RNN’s & Generating Fake Police Reports

My Freshman year in college I had my bike stolen, and when I went to report it I realized that my college provided an in depth ‘Police Blotter’ which contained a crime log that went back several years. I realized the opportunity for a fun project and some analysis, and within a few hours had built a scraper to get me all of the crime reports since 2014.

While my initial project was just some exploratory data analysis and modeling of crime across campus, it wasn’t until a few years later that I decided to try to do some deep learning on the dataset. With each crime report there is a free text portion of the report, containing a short description of the crime. I decided to implement a Recurrent Neural Network to try and generate fake reports that were similar to the real reports written by the Police Department. While there were only ~1800 text reports for me to train on, I couldn’t bring myself to think that “having more data” would necessarily be a good thing.

The majority of crime on campus relates to bike theft


Real Reports

1871 Total Crimes Reported Since 2014

  • An individual reported an financial fraud incident regarding a supposed part-time employment offer via e-mail.
  • An individual reported a financial fraud incident involving the “BYU Public Directory” information.
  • An individual reported graffiti on a restroom stall.  Custodial was contacted for clean-up.
  • An individual reported a verbal threat made towards an individual.
  • An individual reported an e-mail scam of a company offering a job where an larger than needed amount of money is sent on a fake check.
  • Individuals reported clothing items taken from the hall dryers that had been left over night.

Fake Reports

  • An individual reported a phone calls from the area. An individual reported a paint returned the internet. The officer was located the area and nothing with individuals. The prover indicated the area
  • An individual reported a citation. The individuals were taken from the bike racks. An individual reported a credit card. An individual report a bicycle taken from the area. On officers arrival individ
  • An individual reported a the area was toll. An individual reported a laptop was taken from the bike racks. The individual was located and escorted from the area. An individual reported a secured bicyc
  • An individual reported a secured bicycle taken from the area. The bicycle has been entered and cited and assault area. national theft database. An individual reported an unsecured segim and items cont
  • An individual reported seeing a cable lock taken from the bike racks. The bicycle has been entered on a closed area. An officer responded to a report of an individuals harassment and closed area. On

Other Interesting Trends

While I enjoy math and coding, what I enjoy most is the flexibility it gives me to solve interesting problems and work with so many different people in various fields. The RNN was cool to build, but what was most enjoyable was meeting with the Police Chief and officers about trends in campus crime that I explored. The structured data of crime occurrence, report date, location, etc. lead to interesting conversations with the police force on how to better serve and help our campus community.

Even if my bike was stolen, I’m glad I was able to turn it into a learning opportunity to leverage data to make a difference, and I am grateful for the work of the police to keep our campus safe.

How long does it take a victim to report a crime?

Very telling story, this brought a lot of discussion on how one could help those suffering from abuse or sexually related crimes feel comfortable to come forward faster

When do crimes happen most often during the week?

The x-axis is day of week, with Monday being number 1, thefts (particularly bike theft) seems to have the worst day on Monday and then it slowly decays.