Method and device of transfer learning classification

Abstract

The invention discloses a method and a device of dynmaic-TrAdaboost transfer learning classification based on clustering. An error and a dynamic factor of each iteration process are dynamically calculated through a different data set and weight values obtained by each iteration; then weights of the data set are dynamically adjusted in each iteration process according to the weights and the dynamic factor, data with small weights is deleted from the data set, and data with large weights is reserved; and the next iteration is carried out, and a final target classifier can be obtained when updating reaches a standard. Therefore, the problem of weight updating polarization can be avoided through dynamically adjusting the weights by the dynamic factor, and the precision of classification is improved. The invention also provides the device of dynmaic-TrAdaboost transfer learning classification based on clustering, and the device can also realize the above-mentioned technical effects.

Description

technical field [0001] The present invention relates to the field of classification learning, more specifically, to a method and device for clustering-based dynmaic-TrAdaboost transfer learning classification. Background technique [0002] The concept of classification is to learn a classification function or construct a classification model based on existing data, which is what we usually call a classifier. In the traditional classification learning, the classification model is mainly obtained by training the data set composed of labeled data, but in fact, the work of collecting labeled data is very difficult and requires a lot of energy and resources. Migration learning can process the source data set so that it can be migrated to the target data set, and the required data set can be obtained after migration, thereby solving the problem of lack of labeled data. [0003] In the existing transfer learning, the main idea is to automatically iteratively increase the weight of...

AI Technical Summary

Problems solved by technology

However, as the iteration progresses, due to some error interference in the iteration, the weight in the source domain will only decrease, and the weight in the target domain will only increase. Therefore, there will be polarization between the target domain and the source domain. , so that the weight of the source domain drops too fast in the process, and the classifier tends to over-focus on samples that are difficult to classify, but cannot focus on samples that are easy to classify

Images