Scalable Learnability Measure for Hierarchical Learning in Large Scale Multi-Class Classification

Overview & Synopsis

The increase in computational and storage capacities leads to an increasing complexity of the data to be treated: data can be represented in much more detail (many features) and in very large amounts : in the context of text categorization or image classification, the number of labels can scale from $10^2$ to $10^5$, and features range from $10^4$ to $10^6$. The main trade-off is generally between the accuracy of the predictions and the inference time. A usual methodology consists in organizing multiple classifiers in a hierarchical structure in order to reduce the computation cost of the inference. A popular category of algorithms is to iteratively build the structure. Inspired by clustering, the iteration scheme is a splitting (top-down lgorithms) or aggregating (bottom-up algorithms) process. This step uses measures to determine the split/aggregation rule (like entropy, similarity between classes, separability ...). These kinds of measures are often computationaly heavy and can not be used in a large scale context. In this paper, we propose to use a reduced projected space of the input space to build measures of interest. Preliminary experiments on real dataset show the interest of such methods. We propose preliminary experiments which integrate a ''learnability'' measure in hierarchical approaches.