Graph-Based Recommendation System for an eCommerce website

Overview

This project developed a Graph-based recommendation system for an eCommerce platform, leveraging user purchase and search history to recommend products. By transforming user interaction data into graph embeddings and using ANN-FAISS model, the system aimed to enhance product discoverability and user experience.

Data Overview

The dataset comprised user interactions on the eCommerce website, spanning 9 attributes. This was transformed into a graph format for analysis and modelling.

Technology Stack

• Programming Language: Python
• Libraries: pandas, DuckDB, numpy, pecanpy, gensim, plotly, UMAP, ANN, FAISS
• File Management: Parquet files for efficient data storage
• Database Management: SQL for data querying

Approach

1. Problem Understanding

Researched the requirements for creating a recommendation system that uses graph theory to suggest products based on historical user data.

2. Data Preparation

Extracted relevant data from the provided dataset and optimised it for the analysis.

• Libraries Used: pandas, numpy, DuckDB

3. Deepwalk and Node2vec Model Training

• Libraries Used: Pecanpy, gensim
• Main Steps:
  • Performed graph random walks to generate sequences for embeddings.
  • Trained the Deepwalk and Node2Vec models using the gensim library.

4. Result Analysis and Visualization

• Libraries Used: UMAP, plotly
• Main Steps:
  • Categorized products into hierarchical levels for detailed analysis.
  • Applied UMAP for dimensionality reduction of embeddings to visualize data clusters.
  • Used Plotly for interactive 2D and 3D visualization of product clusters.

5. Embedding Vector Search with FAISS

• Libraries Used: FAISS
• Main Steps:
  • Initialized the FAISS library for efficient similarity search.
  • Used FAISS to find similar products based on embedding proximity.
  • Generated product recommendations for users based on similarity scores.

Folder Structure

• Model: Stores the trained Deepwalk and Node2Vec models.
• Notebooks: Contains Jupyter notebooks used in the project:
  1. Data Optimisation
  2. Data Exploration and Data Analysis
  3. Deepwalk and Node2Vec Model Training
  4. Graph Construction
  5. Result Analysis
  6. Embedding Vector Search and Recommendation with FAISS
• Scripts: Python scripts are available for reproducibility and automation.
• requirements.txt: Lists all necessary libraries with versions. Install these requirements to ensure compatibility.