| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 4.23 MB | Adobe PDF |
Advisor(s)
Abstract(s)
Nesta era tecnológica em que nos encontramos há cada vez mais informação disponível na internet, mas grande parte dessa informação não é relevante. Isto leva à necessidade de criar maneiras de filtrar informação, de forma a reduzir o tempo de recolha de informação útil.
Esta necessidade torna o uso de sistemas de recomendação muito apelativo, visto estes personalizarem as pesquisas de forma a ajudar os seus utilizadores a fazer escolhas mais informadas. Os sistemas de recomendação procuram recomendar os itens mais relevantes aos seus utilizadores, no entanto necessitam de informação sobre os utilizadores e os itens, de forma a melhor os poder organizar e categorizar.
Há vários tipos de sistemas de recomendação, cada um com as suas forças e fraquezas.
De modo a superar as limitações destes sistemas surgiram os sistemas de recomendação híbridos, que procuram combinar características dos diferentes tipos de sistemas de recomendação de modo a reduzir, ou eliminar, as suas fraquezas.
Uma das limitações dos sistemas de recomendação acontece quando o próprio sistema não tem informação suficiente para fazer recomendações.
Esta limitação tem o nome de Cold Start e pode focar-se numa de duas áreas: quando a falta de informação vem do utilizador, conhecida como User Cold Start; e quando a falta de informação vem de um item, conhecida como Item Cold Start.
O foco desta dissertação é no User Cold Start, nomeadamente na criação de um sistema de recomendação híbrido capaz de lidar com esta situação.
A abordagem apresentada nesta dissertação procura combinar a segmentação de clientes com regras de associação. O objetivo passa por descobrir os utilizadores mais similares aos utilizadores numa situação de Cold Start e, através dos itens avaliados pelos utilizadores mais similares, recomendar os itens considerados mais relevantes, obtidos através de regras de associação.
O algoritmo híbrido apresentado nesta dissertação procura e classifica todos os tipos de utilizadores. Quando um utilizador numa situação de Cold Start está à procura de recomendações, o sistema encontra itens para recomendar através da aplicação de regras de associação a itens avaliados por utilizadores no mesmo grupo que o utilizador na situação de Cold Start, cruzando essas regras com os itens avaliados por este último e apresentando as recomendações com base no resultado.
Recommender systems, or recommenders, are a way to filter the useful information from the data, in this age where there is a lot of available data. A recommender system’s purpose is to recommend relevant items to users, and to do that, it requires information on both, data from users and from items, to better organise and categorise both of them. There are several types of recommenders, each best suited for a specific purpose, and with specific weaknesses. Then there are hybrid recommenders, made by combining one or more types of recommenders in a way that each type supresses, or at least limits, the weaknesses of the other types. A very important weakness of recommender systems occurs when the system doesn’t have enough information about something and so, it cannot make a recommendation. This problem known as a Cold Start problem is addressed in this thesis. There are two types of Cold Start problems: those where the lack of information comes from a user (User Cold Start) and those where it comes from an item (Item Cold Start). This thesis’ main focus is on User Cold Start problems. A novel approach is introduced in this thesis which combines clients’ segmentation with association rules. The goal is first, finding the most similar users to cold start users and then, with the items rated by these similar users, recommend those that are most suitable, which are gotten through association rules. The hybrid algorithm presented in this thesis finds and classifies all users’ types. When a user in a Cold Start situation is looking for recommendations, the system finds the items to recommend to him by applying association rules to the items evaluated by users in the same user group as the Cold Start user, crossing them with the few items evaluated by the Cold Start user and finally making its recommendations based on that.
Recommender systems, or recommenders, are a way to filter the useful information from the data, in this age where there is a lot of available data. A recommender system’s purpose is to recommend relevant items to users, and to do that, it requires information on both, data from users and from items, to better organise and categorise both of them. There are several types of recommenders, each best suited for a specific purpose, and with specific weaknesses. Then there are hybrid recommenders, made by combining one or more types of recommenders in a way that each type supresses, or at least limits, the weaknesses of the other types. A very important weakness of recommender systems occurs when the system doesn’t have enough information about something and so, it cannot make a recommendation. This problem known as a Cold Start problem is addressed in this thesis. There are two types of Cold Start problems: those where the lack of information comes from a user (User Cold Start) and those where it comes from an item (Item Cold Start). This thesis’ main focus is on User Cold Start problems. A novel approach is introduced in this thesis which combines clients’ segmentation with association rules. The goal is first, finding the most similar users to cold start users and then, with the items rated by these similar users, recommend those that are most suitable, which are gotten through association rules. The hybrid algorithm presented in this thesis finds and classifies all users’ types. When a user in a Cold Start situation is looking for recommendations, the system finds the items to recommend to him by applying association rules to the items evaluated by users in the same user group as the Cold Start user, crossing them with the few items evaluated by the Cold Start user and finally making its recommendations based on that.
Description
Keywords
Sistema de recomendações Data mining Sistema híbrido Cold start Novo utilizador Recommender system Data mining Hybrid system New user