Multi-view classifier based on information enhancement and design method thereof

Abstract

The invention discloses a multi-view classifier based on information enhancement and a design method thereof. The multi-view classifier comprises a multi-view data collection module, a missing sampleinformation restoration module and an effective sample information enhancement module which are connected in sequence. According to the method, through two aspects of missing sample information restoration and effective sample information enhancement, the method is applied to the fields of ports and the like under the action of related interfaces, and the classification performance of a multi-viewdata set in an actual scene can be effectively improved.

Description

technical field [0001] The invention relates to the technical field of multi-view learning, in particular to an information-enhanced multi-view classifier and a design method thereof. Background technique [0002] In the process of building a "smart city" in an all-round way, the data sets that people need to deal with often have multiple manifestations or sources. This type of data set is called a multi-view data set, a form or source is a view (such as text, image, video in a web data set), and the different types of information contained in any view are called Features (such as text color, text size, text weight in text perspective). Different from single-view data sets with a single form or source, due to the relatively complex structure of multi-view data sets, processing is more difficult, and generally requires a multi-view classifier based on this type of data set to solve. In addition, as far as the features of multi-view datasets are concerned, they are further d...

AI Technical Summary

Problems solved by technology

[0007] (2) For multi-view datasets where the proportion of labeled samples is too small: Due to the limitation of labor costs, for most multi-view datasets used in real scenes, the proportion of labeled samples obtained before classifier training The ratio is very small, which makes the relevant multi-view classifiers have extremely limited prior knowledge at the beginning of training.

For example, the role of original unlabeled samples is not considered (although they provide little prior knowledge, it does not mean that these samples are worthless), and the role of different perspectives or features on classifier design is not reflected.

Images