Adaptive Data Embedding Framework for Multiclass Classification

摘要: The objective of this paper is the design an engine for automatic generation supervised manifold embedding models. It proposes a modular and adaptive data framework classification, referred to as DEFC, which realizes in different stages including initial preprocessing, relation feature computation. For computation embeddings, concepts friend closeness enemy dispersion are introduced, better control at local level relative positions intraclass interclass samples. These shown be general cases global information setup utilized Fisher criterion, employed construction optimization templates drive DEFC model generation. identification, we use simple but effective bilevel evolutionary optimization, searches optimal its best parameters. effectiveness demonstrated with experiments using noisy synthetic datasets possessing nonlinear distributions real-world from application fields.