CrowdStart: Warming up cold-start items using crowdsourcing

The cold-start problem is one of the critical challenges in personalized recommender systems. A lot of existing work has been studied to exploit a user-item rating matrix as well as additional information for users/items, e.g., user profiles, item contents, and social relationships among users. However, because existing work is primarily biased to the auxiliary information for users/items, it is difficult to identify various and reliable item neighbors that are relevant to cold-start items. To alleviate this limitation, we propose a new crowd-enabled framework, called CrowdStart, which is an integrated human-machine approach for new item recommendation. The main contributions of the CrowdStart framework are twofold: (1) To find various and reliable item neighbors for new items, we design two-step crowdsourcing tasks that harness not only machine-only algorithms but also the knowledge of crowd workers (including a few experts and a large number of non-expert workers in a crowdsourcing platform). (2) We develop a novel hybrid model to exploit the user-item rating matrix, the content information about items, and the crowd-based item neighbors from human knowledge into new item recommendation. To evaluate the effectiveness of the CrowdStart framework, we conduct extensive experiments including both a user study and simulation tests. Through the empirical study, we found that the CrowdStart framework provides relevant, diverse, reliable, and explainable crowd-based neighbors for new items and the crowd-based neighbors are meaningful for improving the accuracy of new item recommendation. The datasets and detailed experimental results are available at https://goo.gl/1iXTUE. (C) 2019 Elsevier Ltd. All rights reserved.