กก
This project proposes three algorithms for determining the optimal network structure of multilayer bootstrap networks automatically:
1. MBN-E: an Ensemble of MBN base models that have different network structure;
2. fMBN-E: A fast version of MBN-E that accelerates MBN-E by over hundreds of times without losing accuracy. The core idea of fMBN-E is to delete the random feature selection step in MBN, and replace the random data resampling by random similarity score resampling.
3. MBN-SO: it first conducts clustering on the output of MBN-E, and then uses the clustering result as a guidance to select a small number of base models whose output representation is the most discriminant. Finally, it uses the selected base models to group into a new MBN-E.
4. MBN-SD: It uses the meta-representation produced by MBN-E as a reference for selecting a small number of base models whose output representation is the most discriminant. Finally, it uses the selected base models to group into a new MBN-E.
The source code (still dirty) with demos and datasets is downloadable here: [fMBN-E.rar] (51.55MB)
References:
Xiao-Lei Zhang. fMBN-E: Efficient Unsupervised Network Structure Ensemble and Selection for Clustering. arXiv preprint arXiv:2107.02071. 2021. [supplement]