| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 5.67 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Atualmente, no setor do retalho, as estratégias inovadoras são fundamentais para apoiar os
retalhistas a destacaram-se da concorrência e a atender às necessidades dos clientes. A integração
eficaz de sistemas de recomendação baseados em regras de associação no retalho, não só permite
os retalhistas conhecerem melhor os seus clientes, permitindo criar recomendações
personalizadas, mas também fazer uma gestão mais otimizada dos seus inventários, garantindo que
os produtos estejam disponíveis no tempo e lugar certo, quando os clientes os desejam.
No mercado atual, existe uma enorme diversidade de opções disponíveis para atender às
necessidades dos clientes, tornando imperativa a necessidade de os retalhistas estabelecerem
estratégias precisas e ponderadas que os permitem destacar da concorrência. Deste modo, nesta
presente dissertação, conduziu-se uma investigação com o intuito de identificar padrões de
consumo baseados em regras de associação e sequências frequentes, a partir de um conjunto de
dados disponibilizado pela plataforma kaggle, denominado de Instacart Market Basket Analysis.
Além disso, foi desenvolvido um sistema de recomendação baseado nas regras de associação
identificadas e diversos modelos de Machine Learning foram aplicados de forma a selecionar o mais
adequado para prever a recompra de produtos por parte dos clientes.
Numa primeira fase, o processo para encontrar as regras de associação presente no conjunto de
dados começou com uma definição clara dos objetivos de investigar os padrões de consumo e
relacionamento entre os produtos. De seguida, a ferramenta selecionada para realizar a análise foi
o Knime, que possui um nó designado de “Association Rule Learner”, baseado no algoritmo Apriori,
para descobrir a relação entre os produtos. Diferentes limites mínimos de suporte foram utilizados
para identificar associações significativas. O foco recaiu sobre as regras de associação que
apresentaram elevados valores das métricas de suporte, confiança e lift, uma vez que tais
associações se demonstraram mais significativas do que as que ocorriam ao acaso. Para o
desenvolvimento do sistema de recomendação baseado em regras de associação foram realizados
dois estudos de precisão, um com base em altos níveis de confiança e suporte e outro com base em
altos níveis de lift e suporte. Para ambos os estudos foram removidos produtos de forma aleatória
para testar a eficácia do sistema. Por último, diversos modelos de Machine Learning foram
aplicados com o objetivo de identificar o que oferecia melhor precisão, a fim de ser selecionado
para prever a recompra de produtos.
Com base nas regras de associação encontradas, foram sugeridas diversas estratégias de negócio,
como promoções e posicionamento estratégico dos produtos com o objetivo de incentivar os
clientes a comprar determinado produto e, por consequência, a impulsionar o volume de vendas.
O sistema de recomendação baseado em métricas de confiança e suporte de valores altos
demonstrou ser o mais eficaz na previsão correta dos produtos, pelo que seria o selecionado para
entrar em produção. O modelo Stacking revelou ser o modelo mais eficaz, com uma precisão de
75%, tornando-o a escolha preferencial para prever se um cliente irá adquirir novamente um
determinado produto.
In today's retail sector, innovative strategies are key to helping retailers stand out from the competition and meet their customers' needs. The effective integration of recommendation systems based on association rules in retail not only allows retailers to get to know their customers better, enabling them to create personalized recommendations, but also to manage their inventories more optimally, ensuring that products are available at the right time and place, precisely when customers require them. In today's market, there is a huge diversity of options available to meet customer needs, making it imperative for retailers to establish precise and thoughtful strategies that allow them to stand out from the competition. In this way, this dissertation conducted an investigation with the aim of identifying various association rules and frequent sequences, based on a set of data made available by the kaggle platform, called Instacart Market Basket Analysis. In addition, a recommendation system was developed based on the association rules identified, and several Machine Learning models were applied in order to select the most appropriate one to predict product repurchase by customers. Initially, the process of finding the association rules present in the data set began with a clear definition of the objectives of investigating consumption patterns and relationships between products. Next, the tool selected to carry out the analysis was Knime, which has a node called "Association Rule Learner", based on the Apriori algorithm, to discover the relationship between products. Different minimum support thresholds were used to identify significant associations. The focus was on association rules that had high values of the support, confidence and lift metrics, since these associations proved to be more significant than those that occurred randomly. To develop the recommendation system based on association rules, two accuracy studies were carried out, one based on high levels of confidence and support and the other based on high levels of lift and support. For both studies, products were randomly removed to test the effectiveness of the system. Finally, several Machine Learning models were applied with the aim of identifying the one that offered the best accuracy to be selected to predict product repurchase. Based on the association rules found, various business strategies were suggested, such as promotions and strategic positioning of products with the aim of encouraging customers to buy a particular product and, consequently, boosting sales volume. The recommendation system based on confidence metrics and high-value support proved to be the most effective in correctly predicting products. Therefore, was the one selected to go into production. The Stacking model proved to be the most effective model, with an accuracy of 75%, making it the preferred choice for predicting whether a customer will purchase a particular product again.
In today's retail sector, innovative strategies are key to helping retailers stand out from the competition and meet their customers' needs. The effective integration of recommendation systems based on association rules in retail not only allows retailers to get to know their customers better, enabling them to create personalized recommendations, but also to manage their inventories more optimally, ensuring that products are available at the right time and place, precisely when customers require them. In today's market, there is a huge diversity of options available to meet customer needs, making it imperative for retailers to establish precise and thoughtful strategies that allow them to stand out from the competition. In this way, this dissertation conducted an investigation with the aim of identifying various association rules and frequent sequences, based on a set of data made available by the kaggle platform, called Instacart Market Basket Analysis. In addition, a recommendation system was developed based on the association rules identified, and several Machine Learning models were applied in order to select the most appropriate one to predict product repurchase by customers. Initially, the process of finding the association rules present in the data set began with a clear definition of the objectives of investigating consumption patterns and relationships between products. Next, the tool selected to carry out the analysis was Knime, which has a node called "Association Rule Learner", based on the Apriori algorithm, to discover the relationship between products. Different minimum support thresholds were used to identify significant associations. The focus was on association rules that had high values of the support, confidence and lift metrics, since these associations proved to be more significant than those that occurred randomly. To develop the recommendation system based on association rules, two accuracy studies were carried out, one based on high levels of confidence and support and the other based on high levels of lift and support. For both studies, products were randomly removed to test the effectiveness of the system. Finally, several Machine Learning models were applied with the aim of identifying the one that offered the best accuracy to be selected to predict product repurchase. Based on the association rules found, various business strategies were suggested, such as promotions and strategic positioning of products with the aim of encouraging customers to buy a particular product and, consequently, boosting sales volume. The recommendation system based on confidence metrics and high-value support proved to be the most effective in correctly predicting products. Therefore, was the one selected to go into production. The Stacking model proved to be the most effective model, with an accuracy of 75%, making it the preferred choice for predicting whether a customer will purchase a particular product again.
Description
Keywords
Retail Association Recommendation Models Machine Learning Accuracy