Academic Literature Recommendation using Semantic Feature Analysis

Abstract

Researchers face an ever-increasing volume of academic literature. At the current publishing rate, the total number of academic contributions is doubling approximately every nine years. Staying up to date with the latest and most important developments – even within a narrow research field – can be an overwhelming and time-consuming task in today’s fast-paced research landscape. Moreover, discovering niche publications that address a unique or narrowly defined information need is akin to finding the proverbial needle in a haystack. Fortunately, the advent of digital libraries and digital publishing has afforded researchers some support in this daunting task through automated information retrieval tools, such as paper search and recommendation systems. However, today’s automated search and recommendation solutions do not support the full spectrum of researchers’ needs. One specific need that is not being met by existing solutions is the ability to effectively discover non-textual features of interest contained in the full texts of scientific publications. Current content-based literature recommendation systems do not consider the entire range of semantic features, including citation-based similarity, mathematical notation-based similarity, or image-based similarity. Researchers seeking to identify these text-independent semantic features of interest must invest a painstaking effort into a manual inspection and comparison of potentially relevant articles. Especially in the STEM fields (science, technology, engineering, and mathematics), text-independent features, such as in-text citations, mathematical formulae, as well as images, figures, and charts, often communicate valuable semantic information to the reader. This thesis is dedicated to addressing this largely unexplored research area in academic literature recommendation systems. The contributions of the thesis include conceiving and evaluating several specialized feature-based similarity measures for literature recommendation and combining these with existing approaches into a hybrid content-based recommendation approach. A second contribution is the RecVis paper recommendation prototype, which provides the first feature-aware recommendation interface for academic literature. Its us-er-customizable interface and feature visualizations support STEM researchers in the filtering and decision-making process to discover and understand in-stances of text-based and text-independent similarity contained within the recommended literature. Three evaluations, in which semantic feature-based similarity measures were individually applied to large-scale datasets, demonstrated their feasibility for literature recommendation. Subsequently, three user studies conducted with the introduced web-based RecVis prototype yielded several findings. First, a study of the novel force-directed graph-based layout introduced in the RecVis prototype demonstrated low cognitive load and good usability, with customization options adding value and increasing users’ feelings of control when narrowing down recommendations. Second, dividing feature-based similarity into separate views improved understanding and helped experts quickly identify research papers of interest, with the interactive interface addressing specialized information needs not supported by existing recommendation exploration interfaces. Finally, the third study found no significant difference in user-perceived relevance between RecVis recommendations and a purely text-based baseline. However, RecVis recommendations received higher ratings for novelty, diversity, user control, and explanations. An analysis of user log data showed time savings in task completion for specialized information retrieval tasks, which were largest for the assessment of citation and mathematical relevance. Customizability was also rated higher, but users required more time to become familiar with the RecVis interface and its interaction capabilities, which is expected for a specialized recommendation and decision support tool. The research performed in this thesis has made an impact on the field of Human-Computer Information Retrieval (HCIR) by addressing the unexplored area of text-independent semantic features in academic recommendation systems. The development and evaluation of the RecVis prototype demonstrated its potential to better support STEM researchers in discovering and under-standing text-based and text-independent relevance in the recommended literature. By offering a user-customizable interface and visualizing semantic features, this research has contributed to the improvement of academic literature recommender systems, ultimately enhancing the process of information retrieval and decision-making for STEM researchers in the fast-paced research landscape.