Fast Online Recommendation

Abstract

Based on the spatiotemporal recommender models developed in the previous chapters, the top-k recommendation task can be reduced to an simple task of finding the top-k items with the maximum dot-products for the query/user vector over the set of item vectors. In this chapter, we build effective multidimensional index structures metric-tree and Inverted Index to manage the item vectors, and present three efficient top-k retrieval algorithms to speed up the online spatiotemporal recommendation. These three algorithms are metric-tree-based search algorithm (MT), threshold-based algorithm (TA), and attribute pruning-based algorithm (AP). MT and TA focus on pruning item search space, while AP aims to prune attribute space. To evaluate the performance of the developed techniques, we conduct extensive experiments on both real-world and large-scale synthetic datasets. The experimental results show that MT, TA, and AP can achieve superior performance under different data dimensionality.