Hierarchical manifold sample envelope transformation model for ensemble classification

Abstract

Ensemble classification is an important branch and research focus in machine learning and pattern recognition. The current main paradigm of ensemble classification algorithms is based on same original samples, resulting in limited diversity between subsets. Therefore, it is particularly important to mine diverse and effective information from the original samples to build a multi-layer samples. However, samples are input into classifier for training one by one, or batch by batch in the existing ensemble algorithms. This ignores the potential value of correlation information among samples during base classifier training. To solve these problems, a new sample transformation model for ensemble classification - Hierarchical Manifold Sample Envelope Transformation Model (HMSET) is proposed. The model consists of three main parts. The first part is manifold sample enveloping model. It extracts local correlation among samples, thereby constructing manifold envelope samples. The second part is hierarchical sample envelope transformation model, which uses a variety of transformation operators and interlayer consistency to mine and constrain the correlation information among samples to enhance the diversity. The third part is two-dimensional fusion mechanism which fuse the final prediction results of the base classifiers. The 19 UCI datasets and several representative algorithms are used for validation. The results show that compared with the original samples, the proposed model improves the diversity of the sample subsets significantly. Compared with related ensemble classification algorithms, the proposed model has significantly better performance of ensemble classification.

Data and code are available in: https://github.com/acceptthisjj/HMSET