• No results found

Tag Recommendation

2.2 Tag Recommender Systems

2.2.2 Tag Recommendation

In [Symeonidis et al., 2008] Symeonidis et al. present a tag recommender system, whose main steps can be recapped as follows:

• The algorithm models the entities of the social tagging systems, users, items and tags by a 3-order tensor;

• a Higher Order Singular Value Decomposition algorithm is applied in 3-order tensors to reveal the latent semantic associations between users, resources and tags in order to perform the recommendations;

• a comparison with two state of the art algorithms is performed [Xu et al., 2006, Hotho et al., 2006b]

2.2. Tag Recommender Systems 31

The experimental results show significant improvements with respect to the state of the art algorithm in terms of e↵ectiveness measured through precision and recall. The proposed approach di↵ers from the one just described because it does not consider which items the users interact with. In [Rendle and Schmidt-Thieme, 2010], authors describe a tag recom- mendation system based on PITF model (Pairwise Interaction Tensor Fac- torization), which is a particular case of the Tucker Decomposition (TD) model with linear runtime, both for learning and prediction. The advan- tage of this model is that the complexity of the model equation is linear in the number of factorization dimensions, which makes it feasible for high dimensions. The proposed system operates in two steps: first, the system models interaction between users, items and tags, then it uses a Bayesian Personalized Ranking criterion to produce the recommendations. The ap- proach of tag recommendation proposed in this thesis does not use any probabilistic model.

In [Carmel et al., 2010], authors present a framework for social book- mark weighting, which allows to estimate the e↵ectiveness of each of the bookmarks individually for several Information Retrieval (IR) tasks. They consider each bookmark as an indivisible triplet (document, user, tag) and each bookmark is weighted by the framework, according to its predicted e↵ectiveness in describing the content of the document it is associated with, given that it was annotated by a specific user with a specific tag. This framework is able to perform tag recommendations, user recommenda-

tions, and document recommendations. The tag recommendation process is done by computing the similarity between each tag and other tags previ- ously used by the user and between each tag and the documents that have to be tagged. The study proposed in this thesis does not consider neither the tags previously used by the users, nor the similarity between tags and documents; in fact this system does not su↵er from cold start problem.

Inspired by the PageRank algorithm [Brin and Page, 1998], Hotho et al. present FolkRank algorithm [Hotho et al., 2006a]. The basic idea of this algorithm is that a resource tagged by important users with important tags is important itself. In this work authors represent the system through a undirected graph (while, in the PageRank algorithm, the edges of the graph have a direction), where the nodes represent users, resources, and tags and the edges represent the connection between tags and users, users and resources or tags and resources. In order to assign a weight to each node, the algorithm executes a random walk algorithm on the graph and recommendations are built by choosing the top ranked tags associated to a given tag. RATC di↵ers from this approach because it does not use a random walk algorithm to make associations between tags and resources and, moreover, the work proposed in this thesis updates these associations continuously and not only when new resources are added to the system.

In [Givon and Lavrenko, 2009], authors describe a system that recom- mends tags for full text books. They use a dataset composed only by books written in English, and that belong to the fiction/literature domains, which

2.2. Tag Recommender Systems 33

are split into a training set and a test set. Furthermore, they collect a set of social tags, which they pre-process by means of a stemming task, dupli- cate removing task. Then, each book is represented as a TF-IDF vector and each tag is associated to a given book by using a Relevance Model, which is a method adopted from Information Retrieval to match documents to a given query. Through Relevance Model, the system selects a set of tags to recommend for each book.

In [Sigurbj¨ornsson and van Zwol, 2008], authors present an approach to support the user during the tagging process of a photo in Flickr. Given a photo with user-defined tags, a set of candidate tags is derived for each of the user-defined tags, by using a “promotion function” based on tag co-occurrence. The lists of candidate tags are then used as input for tag aggregation and ranking, which ultimately produces the ranked list of n recommended tags. Tag co-occurrence identification is a crucial task in the presented tag recommendation approach. Co-occurrences between two tags is defined as the number of photos in the system for which both tags are used in the same annotation. It is common to normalize the co-occurrence count with the overall frequency of the tag. Once the lists of candidate tags for each of the user defined tags are built, a tag aggregation step is needed to merge the lists into a single ranking of tags to recommend. Even if co-occurrences of tags in resources are considered (like RATC does), the proposed system continuously and implicitly monitors the tagging behavior of users. Similarity between tags is not calculated

using a promotion function, which is built with an observation of the tagging system at a certain time, but considering the use of the tags at the moment in which similarities are calculated.

Related documents