Projet_06_OCR_Data_Scientist

Automatic Consumer Goods Classification

Overview

This project explores how to automatically classify consumer products from marketplace data using both text and images.

The work combines exploratory NLP, image-based analysis, supervised deep learning with transfer learning, and a small API extraction step to illustrate how external data can enrich a product catalog.

---

Dataset

The project uses a sample e-commerce dataset based on Flipkart product data, including:

• product titles
• product descriptions
• category labels
• and associated product images

The text notebook explicitly works on the file flipkart_com-ecommerce_sample_1050.csv and linked product images.

---

Method

The workflow of the project was:

1. study the feasibility of classification from text data
2. preprocess product descriptions with standard NLP cleaning
3. compare text representations such as Bag of Words, TF-IDF, and embedding-based approaches
4. study the feasibility of classification from images
5. build a supervised image classifier with CNN transfer learning
6. test an external API extraction example to collect complementary product-related data

---

Tools

• Python
• Pandas
• NumPy
• Matplotlib
• Seaborn
• NLTK
• Scikit-learn
• Gensim
• Transformers
• TensorFlow / Keras
• Requests
• Jupyter Notebook

---

Repository Structure

.
├── notebooks/
│   ├── 0_text_feasibility analysis.ipynb
│   ├── 1_images_feasibility analysis.ipynb
│   ├── 2_images_supervised_classification.ipynb
│   └── 3_extraction_by_API.ipynb
├── 4_extraction_API_102024.csv
├── 5_presentation_102024.pdf
└── README.md

---

Main Takeaways

This project helped me practice:

• text preprocessing for NLP
• exploratory clustering for classification feasibility
• image-based classification workflows
• transfer learning with CNNs
• and basic API-based data extraction

It was also a good introduction to multimodal product classification.

---

Limitations

This project has a few important limitations:

• the dataset is relatively small
• the work remains partly exploratory
• and the image/API parts could be pushed further for production use

A natural next step would be to build a more robust multimodal classifier combining text and image features in a single pipeline.

---

Author

Jérôme Le Gal
Data Science student