Machine Learning Approaches to Predict Crime Rate in the City of Los Angeles
Published in Summer Undergraduate Research Fellowship (SURF), 2023
Abstract
This research poster explores the use of machine learning (ML) algorithms to predict the crime rate in the City of Los Angeles.
The study compares the performance of four ML models — Clustering, Linear Regression, Support Vector Classifier (SVC), and CatBoost — to identify the most effective predictive framework for understanding crime dynamics.
Methodology
- Clustering: The K-Means algorithm was used to group similar crime data points. The optimal number of clusters (K=5) was determined using the elbow method, balancing model simplicity and interpretability.
- Linear Regression: A classical regression approach was employed to capture linear relationships between crime rates and socioeconomic predictors, validated through diagnostic testing.
- SVC (Support Vector Classifier): This supervised learning model aimed to find the optimal hyperplane for classifying crime patterns. Dimensionality reduction and feature selection techniques were applied to enhance generalization.
- CatBoost: A gradient boosting algorithm optimized for categorical data. It achieved the highest predictive accuracy among all models, demonstrating robustness to overfitting and superior interpretability.
Key Results
- The CatBoost model achieved an accuracy of 91.37%, outperforming SVC, Linear Regression, and Logistic Regression.
- The confusion matrix revealed high accuracy in negative-class detection and strong consistency in prediction across validation sets.
- The findings show that machine learning can effectively support urban safety analytics and guide evidence-based policing strategies.
Discussion
This study demonstrates the potential of machine learning to enhance crime prediction and prevention strategies.
By combining supervised and unsupervised learning techniques, the research highlights how advanced analytics can help law enforcement agencies identify risk patterns, allocate resources efficiently, and strengthen data-driven policy decision-making.