An Ensemble Approach for Bundle Recommendation Based on Performance Evaluation

Aiming at the problems that an individual bundle recommendation approach is prone to producing recommendation bias and has poor generality, we propose an ensemble approach for bundle recommendation based on performance evaluation. First, we propose an ensemble bundle recommendation framework, which integrates the predictive power of multiple base recommenders to improve the bundle recommendation performance. Second, we train the base recommenders in the ensemble on the training set and endow them with importance weights according to their performance evaluation on the validation set. Finally, we propose two bundle recommendation strategies for the ensemble and aggregate the recommendation lists generated with these two strategies in a weighted manner. Experiments show that the average improvement of recall@K and NDCG@K for our ensemble approach over state-of-the-art methods is 3.82% and 7.39%, respectively.