Research on feature engineering for time series data mining

Abstract

Today Internet and information technology are flourishing rapidly. Data collected from network and mobile devices can bring us huge opportunities to understand some significant characteristics of users and merchants. Time series analyzing is an extremely important topic in data mining that help users and merchants use data to do forecasting. Support Vector Machine (SVM) and Long Short-Term Memory (LSTM) both are excellent algorithms in time series forecasting. Nevertheless, there are few studies focused on feature engineering of LSTM and SVM in time series forecasting. In this paper, we focus on the influence of multi-feature fusion and feature ordering. Our experimental results show that LSTM is more sensitive about feature fusion than SVM, and the position of feature in the feature sequence could affect forecasting results obviously.