Overview

This project aims to develop a modular recommender system using multiple algorithms. It includes data loading and preprocessing, model training, and evaluation using various collaborative and content-based filtering techniques. Futher more, it includes a web app that allows users to interact with the recommender system.

Dataset downloaded from: Amazon-reviews-2023

Table of Contents

• Overview
• Project
• Setup
• Usage
• Models
• Evaluation

Project

Project Structure

recommender_system/
├── data/
│   └── datasets_amazon/
│       └── Automotive_5_core.csv
├── models/
│   ├── __init__.py
│   ├── recommender.py              # Models and Content-based recommender
│   ├── data_loader.py              # Data loading and preprocessing
│   └── evaluator.py                # Model evaluation
├── config/
│   └── config.yaml                 # Configuration file
├── streamlit_app/
│   └── app.py
├── main.py                         # Main script to run the project
├── pyproject.toml                  # Poetry configuration file
├── README.md                       # Project documentation
└── mlflow.db                       # MLflow tracking database

Setup

1- Clone Repository:

git clone https://github.com/cristianBMJ/Recommendation_System_MLflow.git
cd Recommendation_System_MLflow

2- Install Poetry if you don't have it:

curl -sSL https://install.python-poetry.org | python3 -

3- Install the dependencies:

poetry install

4- Activate the virtual environment:

poetry shell

Usage

Run the Tracking Server

Run the MLflow tracking server using the following command:

mlflow server --backend-store-uri sqlite:///mlflow.db --default-artifact-root ./artifacts

Run

Run the main script to load data, train models, and evaluate their performance:

python Main.py

Run Data Web apps with Streamlit

Run the app.py script from folder streamlip_app:

export PYTHONPATH=$(pwd)
streamlit run streamlit_app/app.py

Models

This project implements the following recommendation algorithms:

SVD (Singular Value Decomposition), KNNBasic (k-Nearest Neighbors), SlopeOne, CoClustering. Additionally, a content-based recommender system is included.

Evaluation

Model performance is evaluated using the functions provided in models/evaluator.py. The results are printed to the console and can be tracked using MLflow.

Troubleshooting

Splitting Data:

The provided data splitting methods (Leave-One-Out and Absolute Timestamp) are not directly supported by the Surprise library due to its data format requirements. In the next step, we will explore preprocessing techniques to make your data compatible with Surprise.