Structured SVM-based unbalanced evaluation criterion direct optimization algorithm

Abstract

The invention discloses a structured SVM-based unbalanced evaluation criterion direct optimization algorithm. The algorithm is characterized by being implemented by the following steps of 1, selecting a data set and performing positive and negative sorting on all training samples in the data set; 2, defining a structured SVM framework-based target function; 3, defining an unbalanced evaluation criterion-oriented loss function delta GTP / PR(y,y) and an associated function psi (x,y) according to a loss function delta (y,y); 4, initializing a trade-off parameter C and an error allowable value epsilon; and 5, performing iterative solving on the target function by utilizing a cutting plane algorithm to obtain an institutionalized SVM framework-based target function. According to the algorithm, the classification precision of minority classes is improved; and the algorithm can be applied to classification of unbalanced data.

Description

technical field [0001] The invention relates to the technical field of statistical learning classification, more specifically, a direct optimization algorithm based on structured SVM imbalance evaluation criteria. Background technique [0002] With the advent of the data age and the rapid development of information technology, people will continue to generate massive amounts of data information in their daily life, production and business operations, and these data contain data information that is of great significance to people, but in There is often only a small part of information in such massive data that we really need to analyze and process. [0003] There are also many important applications in real life. For example, mobile phones and computers used in daily life provide telecom operators with a large amount of life data every day. Operators can obtain the preferences of different users by analyzing the effective information in these massive data, and launch suitabl...

AI Technical Summary

Problems solved by technology

[0005] Facing the learning problem of unbalanced data sets, the difficulty of research mainly comes from the characteristics of the unbalanced data itself: the minority class samples in the unbalanced data set are insufficient, and the distribution of samples cannot well reflect the actual distribution of the entire class; the majority class Noisy data is usually mixed in, so that the two types of samples tend to overlap to varying degrees

When the traditional classification methods in the field of machine learning are directly applied to the classification problem of unbalanced data, it is easy to misclassify the samples of the minority class, which makes the classification accuracy of the minority class very low.

Images