Recommender Systems for Social Tagging Systems: A review by Epaminondas Kapetanios
The book, authored by Leandro Balby Marinho, Andreas Hotho, Robert Jäschke, Alexandros Nanopoulos, Steffen Rendle, Lars Schmidt, Thieme Gerd Stumme, Panagiotis Symeonidis, and published by Springer in 2012, discusses the role of recommender systems in order to serve social tagging systems.
With the emergence of Web 2.0 application, or the participatory Web, social tagging systems, such as Bibsonomy and CiteULike, gained momentum as a response to the problem and challenges posed by the creation of adequate descriptions of Web resources, be them documents, articles, Web pages, URL’s, images, etc., in order to foster sharing, effecting search, reusability of resources on the Web. Though social tagging systems are promising, the book is also an attempt to respond to the challenge imposed by crowd sourced descriptions with its many facets such as variety in users’ conceptions, disagreements, innate difficulty to converge towards common agreed upon vocabularies, lack of gold standards, just to name a few.
In this context, the book is a useful contribution towards harnessing crowd sourced descriptions by using recommender systems as it epitomises the long experience of the majority of the authors in social tagging systems, particularly, the experience and long research activities via the development of Bibsonomy at the University of Kassel, Germany, a social tagging system for BibTex based bibliographies for the sake of sharing and finding of research articles on the Web, which are related with specific users’ interests.
To this extent, the book is structured in such a way that brings together the two worlds of social tagging systems and recommender systems. In part I, an introduction into these two worlds is presented. The discussion of social tagging systems primarily focuses on folksonomies, a long standing model for users’ based tagging and retrieval of resources on the Web, together with their underpinning mathematical definitions such as tensors and hypergraphs. A reference to existing systems, e.g., Bibsonomy and CiteULike, rounds up the discussion. The discussion of recommender systems primarily focusses on those aspects related with social tagging, since recommender systems are considered as a huge topic, which is already covered extensively by the Recommender Systems Handbook, published in 2011, and includes aspects related with the ones covered in the book under review. These are, for instance, rating and item prediction, context-aware and multi-mode recommendations, as well as the associated ranking, regression, new user/item problems.
In part II, however, the bridge between these two worlds is being established in terms of baseline and advanced techniques, together with a follow-up discussion of off-line evaluation techniques. Part III concludes the book by discussing implementation aspects of recommender systems for social tagging ones. It is no surprise that Bibsonomy system architecture dominates the discussion in this part of the book as well, together with an online evaluation challenge for such systems, in conjunction with the European Conference on Machine Learning (ECML), Principles and Practices of Knowledge Discovery in Databases (PKDD).
After having finished reading the book, and from a birds-eye perspective, the entire book resembles the writing style of a survey paper with an introduction, main thesis, implementation and evaluation, as reflected by collections of smaller survey papers. To this extent, the book has definitely a strong academic and learning character, which is also strengthened by the mastery of mathematical approaches in the area, though this will probably limit its choice from the interested readership as it sets the barrier high in regards with the assumption of a readership already in knowledge of the advanced mathematics used in the book. For the active researcher in the area, or those who aspire to become, it provides a nice survey of mathematical techniques. Another aspect, which may limit the potential readership, is the highly specialised thematic area addressed by the authors. Finally, the strong overlap with contents from the Recommender Systems Handbook, 2011, may further handicap this book as first choice among the interested readership.
Nevertheless, the engaged researcher in the area will benefit from reading this book in that it provides a good orientation over state-of-the-art approaches and techniques for building recommender systems in order to harness the innate variety of crowd sourced annotations and tags. As an active researcher in semantic computing, the reviewer enjoyed the discussions about folksonomies and factorisation methods, as well as the discussions of algorithms highly related with implications for information retrieval aspects, thereof. In addition, practicing named entity recognition and evaluation in a business and industrial context, whilst facing the lack of gold standards, the discussion of alternative evaluation techniques such as involving users by randomly picking tagged resources, or the separation of an existing data set into dynamically changing two parts, the first to be used as an increasing and learning reinforcing training data set, the second to be used as decreasing test data set, or the gold-start problem, where no previous users’ actions are recorded, have been a valuable lesson learned and eyes opener.