I made many models, all of which I tried to accurately predict whether a crime would be attempted or completed, based on the features. Of the models that worked well for this dataset, I hyperparameter tuned and tried to optimize on recall. The reason for this was because we wanted to make sure the attempted crimes would be accurately predicted. It was very hard to accurately predict attempted crimes, however, and much easier to predict the completed crimes. This was partly because this dataset was very imbalanced and there were 719,618 completed crimes, compared to 9,550 attempted crimes. I took a sample of this dataset, of 100,000. Then, I made a copy of the cleaned dataset, and increased the sample size to 500,000 and saw how the classification report changed. The results were better overall with the increased sample size, as was expected.

To use previous NYC crime data to make accurate predictions about whether crimes were attempted or completed

-Will determine where to invest in resources to help reduce crime in these areas, based whether crimes were attempted or completed

NYPD Open Data Crime Complaints from 2006-2019. After data cleaning, there were 729,168 crimes of which I took samples of 100,000 and 500,000. The sample of 500,000 yielded better results overall.

Complaint Date and Time Crime and Crime Description Suspect and Victim Demographics Police Precinct Location and Location of the Crimes Borough Where Crime Occurred and Police Station Borough

We created new columns for details of the month, year, and duration of the crime, day of the week

We wanted to look at attempted vs completed crimes and see which features were the most accurate predictors of whether the crime was attempted or completed

Logistic Regression with GridSearch Random Forest with Random Oversampler Random Forest with Random Undersampler Random Forest with SMOTE, with GridSearch

-It is very difficult to predict attempted crimes, and more easy to predict completed crimes based on this dataset -Harassment and robbery are two crimes that are very ambiguous as to whether they were attempted or completed. It is particularly difficult to predict these two crimes -A business recommendation is to allocate more resources to the locations with the most completed crimes

We tried the original sample size of 100,000 for the models. Because increasing the models sample size increases the recall score, we reran the models with a sample size of 300,000.

Multinomial Bayes XGBoost with SMOTE KNN with GridsearchCV