Datenbestand vom 04. Dezember 2018

Warenkorb Datenschutzhinweis Dissertationsdruck Dissertationsverlag Institutsreihen     Preisrechner

aktualisiert am 04. Dezember 2018

ISBN 9783843933049

Euro 84,00 inkl. 7% MwSt

978-3-8439-3304-9, Reihe Informatik

Laura Steinert
Beyond Similarity and Accuracy - A New Take on Automating Scientific Paper Recommendations

161 Seiten, Dissertation Universität Duisburg-Essen (2017), Hardcover, B5

Zusammenfassung / Abstract

In the last years, many paper recommender systems have been developed. Yet many of these approaches do not differentiate between different recommendation tasks. As such, the recommendation requirements are not clearly specified. Additionally, many systems select the recommended papers individually and ignore the interplay among the different papers. Therefore, the overall effect of a recommendation set is neglected.

This thesis identifies and discusses problems of paper recommender systems and addresses them by developing several metrics. With these metrics, different recommendation tasks can be clearly defined and differentiated. Thus, the first of the above mentioned problems of the paper recommender system community is approached. The metrics not only concentrate on single recommendations, but the whole set of papers that make up a recommendation. As such, they emphasize the need for the perception of recommendation sets as a whole and hence address the second problem.

By examining the generated recommendations, the defined metrics can additionally be used to characterize different algorithms. By comparing this information with the use case specification, suitable algorithm candidates can be selected.

In this thesis the capabilities of the metrics are shown for a particular use case: recommending scientific papers that give an introductory overview of a scientific field to a researcher. In two studies the suitability of the aspects underlying the metrics for the specific use case characterization is demonstrated. Moreover, the metrics themselves are evaluated. It is also shown that the metrics can distinguish between the recommendations of different algorithms. Therefore, they are suited to characterize algorithms which can be used to select algorithms for a specific recommendation use case. Finally, in this thesis a paper recommender system for the examined use case is developed and evaluated.