Databases Reference
In-Depth Information
recommender systems such as improved accuracy and explanations. The rating dimensions are, however,
not static in our approach and require metrics that are different to those put forward, for example, in
[Adomavicius and Kwon, 2007].
In this respect, this work is also in line with the ideas of Shirky [Shirky, 2005], who was among
the first who argued that using predefined (rating) categories leads to different challenges such as the
following. First, professional experts are needed who design the rating dimensions; in addition, new rating
dimensions may emerge over time that were not covered by the predefined and pre-thought static rating
dimensions designed or foreseen by a domain expert. In collaborative tagging systems, the set of rating
dimensions is not limited which allows users to pick their particular way of stating their preferences. Of
course, this comes at the price of a less homogeneous and more unstructured set of item annotations.
To the best of our knowledge, the concept of tag preference was first introduced by [Ji et al., 2007].
The authors present a tag preference based recommendation algorithm for a collaborative tagging system.
The authors first compute the target user's
candidate tag set
which consists of all tags for which a high tag
preference value was predicted. Afterwards a naıve Bayes classifier is used for making recommendations
by exploiting the user's candidate tag set. The proposed algorithm was evaluated on a data set collected
from the social bookmarking site Delicious
10
. In contrast to the work of [Sen et al., 2009b], the tag
preference predictor in [Ji et al., 2007] does not make use of item ratings at all because the Delicious data
set does not support ratings for items (bookmarks) like the tag-enhanced MovieLens data set.
In [Vig et al., 2009], Vig et al. propose another concept called
tag relevance
which describes “the
degree to which a tag describes an item”. In the example from Section 4.2, tag relevance would measure
how well the tag “Bruce Willis” describes a particular movie. Overall, in previous work
tag preference
was considered a user-specific concept whereas
tag relevance
is considered to be an item-specific concept.
In contrast, in this work our proposed concept of
tag preference
is user- and item-specific which has shown
to be a helpful means to capture the user's preferences more precisely and thus produce more accurate
recommendations.
Beside the relation of the work in [Sen et al., 2009b], which we extend by item-specific tag preferences in
this chapter, our approach is also closely related to the recent work of [Vig et al., 2010], who experimented
with a recommender system interface that allowed users to assign
affect
to the tags of a movie. In [Vig
et al., 2010], the authors introduce the idea of “tag expressions”, which, at its heart, represents the same
idea of rating items by rating tags proposed in our own previous work [Gedikli and Jannach, 2010c].
Users are able to assign a so-called
affect
(preference) - like, dislike or neutral - to tags in order to
measure a user's pleasure or displeasure with the item with respect to the tag. In contrast to this work
the authors are focussing on user interface aspects and how the possibility to express tag preferences
affects the tagging behavior of the community. In particular, Vig et al. also analyze the design space
of tag expressions and focus on three elements: preference dimensions, affect expression, and display of
community affect. In this work, however, we present first algorithms that consider tag preferences (tag
affects) to generate more accurate predictions which is one of the main challenges listed in the future
work section of [Vig et al., 2010]. Additionally we also show how to infer the user opinion regarding a
certain feature (tag) for a given item automatically.
4.7 Summary and outlook
The main new idea presented in this chapter is to incorporate item-specific ratings for tags in the rec-
ommendation process. Based on such an approach, users are able to evaluate an existing item in various
dimensions and are thus not limited to the one single overall vote anymore. In contrast to previous
attempts toward exploiting multi-dimensional ratings, this work aims to follow a Web 2.0 style approach,
in which the rating dimensions are not static or predefined.
The goal was to develop and evaluate different recommendation schemes that take item-specific tag
preferences into account when generating rating predictions. In addition, we proposed a metric to auto-
matically derive user- and item-specific tag preferences from the overall ratings based on item similarities,
in order to demonstrate that quality improvements can be achieved even when the tag preference data
is not explicitly given. The results of the evaluation on two data sets show that a measurable accuracy
10
http://www.delicious.com
