论文部分内容阅读
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-generated tags was proposed. Using the bag-of-words model to obtain the media content representation and the label feature representation, respectively, and learning the mapping model of the text feature space and the content feature space through the modeling of the relevance between the media vocabulary and the text vocabulary. On this basis, the theoretical basis for the representation of equidistant mappings is given before and after the optimization. The proposed method has the linear time complexity relative to the scale of dataset and is suitable for large-scale datasets with multi-view feature fusion ability. Tests on the benchmark dataset show that the optimized features represent stronger discriminating power than methods such as feature splicing and related component analysis.