This is a class of methods that combine both CBF and CF in a single recommender to achieve better results. Several approaches have been tried and can be summarized in the following categories:
- Weighted: The CBF and CF predicted ratings are combined in to some weighted mean.
- Mixed: CF and CBF predicted movies are found separately and then merged in to a single list.
- Switched: Based on certain criteria, the CF predictions or CBF predictions are used.
- Feature combination: CF and CBF features are considered together to find the most similar users or items.
- Feature augmentation: Similar to feature combination, but the additional features are used to predict some ratings and then the main recommender uses these ratings to produce the recommendation list. For example, Content-Boosted Collaborative Filtering learns the ratings of unrated movies by a content-based model and then a collaborative approach is employed to define the recommendations.
As an example, we implement two hybrid feature combination methods merging an item's features CBF method with a user-based CF method. The first method employs a user-based CF to the expanded utility matrix that now also contains the average rating per genre per user. The Python class is as follows:
The constructor generates the expanded utility matrix with the movies' genres average rating features associated to each user, Umatrix_mfeats. The function CalcRatings finds the K-NN using the Pearson correlation comparing the expanded feature vectors of the users. The second method applies and SVD factorization to the expanded utility matrix that contains the genre preferences for each user.
As the SVD method, the ratings are subtracted with the user rating's average, and genre preferences are subtracted from the same user rating's average.