Candidate Strategy for GenericUserBasedRecommender

问题:

In mahout you can define a CandidateItemsStrategy for GenericItemBasedRecommender such that specific items e.g. of a certain category are excluded. When using a GenericUserBasedRecommender this is not possible. How can I accomplish this with GenericUserBasedRecommender? Is the only way to do this using a IDRescorer? If possible I'd like to avoid using a IDRescorer. Thank you for your help!

[Edit]

For the item based recommender I do it like this:

private final class OnlySpecificlItemsStrategy implements CandidateItemsStrategy {
    private final JpaDataModel dataModel;

    public OnlySpecificlItemsStrategy(JpaDataModel dataModel) {
        this.dataModel = dataModel;
    }

    @Override
    public FastIDSet getCandidateItems(long userID, PreferenceArray preferencesFromUser, DataModel dataModel) throws TasteException {
        List<Long> specificlItemIDs  = this.dataModel.getSpecificlItemIDs();
        FastIDSet candidateItemIDs = new FastIDSet();

        for (long itemID : specificlItemIDs)
          candidateItemIDs.add(itemID);

        for (int j = 0; j < preferencesFromUser.length(); j++)
            candidateItemIDs.remove(preferencesFromUser.getItemID(j));

        return candidateItemIDs;
    }

}

For the user based recommender I do it with a Rescorer:

public class FilterIDsRescorer implements IDRescorer {

    FastIDSet allowedIDs;

    public FilterIDsRescorer(FastIDSet allowedIDs) {
        this.allowedIDs = allowedIDs;
    }

    @Override
    public double rescore(long id, double originalScore) {
        return originalScore;
    }

    @Override
    public boolean isFiltered(long id) {
        return !this.allowedIDs.contains(id);
    }

}

and then set it up like this:

List<Long> specificItemIDsList = dataModel.getOtherSpecificlItemIDs();
FastIDSet specificItemIDs = new FastIDSet(specificItemIDsList.size());
for (Long id : specificItemIDsList) {
    specificItemIDs.add(id);
}
this.filterIDsRescorer = new FilterIDsRescorer(specificItemIDs );
userBasedRecommender.recommend(userID, howMany, this.filterIDsRescorer)

For the purpose of filtering/excluding certain items I could also subclass my data model for each type of recommender, but then I cannot share the same data model instance which would have an impact on performance.

回答1:

There isn't a way to do this. The user-based and item-based algorithms are not quite symmetric, and it's mostly on purpose. The user-based system already has a notion of user neighborhood which is kind of like this idea. IDRescorer is unrelated.