Data-Driven Approach for
Rapid Response
to Noise Complaints

Authors:

Duke Sung Hoon Yang, Qinyu Goh, Sam Ovenshine, Siddhanth Shetty, Zoe Martiniak

Supported by:

Center for Urban Science & Progress, New York University
New York City Department of Environmental Protection

In recent years, the growing volume of noise complaints in New York City has increased response times and hindered enforcement of the city’s noise code by the Department of Environmental Protection. In this Capstone project, we used supervised learning to assign complaint priority and produced an interactive heatmap to optimize response to noise complaints.

Noisy New York City

The noise code was created to reduce excessive and unreasonable noises. The city's enforcement of the noise code is primarily complaint-based. Residents register a complaint via the 311 hotline and the complaint is directed to the relevant agency.

NYPD

The New York Polic Department (NYPD) handles noise complaints involving people to people interactions.

Clubs or Bars
Parks
Stores or businesses
Streets or sidewalks
Vehicles

NYC DEP

New York City Department of Environmental Protection (NYC DEP) handles complaints relating to equipment or repeating noise.

Construction
Pets and Animals
Food vending vehicles
Air conditioners
Motor vehicles

The Problem

Noise complaints have increased by 20% since 2014. Enforcement of the noise code has been hampered due to inability to handle sheer volumes.

In 2018, the DEP could not verify noise was occuring in 55% of noise complaints.

The primary motivation of this project is to provide a data-driven approach to:

Increase efficiency of complaints successfully closed

Decrease complaint response times

Data Exploration

Construction Data

Census Demographic

Land Use Data

Weather

311 Noise Complaints

Our Approach

To accomplish our goal, we explored the following methods:

PREDICTION

Use supervised learning to
optimize future scheduling.
PRIORITIZATION

Use supervised learning
to optimize current scheduling.

Prediction

LSTM Neural Network

A Long short-term memory (LSTM) Neural Network model was developed with 49 spatiotemporal features including weather and after-hour variance permits, as well as 28 autoregressive features.

Overall, the NN provided a good estimate of the true standardized mean but failed to capture spikes in complaint variance. Our findings suggest that a NN model can learn the baseline volume (including periodicity) of complaints within targeted spatial regions but struggles with predicting specific complaint clusters.

More Results

Optimization

Using unsupervised learning to cluster complaints

To optimize the inspector's response to current backlog of open complaints, we originally used the ST-DBSCAN method to cluster complaints within fixed temporal and spatial constraints.
With discussion with the spinsors, we discovered that an interactive heat map for visualiation was preferred over this clustering method.

Prioritization

Classifying Priority with Random Forest

We trained a random forest classifier on 311 complaint outcomes from 2016 and tested on July 2019 complaints. A probability score was assigned to each complaint based on its likelihood of enforcement. We optimized for model recall based on sponsor preference.
In evaluation, the model precision was 9.3% and recall was 91.1%.

Interactive Dashboard

Our main goal was to develop a data-driven approach to optimize the DEP's scheduling of noise complaint inspections. The two methods which we explored were to optimize the response to the current backlog of open complaints by prioritization, as well as optimizing future scheduling with prediction. Based off of our research and feedback from the sponsors, we determined that optimizing the response to current complaints by visualizing priorities of current complaint backlog on a heatmap is the most useful and practical approach.

About the Authors

Background and interests of each contributor to the project.

Sung Hoon Yang

Worked on building a forecasting model for DEP noise complaints using PyTorch.

Qinyu Goh

Qinyu Goh is an aspiring urban data scientist who loves a good adventure.

Sam Ovenshine

Sam studied economics at Occidental College and currently works in data at Zocdoc, a healthcare technology startup.

Siddhanth Shetty

With experience as a business analyst, business operations and urban planning intern, I look forward to the challenges in working with different data and expanding my skills as a data scientist.

Zoe Martiniak

Zoe Martiniak is a Junior Business Analyst who worked as an Environmental Engineer in a previous life.

Want to learn more?

Check out our shared repository on GitHub.

Credits

Thank you to our mentors and sponsor!

Charlie Mydlarz

I am a Senior Research Scientist at NYU’s Center for Urban Science and Progress (CUSP), the Music and Audio Research Laboratory (MARL), and an Engineering Lead in the NYU Immersive Audio Group.

Vincent Lostanlen

Vincent Lostanlen is a postdoctoral researcher affiliated to the Cornell Lab of Ornithology, and is currently visiting both the MARL and NYU’s Center for Urban Science and Progress (CUSP).

Mark Cartwright

Mark Cartwright is Brooklyn-based computer scientist and musician. He is currently a Research Assistant Professor in NYU’s Department of Computer Science and Engineering with affiliations to NYU’s Music and Audio Research Lab and the Center for Urban Science and Progress.