Email Classification: a case study

• Autor(a) principal: André Ricardo Azevedo Gonçalves da Silva
• Data de Publicação: 2016
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Texto Completo: https://hdl.handle.net/10216/88362
• topic: Ciências da computação e da informação; Computer and information sciences

Internet dependance on email has been frequent since its early days. In the present days, electronic mail is widely used in a professional and personal context. Although this service was developed as a way of communication, nowadays it serves many other purposes. The majority of services available online will require an email address in order to authenticate or as a bridge of communication between the user and the service. The average number of emails sent and received, by the average user, is in the order of the hundreds per day, and these emails can be of varying categories: social, professional, notifications, marketing, transactional, emails which warrant no response, emails to send files, emails requiring response, among others with different purposes. This originates an information overload problem, that proves difficult to be completely solved manually by the email address owner. Therefore, there is a growing need to develop systems that can automatically learn and recommend users effective ways to organize their email information, which can aggregate emails into smaller groups to be easily interpreted by the user, expediting the process of reading and consulting the mailbox. To alleviate this information overload problem there are several possible approaches and techniques, such as machine learning to help on email classification and clustering, in order to find new subsets of emails in the massive inboxes we all have, now or in the future. After a careful review of the state of the art on email classification and grouping techniques, this work will enumerate and select the most effective approaches for the problem at hand, and will adapt them to a very concrete case study, a desktop email client under development at Mailcube Lda. The approach in mind will follow a competitive learning paradigm, which means that emails will compete with each other in order to find subsets in the inbox. It will also follow a reinforcement learning paradigm to add sensitiveness to user profile and interaction history. At the end, the resulting system is expected to suggest the user to organize his inbox into relevant groups of emails, based on learning users' interactions and continuously adapting to the arrival of new emails, improving the overall user experience and saving precious time for the users.