Ontology-based Content Recommendation
- The framework of our TV program recommender system
- The recommender system has basically content-based framework. In addition to the framework, the system consists of a user clustering part and feature extraction part with ontology.
- User clustering model
- The user cluster is built up using bottom-up approach.
- New user who has no rating history initially belongs to the root cluster.
- According to the user's accumulated rating history, the cluster which is belonged by the user is changed and has more similar pattern to the user's rating pattern.
- To measure the similarity, Cosine similarity and Pearson's correlation is used.
- Item feature supplyment from ontology
- The ontology consists of basic ontology which has TV program information and domain specific ontologies which have detail information in their own domains.
- From the ontologies we can extract userful specific and dynamic features which cannot be obtained from EPG metadata.
- Classifier Algorithm
- The classifier can be implemented using linear regression, SVM, or maximum entropy technique.
We are developing the TV program recommender system which solves ramp-up problems for users and items. To avoid user ramp-up problem, we use the user clustering technique. As an user's rating histories are accumulated, the system can find the cluster which has more similar pattern to the user. With this cluster information and the user's rating histories, the system generates an classifier and any items (including new items) can be scored by the classifier. To make a novel classifier, we add more detail features using the ontology technique. From the ontologies which are previously built, we extract useful features in addition to the basic EPG metadata. Our overall framework can be thought as a hybrid approach which use collaborate filtering technique and content-based technique with ontologies.
