Open-Box Training of Kernel Support Vector Machines: Opportunities and Limitations

In proceedings of Vision, Modeling & Visualization, 2019
 

Abstract

Kernel Support Vector Machines (SVMs) are widely used for supervised classification, and have achieved state-of-the-art performance in numerous applications. We aim to further increase their efficacy by allowing a human operator to steer their training process. To this end, we identify several possible strategies for meaningful human intervention in their training, propose a corresponding visual analytics workflow, and implement it in a prototype system. Initial results from two users, on data from three different domains suggest that, in addition to facilitating better insight into the data and into the classifier’s decision process, visual analytics can increase the efficacy of Support Vector Machines when the data available for training has a low number of samples, is unbalanced with respect to the different classes, contains outliers, irrelevant features, or mislabeled samples. However, we also discuss some limitations of improving the efficacy of supervised classification with visual analytics.

Images

Bibtex

@INPROCEEDINGS{Khatami:VMV2019,
     author = {Khatami, Mohammad and Schultz, Thomas},
      title = {Open-Box Training of Kernel Support Vector Machines: Opportunities and Limitations},
  booktitle = {Vision, Modeling {\&} Visualization},
       year = {2019},
   abstract = {Kernel Support Vector Machines (SVMs) are widely used for supervised classification, and have
               achieved state-of-the-art performance in numerous applications. We aim to further increase their
               efficacy by allowing a human operator to steer their training process. To this end, we identify
               several possible strategies for meaningful human intervention in their training, propose a
               corresponding visual analytics workflow, and implement it in a prototype system. Initial results
               from two users, on data from three different domains suggest that, in addition to facilitating
               better insight into the data and into the classifier’s decision process, visual analytics can
               increase the efficacy of Support Vector Machines when the data available for training has a low
               number of samples, is unbalanced with respect to the different classes, contains outliers,
               irrelevant features, or mislabeled samples. However, we also discuss some limitations of improving
               the efficacy of supervised classification with visual analytics.}
}