Predicting Neighborhood Safety

The accessibility of commercial data has now exceeded that of federal data in certain cities and categories. For example, there are now Airbnb reviews datasets available in cities where no official crime dataset or a crime events map available to the public. Our team saw an opportunity from this observation and decided to create a machine learning model that enables us to generate a crime level map based on Airbnb dataset. Ultimately, we would like to offer the public a crime map indicating safety levels in cities where there isn’t one.

Our final output is the interactive heat density map which allows users to locate areas by inputting zipcodes in the search bar and view safety levels predicted by our best trained model. The map is created with folium package and it is layered on openstreetmap which also provides detailed information of the area such as street names and landmarks. To create the model, we generated crime level measurements and used features related to neighborhood safety. We learned from research papers and created a 1-100 spectrum of crime score normalized by population density. For text review data, we utilized NLP techniques including bag of words and sentimental analysis to select top important key phrases as prediction features. Also, we applied word embedding and google’s word2vec model to measure the distance between reviews text and self-selected words suggesting neighborhood safety. Comparing to baseline model, we saw our best tuned random forest model has a near 20% increase in the cross validation accuracy for LA data and near 50% increase for Austin data. Overall we think our current results can help residents and travellers to make more informed decisions regarding safety when living or exploring cities without public crime data. We hope that our project will be useful or inspiring anyone interested in analyzing the area safety data to make a difference for our community.