MatrixNet is Yandex's Implementation of Gradient Boosted Decision Tree Algorithm (GBRT)
In this talk I'll cover MatrixNet’s differences from the classical approach, its applications and limitations. One of the limitations is that basic MatrixNet is designed for object classification. There are tasks where we need to classify sequences of objects, for example, strings or event sequences. GBRT and the like aren’t directly applicable to this kind of problems. There are many methods to apply regular classification algorithms to sequence classification like string kernels. We will discuss similarities of these methods and I'll present our boosting algorithm for sequence classification. We will compare some of the methods using a model task.