Tag-aware Recommender Systems by Fusion of Collaborative Filtering Algorithms
Karen H. L. Tso-Sutter, Leandro Balby Marinho and Lars Schmidt-Thieme (SAC-08)
The paper “Tag-aware Recommender Systems by Fusion of Collaborative Filtering Algorithms” introduces an approach that allows to integrate tags in recommender systems in order to improve recommendation quality. Users’ rating information to items and content information of items are widely exploited in most of traditional recommender systems. Unlike attributes which are “global” descriptions of items, tags are “local” descriptions of items given by the users. Thus, tags could be interesting and useful information to enhance recommender system algorithms.
The main goal of recommender systems is to predict items or ratings of items that users are interested in so that they can recommend items to the users. Metadata such as content information of items has typically been used as additional knowledge to improve the quality of recommendations. Attribute aware RS algorithms is typically attached to the items and is usually provided by domain experts, which means an item always has the same attributes among all users. In contrast, tags are provided by various users not only associated to the items but also to the users act as additional background knowledge to improve RS algorithms; tags provided the users not solely help to describe the items, but also contain information about the users’ preferences. Karen et al. take use of this information and incorporate them to standard collaboratively filtering algorithms, by reducing the three-dimensional correlations to three two-dimensional correlations and then applying a fusion method to re-associate these correlations.
To extend traditional collaborative filtering (CF) with tags, the authors consider tags as a three-dimensional relation between users, items and tags. They deal with this three-dimensionalities by projecting it as three two-dimensional matrices: user-item matrix, item-tag matrix (called item tags) and user-tag matrix (called user tags).
- User tags: are tags that a user uses to tag items and are viewed as items in the user-item matrix.
- Item tags: are tags that describe an item by users and play the role of users in the user-item matrix
After extending original user-item matrix by including user and item tags, Karen et al. apply both user-based and item-based CF algorithms and fuse the predictions together. They fuse the user-based and item-based predictions by calculating the sum of the two conditional probabilities that are based on user-based and item-based similarities. The similarities are computed using standard user-based and item-based CF. They also use a parameter λ to control the impact of the two predictions.
For the experiment, the authors use the dataset on Last.fm (i.e., a radio and music community website which allows user to tag the music), including 1853 items (artists), 2917 users and 2045 tags. They run 10-fold cross validation and additionally split the training data to optimize the parameters λ and k (the neighborhood size). For the testing set, they randomly select one listened item from every user. The system generates top 10 recommended items for each user, and they use recall metric to evaluate the quality of the recommendations. The results show that:
- The fusion method (i.e., using both user-based and item-based CF), both with and without tags, significantly outperform the standard CF models. With tags, they achieve the highest recall.
- One interesting finding is that incorporating tags to the baseline models does not improve the recommendation quality. Probably, applying user or item tags alone does not exploit the characteristic of tags correctly, because tags present the relationship between users, items and tags.
In this paper, Karen et al. show a generic method that take advantage of social information (i.e., tags provided by users) and integrate them to standard CF algorithms to improve the quality of recommendations. They also present an approach that deals with the 3-dimensional correlation between the users, items and tags. The findings from the experiment have suggested that the adapted fusion method has successfully captured the relationships between users, items and tags. However, the authors do not discuss about the efficiency of their approach. The low performance is an existing issue of user-based CF, but their method uses both user-based and item-based, so the performance in terms of efficiency should not be better, if not worse. Moreover, they just use recall for evaluation, and the testing set is only one item, but the number of recommended items are 10. Therefore, it is not a very strong evidence to be applied to real recommender systems.