Alternatives to Neighborhood-Based CF - Week 1

The goal of my summer project is to improve the CF framework from different aspects, including rating normalization (aka. global effects removal), interpolation of neighborhood weights. I will also work on implmenting more expressive CF models like BiasSVD and SVD++.

For the first week, I have implmented all common-used data normalization methods: OverallMeanNormalization, UserMeanNormalization, ItemMeanNormalization, ZScoreNormalization, and CombinedNormalization<> which can apply several normalization methods in a sequential manner. Explanation on rating normalization and global effects removal can be found in this paper. Data normalization in CF complies with policy-based design, so a user can easily write a customized data normalization class. Normalization class needs to implement Normalize() and Denormalize() methods which are used to process ratings in CF. I have run an experiment with these normalization methods and found that, with normalization added, the performance of CF with the default factorizer can be significantly improved.

For the next week, I plan to add tests for the normalization classes to ensure they produce reasonable prediction accuracy. I will also start to investigate and implment different weight interpolation poclies for aggregation of neighborhood ratings.