Enhanced Principal-Curve based Classifiers for Time-Series Label Prediction
Aref Hakimzadeh, Koorush Ziarati · 2021
In many algorithms that predict time-series labels, time is just used to generate a sequence of data that will be fed to the predictor system. In other words, these algorithms do not consider the concept of time and its characteristics. The unforeseen circumstances of input data structure caused by the passage of time puts the system’s reliability and functionality in danger. These variations in time include increasing or decreasing sampling rate or any unforeseen regular or irregular change in sampling intervals. Source of these unexpected changes can be human or environmental factors. This paper proposes an algorithm which is immune to regular and irregular changes in the time parameter. Proposed algorithm can handle highly imbalanced data and multi-label classification with no change in the algorithm. Results of the proposed algorithm would be compared with RNN-based networks, which are believed to be the most prominent algorithms for time-series prediction.