Recommender Systems for Grocery Retail - A Machine Learning Approach

• Autor(a) principal: Silva, Xavier dos Santos
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.22/17716
• topic: Recommender systems; Grocery retail; Machine learning; Gradient boosted trees

Recommender systems are present in our daily activities in different moments, such as when choosing a song to listen to or when doing online shopping. It is an everyday reality for people to have the help of computer systems in order to simplify regular decision activities. Grocery shopping is an essential part of people’s life and a frequent activity. Despite being a common habit, each customer has unique routines, needs and preferences regarding products and brands. This information is valuable for grocery retailers to know their customers better and to improve their marketing and operational activities. This dissertation aims to apply machine learning algorithms to the development of a recommender system capable of preparing personalized grocery shopping lists. The proposed architecture is designed to allow integration with different grocery retailers and support distinct TensorFlow algorithms. The process of extracting information from the dataset as features was explored, as well as the tuning of the model hyperparameters, to obtain better results. The recommendation engine is exposed via a distributed software architecture designed to allow retailers to integrate the recommender system with different existing solutions (e.g., websites or mobile applications). A case study to validate the implemented solution was performed, integrating it with a public dataset provided by Instacart. A comparison study between different machine learning algorithms over the adopted dataset has lead to the choice of the gradient boosted trees algorithm. The solution developed in the case study was compared against two non-machine learning approaches at predicting the last purchase of 360 arbitrary test customers. A pattern miningbased solution and a SQL-based heuristic were used. Different evaluation metrics (namely, the average accuracy, precision, recall, and f1-score) were registered. The way association rules with different strengths were reflected in the predictions of the developed solution was also analyzed. The gradient boosted trees-based implementation from the case study was capable of outperforming the compared solutions as far as evaluation metrics are concerned, and has shown a higher capability of predicting at least one correct item per customer. Also, it became evident that the strictest association rules were frequently found in the recommendations. The adopted solution and algorithm have shown promising results and a remarkable capability to provide meaningful predictions to the different customers, evidencing its capability to add value to grocery retail. Nevertheless, there is still potential for further expansion.