DEEP LEARNING MODELS FOR RESEARCH PAPER RECOMMENDER SYSTEMS

Abstract

The amount of scholarly publications has been rapidly increasing during the last decades. The need to intelligently process such amount and produce rec ommendations to researchers is becoming urgent. In the last years, Recommender System (RS) approaches have emerged to facilitate finding publications related to the researchers’ area of interest. However, natural language ambiguity is still a challenge to generate accurate recommendations exploiting textual features. The goal of this doctoral thesis is to propose deep learning models, which could learn semantic representations of research papers in order to obtain effective recommendations. In other words, proposing models that help in providing recommendations based on the semantic similarity between research papers. In this study, we make several contributions that address the problem of considering the semantic similarity between research papers: Firstly, we propose a supervised approach that adopts gated recurrent networks with attention mechanism, for aggregating important words and sentences from research paper titles and abstracts, in order to increase the general representation and visualization of the key concepts in research papers. This approach has been exploited for predicting social tags from research papers, since tags help in organizing, sharing and even recommending research papers. Secondly, we propose a tag-aware research paper recommendation approach, that utilizes the same model proposed for tag prediction, in extracting tag-based document repre sentations. We show how semantic document representations based on social tags can be combined with the traditional collaborative filtering methods to yield superior performance with any number of ratings. Finally, we propose an unsupervised approach for non-personalized research paper recommendations, which leverages pre-trained sentence encoders based on deep learning models.