关键词: 标签推荐, 多任务学习, 分组推荐, 联合推荐, 偏序学习, 特征融合

Abstract: With the rapid development of multimedia sharing websitesthe social image recommendation has become a hot research topic recently.It is much easier to manage social images by taging and grouping them.Traditional image recommendation methods tend to focus on a specific task just like tag or group recommendationwhich ignore the correlation between the two tasks.By fusing image features extracted from multiple tasksmulti-task learning method exploit image sharing or correlation representation among different tasks to improve the accuracy of the single task.Thereforethis paper proposes a novel social image tag and group joint recommendation model based on deep multi-task learning.In the signal taskthe tag and group recommendation are solved using comparison-based partial order deep learning method respectively for alleviating the data sparsity.Furthermorethe features extracted from their intermediate layer are saved for multi-task learning.In the convolutional neural network processing the visual features of social imagesthe features of two recommendations are connected.Then the dimension-reduction and automatic fusion of the features are realized by convolution.Hence image features extracted from different recommendations are shared.Moreoverthe size of processed features is suitable for the next layer of the convolutional neural network so that the network architecture of single recommendation task can be maintained.The experimental results on a real Flickr dataset show that compared with the traditional methodsthe accuracy and recall rate of the proposed algorithm are greatly improvedwhich proves the effectiveness and feasibility of the proposed method.

Key words: Feature fusion, Group recommendation, Joint recommendation, Multi-task learning, Partial order learning, Tag recommendation