Brandon Alves

Context

This project is part of the Shopee - Product Matching Kaggle competition. The goal of the competition is to develop a model that can identify duplicate products in an e-commerce setting. The model is evaluated on the F1 score at various thresholds. The project was part of the final project for the CS 7643 Deep Learning course at Georgia Tech.

Actions

• Image Embeddings and Loss Functions: I implemented image embeddings using the ArcFace loss function and CurricularFace loss function.
• Training and Fine-Tuning: I dedicated efforts to training and fine-tuning the image embeddings with the ArcFace loss. This involved experimenting with various hyperparameters, model architectures (ResNet-18 and DenseNet-121), and training settings to achieve the best possible performance.
• Results Analysis: After training the models, I conducted thorough analysis of the results. This included evaluating the F1 scores on both the training and validation sets. T
• Exploratory Data Analysis: I performed the exploratory data analysis, which involved gaining insights from the dataset to better understand its characteristics. This analysis informed decisions about data preprocessing, model selection, and hyperparameter tuning.
• Collaboration and Teamwork: Throughout the project, I collaborated with my team members to ensure a holistic approach. We discussed strategies, shared insights, and collaborated on the overall project structure.

Results

The ArcFace loss for image embeddings achieved impressive F1 scores, particularly on the training set, with a maximum F1 score of 0.91. The F1 scores on the validation set were slightly lower, suggesting potential overfitting or challenges in generalizing to new data.

By effectively integrating image embeddings with the ArcFace loss, I helped create a strong foundation for the project's image-based duplicate product detection capabilities.

Suggestions for improvement could include OCR captions, YOLOv4 labels, and generated image captions. These additional image features could be extracted to improve the system’s performance followed by a more sophisticated ranking model.