Software library for stream-based recommender systems

• Autor(a) principal: Fernando André Bezerra Moura Fernandes
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: https://hdl.handle.net/10216/129626
• topic: Engenharia electrotécnica, electrónica e informática; Electrical engineering, Electronic engineering, Information engineering

Traditionally, a machine learning algorithm is able to learn from data, given a previously built and treated data set. One can also analyze that data set, using data mining techniques, and draw conclusions from it. Both of these concepts have numerous world-wide applications, from medical diagnosis to speech recognition or even search engine queries. However, traditionally speaking, it is being assumed that the data set is available at all times. That is not necessarily true with modern data, as distributed systems applications receive and process millions of data streams on a limited time fraction. Therefore, there is a need for techniques to mine and process these data streams,on a limited time period with good results and effective scaling as data grows. One specific use case of analyzing and predicting future conclusions from given data, are recommendation systems.Several online services use recommender systems to deliver personalized content to their users.In many cases, recommendations are one of the most effective traffic generators in such services.The problem lies in finding the best small subset of items in a system that matches the personal preferences of each user, through the analysis of a very large amount of historical data. This problem gets more attention if it is considered as a generic problem, not as a specific one, that is,if a library is built so that it can be extended and used as a tool to help build a system for a specific use case. One can distinguish solutions between perfect ones or statistically similar ones. Due to the large amount of data available, the decision to reprocess all the data every time new data arrives, would not be feasible so, incremental algorithms are used to process incoming data and keeping the recommendation model updated. The real purpose of this work is to implement such a library which contains, and evaluates these incremental approaches to recommendation since current ones are task-specific.