LAPD Crime ML Models — Diego Cabezas

Data science thesis / portfolio study predicting whether an LAPD crime report looks violent or non-violent, using public Los Angeles crime data (roughly 2020–2024).

Start here if you’re not a Python person:
Friendly study explainer (open in browser) — plain-English summary of the question, data, pipeline, and verified results.

Results teaser (from the saved notebook)

Model	Accuracy	Notes
Random Forest (n_estimators=100)	~87.6%	Best overall; violent-class F1 ≈ 0.69
Neural net (Keras MLP)	val ≈ 87.4%	Close to the forest on overall accuracy
Logistic Regression	~81.9%	Strong on non-violent; weak violent recall (~0.20)

Unsupervised (2024 subset): K-Means silhouette 0.554 vs DBSCAN 0.443 on UMAP embeddings.

Full tables, caveats, and how to read the notebook: docs/study-explained.html.

Source of truth

• Notebook: LAPD_Crime_ML_Models.ipynb
• This docs pass did not change the notebook. The .ipynb remains the canonical code + metrics source.

Links

• Open in Colab (optional — run in the browser)
• Presentation
• LAPD Crime Data (City of LA)

What’s inside the study

1. Clean & explore ~963k public incident rows
2. Build a violent / non-violent label and engineered time/area/victim/weapon features
3. Unsupervised structure discovery on the 2024 slice (PCA, UMAP, K-Means, DBSCAN)
4. Supervised models + a small deep-learning network
5. Interactive prediction playgrounds in the notebook

Author

Diego Cabezas — data science thesis / portfolio piece.