Individual Submission Summary
Share...

Direct link:

Interpretable Machine Learning Algorithm to Forecast Recidivism

Thu, Nov 17, 11:00am to 12:20pm, L404, Lobby Level

Abstract

The prediction power of machine learning algorithms has proven to be highly accurate for different kinds of settings. For the NIJ Recidivism Challenge, I performed Logistic regression, Random Forest Classifier, XGBoost, LightGBM, and Catboost algorithms and evaluated the performances of each algorithm. After this, stacking was used to make one generalized prediction.

To understand the mechanics of the training model, I used the SHAP Values technique, which breaks down a prediction to show the impact of each feature. This analysis revealed that job-related variables like the number of jobs per year and percent of days employed have the highest correlation to recidivism. The idea comes together with Kevin Schnepel's paper (2018), where he examines "good jobs" and asks how their presence influences recidivism. Prior criminal history records such as prior misdemeanor and violation charges were important in the XGboost model, while the average number of days between drug tests helped achieve good predictability in the Catboost and LightGBM models.

Author