Connecting with the collective
Yue Shi, Martha Moon Larson, Alan Hanjalić · 2010
Collaborative recommendation (CR) approaches have proven effective for the Top-N recommendation task. We introduce a novel approach, Rerank-CR, that further improves the Top-N results of an arbitrary CR algorithm using a post-processing step involving Bayesian reranking. The defining characteristic of Rerank-CR is that reranking is self contained, meaning that it requires no external resources, but rather makes use of information derivable from the original user-item matrix. Rerank-CR achieves top performance when used for incorporating collection-level information reflecting global tendencies as constraints into conventional CR, which we refer to as 'connecting with the collective'. Because information about the preferences of the collective is derived directly from the dataset, Rerank-CR has no need of an explicit model of rating styles within a certain community. Further, it is possible to adapt the domain of application (e.g., change to a different cultural setting) without explicit intervention. We evaluate Rerank-CR with experiments that demonstrate the ability of the basic Rerank-CR concept to improve an initial Top-N recommendation list and also the additional improvement achieved by 'multimodal' Rerank-CR, which integrates the collective modality. Additional experiments confirm that the performance of Rerank-CR is significant across different datasets.