Matrix Factorization with e-commerce data example

Contents

1. Introduction
2. Dataset
3. Results
4. Todos

Introduction

---

A simple example to illustrate how the Matrix Factorization model can be used to build a Recommender System.

• Matrix_Factorization.ipynb: preprocessing data, training two kinds of Matrix Factorization models(SVD & NMF), and evaluating trained model
• recsystools.py: some tools that are useful to build and evaluate Recommender System models

In this example, we are going to predict which product a customer might be interested in based on past purchase records from an e-commerce site. Different from the usual movielens example, we are trying to predict user preference using implicit feedback data.

for a simple and intuitive explanation about Matrix Factorization, I recommend this amazing video of Luis Serrano:

https://www.youtube.com/watch?v=ZspR5PZemcs&t=5s

Dataset

---

Download dataset via URL below:

https://www.kaggle.com/carrie1/ecommerce-data

This dataset contains 540k transaction records of 4372 customers and 4070 products.

Results

---

parameter k refers to the dimension of the latent feature space. values are rounded to 2 decimal places.

methods	NDCG@10	Recall@10	training time
SVD (k=200)	0.120	0.104	6.5s
NMF (k=50)	0.081	0.071	16.8s

Todos

---

I am planning to make more examples treating Recsys models using VAE(Variational AutoEncoder), FM(Factorization Machine) algorithms.

References

https://github.com/dawenl/vae_cf/blob/master/VAE_ML20M_WWW2018.ipynb