Heterogeneous classifier aggregation method for improving classification fairness

Abstract

The invention provides a heterogeneous classifier aggregation method for improving classification fairness, aiming at the problem of category prejudice caused by unfair supervision in multi-source heterogeneous classifier aggregation, firstly, the unfair supervision level of each category in the aggregation process is quantitatively analyzed, the category relevancy is measured, and then misclassification cost is customized for each category according to the misclassification cost; and the importance of each category in the classifier aggregation process is adjusted by taking misclassification cost as a penalty parameter, so that the discrimination of the aggregation classifier on the category with relatively weak supervision and the preference of the aggregation classifier on the category with relatively strong supervision are relieved, the classification fairness is improved, and better performance is achieved.

Description

technical field [0001] The invention relates to the field of terminal intelligence and model aggregation, in particular to a heterogeneous classifier aggregation method for improving classification fairness. Background technique [0002] With the explosive growth of smart devices and the maturity of deep learning, on-device model training has become more and more common. This can be applied in many ways, such as facial recognition. However, due to limited local training data and computing resources on devices, the performance and classification ability of locally trained classifiers are usually not too strong. Heterogeneous classifier aggregation has become a paradigm that can integrate local classifiers trained on multi-source heterogeneous devices with heterogeneity in architecture and target tasks into a more capable or better-performing comprehensive classification device. [0003] Existing work proposes a heterogeneous classifier aggregation method based on knowledge...

AI Technical Summary

Problems solved by technology

However, in this process, since the local classifiers on multi-source heterogeneous devices have heterogeneous target tasks, the quantity and quality of the knowledge of each category transferred to the aggregated classifiers in the knowledge distillation-based classifier aggregation process are not Unbalanced, which makes the training of the aggregated classifier under unfair supervision, and the final generated aggregated classifier may have category bias (that is, the classification performance on each category is unfair, and samples belonging to weakly supervised categories are more likely to belong to supervised Samples of stronger categories are more likely to be misclassified), affecting the performance of the aggregation classifier

Images