User / item side information

[rburke2233]

We are working on a number of projects related to fairness-aware recommendation. In our implementations, the system needs to know the status of a user (or an item) -- whether part of a protected group or not. This is a particular case of an algorithm needing side information about users (or items) beyond user ratings: in this case, a "group membership" feature. In general, you would want to have the capability of loading a sparse user-feature (or item-feature) matrix.

I have looked at the Appender classes, but these do not quite meet our needs -- they seem to be intended to store user-user or item-item associations, not user-feature or item-feature associations.

Two questions:

• Is there work already underway to integrate user-feature and/or item-feature associations into LibRec? The DocumentDataAppender class is suggestive, but there is no associated implementation.
• If not, would the DataAppender interface be the place to start? Its signature seems somewhat limited. Is there another place in the framework where it would be appropriate to integrate a general capability for loading such features?

[SunYatong]

You can implement a data appender to read your data refer to the existing ones.

[rburke2233]

Do you recommend that we follow the pattern of SocialRecommender and its subclasses? We could create an abstract class titled FairRecommender and then subclasses for the specific implementations. Should we create a subdirectory of recommender called "fair" or would some other structure be preferable?

[SunYatong]

Yes, that's exactly what I mean and your structure is good to work.

• For example, you can implement a class called FairDataAppender under librec/data/convertor/appender directory to read and store users' fairness data. Then get these data in FairRecommender's setup function ((FairDataAppender) getDataModel.getDataAppender).getUserFairnessData().

• And the new appender need to be configured in driver.class.props just like adding new recommenders.