frkornet / Seattle_Terry_Stops

Analysis of Seattle Terry Stop data set, and a classification model for predicting an arrest.

Seattle_Terry_Stops

Analysis of Seattle Terry Stop data set, and a classification model for predicting an arrest.

The file Coronet_Consultants_Seattle_Terry_Stops.pdf contains the presentation for the projectr. It summarizes the main findings and recemmended next steps.

The Python notebooks are in the directory notebooks. A short description of the notebooks in this directory is given below:

1. data_exploration.ipynb: This notebook analyzes the Terry Stop data to get a sense for the data we are dealing with. It was the first notebook we created

2. log_reg_part_2.ipynb: This notebook is similar to logistic_regression.ipynb except that it uses a random forest classifier on a subset of the data (false positives) to see if we can improve the performance on this data subset. The best accuracy we get is 77 % (with a random forest classifier).

3. logistic_regression.ipynb: This notebook runs the scorecard model on the data with added features. It contains a third degree polynomial scorecard on 4 features: Officer Gender, Reported Time, weapon_present, and Initial Call Type. The most important feature for the model is Initial Call Type. The accuracy of the model is 71% (AUC = 86%). Note that a naive version would score around 97 % accuracy. So, this is not a very successful model. Also note that the trarget is the physical arrest.

4. model_new_target.ipynb: implements a general arrest as target instead of physical arrest. The accuracy achieved is 80% and that is in line with the result a naive prediction would have by predictiong no arrest.

5. trees.ipynb: This notebook implements a decision tree in a scorecard for the physical arrest target. The achieved accuracy is 91 %. Better than a scorecard based on logistic regression alone, but still well below a naive implementation by predicting no arrest which will achieve an accuracy of around 97 %.

6. umap_model.ipynb: this notebook uses UMAP to draw a 2D version of the data. It visualizes overlap of the zero and one target clusters. This explains why it is hard to build a good model. The notebook also implements a threshold for the scorecard model. By using a higher threshold we are able to achieve slightly under 84 % better than the naive prediction baseline.

The data files are in the sub-directory data.

1. Terry_Stops_added_features.csv: The same as Terry_Stops_raw.csv data set but augmnented with features.

2. Terry_Stops_raw.csv: The original data setr that we used for the project. The data (including a desc ription of the fields) can be found at https://catalog.data.gov/dataset/terry-stops

3. subset.csv: The false positives of Terry_Stops_added_features.csv.