Loading...
Publication
Resolver o problema de cold start e escassez de dados com algoritmos híbridos
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 3.22 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Os sistemas de recomendação tornaram-se indispensáveis no panorama digital, proporcionando personalização em setores como comércio eletrónico, saúde, media e retalho. No entanto, estes sistemas enfrentam desafios como o cold start, a escassez de dados, a privacidade e a transparência, que comprometem a sua eficácia e adoção generalizada. O projeto Environment Adaptive Recommendation System (EARS) aborda estas limitações
através de uma abordagem inovadora baseada em aprendizagem federada. Este método permite o processamento descentralizado de dados, reduzindo os riscos associados à centralização e promovendo maior segurança e privacidade para os utilizadores. Para superar os problemas de cold start e escassez de dados, o EARS integra algoritmos híbridos que combinam filtragem colaborativa e baseada em conteúdo. Esta combinação permite oferecer recomendações mais precisas e personalizadas, mesmo em cenários onde
os dados são limitados ou inexistentes. A solução foi validada através de cenários experimentais que demonstraram melhorias significativas na personalização, precisão e cobertura das recomendações, tanto em situações com abundância de dados como em contextos de cold start e escassez de dados. Os resultados
confirmam a eficácia dos algoritmos híbridos em comparação com abordagens tradicionais, evidenciando ainda a escalabilidade e a robustez do sistema. Esta dissertação contribui, assim, para a evolução dos sistemas de recomendação ao propor uma solução prática e inovadora, capaz de responder a limitações destes sistemas. Para além disso, abre caminho a desenvolvimentos futuros, como a aplicação em diferentes setores e a realização de testes em ambientes mais amplos e realistas, de forma a consolidar o impacto e utilidade do EARS.
Recommendation systems have become indispensable in the digital landscape, providing personalization in sectors such as e-commerce, healthcare, media, and retail. However, these systems face challenges such as cold start, data scarcity, privacy, and transparency, which compromise their effectiveness and widespread adoption. The EARS project addresses these limitations through an innovative approach based on federated learning. This method enables decentralized data processing, reducing the risks associated with centralization and promoting greater security and privacy for users. To overcome cold start and data scarcity problems, EARS integrates hybrid algorithms that combine collaborative and content-based filtering. This combination makes it possible to offer more accurate and personalized recommendations, even in scenarios where data is limited or nonexistent. The solution was validated through experimental scenarios that demonstrated significant improvements in personalization, accuracy, and recommendation coverage, both in situations with abundant data and in contexts of cold start and data scarcity. The results confirm the effectiveness of hybrid algorithms compared to traditional approaches, further highlighting the scalability and robustness of the system. This dissertation thus contributes to the advancement of recommendation systems by proposing a practical and innovative solution capable of addressing their limitations. Furthermore, it paves the way for future developments, such as application in different sectors and testing in broader and more realistic environments, in order to consolidate the impact and usefulness of EARS.
Recommendation systems have become indispensable in the digital landscape, providing personalization in sectors such as e-commerce, healthcare, media, and retail. However, these systems face challenges such as cold start, data scarcity, privacy, and transparency, which compromise their effectiveness and widespread adoption. The EARS project addresses these limitations through an innovative approach based on federated learning. This method enables decentralized data processing, reducing the risks associated with centralization and promoting greater security and privacy for users. To overcome cold start and data scarcity problems, EARS integrates hybrid algorithms that combine collaborative and content-based filtering. This combination makes it possible to offer more accurate and personalized recommendations, even in scenarios where data is limited or nonexistent. The solution was validated through experimental scenarios that demonstrated significant improvements in personalization, accuracy, and recommendation coverage, both in situations with abundant data and in contexts of cold start and data scarcity. The results confirm the effectiveness of hybrid algorithms compared to traditional approaches, further highlighting the scalability and robustness of the system. This dissertation thus contributes to the advancement of recommendation systems by proposing a practical and innovative solution capable of addressing their limitations. Furthermore, it paves the way for future developments, such as application in different sectors and testing in broader and more realistic environments, in order to consolidate the impact and usefulness of EARS.
Description
Keywords
Recommendation Systems Federated Learning Hybrid Algorithms Cold Start Data Scarcity Sistemas de recomendação Aprendizagem federada Algoritmos híbridos Cold start Escassez de dados