ML Projects

A collection of course and hobby experiments exploring a variety of machine-learning techniques, from classic NLP to modern deep learning and reinforcement learning. Each project lives in its own directory and can be run independently.

Repository Table of Contents

• Repository Overview
• Projects
  • Author Classification – Naive Bayes NLP
  • Credit Card Fraud Detection – Custom Neural Network
  • Joke Recommendation – Collaborative Filtering
  • Raccoon Detection – Keras YOLOv2
  • Mario AI Agent – Monte Carlo Reinforcement Learning

Repository Overview

The repository root contains individual project folders plus a standalone Java agent:

Path	Description
Author Classification - naivebayLearning - NLP/	Naive Bayes text classifier that distinguishes between disputed Federalist Papers.
CreditCardFraudDetection-NN/	Scratch-built feed-forward neural network for transaction fraud detection.
Joke Recommentation Unsupervised Learning/	Matrix factorisation recommender trained on the Jester joke ratings dataset.
keras-yolo2-master-Raccoon Classification/	YOLOv2-based object detector fine-tuned for raccoon imagery.
ExMarioAgent.java	Reinforcement-learning Mario agent submitted to the 2009 RL-Competition.

Each project section below explains its layout, dependencies, configuration, and how to launch the code.

Projects

Author Classification – Naive Bayes NLP

• Overview
• Directory Layout
• Dependencies
• Configuration
• Running the Classifier
• Key Implementation Notes

Author Classification – Naive Bayes NLP Overview

A Python 3 implementation that tokenises Federalist Papers essays and trains a unigram Naive Bayes classifier to attribute disputed essays to Hamilton, Madison, or Jay. The pipeline covers tokenisation, stop-word removal, stemming, dictionary smoothing, and log-probability scoring.

Author Classification – Naive Bayes NLP Directory Layout

Path	Purpose
main.py	Entry point that orchestrates training and evaluation on the chosen author folders.
tokenizer.py	Handles tokenisation, stop-word removal, and stemming logic.
dictionary.py	Builds smoothed word-frequency dictionaries and computes log-probabilities.
fileGet.py	Splits documents into training/unknown sets and loads corpora.
Hamilton/, Madison/, Jay/, Disputed/	Default corpora of essays grouped by author. Binary subsets (bin*) are also provided for experiments.

Author Classification – Naive Bayes NLP Dependencies

Install Python 3 plus the Natural Language Toolkit. The project requires the Punkt tokenizer and stop-word corpus:

pip install nltk pandas
python - <<'PY'
import nltk
nltk.download('punkt')
nltk.download('stopwords')
PY

Author Classification – Naive Bayes NLP Configuration

The dataset folders are configured at the top of main.py. Adjust the variables if you place corpora in different directories, or switch to the binary subsets by uncommenting the provided examples inside the script.

Author Classification – Naive Bayes NLP Running the Classifier

Run the default experiment with:

python3 main.py

By default the script analyses the Hamilton, Madison, Jay, and Disputed folders. Update the folder names (line 14 of main.py) to evaluate alternative corpora.

Author Classification – Naive Bayes NLP Key Implementation Notes

• Uses log-probabilities with add-one smoothing to avoid underflow.
• Unknown-token handling reserves a random split of each author’s texts to measure distinctive vocabulary usage.
• Outputs include per-author probabilities that incorporate prior publishing frequency.

Credit Card Fraud Detection – Custom Neural Network

• Overview
• Directory Layout
• Dependencies
• Running the Trainer
• Experiment Tips

Credit Card Fraud Detection – Custom Neural Network Overview

Implements a neural network from first principles (no deep-learning frameworks) to classify credit-card transactions as fraudulent or legitimate. The network architecture, activation functions, and training loop are defined in pure Python.

Credit Card Fraud Detection – Custom Neural Network Directory Layout

Path	Purpose
neural_net.py	Command-line interface for training and evaluating the model.
neuron.py	Defines the neuron and network classes powering forward/backward passes.
trainCredit.csv, testCredit.csv	Default Kaggle-style credit card datasets. Binary subsets (trainBin.csv, testBin.csv) are available for quick experimentation.

Credit Card Fraud Detection – Custom Neural Network Dependencies

Requires Python 3 with NumPy for numerical operations.

Credit Card Fraud Detection – Custom Neural Network Running the Trainer

The script expects training and evaluation files, neuron count, and a decision threshold:

python3 neural_net.py trainCredit.csv testCredit.csv 28 0.5

• 28 specifies the neuron count per layer and must match the feature count in the CSVs.
• 0.5 sets the classification threshold; adjust as needed for new datasets.

Credit Card Fraud Detection – Custom Neural Network Experiment Tips

• The script skips the first row and column to match the supplied dataset formatting.
• Tune the iterate variable inside neural_net.py to change the number of passes over each training row.
• When experimenting with other datasets, start by increasing the threshold toward 0.9 to compensate for different class imbalances.

