Predicting purchasing intention through a single stage siamese deep learning models
Understanding consumer buying behavior in the context of e-commerce is a recent trend at large retail stores. It can be very attractive for retail companies to know which users will buy in their market and what products they will buy. Through the study of online user behavior, models can be created...
| Autor principal: | Takano, Kevin Kimiya |
|---|---|
| Outros Autores: | http://lattes.cnpq.br/8375340697156770 |
| Grau: | Dissertação |
| Idioma: | por |
| Publicado em: |
Universidade Federal do Amazonas
2022
|
| Assuntos: | |
| Acesso em linha: |
https://tede.ufam.edu.br/handle/tede/9034 |
| id |
oai:https:--tede.ufam.edu.br-handle-:tede-9034 |
|---|---|
| recordtype |
dspace |
| spelling |
oai:https:--tede.ufam.edu.br-handle-:tede-90342022-08-27T05:03:37Z Predicting purchasing intention through a single stage siamese deep learning models Takano, Kevin Kimiya Carvalho, André Luiz da Costa http://lattes.cnpq.br/8375340697156770 http://lattes.cnpq.br/4863447798119856 Santos, Eulanda Miranda dos Ferreira, Raoni Simões Piedade, Márcio Palheta Comportamento do consumidor Comércio eletrônico - Programas de computador Negócios - Recursos de rede de computador CIENCIAS EXATAS E DA TERRA: CIENCIA DA COMPUTACAO Session-based modeling Predicting purchase intention Customer behavior Siamese-networks Triplet-loss Understanding consumer buying behavior in the context of e-commerce is a recent trend at large retail stores. It can be very attractive for retail companies to know which users will buy in their market and what products they will buy. Through the study of online user behavior, models can be created to improve marketing personalizing, and build digital products. Through the historical data of user events, such as clicked items, it is possible to use them as resources to forecast purchases. Despite how valuable is this data, it is not so simple to create machine learning models using them. A very small number of user sessions are buyers of items, and, in general, models have greater learning difficulties with unbalanced classes. Moreover, there are a large number of products in a store, making the problem even more complex. Previous works in the literature show that it is better to solve the problem in two stages, i.e., using two models: one model to predict which customers will be buying items and another to predict which products will be purchased among these consumers. Solving problems in two stages makes the problem simpler since it divides the model's complexity. However, creating two models the second model does not use information from non-buyer-sessions to solve the item classification. Furthermore, if the first model fails to classify a session as a buyer-session, the second model may have its results negatively impacted. Therefore, for this work, our objective is to develop a model that solves the problem with just a single model, a \emph{single-stage model}. We deployed Siamese neural networks to extract features to deal with imbalances. In our single-stage framework, we had several contributions. First, is the creation of a new loss function, the quartet-loss, which optimizes the parameters differently from the triplet-loss. Second, is the development of two different strategies for modeling user click sessions. Third, is the creation of metrics that evaluate the results of e-commerce models in online sessions. And finally, we developed machine learning methods using this project framework that reached the state-of-the-art for this problem. Compreender o comportamento de compra do consumidor no contexto do e-commerce é uma tendência recente nas grandes lojas de varejo. Pode ser muito atraente para as empresas de varejo saber quais usuários comprarão em seu mercado e quais produtos comprarão. Através do estudo do comportamento do usuário online, modelos podem ser criados para melhorar a personalização de marketing e construir produtos digitais. Através dos dados históricos de eventos do usuário, como itens clicados, é possível utilizá-los como recursos para previsão de compras. Apesar do valor desses dados, não é tão simples criar modelos de aprendizado de máquina usando-os. Um número muito pequeno de sessões de usuários são compradores de itens e, em geral, os modelos apresentam maiores dificuldades de aprendizado com classes desequilibradas. Além disso, há um grande número de produtos em uma loja, tornando o problema ainda mais complexo. Trabalhos anteriores na literatura mostram que é melhor resolver o problema em duas etapas, ou seja, usando dois modelos: um modelo para prever quais clientes estarão comprando itens e outro para prever quais produtos serão adquiridos entre esses consumidores. Resolver problemas em duas etapas torna o problema mais simples, pois divide a complexidade do modelo. No entanto, ao criar dois modelos, o segundo modelo não usa informações de sessões não-compradoras para resolver a classificação do item. Além disso, se o primeiro modelo deixar de classificar uma sessão como sessão do comprador, o segundo modelo poderá ter seus resultados impactados negativamente. Portanto, para este trabalho, nosso objetivo é desenvolver um modelo que resolva o problema com apenas um modelo, um \emph{modelo de estágio único}. Implantamos redes neurais siamesas para extrair recursos para lidar com desequilíbrios. Em nossa estrutura de estágio único, tivemos várias contribuições. Primeiro, é a criação de uma nova função de perda, a perda quarteto, que otimiza os parâmetros de forma diferente da perda tripla. Em segundo lugar, está o desenvolvimento de duas estratégias diferentes para modelar as sessões de clique do usuário. Terceiro, é a criação de métricas que avaliam os resultados dos modelos de e-commerce em sessões online. E por fim, desenvolvemos métodos de aprendizado de máquina usando este framework de projeto que atingiu o estado da arte para este problema. 