Mining Abstract Highly Correlated Pairs

Recommendation systems are essentially solving a prediction problem where, given that p items have already been selected or rated by a user, the goal is to propose k target items most likely to be appreciated by her/him. Many models have been proposed to identify these target items but the results are not always satisfactory in practice because they often only include the most popular items and ignore the "long tail" of items that are either less popular or new ones. This paper investigates the use of a type of domain abstraction to search for highly correlated pairs of abstract items that are then used to infer other target items of interest. The advantage of this approach is evaluated on the basis of real data showing better results compared to an approach only based on the concrete pairs. Basing on an empirical study we confirm that the accuracy improvement is linked to the relevance of the domain abstraction.