Databases Reference
In-Depth Information
algorithms like FolkRank [Hotho et al., 2006] and PageRank [Brin and Page, 1998]. However, the TF-
based model comes with the problem of a cubic runtime in the factorization dimension for prediction
and learning. This problem is addressed in the work of [Rendle and Schmidt-Thie, 2010]. Rendle and
Schmidt-Thie present a Pairwise Interaction Tensor Factorization (PITF) model with a linear runtime in
the factorization dimension. The PITF model explicitly models the pairwise interactions between users,
items, and tags.
In [Liang et al., 2008], [Liang et al., 2009a], and [Liang et al., 2009b] extended standard collaborative
filtering approaches are presented. In [Liang et al., 2009b], for example, the authors present a tag-based
similarity measure to improve standard collaborative filtering approaches. The idea is to cluster users
with similar tagging behavior instead of similar rating behavior. In [Liang et al., 2009a], another tag-
based method is presented to accurately determine the nearest neighbors of the target user. Liang et al.
address the limited tag quality problem [Sen et al., 2007; Bischoff et al., 2008] by building user profiles
based on popular tags since according to the authors tag quality is related to tag popularity. Therefore,
the authors suggest to map the individual tags of each user and item to these commonly used tags to
build a better understanding of the users and the items. Popular tags are referred to as “tags that are
used by at least
θ
users, where
θ
is a threshold”. An experimental evaluation was conducted using a book
data set crawled from Amazon.com. The precision and recall results of the compared approaches show
that their popularity-based approach outperforms both Tso-Sutter's approach [Tso-Sutter et al., 2008]
and Liang's approach [Liang et al., 2008]. The results also indicate that their approach performs better
than a recommendation approach based on Singular Value Decomposition (SVD) [Funk, 2006].
In [Wang et al., 2010], a tag-based neighborhood method is incorporated into a traditional collabora-
tive filtering approach. Tag information generated by online users is used to retrieve closer neighbors for
the target user and the target item. The underlying idea of their neighborhood method is to combine both
usual rating neighbors and tag neighbors to find the best neighbors. The evaluation on the tag-enhanced
MovieLens data set shows that such an approach can produce more accurate rating predictions compared
to other algorithms based on non-negative matrix factorization and Singular Value Decomposition. In
particular, the observed improvements in predictive accuracy were comparably strong for sparse data sets
as more data sources are used.
In [Sen et al., 2009b], Sen et al. propose tag-based recommender algorithms which they call “tagom-
menders”. The idea is to utilize tag preference data in the recommendation process in order to generate
better recommendation rankings than state-of-the-art baseline algorithms. Since no tag preference data
is available, the tag preferences of the target user have to be estimated before the algorithm can predict
a user's preference for the target item. To that purpose, the authors evaluate a variety of tag preference
inference algorithms. Such algorithms estimate the user's attitude toward a tag, that is, if and to which
extent a user likes items that are annotated with a particular tag. Their results show that a linear com-
bination of all preference inference algorithms performed best, that is, algorithms that exploit a variety
of signals such as implicit and explicit user data work best.
After that, the rating prediction for an item is based on the aggregation of the inferred user prefer-
ences for the tags assigned to that item. Again a hybrid approach achieved the best accuracy results,
followed by an SVM-based method. Overall, the evaluation on a tag-enhanced MovieLens data set shows
that tag-based recommender algorithms utilizing users' estimated tag preferences lead to more precise
recommendations than the best tag-agnostic collaborative filtering algorithm.
In [Harvey et al., 2010], a tag recommender is presented which is based on a new probabilistic latent
topic model influenced by LDA [Blei et al., 2003]. The authors extend the LDA model such that infor-
mation about the users who provided each annotation is taken into account leading to a tripartite topic
model (TTM) which covers the whole tripartite structure of a folksonomy. The results show that their
TTM approach outperforms the basic LDA approach and other popularity-based approaches on different
accuracy metrics. In particular, the results are strong for sparsely annotated items which are often the
case in real-world tagging systems.
In line with previous work, we present in the Chapters 3 and 4 new tag-based algorithms which
are able to outperform other state-of-the-art recommendation algorithms on various dimensions such as
predictive accuracy or prediction time.
