This repository contains the code for project Rakuten based on data issued by Rakuten Challenge and developed during Data Scientist training at DataScientest.
The cataloging of product listings through title and image categorization is a fundamental problem for any e-commerce marketplace. The traditional way of categorizing is to do it manually. However, this takes up a lot of employees’ time and can be expensive.
The goal of this project is to predict a product’s type code through its description and image for the e-commerce platform Rakuten.
This project was developed by the following team :
You can browse and run the notebooks. You will need to install the dependencies (in a dedicated environment) :
pip install -r src/requirements.txt
You can also see a summary presentation of the project in the formats pdf presentation.pdf and pptx presentation.pptx.
The application has the following elements :
- Frontend that contains the web interface created with application Streamlit
- Backend that contains the API created with FastAPI
- Monitoring that contains the application Tensorboard
The application should be run within a Docker container [Streamlit + FastAPI + Docker = ♥].
To run the Docker containers use the following command:
docker-compose up --build The app should then be available at localhost:8501 and API documentation should be available at localhost:8111/docs.
Rakuten Challenge contains :
84 916observations27categories to be determined0duplicate data- One color image per product
- Image size is
500x500pxin JPG format
The sample of the data:
The challenge presents several interesting research aspects due to :
- the intrinsic noisy nature of the product labels and images
- the typical unbalanced data distribution
- the large size of the data
- the description of the product in different languages
We use a supervised approach for the one-label classification problem with an imbalanced distribution of labels. Therefore, the metric used in this challenge to rank the model's performance is the weighted-F1 score.
The development of the best model contains the following steps :
- Creation Text Classifier
- Creation Image Classifier
- Fusion Text Classifier and Image Classifier
Text Classifier :
- Contains text prepprocessing and text vectorization using Natural Language Processing (NLP).
- Based on the Neural_Embedder text model.
Image Classifier :
- Bases on the CNN image model
- Use transfer learning with the MobileNetV2 model loaded with pre-trained weights on ImageNet.
The fusion model has the following architecture :
The fusion model uses the image model to categorize products where the text model underperformed. The global weighted-F1 score is 82.2% and all categories exceed the 55% score.
Model density analysis :
