A predictive model for Chicago crime statistics.
Chicago is a big city and is known in part for it's crime. Knowing Chicago has crime is one thing, but being able to dig deeper to see the correlations in crime statics, and better yet, forecast future crime statistics is another feat entirely. Our goal was to provide useful insights in Chicago crime statistics and make predictions on those statistics. Our hope is that these insights and predictions will spread awareness and ultimately keep people safe.
What it does
AccuCrime predicts the amount of crime on any given day and can forecast up to ten days worth of crime statistics given any past or present date to start from.
How we built it
The first step in our process was to collect as much useful data as we could. To do this we collected and combined multiple open online data sets. This master dataset was then divided into separate datasets for each Chicago district. Once our data was cleansed and ready for use, we built two separate prediction models for each district. The first uses a Deep Neural Network and can predict the total amount of crime on any day given the number of abandoned cars, temperature, alley lights out, street lights out, and number of vacant buildings on that day. The second model is a Long Short Term Memory, Recurrent Neural Network that looks at past data up to a specific point and generates / predicts / forecasts future crime statistics.
Challenges we ran into
We originally planned to only use the deep neural network to predict the amount of crime on any day given the input labels. Like with most AI models, optimizing for good accuracy wasn't easy. We got the deep neural network to a passable accuracy, but wanted to do better. Our solution was to use a recurrent (LSTM) neural network. Unlike the deep neural network, the recurrent neural network is able to forecast crime statistics based on previous data. This model proved to be much more accurate.
Accomplishments that we're proud of
The machine learning models and web app are (not surprisingly) our proudest accomplishments. Other than that, and the small accomplishments we made along the way, solid teamwork is something we're proud of.
What we learned
We learned a lot about data collection, data cleansing, and data preprocessing. Data cleansing is like front end development of data science - it's tedious, frustrating, but still needs to be done. At times we wanted to pull our hair out at not being able to figure out the simplest of things, but we eventually got it working. We learned tons about deep neural networks, recurrent (LSTM) neural networks, and how to put it all together and make it easily accessible through a nice web app.
What's next for AccuCrime
AccuCrime will live on as the #1 hackathon project of Boilermake 2017. We hope others find it's predictive capabilities useful and at the least hope it spreads awareness of the usefulness of machine learning.