Joke Recommendation – Collaborative Filtering

• Overview
• Directory Layout
• Dependencies
• Running the Recommender
• Experiment Tips

Joke Recommendation – Collaborative Filtering Overview

A matrix-factorisation recommendation engine that predicts joke ratings using collaborative filtering on the Jester dataset. The implementation evaluates reconstruction accuracy across multiple optimisation iterations.

Joke Recommendation – Collaborative Filtering Directory Layout

Path	Purpose
project3.py	Entry point implementing stochastic gradient descent for latent-factor learning.
testJ.csv	Lightweight sample dataset for quick experiments.
jester_dataset_3.zip	Full Jester dataset archive; unzip to access the complete ratings matrix.

Joke Recommendation – Collaborative Filtering Dependencies

Install Python 3 along with the required scientific stack:

pip install pandas numpy matplotlib

Joke Recommendation – Collaborative Filtering Running the Recommender

Provide the dataset path and optional hyperparameters. Missing arguments fall back to sensible defaults.

python3 project3.py testJ.csv 0.005 0.001 10 500

Arguments correspond to [dataset] [learning_rate] [regularisation_lambda] [latent_features] [iterations].

Joke Recommendation – Collaborative Filtering Experiment Tips

• Keep the learning rate small (≈0.005) to avoid divergence.
• Use testJ.csv for rapid iteration, then switch to the unzipped Jester dataset for full-scale evaluation.
• Visualisation hooks rely on Matplotlib; ensure you have a display backend if plotting on a remote server.

Raccoon Detection – Keras YOLOv2

• Overview
• Directory Layout
• Dependencies
• Configuration
• Training
• Inference
• Notes

Raccoon Detection – Keras YOLOv2 Overview

Forks the keras-yolo2 project to train an object detector that finds raccoons in images. The workflow covers anchor generation, training on annotated images, and running inference with the trained weights.

Raccoon Detection – Keras YOLOv2 Directory Layout

Path	Purpose
config.json	Central configuration for dataset paths, training hyperparameters, and model architecture.
train.py / predict.py	CLI scripts for training and inference, respectively.
gen_anchors.py	Utility for recomputing anchor boxes from the training annotations.
images/	Sample images for inference sanity checks.
experimental/, examples/, backend.py, frontend.py, utils.py	Core YOLOv2 implementation components.

Raccoon Detection – Keras YOLOv2 Dependencies

Install Python 3 with the libraries listed in requirements.txt (Keras, OpenCV, imgaug, tqdm, h5py, etc.). Use a virtual environment for isolation:

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Raccoon Detection – Keras YOLOv2 Configuration

Edit config.json before training:

• Set model: "Tiny Yolo" to match the supplied lightweight backbone.
• Update train_image_folder and train_annot_folder to point at the extracted raccoon dataset (images + Pascal VOC XML annotations).
• Leave pretrained_weights blank unless resuming from existing weights.
• Choose a descriptive saved_weights_name for the output model file.

Raccoon Detection – Keras YOLOv2 Training

Generate anchors (optional) and train the network:

python gen_anchors.py -c config.json
python train.py -c config.json

Use the provided hyperparameters (e.g., 2 training iterations, 10 epochs, 3 warm-up epochs) as a starting point.

Raccoon Detection – Keras YOLOv2 Inference

After training, run detection on an image or video:

python predict.py -c config.json -w path/to/best_weights.h5 -i path/to/input

Detections are written alongside the input media with a -Detected suffix.

Raccoon Detection – Keras YOLOv2 Notes

• Ensure the raccoon dataset is resized consistently; YOLO performs best when training images share similar dimensions.
• The project is optimised for live video streams, so highly variable still-image sizes may reduce accuracy.

Mario AI Agent – Monte Carlo Reinforcement Learning

• Overview
• Dependencies
• Running the Agent
• Implementation Notes

Mario AI Agent – Monte Carlo Reinforcement Learning Overview

ExMarioAgent.java customises the Rutgers RL-Competition 2009 Mario agent. It augments the baseline heuristic controller with Monte Carlo state-value updates, episodic memory, and JSON-driven policy initialisation to solve Mario levels via reinforcement learning.

Mario AI Agent – Monte Carlo Reinforcement Learning Dependencies

• Java 8 or newer.
• RL-Glue codec (org.rlcommunity.rlglue) to interface with the Mario environment.
• JSON processing library available on the classpath to parse ref_export.json.

Mario AI Agent – Monte Carlo Reinforcement Learning Running the Agent

Compile the agent alongside the RL-Competition environment and launch via the RL-Glue agent loader:

javac ExMarioAgent.java
java org.rlcommunity.rlglue.codec.util.AgentLoader edu.rutgers.rl3.comp.ExMarioAgent

Ensure ref_export.json and the RL environment resources are accessible on the classpath or working directory.

Mario AI Agent – Monte Carlo Reinforcement Learning Implementation Notes

• Tracks 13 binary state features covering monsters, pits, power-ups, and terrain topology.
• Maintains policy, reward, and iteration tables sized at 2^13 × 12 for Monte Carlo updates.
• Stores per-episode state/action/reward sequences to refine the policy after each run.

---

Have fun exploring the projects! Contributions and experiments are welcome—feel free to open issues or pull requests with improvements.