Multi-label classification method and system combining matrix decomposition and bidirectional mapping network

Abstract

The invention belongs to the technical field of computer application, and discloses a multi-label classification method and system combining matrix decomposition and a bidirectional mapping network, and the method comprises the steps: constructing an instance feature matrix and a category label matrix, carrying out the normalization processing, and constructing a standard multi-label data set; decomposing the category label matrix into a semantic label matrix and a potential label incidence matrix; constructing a multi-label classification model based on the semantic label matrix, the potential label incidence matrix, the model coefficient matrix and the label correlation matrix; respectively carrying out optimization solution on the model coefficient matrix, the semantic tag matrix and the potential tag incidence matrix to obtain a trained model; taking instances of invisible labels as a test set to be input into a label matrix predicted by the model; and comparing the predicted label matrix with the real label matrix under five evaluation indexes, and evaluating the performance and effectiveness of the model. Related experiments are carried out on multi-label data sets in multiple fields, and the experiments prove the effectiveness and competitiveness of the method.

Description

technical field [0001] The invention belongs to the technical field of computer applications, and in particular relates to a multi-label classification method and system of joint matrix decomposition and bidirectional mapping network. Background technique [0002] Currently, multi-label classification research aims to predict multiple possible labels for unseen instances, that is, an instance sample may match multiple label information at the same time. With the development of Internet technology, modern society has entered the era of big data. Therefore, large-scale data is generated every moment, and the amount is huge, and the data that needs to be labeled increases accordingly. Effectively mining valuable information from a huge amount of data lies in how to make the most appropriate classification, that is, classifying large-scale data. Traditional single-label classification is a one-to-one correspondence between instances and labels, while multi-label classification...

AI Technical Summary

Problems solved by technology

[0007] (1) Traditional single-label classification is a one-to-one correspondence between instances and labels, while multi-label classification is a "one-to-many" relationship between instances and labels, which brings huge challenges to multi-label classification

[0008] (2) Most of the existing multi-label classification algorithms are designed for multi-label classification, but most multi-label data sets have the characteristics of high dimensionality and large capacity

[0009] (3) The existing specific feature learning based on the subset feature selection method only considers the one-way projection from the instance space to the label space. This feature extraction method can be considered as a kind of data compression, and some instance information is likely to be lost. lost during compression, leading to inaccurate predicted labels

Images