Court Judgments Prediction using Machine Learning

Supreme Court of Nigeria (SCN) appeal-case outcome prediction with an interpretable machine learning workflow.

Table of Contents

• 🚀 Project intro
• 📊 Dataset
• 🧰 Tools
• ⚙️ Install methods
• 🧪 Methodology
• 📈 Results
• 🔄 Project flow
• 📁 Project structure
• 📄 License

🚀 Project intro

This repository demonstrates an end-to-end machine learning workflow for predicting the outcome of Supreme Court of Nigeria appeal cases. The project combines dataset profiling, exploratory analysis, feature preprocessing, and classifier comparison in a single reproducible notebook-based pipeline.

The main goals are to:

• profile the legal case dataset and inspect its structure
• prepare categorical and numerical features for model training
• compare multiple classification models for case outcome prediction
• generate visual artifacts such as confusion matrices and correlation plots

📁 Project structure

Court-Judgments-Prediction-using-Machine-Learning/
├── CSE 445.ipynb
├── scn_appeal_cases_data.csv
├── datasetprofiling.html
├── Dataset description.pdf
├── IMAGES/
├── scripts/
│   └── download_judgement_flaticon.ps1
├── LICENSE
└── README.md

📊 Dataset

Dataset characteristics

• Source: Primsol Law Pavilion archive, distributed via Mendeley
• Records: 5,585 appeal cases
• Scope: Criminal and civil appeal matters from the Supreme Court of Nigeria
• Dataset link: https://data.mendeley.com/datasets/ky6zfyf669/1
• Reference material: A detailed data description is available in Dataset description.pdf

Features and target variable

The dataset contains both categorical and numerical attributes such as:

• appeal district and trial district
• offence type and sentence category
• counts of complainants, appellants, and witnesses
• the target label: scn_decision

These fields make the task suitable for supervised classification with interpretable preprocessing and evaluation.

🧪 Methodology

The notebook is organized into four clear stages that support repeatable model development and interpretation.

1. Data ingestion and profiling

• Load scn_appeal_cases_data.csv into pandas
• Validate data quality and missing values
• Generate an automated profile report with pandas-profiling
• Inspect class balance and feature distributions

2. Exploratory analysis and feature engineering

• Visualize distributions for categorical and numerical attributes
• Plot correlation matrices and pairwise relationships
• Identify informative legal attributes and reduce noise
• Encode categorical features and scale numeric values when needed

3. Training workflow

• Split the dataset into training and testing subsets
• Train a set of classifiers on the processed feature set
• Compare baseline, linear, tree-based, and neural models
• Models evaluated include:
  • DummyClassifier
  • LogisticRegression
  • DecisionTreeClassifier
  • RandomForestClassifier
  • SVC
  • GaussianNB
  • MLPClassifier
  • AdaBoostClassifier
  • BaggingClassifier
  • VotingClassifier

4. Evaluation and visualization

• Compute classification metrics: accuracy, precision, recall, F1 score
• Generate confusion matrices for each classifier
• Visualize model performance with graphs and heatmaps
• Compare model behavior across the same test split

🔄 Project flow

flowchart TD
    A[Load SCN appeal-case dataset] --> B[Profile and inspect data]
    B --> C[Engineer features and transform data]
    C --> D[Train classifiers on training data]
    D --> E[Evaluate results and visualize graphs]

Loading

🧰 Tools

The project is implemented using a Python-based data science stack:

• Python 3.x for data processing and modeling
• Jupyter Notebook for interactive experimentation
• pandas and NumPy for data handling
• matplotlib and seaborn for visualization
• scikit-learn for preprocessing, modeling, and evaluation
• pandas-profiling for automated dataset profiling
• VS Code as the development environment

⚙️ Install methods

Option 1: Python environment with pip

git clone https://github.com/shakil-sheikh/Court-Judgments-Prediction-using-Machine-Learning.git
cd Court-Judgments-Prediction-using-Machine-Learning
python -m venv .venv

# Windows
.venv\Scripts\activate

# macOS / Linux
source .venv/bin/activate

pip install --upgrade pip
pip install pandas numpy matplotlib seaborn scikit-learn pandas-profiling jupyter

Option 2: Conda environment

conda create -n court-judgment-prediction python=3.10 -y
conda activate court-judgment-prediction
pip install --upgrade pip
pip install pandas numpy matplotlib seaborn scikit-learn pandas-profiling jupyter

Run the notebook

jupyter notebook "CSE 445.ipynb"

📈 Results

The repository includes a rich set of model evaluation outputs and visual artifacts in the IMAGES folder. These include:

• confusion matrices for multiple classifiers
• feature correlation heatmaps
• count plots for case-related attributes
• distribution and relationship plots for the legal features

Model comparison summary

Model	Typical metrics	Analysis
Logistic Regression	Balanced accuracy, precision, recall, F1	Provides a stable baseline with interpretable coefficients for legal feature impact.
Decision Tree	Precision, recall, support	Captures non-linear relationships and highlights the most discriminative case attributes.
Random Forest	Accuracy, F1 score, ROC-AUC	Offers the best tradeoff between generalization and robustness across dataset splits.
SVM	Precision, recall, F1	Useful for higher-dimensional feature spaces and supports margin-based decisions.
MLP Classifier	Accuracy, training loss, validation score	Demonstrates the potential of a simple neural classifier for more complex feature interactions.

Key findings

• The notebook evaluates several classifiers rather than relying on a single model.
• Ensemble methods such as Random Forest generally show stronger overall performance and stable behavior.
• Interpretable models like Logistic Regression and Decision Tree make it easier to explain predictions in the legal context.
• Graphs and heatmaps provide visual evidence of feature importance, class separation, and model mistakes.
• Confusion matrix graphs are used to compare false positives and false negatives across classifiers.

Graphs and visual analysis

• Confusion matrix heatmaps: show where models confuse appeal outcomes
• Feature correlation graphs: reveal strong and weak relationships among legal attributes
• Class distribution plots: compare counts of outcome labels and reveal imbalance
• Model comparison figures: aggregate performance metrics across classifiers for easy visual ranking

In summary, the project provides a reproducible comparison of multiple machine learning models on the SCN appeal-case dataset and supports selecting the right classifier based on accuracy, interpretability, and robustness.

📄 License

This project is licensed under the MIT License. See the LICENSE file for details.