orianrivlin/chicago-crime-latitude-model

Assignment #2 – Classification, Regression, Clustering & Evaluation

Video Presentation

https://www.youtube.com/watch?v=ASl7FbvRaSg

Author: Orian Rivlin
Course: Data Science – Assignment #2
Goal: Build a complete ML pipeline – EDA, regression, feature engineering, clustering, improved models, classification, HF models, and insights.

---

1. Project Overview

This project uses a cleaned subset of the Chicago Crimes dataset (~19,500 rows) to build:

• A regression model predicting crime latitude.
• An improved model using feature engineering and clustering.
• A classification pipeline converting latitude into 3 geographic regions.
• Full evaluation, insights, HF deployment, and video walkthrough.

Target variables:

• Regression: Latitude
• Classification: Region class (0=South, 1=Central, 2=North)

---

2. Dataset Description

The dataset includes:

• Crime type
• Location description
• Police district, ward, community area
• Coordinates (X, Y, Latitude, Longitude)
• Time information (Date parsed to Month, DayOfWeek, Hour)
• Flags (Domestic, Arrest)

Final shape after cleaning: 19,496 rows × 22 columns

EDA Summary

Key Steps

• Removed missing coordinates, duplicates.
• Parsed date column into time features.
• Removed extreme latitude outliers (IQR-based).

Insights

Latitude by Police District

Domestic vs Non-Domestic Locations

Latitude Stability Over Years

Crime Type Frequency

Latitude by Crime Type

---

3. Baseline Regression Model

Model: Linear Regression with one‑hot encoded categorical features.

Performance

• MAE: ~1.20e‑05
• RMSE: ~1.76e‑05
• R²: 0.99999996

Latitude is nearly a linear function of the Y coordinate → high accuracy expected.

Plots

---

4. Feature Engineering

Added:

Temporal Features

• Month
• DayOfWeek
• Hour
• IsWeekend
• YearsSince2012

Geographic Feature

• Distance from city center (DistFromCenter)

Clustering Feature

Applied K-Means (k=6) on scaled spatial & time data → added ClusterID.

---

5. Extended Regression Model

Includes all engineered features + clustering.

Performance

• MAE: 9.5e‑06
• RMSE: 1.31e‑05
• R²: 0.99999997

An improvement over the baseline.

Plot

---

6. Model Comparison

Model	MAE	RMSE	R²
Baseline Linear Regression	1.20e‑05	1.76e‑05	0.99999996
Improved Linear Regression	9.53e‑06	1.31e‑05	0.99999997
Random Forest	7.8e‑05	1.21e‑04	0.999998
Gradient Boosting	5.19e‑04	6.84e‑04	0.999938

Winner (Regression):

Improved Linear Regression – simplest, fastest, most accurate.

---

8. Classification Pipeline

8.1 Creating Classes

Latitude divided into 3 balanced regions using 33% and 66% quantiles.

Classes:

• 0 – Southern region
• 1 – Central region
• 2 – Northern region

Balanced distribution → accuracy is meaningful.

8.2 Models Trained

• Logistic Regression
• Random Forest Classifier
• Gradient Boosting Classifier

8.3 Evaluation

Logistic Regression

Accuracy: ~97%
Errors are between adjacent regions.

Random Forest

Accuracy: 100%

Gradient Boosting

Accuracy: 100%

8.4 Classification Winner

Gradient Boosting Classifier – perfect accuracy, compact, efficient.

A pickle file was exported:
best_classification_model.pkl

---

9. HuggingFace Repository

HF Repo Link

Contains:

• README
• Notebook
• Regression pickle model
• Classification pickle model
• materials/

---

Final Notes & Insights

• Latitude is extremely predictable from spatial coordinates.
• Feature engineering provided meaningful improvements despite high baseline accuracy.
• Clustering added interpretable structure for classification models.
• Both regression and classification tasks achieved near-perfect performance.