Sistema de Recomendações Multi-Objectivo Multi-Model para Previsão de Ações e-Commerce

• Autor(a) principal: Costa, Nuno
• 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/10400.22/24056
• topic: E-Commerce; Recommender Systems; Model Ensemble; Multi-Objective; Sequence Data; Session; Domínio/Área Científica::Engenharia e Tecnologia

The thriving retail e-commerce sector, driven by the surge in digital transactions and con sumer engagement, emphasizes the imperative for enterprises to optimize revenue and prof itability. To achieve this, online stores are increasingly turning to advanced recommender systems. These systems strategically target multiple objectives, focusing on factors that boost user interaction and value extraction, such as increased item viewing and cart addi tions. By prioritizing a set of objectives, recommender systems aim to cater to immediate user preferences and cultivate a personalized user experience, fostering loyalty and continu ous engagement in the dynamic landscape of e-commerce. In response to this evolving landscape, e-commerce enterprise OTTO initiated a Kaggle competition, calling upon the global community of data scientists and machine learning enthusiasts to model and predict a set of events within their products. This collaborative effort not only propelled advancements in the field but also underscored the significance of community-driven initiatives in shaping the future of personalized online shopping experi ences. This project directly addresses the challenges posed by the OTTO Kaggle competition, aiming to evaluate the individual and collective performance of diverse recommendation models within e-commerce recommender systems. Utilizing the Design Science Research (DSR) methodology, the project underwent iterative design and development, aligning with specific goals derived from a comprehensive review of existing literature and state-of-the-art recommender systems as well as goals and requirements extrapolated from the mentioned competition. The implemented system integrates Gradient Boosting Decision Trees (GBDT), Dropouts meet Multiple Additive Regression Trees (DART), Gated Recurrent Unit for Recommender Systems (GRU4Rec), and Random Forest models into an ensemble framework. Evaluated using predefined metrics from the Kaggle competition, the system leverages user session data to predict user actions across various event types. While the performance analysis demonstrates the system’s competency, there is room for improvement to provide enhanced value in real-world e-commerce scenarios. The project highlights the continuous evolution of recommender systems, emphasizing the need for ongoing research and refinement.