基于用户生成标签的多视角特征学习方法

Multi-View Feature Learning Based on User Contributed Tag

提出了一种基于用户生成标签的多视角特征学习方法。采用词袋模型分别得到媒体的内容特征表示和标签特征表示;通过媒体词汇和文本词汇的相关性建模,学习文本特征空间和内容特征空间的映射模型。在此基础上,给出了优化前后的特征表示具备近似等距映射保持的理论依据。该方法相对数据集规模具备线性时间复杂度,适用于大规模数据集,具备多视角特征融合能力。基准数据集上测试表明,优化后的特征表示较特征拼接和相关成分分析等方法鉴别力更强。

A multi-view feature learning method based on user contributed tag was proposed. Bag-of-words representation for content feature and textual feature was learned. A multi-view feature learning framework was proposed to explicitly model the relevance between multimedia object and tags by learning a linear mapping from textual representation to visual representation. The learned feature encoded the information conveyed by original feature, and inner products of leaned features were preserved with a high probability with visual features and textual features. The complexity of the method is linear with respect to the size of dataset. Furthermore, the method can be extended to deal with more than two views. The performance of the proposed method indicts it＇s superiority over other representative method.