Optimizing product recommendations for a try-before-you-buy fashion e-commerce sit

• Autor(a) principal: Ruh, Paul Johan
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.14/41433
• topic: Machine learning; Recommendation systems; Deep learning; Data sparsity; Fashion; e-Commerce; Product recommendations; Aprendizagem de máquinas; Sistemas de recomendação; Aprendizagem profunda; Espaçamento de dados; Moda; Recomendações de produtos; Domínio/Área Científica::Ciências Sociais::Economia e Gestão

The fashion e-commerce market has experienced a significant growth and more and more customers tend to buy products online, rather than in physical stores. However, after a customer buys a product online, only a fraction of the garments stay in their wardrobe as many items are being returned to the vendor. Due to the absence of physical examination and misleading product descriptions customers struggle to find the right product suitable to their personal preferences. Especially the category of women’s lingerie suffers to a great extend from high return rates. Different sources report that between 70 up to 100% of women wear wrong sized bras. Personalized recommendations through so called recommendation systems play an essential role in e-commerce. This thesis aims to optimize the current product recommendations of a Belgium start-up called CurveCatch that sells women’s lingerie articles online and relies on a try-before-you-buy concept. To predict which products a customer is likely to buy two different personalized deep learning approaches were introduced. Data sparsity was addressed by labeling each unique product per customer and minority classes were synthetically oversampled. The findings demonstrated that recommendation systems are not only relevant for companies operating on a large scale. Rather, they also can be a valuable source of accurate recommendations for start-ups with sparse data. However, results also underlined well-known limitations of recommendation systems. Both models struggled especially when identifying products a customer is likely to buy, while it was rather easy to identify products a customer is not likely to buy.