Before the introduction of multi-view learning, let's review the single view of learning. Single view of learning can be seen in traditional spam filtering. For example, we train a classifier to filter junk e-mail, we can judge by natural language processing (NLP) to go through the contents of the message whether it is spam. We usually add some features, such as document length ratio punctuation mail. The new features and content of the document placed under the same view, we view here can be understood as a pd.dataframe. Then through supervised learning (supervised-learning) to train a model.
The multi-view learning can solve more complex problems. We come to understand through a simple example. We are now building a firm defense system to detect abnormal behavior of employees. This defense system employees can see the network mail exchanges, as well as company employees to access the log file. File access logging systems learned employee A, B, C I only access to resources, personnel D, E, F, G and I are simultaneously access resources resources II. In this case no abnormality information may be detected. Next, the system network by e-mail learned: Employee {A, B, C, D} in a team of employees {E, F, G} at another team. By these two views, we can detect abnormal behavior of D staff. Because the first team {A, B, C, D} he is the only one person at the same time I access resources and resource II, where we can put this team understanding for the development of the sector projects I. {E, F, G} are simultaneously responsible for the project, and the project I II (for example, management roles). For the second team, while access to resources I and II resource is normal, and the first team of D should not access resources II. We can see that it is necessary from two different views to determine abnormal behavior D staff.
例子来源: Liu, A., & Lam, D. (2012). Using Consensus Clustering for Multi-view Anomaly Detection. In 2012 IEEE Symposium on Security and Privacy Workshops (pp. 117–124). IEEE. https://doi.org/10.1109/SPW.2012.18）
The multi-view learning as well as do semi-supervised learning (semi-supervised learning) to save labor costs, as well as good at coping mechanism to bypass the detection of malicious users. To be continued. . .