jingxuanyang/Shopee-Product-Matching

Combines multimodal embedding approaches leveraging both image and text features through ensemble methods, utilizing pretrained sentence transformers (paraphrase-xlm-r-multilingual) and deep learning models trained on PyTorch. The pipeline extracts BERT embeddings for product descriptions alongside visual features, then applies dynamic threshold tuning to optimize F1-score for product matching classification. Includes comprehensive post-processing utilities and Jupyter notebooks for model training, inference, and cross-validation across isolated image, text, and hybrid architectures.

No commits in the last 6 months.