One-pass-throw-away learning of temporal class-shift by multi-stratum network

Abstract

The problem of learning streaming non-expired and temporally expired data occurring in various business applications was studied. If there exists a class whose all data are eventually expired with some unknown reasons, then the relevant neurons and their links must be entirely removed. A new learning based on dynamic network structure called Multi-Stratum Network Learning (MSNL) was proposed to cope with this problem of data life change. Furthermore, to speed up the learning time and to maintain a minimum space complexity for streaming data, the new concept of one-pass-throw-away learning in forms of recursive functions for handling the expiration class was introduced. The experimental results signified that the proposed algorithm outperformed the other incremental-like learning algorithms in terms of time and space complexities.