Hierarchical, multi-label classification: exploiting class knowledge in examination of geospatial data

Abstract

Classification spans a broad range of chemical analysis tasks. The majority of currently available classification techniques rely solely on mathematical transformations of data collected from laboratory measurements. Often additional information is available regarding the relationships between classes that is possibly not collected in those measurements. This work seeks to develop a new method of identifying and incorporating that additional information to create more accurate and more reliable classification models. ☐ First, a method to incorporate class relationships in classification beyond the standard hierarchical classification is presented. The model uses class information for sample selection via policy and submodel construction via classification mode. Certain policy and mode combinations are shown to out-perform non-hierarchical classification on some datasets. ☐ Second, the choice of hierarchy for incorporation in HMLC is examined. The effects of hierarchy structure, class location in the hierarchy, and the ease of separation of a class by flat classification are examined. Only four policy and mode combinations are found to be unique. The “best” hierarchy is found to depend upon the desired application of the classification model. Models with few false negative predictions require a comparable-based model and a hierarchy that maximizes separation between classes. Models with few false positive predictions require a hierarchically-based model, but are independent of the particular hierarchy used.