Information Technology Reference
In-Depth Information
cluster. The computed clusters are treated as nodes. Thus, path-based search strategies
can be used for searching relevant entities.
The advantages of model-based recommenders are that the complexity of the
dataset can be effectively reduced to speed-up the computation of relevant entities.
Furthermore, the reduction of noise in large datasets often improves the result quality.
The algorithm applied for reducing the graph complexity highly depends on the
domain. We decided to focus on Hierarchical Agglomerative Clustering since it
enabled us to choose similarity measures and clustering parameters optimized for
each relationship set. Moreover, for recommendations computed based on clustered
entity sets, path-based explanation can be provided. The disadvantages of model-
based recommenders are that additional effort is needed for calculating and updating
the model. A prediction based on clustering is presented in Fig. 7.13 . As the results
of the cluster algorithm are most of the time only loosely related to the input node,
the results from the clustering are not considered in the evaluation.
7.5 Evaluation
The goal of evaluation is to research the impact of an enriched user profile on the
cold start problem for CF. We therefore consider two evaluation scenarios:
New user and new application
The first scenario covers the cold start problem for a new music recommendation
application with few users. In this scenario, we want to analyze the effect of the
enriched user profiles for a new music recommendation application that has a small
number of users and how recommendation quality is affected for new users.
Fig. 7.13 Cluster-based prediction: Explanation of cluster-based enrichments using automatically
generated genre cluster
Search WWH ::




Custom Search