Papers at RecSys 2008

A lot of interesting papers have been discussed during RecSys conference. Here are a few list of interesting ones.

Panagiotis Symeonidis, Alexandros Nanopoulos and Yannis Manolopoulos. “Tag Recommendations based on Tensor Dimensionality Reduction” Aristotle University, Department of Informatics, Thessaloniki, Greece

The goal of the authors is to provide better tag recommendation by performing a 3-dimensional analysis on the social tagging data (users which add tags to a set of resources) and by modeling them using a 3-order tensor and Singular Value Decomposition techniques. With such approach, they want to face two challenges which remain open by using a standard collaborative filtering algorithm to provide recommendation, which are:

  1. Capturing the 3-dimensional relationship between user-tags-resources
  2. Robustness to data sparsity

They perform an experimental comparison on Last.fm and the BibSonomy data set and they show that with their approach they can get better precision/recall performance compared to the FolkRank and PR algorithm.

Werner Geyer, Casey Dugan, David Millen, Michael Muller and Jill Freyne. “Recommending Topics for Self-Descriptions in Online User Profiles” IBM T.J. Watson Research, Cambridge, USA; University College Dublin, School of Computer Science and Informatics, Belfield, Dublin, Ireland

In this paper, the authors focus their attention on a specific ‘About you’ recommender system, which allows users to create their own profile by crafting a list of their own questions/topics. Such system differs from the traditional recommender system ones, since it recommends content for users to create, rather then consume. They deploy two different algorithms (Network-based and Content-based), with the aim of recommending a set of meaningful questions to the user by looking at the behaviour of the users which are similar to the target one.

They performed an experiment followed by a user study to evaluate the effectiveness of their algorithms and they found out that Network-based recommendations were liked most by users, followed by Content-based.

Nikhil Garg and Ingmar Weber. “Personalized, Interactive Tag Recommendation for Flickr”
Ecole Polytechnique Federale de Lausanne, Lausanne, Switzerland

The goal of the authors is to provide better tag reccommendation to users that want to annotate images belonging to Flickr dataset. They propose a new method, called Hybrid, capable of suggesting meaningful tags according to the hystory of the selected user (the list of tags he used in the past to tag all the resources), the hystory of the selected image and tag (how many tags have been attached to the image and how many images have been tagged by each tags in the past) and the relevance of each tag (a tag is considered relevant for an image if it has been given by the author of the selected picture).

They evaluated the effectivness of their algorithm on a selected set of images taken from Flickr dataset. They demonstrate that their algorithm improves performance expecially for cases of low-coverage, so when both the user tagged only a small subset of images in the past and the selected image was tagged only by using a small subset of tags.

Poster session: Andriy Shepitsen, Jonathan Gemmell, Bamshad Mobasher, and Robin Burke. “Personalized Recommendation in Social Tagging Systems Using Hierarchical Clustering”
Center for Web Intelligence, School of Computing, DePaul University, Chicago, Illinois, USA

The goal of the authors is to present a new personalization algorithm for recommendation in folksonomies which relies on hierarchical tag clusters. In order to give recommendations, they first perform an off-line jerarchical (k-means) clustering step to group tags according to the resources they have been attached to.
They infer then the interest of each user in each resource according to:

  1. The relevance of each resource for all the existing tag clusters
  2. The interest of the selected user in each tag cluster

They evaluated the effectiveness of their algorithm on del.icio.us and Last.fm data sets. The provided results show that Last.fm does not really benefit from applying their technique, while del.icio.us, which suffers from greater sparsity and which deals with tags that can have a variety of meaning across different topic areas, gets a boost from selecting clusters specific to the user’s query.

Yoon-Joo Park, Alexander Tuzhilin. “The Long Tail of Recommender Systems and How to Leverage It” Stern School of Business, New York University

The paper studies the Long Tail problem of recommender systems, such as the existence of a large set of items which have only few ratings. Since there is only a little information to exploit about these items, the quality of the recommendations involving them is very poor. To overcome the problem, they decided to split the data set into two parts (items heavily rated -head- and items rated only by a small subset of users -tail-) and performed a clustering phase (using the tool Weka) for data belonging to the tail. Clusters are used to infer unknown ratings for items belonging to the cluster based on the known ratings (using a Support Vector Machine model for each cluster).

They evaluated the effectivness of their algorithm on MovieLens and BookCrossing datasets. They computed for both dataset the prediction error and they show that their method outperforms the standard one (without the pre-clustering phase).

One Response to “Papers at RecSys 2008”

  1. [...] an interesting blog post by @HDrachsler, who I started following on twitter after this year’s RecSys conference. The post contains a recording of the question/answer time at the RecSys doctoral symposium (which [...]