Skip to content

비타민 9기 컨퍼런스 (NLP/추천시스템)

Notifications You must be signed in to change notification settings

bokyung124/bitamin-conference

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

91 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Title

Abstract

NLP Recommendation System Based Fashion E-Commerce Platform Demo built with Pytorch, Huggingface, React, Express, FastAPI.



Introduction

This GitHub repository contains the source code for our fashion e-commerce website demo, which serves various AI technologies that could be implemented in the web fashion platform. Our website features recommendation systems based on user likes, the ability to find similar items, and review summaries. These AI models are served with FastAPI, while the frontend is built with React and the backend with ExpressJS. We utilized Selenium to gather data such as images and item details, ensuring our AI models are provided with relevant information. The website has an intuitive and user-friendly interface that allows users to easily navigate and find their desired items.

Data

We extracted 45,000 product information and photos, and 700,000 product review informations using Selenium. For more information, visit this repository.
[Data column] Name, Brand, Likes, Sales, Review Score, Price

Recommendation System

Our recommendation system consist of three parts.

  • Image + Category Created an image feature (embedding vector) extractor that incorporates category labels using a MobileNet, whose purpose is to speed up the training and inference.
  • Image + Tag Created an image feature (embedding vector) extractor that incorporates tags using the Show and Tell model. To process the Korean, which is a agglutinative language, we used skip-gram based embedding.
  • Similarity Concatenated the Image + Category and Image + Tag models and calculated Euclidean distance to recommend the 5 most similar items.

[Fig. 1] Model Structure

[Fig. 2] Similarity Result

Review Summarization

Our review summarization system consist of two parts.

  • Text Review to Score : Using koBERT's cls tokens, we finetuned the model to classify review text of an item to score, 1 to 5.
  • Review Keyword Extraction : Using Khaiii tokenizer, seed word, Skip-gram, We finetuned koBERT with 5M sentence with 54 words.

Web Development

  • Frontend : React JS with Typescript
  • Backend : Express JS with MongoDB
  • AI Model : FastAPI

About

비타민 9기 컨퍼런스 (NLP/추천시스템)

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 43.5%
  • Jupyter Notebook 33.1%
  • TypeScript 17.6%
  • JavaScript 3.6%
  • Shell 1.3%
  • CSS 0.6%
  • HTML 0.3%