Background

Nowadays, the retail e-commerce share has been growing with a incredibly fast pace around the world. For US, in 2016, online sales of physical goods amounted to 360.3 billion US dollars and are projected to surpass 603.4 billion US dollars in 2021. Despite this fast growth, United States actually rank behind several countries in terms of e-commerce sales as percentage of total retail sales - in 2016, almost a fifth of China’s retail sales occurred via the internet, compared to only 8.1 percent in the United States. The UK, South Korea, and Denmark are also ahead of the U.S. in terms of retail e-commerce share.

Cdiscount.com, France’s largest non-food e-commerce company, generated nearly 3 billion euros last year. While the company already sells everything from TVs to trampolines. The list of products is still rapidly growing. By the end of this year, Cdiscount.com will have over 30 million products up for sale. This is up from 10 million products only 2 years ago. Ensuring that so many products are well classified is a challenging task. Currently, Cdiscount.com applies machine learning algorithms(NLP) to the text description of the products in order to automatically predict their category. As these methods now seem close to their maximum potential, Cdiscount.com believes that the next quantitative improvement will be driven by the application of data science techniques to images.

Data Source and size

In this project, the original image dataset is provided on Kaggle by Cdiscount.com. It consist of more than 15 million images at 180x180 resolution of almost 9 million products (half of the current catalogue) which fall under more than 5000 categories (detailed categories).

Since I only have about 3 weeks to finish the project and don’t have access to the powerful GPUs, I have sampled two sub-datasets with size of 100k images (about 1% of the original dataset). One subset was randomly sampled while the other one was stratified sampled based on the top level categories (I will explain why I did this later). By doing this, I have simplified the original problem to categorizing 100k images into around 50 top level categories ( about 1% of the scope of original problem in terms of number of images and number of categories).

I hosted in AWS,

Modeling Steps

• Transfer Learning:

1. To solve this

• Ensemble Modeling:

1. WIP

2. WIP

Product Classification Examples

Future Works

In the future, I would like to:

1. Ensemble the NLP and Image models together to improve the classification accuracy.
2. Given more time and resource, I would train the model using the whole original dataset and predict on more product categories.

Project Presentaton Slides