Support vector machine semi-supervised learning method in time-frequency joint

Abstract

The invention discloses a support vector machine semi-supervised learning method in time-frequency joint. The method comprises the specific steps as follows: step one, training initial SVW classifiers; step two, searching high-confidence samples using the SVW classifier C1 and the SVW classifier C2 to form a high-confidence sample set S; step three, automatically annotating the samples in the high-confidence sample set, placing in a annotated sample set L of the SVW classifier C; step four, repeatedly training the SVW classifier C using the updated annotated sample set L; step five, judging whether to quit the circulation or continuously iterate according to a stopping rule. The confidence of the sample is judged by uniting the feature spaces of the time domain and frequency domain, the judgment to the sample confidence is more accurate in comparison with the traditional judgment based on the single feature space; since the judgment to the sample confidence is more accurate, the classification performance reduction of the classifier caused by the false annotating can be reduced, and the workload of manual annotation is greatly reduced when the method is used for the training of the SVW classifier.

Description

technical field [0001] The invention relates to the field of machine learning, in particular to a time-frequency joint support vector machine semi-supervised learning method. Background technique [0002] Classification problems widely exist in many disciplines, such as speech recognition, image recognition, audio classification, text classification and so on. In order to obtain a classifier with good classification performance, it is usually necessary to use a large number of labeled samples to participate in the training of the classifier. However, the cost of obtaining labeled samples is relatively expensive. For example, in the field of speech recognition, phonemes need to be labeled one by one, and labeling is particularly time-consuming and energy-intensive. The high cost of sample labeling makes it impractical to completely rely on manual labeling to achieve training sample labeling. Based on this, semi-supervised learning techniques emerged as the times require. Se...

AI Technical Summary

Problems solved by technology

At present, there have been many related studies on finding high-confidence samples. In the semi-supervised learning method using Support Vector Machines (SVM) as the training model, the existing algorithms are generally after feature extraction. Here Judging the confidence of a sample in a feature space, this method of judging the confidence of a sample in a certain feature space is generally ineffective, and the method of combining two or more feature spaces to judge the confidence of a sample, its confidence in the sample The degree of judgment will be more accurate, so the present invention combines the two feature spaces of time domain and frequency domain to judge the confidence of samples, so as to find samples with real high confidence to participate in training

Images