A Multi-clustering Unbiased Relative Prediction Recommendation Scheme for Data with Hidden Multiple Overlaps

This paper presents a recommendation system designed for shortening and streamlining hiring processes on both candidate and recruiting company parts, which is specialized in Hi-Tech companies. In this case, the hidden multiple overlaps challenge arises, where the trivial matching of data items to types does not necessarily reflect a ground truth labelling and other hidden multiple matches may exist. To this end, we propose a multi-clustering unbiased relative prediction recommendation scheme. Our mechanism enables to efficiently deal with data containing hidden multiple overlaps. The multi-clustering allows a many-to-many matching of items to types, enabling to expose hidden item-user matches. The mechanism’s efficiency is vital for supporting a dynamic modeling implementation, which is required for the Hi-Tech recruiting application. Moreover, we design an unbiased relative prediction scheme for providing recommendations. Our prediction scheme is two-sided (but asymmetric): ranking the multiple items for a given user and (differently) ranking the multiple users for a given item. Though the underlying problem of overlapping clustering and multi-label learning is complicated, our scheme enables to bypass this difficulty in a conceptually exquisite and efficient way for recommender systems, in particular, for the Hi-Tech recruiting process.