorities of where intervention might be needed.
