Graph-based Analysis of Hierarchical Embedding Generated by Deep Neural Network

• Main Authors: Rysbayeva, Korlan, Giot, Romain, Journet, Nicholas
• Other Authors: Laboratoire Bordelais de Recherche en Informatique (LaBRI), Université de Bordeaux (UB)-École Nationale Supérieure d'Électronique, Informatique et Radiocommunications de Bordeaux (ENSEIRB)-Centre National de la Recherche Scientifique (CNRS)
• Format: Conference Object
• Language: English
• Published: HAL CCSD 2022
• Subjects: Graph analysis; Hierarchical embeddings; eXplainable Artificial Intelligence; [INFO]Computer Science [cs]; [INFO.INFO-AI]Computer Science [cs]/Artificial Intelligence [cs.AI]; Canada; DML
• Online Access: https://hal.science/hal-03981883; https://hal.science/hal-03981883/document; https://hal.science/hal-03981883/file/W18P05.pdf

International audience In a previous work, we have developed a framework for the multimodal and hierarchical classification of images from soil remediation reports. We extended this work using Deep Metric Learning (DML) as an additional training step to improve embeddings quality and obtained 84.24% of weighted F1 score for the level 5th hierarchical level. However, the standard classifier performance metrics are insufficient to explain the decision process reasoning. So far of our knowledge, there are no methods to analyze hierarchical classification algorithms. In this work, we propose a method of graph analysis to describe the embeddings that represent the extended classifier, which we believe properly interprets the obtained results than classification metrics. We illustrate the method of analyzing hierarchical classification algorithms on private dataset, but the method remains generic enough to be used in other contexts.