Automatic time series forecasting model design based on pruning

Abstract

Automatic Time Series Forecasting (TSF) model design which aims to help users to efficiently design forecasting models for the given time series data scenarios, is a novel research topic to be urgently solved. In this paper, we propose AutoTS algorithm trying to utilize the existing design skills and design efficient search methods to effectively solve this problem. In AutoTS, we extract effective design experience from the existing TSF works. We allow the effective combination of design experience from different sources, to create an effective search space containing a variety of TSF models to support different TSF tasks. Considering the huge search space, in AutoTS, we propose a two-stage pruning strategy to reduce the search difficulty and improve the search efficiency. Specifically, at the beginning of the search phase, we apply the vertical pruning method to quickly optimize each module of the TSF model. In the middle of the search, we turn to apply the horizontal pruning method to filter out less effective options for each module according to the learned experience and optimize modules. Moreover, in AutoTS, we introduce the knowledge graph to reveal associations between module options. We make full use of this relational information to learn higher-level features of each module option, to further improve the search quality. Experimental results show that AutoTS is well-suited for the TSF area. On some general datasets, our algorithms discover the assembled model with better performance. Additionally, compared to existing NAS algorithms, they exhibit higher search efficiency and greater search potential than the manually designed ones.