Análise e previsão de comportamento de clientes de um e-commerce de peças de reposição para máquinas pesadas
Arquivos
Data
2022-06-16
Autores
Matsui, Natália
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
No contexto atual, milhões de dados são obtidos diariamente através dos sites de e-commerce. O preço atrativo, por si só, não é mais uma vantagem competitiva, os clientes estão em busca de personalização, ofertas direcionadas e ambientes que seja adaptáveis ao seu estilo de compra. Para que seja possível oferecer essa experiência aos clientes faz-se necessária a compilação e análise desses dados, a fim de encontrar padrões e gerar classificações. Métodos de Aprendizado de Máquina têm apresentado resultados relevantes na execução automática dessas tarefas. Por outro lado, a análise de comportamento de clientes de uma empresa podem fornecer outros indicativos valiosos para direcionar aquisições, produções e contratações internas da empresa. Este trabalho visa analisar as aplicações de algoritmos de Aprendizado de Máquina sobre a base de dados de uma empresa dentro de um universo de e-commerce de peças de reposição para máquina pesadas. O objetivo é descobrir indicativos úteis a respeito do comportamento do usuário para direcionar planos estratégicos da empresa, como por exemplo ações de marketing. Para tanto foram analisados os algoritmos de Classificação - Random Forest, KNN e Naïve Bayes - com a utilização da linguagem Python no ambiente Jupyter Notebook. Duas fontes principais de informação foram empregadas para compor a base de dados para treinamentos e testes: os dados referentes a vendas para o cliente e os dados históricos de comportamento do usuário. Os targets selecionados para os treinamentos e testes foram Produto e Categoria, pois é de suma importância para o comércio eletrônico conhecer o hábito de consumo dos seus clientes de maneira a criar campanhas e ofertas personalizadas, o que gera aumento não só no ticket médio como também da conversão. Após treinamento, testes de predição com os algoritmos foram realizados para os targets Produto e Categoria de Produtos. Resultados indicaram desempenho abaixo de 30% para predizer o target Produto. Para o target Categoria foi considerada a realização de testes com os algoritmo que apresentou melhor desempenho para predição do Produto: Random Forest Classifier além do Naïve Bayes. Em análise aos resultados obtidos, foi concluído que o baixo desempenho para predição do Produto justifica-se pela insuficiência de dados na base utilizada para o treinamento do aprendizado de máquina, havia poucas ocorrências de vendas para um mesmo Produto. No entanto, para a aplicação de predição de Categoria o algoritmo Naïve Bayes se mostrou eficaz conseguindo acertar 81,4% das saídas previstas.
Nowadays, millions of data are flowing daily in the e-commerce websites. Prices by themselves are not a competitive advantage anymore, clients are looking for personalization, targeted offers and environments that are adaptive of their purchase behavior. Furthermore, to make that kind of user experience possible it is important to compile and analyze all that data, along with the application of Machine Learning techniques so that patterns and classifications can be found more easily. Machine Learning Algorithms presented relevant results in the automatic execution of these tasks. Besides that, a behavioral analysis of a company’s customer can indicate valuable elements in order to propose acquisitions, production and even company new hirings. This paper aims to analyze the different approaches using Machine Learning algorithms applied in the data base of a company inside the universe of an e-commerce of spare parts for heavy machinery. The goal is to discover useful indicators related to user’s behavior to orientate the company’s strategic plan, as marketing campaigns for example. To do so, classification algorithms were analyzed, Random Forest, KNN and Naïve Bayes. Those algorithms were developed using Python language within Jupyter Notebook. Two main sources of data were used to compose the dataset for training and testing: data related to client’s sales and the user’s historical behavioral data. The targets selected for training and tests were, product and category, since it is highly important for an e-commerce to be aware of its client’s consumption habits, in this way personalized campaigns and promotions can be created leading not only to an increase of the average ticket but also the conversions. Tests were performed for the following targets, Products and Categories. Within the tests it was possible to notice that none of them had a performance which was higher than 30%. For the target Categories, the algorithm that showed the best performance in product prediction (Random Forest Classifier) and also Naïve Bayes were tested. In a deep analysis of the achieved result, the bad performance for product prediction can be linked to the lack of data since the available database did not have many sales of the same product. On the other hand, for Category prediction, the algorithm Naïve Bayes have proven to be efficient, being able to correctly guess 81,4% of the predicted outcomes.
Nowadays, millions of data are flowing daily in the e-commerce websites. Prices by themselves are not a competitive advantage anymore, clients are looking for personalization, targeted offers and environments that are adaptive of their purchase behavior. Furthermore, to make that kind of user experience possible it is important to compile and analyze all that data, along with the application of Machine Learning techniques so that patterns and classifications can be found more easily. Machine Learning Algorithms presented relevant results in the automatic execution of these tasks. Besides that, a behavioral analysis of a company’s customer can indicate valuable elements in order to propose acquisitions, production and even company new hirings. This paper aims to analyze the different approaches using Machine Learning algorithms applied in the data base of a company inside the universe of an e-commerce of spare parts for heavy machinery. The goal is to discover useful indicators related to user’s behavior to orientate the company’s strategic plan, as marketing campaigns for example. To do so, classification algorithms were analyzed, Random Forest, KNN and Naïve Bayes. Those algorithms were developed using Python language within Jupyter Notebook. Two main sources of data were used to compose the dataset for training and testing: data related to client’s sales and the user’s historical behavioral data. The targets selected for training and tests were, product and category, since it is highly important for an e-commerce to be aware of its client’s consumption habits, in this way personalized campaigns and promotions can be created leading not only to an increase of the average ticket but also the conversions. Tests were performed for the following targets, Products and Categories. Within the tests it was possible to notice that none of them had a performance which was higher than 30%. For the target Categories, the algorithm that showed the best performance in product prediction (Random Forest Classifier) and also Naïve Bayes were tested. In a deep analysis of the achieved result, the bad performance for product prediction can be linked to the lack of data since the available database did not have many sales of the same product. On the other hand, for Category prediction, the algorithm Naïve Bayes have proven to be efficient, being able to correctly guess 81,4% of the predicted outcomes.
Descrição
Palavras-chave
Predição, Aprendizagem de máquina, Aprendizado supervisionado, E-connerce, Prediction, Machine learning, Supervised learning