2022-08-26T17:14:54Z 2022-07-06 Dissertação TAKANO, Kevin Kimiya. Predicting purchasing intention through a single stage siamese deep learning models. 2022. 100 f. Dissertação (Mestrado em Informática) - Universidade Federal do Amazonas, Manaus (AM), 2022. https://tede.ufam.edu.br/handle/tede/9034 por Acesso Aberto http://creativecommons.org/licenses/by/4.0/ application/pdf Universidade Federal do Amazonas Instituto de Computação Brasil UFAM Programa de Pós-graduação em Informática |
| institution |
TEDE - Universidade Federal do Amazonas |
| collection |
TEDE-UFAM |
| language |
por |
| topic |
Comportamento do consumidor Comércio eletrônico - Programas de computador Negócios - Recursos de rede de computador CIENCIAS EXATAS E DA TERRA: CIENCIA DA COMPUTACAO Session-based modeling Predicting purchase intention Customer behavior Siamese-networks Triplet-loss |
| spellingShingle |
Comportamento do consumidor Comércio eletrônico - Programas de computador Negócios - Recursos de rede de computador CIENCIAS EXATAS E DA TERRA: CIENCIA DA COMPUTACAO Session-based modeling Predicting purchase intention Customer behavior Siamese-networks Triplet-loss Takano, Kevin Kimiya Predicting purchasing intention through a single stage siamese deep learning models |
| topic_facet |
Comportamento do consumidor Comércio eletrônico - Programas de computador Negócios - Recursos de rede de computador CIENCIAS EXATAS E DA TERRA: CIENCIA DA COMPUTACAO Session-based modeling Predicting purchase intention Customer behavior Siamese-networks Triplet-loss |
| description |
Understanding consumer buying behavior in the context of e-commerce is a recent trend at large retail stores. It can be very attractive for retail companies to know which users will buy in their market and what products they will buy. Through the study of online user behavior, models can be created to improve marketing personalizing, and build digital products.
Through the historical data of user events, such as clicked items, it is possible to use them as resources to forecast purchases. Despite how valuable is this data, it is not so simple to create machine learning models using them. A very small number of user sessions are buyers of items, and, in general, models have greater learning difficulties with unbalanced classes. Moreover, there are a large number of products in a store, making the problem even more complex.
Previous works in the literature show that it is better to solve the problem in two stages, i.e., using two models: one model to predict which customers will be buying items and another to predict which products will be purchased among these consumers. Solving problems in two stages makes the problem simpler since it divides the model's complexity. However, creating two models the second model does not use information from non-buyer-sessions to solve the item classification. Furthermore, if the first model fails to classify a session as a buyer-session, the second model may have its results negatively impacted. Therefore, for this work, our objective is to develop a model that solves the problem with just a single model, a \emph{single-stage model}. We deployed Siamese neural networks to extract features to deal with imbalances.
In our single-stage framework, we had several contributions. First, is the creation of a new loss function, the quartet-loss, which optimizes the parameters differently from the triplet-loss. Second, is the development of two different strategies for modeling user click sessions. Third, is the creation of metrics that evaluate the results of e-commerce models in online sessions. And finally, we developed machine learning methods using this project framework that reached the state-of-the-art for this problem. |
| author_additional |
Carvalho, André Luiz da Costa |
| author_additionalStr |
Carvalho, André Luiz da Costa |
| format |
Dissertação |
| author |
Takano, Kevin Kimiya |
| author2 |
http://lattes.cnpq.br/8375340697156770 |
| author2Str |
http://lattes.cnpq.br/8375340697156770 |
| title |
Predicting purchasing intention through a single stage siamese deep learning models |
| title_short |
Predicting purchasing intention through a single stage siamese deep learning models |
| title_full |
Predicting purchasing intention through a single stage siamese deep learning models |
| title_fullStr |
Predicting purchasing intention through a single stage siamese deep learning models |
| title_full_unstemmed |
Predicting purchasing intention through a single stage siamese deep learning models |
| title_sort |
predicting purchasing intention through a single stage siamese deep learning models |
| publisher |
Universidade Federal do Amazonas |
| publishDate |
2022 |
| url |
https://tede.ufam.edu.br/handle/tede/9034 |
| _version_ |
1831970100153942016 |
| score |
11.753735 |