Research On Hierarchical Multi-instance Multi-Label Learning Algorithms

Multi-Instance Multi-Label(MIML)learning is a general framework for solving practical tasks in machine learning.Formerly,the object is denoted by an instance(ie,attribute vector)and owns a category label.But in some complex tasks a object would be presented by multi-instances,and simultaneously belongs to more than one category labels.That is defined as the multi-instance multilabel learning framework.As the rapid development of science and technology,the era of big data is getting closer and closer,and it is great significance to study the multi-instance multi-label learning framework.Usually,there are label relationships in many MIML applications.Moreover,the labels present hierarchy tree(TREE)or Directed Acyclic Graph(DAG)structure in many scenarios.However,many MIML works often neglect the hierarchical dependency between labels,and it is urgent to develop a new multi-instance multi-label learning method that can consider this hierarchical label relationship.Therefore,the paper has proposed two novel MIML learning algorithms which consider the hierarchy structures of labels.One is to consider the tree structure,and the other is to take the the directed acyclic graph structure into consideration.The algorithms have taken into consideration the hierarchy structures among labels,which improve the performance of models and expand the application scope of the MIML learning.In this paper,our proposed hierarchical MIML learning alogrithms is adopted to predict the GO(Gene Ontology)biological functions of G-protein coupled receptors.Usually GO labels include,Tree and Directed Acyclic Graph,two hierarchical structures.So two algorithms,TreeMIML and DAGMIML,were proposed and simulated respectively in this data set.Multiple expermiments prove that the two hierarchy MIML learning algorithms present good predictive performance.