The human mind perceives the world around it in hierarchical structures. The input for the models built by humans are, however, flat. This creates dissonance. For instance, humans find the class-hierarchy tree for animals, intuitive. Nevertheless, they find presenting the relationships to a machine-learning algorithm a challenge. So, how to work with taxonomic data, without creating a dissonance?
A few approaches can be effectively used to overcome the challenge of presenting relationships to a machine-learning algorithm. One among them is the flat classification approach where each example is classified to its final, leaf-level label without dwelling much on parent categories. This approach, which can be implemented with a single out-of-the-box classifier, comes with a disadvantage. The natural hierarchy of the data could have highly valuable classification information, and hence ignoring parent-child class relationships could reduce performance. The straightforward flat classification approach can be employed for quick and dirty projects, which have constraints or where accuracy is not a concern.
Another method is the global classifiers approach. Some of these global classifiers use-clustering methods, some reframe the problem as a multi-label one, and some use existing algorithms customized to suit the data hierarchy at hand. Common to this varied bunch is a single relatively complex model, which considers the entire class hierarchy as a whole during a single run. This approach is highly complex and hence, rarely deployed. However, it may be considered a good option where a model to provide fast predictions without giving up on accuracy is needed.
The hierarchically structured local classifiers approach uses predefined data taxonomy to create a hierarchy of classifiers. This approach is highly intuitive and uses the hierarchy information in the data while retaining simplicity and generality. However, depending on the taxonomy and the chosen method, the final model could turn out to be rather unwieldy. There is also the problem of error propagation, where an error at one level could affect all of the following levels. The intuitive, clever approach that preserves the natural information on data hierarchy can be used for everything hierarchical.
Click here to read the original article published in Towards Data Science.
Please give your feedback on this article or share a similar story for publishing by clicking here.
