Multi-relational matrix factorization is an effective technique for incorporating heterogeneous data into prediction tasks, such as personalized recommendation. Recent research has extended the set of relations that can be applied within heterogeneous network settings by composing non-local relations using network meta-paths. One of the key problems in applying this technique is that the set of possible non-local relations is essentially unbounded. In this paper, we demonstrate that an information gain based technique for heuristic pruning of relations can enhance the performance of multi-relational matrix factorization recommenders.
展开▼
