Interactive online learning for graph matching using active strategies

• Published in: Knowledge-based systems 2020-10, Vol.205, p.106275, Article 106275
• Main Authors: Conte, Donatello, Serratosa, Francesc
• Format: Article
• Language: English
• Subjects: Active learning; Algorithms; Cognitive science; Computation; Computer Science; Computer Vision and Pattern Recognition; Cost function; Costs functions; Distance learning; Graph matching; Graphical representations; Graphs; Human interaction; Image Processing; Interactive learning; Linear equations; Machine learning; Mapping; Nodes; Object recognition; Online learning; Pattern recognition; Signal and Image Processing
• ISSN: 0950-7051; 1872-7409
• DOI: 10.1016/j.knosys.2020.106275

In some pattern recognition applications, objects are represented by attributed graphs, in which nodes represent local parts of the objects and edges represent relationships between these local parts. In this framework, the comparison between objects is performed through the distance between attributed graphs. Usually, this distance is a linear equation defined by some cost functions on the nodes and on the edges of both attributed graphs. In this paper, we present an online, active and interactive method for learning these cost functions, which works as follows. Graphs are provided to the learning algorithm by pairs in a sequential order (online). Then, a correspondence between them is computed, and there is a strategy that, given the current pair of graphs and the computed correspondence, proposes which node-to-node mapping would most contribute to the learning process (active). Finally, the human can correct some node-to-node mappings if the human thinks they are wrong (interactive). This is the first learning method applied to graph matching that has the following two features: Being an online method and being active and interactive. These properties make our method useful in the cases that data does not arrive at once and when the human can interact on the system. Thus, given some human interactions the method would have to tend to gradually increase its accuracy. The results show that with few interactions, we achieve better results than the offline learning state of the art methods that are currently available.