el-bahnasawi/analyzing-u.s.-crime-data

Project Overview 📝

This repository contains two separate notebooks focusing on exploratory data analysis (EDA) and regression analysis using machine learning (ML) models.

Notebook 1: Exploratory Data Analysis (EDA) 📊

• Addressing questions through visualizations and hypothesis testing.
• File: Project.ipynb

Notebook 2: Regression Analysis and ML Models 🧠

This notebook is divided into several key steps:

1. Data Validation and Preparation 📝

   • Ensuring data availability.
   • Reading the dataset.
   • Handling NaN values by categorizing them into numerical and categorical features. Then using KNN imputer for numerical features, Simple Imputer with most frequent for categorical features.

2. Feature Engineering 🔧

   • Identifying categorical features and further categorizing them into boolean, nominal, and ordinal.
   • Encoding nominal features using OneHotEncoder and ordinal features using OrdinalEncoder.

3. Regression Analysis 📈

   • Investigating feature correlations via heatmap.
   • Conducting linear regression, assessing the model's performance with R2 score, and visualizing coefficients.

4. Machine Learning Models 🤖

   • Checking dataset balance.
   • Establishing a baseline accuracy.
   • Implementing various ML models, evaluating them with training and testing accuracies, confusion matrices, precision, recall, and AUC-ROC score.

Enhanced Model Building Process 🛠️

Outlined in models.py, the following comprehensive approach is undertaken:

Model Categories:

• Dimensionality Reduction Models: These models employ techniques to reduce feature dimensionality while retaining crucial information.
• Feature Selection Models: Focused on selecting pertinent features while discarding redundant or correlated ones.

1. Dimensionality Reduction plus Classifier:

   • Pipeline Setup: Begins with StandardScaler for feature normalization, ensuring uniformity across features.
   • Dimensionality Management: using PCA and UMAP to reduce dimensionality while keeping valuable information
   • Classifier Integration: Incorporating Random Forest and Gradient Boosting classifiers to leverage ensemble learning for robust predictions.
   • Hyperparameter Tuning: RandomizedSearchCV facilitates efficient hyperparameter optimization, complemented by cross-validation to improve model generalization.

2. Feature Selection Methods plus Classifier:

   • Recursive Feature Elimination (RFE): Prioritizes feature selection based on importance to model performance, further refining the feature set.
   • Correlation-based Feature Elimination: Subsequent utilization of Variance Inflation Factor (VIF) identifies and eliminates correlated features, enhancing model stability and interpretability.
   • Efficiency and Simplicity: These models offer lower computational complexities compared to their dimensionality reduction counterparts, ensuring suitability for scenarios prioritizing interpretability and computational efficiency.

Models Comparison 📊

Model	Technique	features	Estimators/Neighbors	Min Samples Split	Max Depth	Training Accuracy	Test Accuracy
Model 1	PCA with RF	20	100	5	30	99.96%	97.12%
Model 2	PCA with GB	20	5	-	20	98.51%	92.55%
Model 3	UMAP with RF	20	-	2	20	100%	93.28%
Model 4	RFE, VIF, RF	40	50	-	4	99.26%	99.45%
Model 5	RFE, VIF, XGB	40	12	-	1	100%	100%