Presenting our research on ROC hierachical classifiers at NVPHBV meeting
We have presented our research on optimization of hierarchical classifiers at the spring meeting of NVPHBV (Dutch society for pattern recognition and image processing).
Complex problems are often easier to handle if decomposed into sub-problems and tackled independently. Hierarchical classifiers offer a great tool for such decomposition but are difficult to optimize according to application requirements. This is a serious problem we encounter daily in our industrial projects. In our talk, we described our approach allowing the designer to perform cost-sensitive ROC optimization for apriori-defined hierarchical classifiers.
In our projects, we encounter complex pattern recognition problems that are difficult to tackle in one step. Building a single classifier would often result in highly complex model using many features. Complex models are, however, not easy to train with limited amount of training examples that are typically available. One possible solution is decomposition of the complex problem into simpler sub-problems. This approach allows us to leverage our prior knowledge in classifier design. For example, we may know that sophisticated features are needed only for specific sub-classes. Or that classes form a natural hierarchy and we may first identify the meta-class and only for these examples separate multiple low-level classes.
Although the construction of hierarchical classification systems is a standard engineering practice, their optimization is very difficult. The most of classifiers we design need to be optimized to provide very specific performances and errors depending on an application. The optimization of a single classifier is done with ROC analysis. In our research, we developed a technique of performing ROC analysis and cost-sensitive optimization on the entire classifier hierarchy.
The results of our research allow the system designers to:
* construct and optimize detector/classifier cascades
* combine multiple detectors
* protect your data from outliers by robust models