Classification using hierarchical mixture of discriminative learners: How to achieve high scores with few resources?

In many real-world classification tasks, discriminative models are widely applied since they achieve good predictive scores. In this paper, we propose a generalized framework for the well-studied Hierarchical mixture of experts (HME) model. HME combines hierarchically several discriminative models through a set of input-dependent weights. We derive from our generalized framework, two models as examples of how we can reduce the number of experts used. Those two examples are based on the choice of the weights functions. We choose a Gaussian-based and a linear softmax as weights, restricting our study to a two-level tree. Experiments on synthetic and real-world datasets show that our models can efficiently reduce the number of experts and outperform some state-of-art algorithms.