Heterogeneous information fusion: Combination of multiple supervised and unsupervised classification methods based on belief functions

In real-life machine learning applications, a common problem is that raw data (e.g. remote sesning data) is sometimes inaccessible due to confidentiality and privacy constrains of corporations, making classification methods arduous to work in the supervised context. Moreover, even though raw data is accessible, limited labeled samples can also seriously affect supervised methods. Recently, supervised and unsupervised classification (clustering) results related to specific applications are published by more and more organizations. Therefore, combination of supervised classification and clustering results has gained increasing attention to improve the accuracy of supervised predictions. Incorporating clustering results with supervised classifications at the output level can help to lessen the recline on information at the raw data level, so that is pertinent to improve the accuracy for the applications when raw data is inaccessible or training samples are limited. We focus on the combination of multiple supervised classification and clustering results at the output level based on belief functions for three purposes: (1) to improve the accuracy of classification when raw data is inaccessible or training samples are highly limited; (2) to reduce uncertain and imprecise information in the supervised results; and (3) to study how supervised classification and clustering results affect the combination at the output level. Our contributions consist of a transformation method to transfer heterogeneous information into the same frame, and an iterative fusion strategy to retain most of the trustful information in multiple supervised classification and clustering results. (C) 2020 Elsevier Inc. All rights reserved.

Automated summary

In real-life machine learning applications, combining supervised classification and clustering results at the output level is gaining attention to improve accuracy when raw data is inaccessible or training samples are limited. This approach helps reduce dependency on raw data and uncertainty in supervised results, while studying the impact of supervised classification and clustering results on output combination.