Análise e Avaliação de Performance de Técnicas de Machine Learning para Product Matching

Abstract

Product Matching is the task of identifying pairs that refer to the same product from distinct sources. Traditional rule-based approaches exhibit limitations when handling incomplete, non-standardized data or subtle variations. To address these constraints, Machine Learning methods have emerged as more effective alternatives. This study investigates these techniques through a Systematic Literature Review to identify the most promising current methods. Additionally, a series of experiments are conducted to evaluate the performance of the proposed solutions using product datasets from both The WDC Product Data Corpus (LSPM) and The Magellan Data Repository. The findings support the evidence presented in the selected sample of studies from the review. Strategies based on fine-tuning pre-trained Language Models, such as BERT and its variants, demonstrate their superiority in the task of distinguishing identical products. Multi-objective Learning and Supervised Contrastive Learning (SupCon) approaches stand out, achieving F1 scores exceeding 95%.