Multi-view data commonly exists in real applications. Take a user-tagged image on the web, for instance. One of its views is represented by the image' s low level features, the other by its textual features. It is a big challenge for people doing data mining and retrieval how to effectively dig out valuable information from such kind of data. Multi-view Prediction ( MVP) algorithm, proposed by the authors, aims to obtain a subspace, where the mutual correlation between two views is maximized via canonical correlation analysis (CCA). In the training stage, it is expected to obtain, along with learning, the subspace composed by canonical vectors and the correlation coefficients with respect to the canonical vectors; in the prediction stage, the multi-view data score vector is generated by projecting the data to the subspace, then according to the score vector, through repeated regression, it effectively judges whether there is mutual correlation between two views of the testing sample. Experiments based on images with textual tags validate the effectiveness of the algorithm.%多视图的数据广泛存在于真实的应用中.比如说网络上用户标注的图像,一个视图是由图像的底层特征去表征,而另一个则由文本特征去表征.如何从这种类型的数据中有效地挖掘出有价值的信息对于做数据挖掘和数据检索的人来说具有很大的挑战性.提出多视图的预测算法(MVP)去获取一个子空间,在这个子空间上,通过典型相关分析使得两个视图之间的相互关系最大化.在训练步,期望能通过学习同时得到典型向量组成的子空间和对应典型向量的相关系数;在预测步,通过把数据投影到子空间上产生多视图数据的得分向量.再由得分向量通过多重回归有效地判断出测试样本两个视图之间是不是存在相互关系.基于文本标注图像的实验表明了算法的有效性.
展开▼
