Fully Automated Correlated Time Series Forecasting in Minutes

Xinle Wu,Xingjian Wu,Dalin Zhang,Miao Zhang,Chenjuan Guo,Bin Yang,Christian S. Jensen

2024-11-06

Abstract:Societal and industrial infrastructures and systems increasingly leverage sensors that emit correlated time series. Forecasting of future values of such time series based on recorded historical values has important benefits. Automatically designed models achieve higher accuracy than manually designed models. Given a forecasting task, which includes a dataset and a forecasting horizon, automated design methods automatically search for an optimal forecasting model for the task in a manually designed search space, and then train the identified model using the dataset to enable the forecasting. Existing automated methods face three challenges. First, the search space is constructed by human experts, rending the methods only semi-automated and yielding search spaces prone to subjective biases. Second, it is time consuming to search for an optimal model. Third, training the identified model for a new task is also costly. These challenges limit the practicability of automated methods in real-world settings. To contend with the challenges, we propose a fully automated and highly efficient correlated time series forecasting framework where the search and training can be done in minutes. The framework includes a data-driven, iterative strategy to automatically prune a large search space to obtain a high-quality search space for a new forecasting task. It includes a zero-shot search strategy to efficiently identify the optimal model in the customized search space. And it includes a fast parameter adaptation strategy to accelerate the training of the identified model. Experiments on seven benchmark datasets offer evidence that the framework is capable of state-of-the-art accuracy and is much more efficient than existing methods.

Machine Learning

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the limitations of existing automated time - series prediction methods in practical applications. Specifically, these limitations include:

1. **Manually - designed search space**: Existing automated methods rely on human experts to manually construct the search space, which may not only lead to sub - optimal performance but also violates the goal of AutoML (Automated Machine Learning), that is, the automation of the entire process. Due to the existence of many heterogeneous spatio - temporal (S/T) operators and topological connection methods in time - series prediction tasks, it is very difficult to construct a general search space.

2. **High search cost**: Existing automated methods use gradient - based, comparator - based or random search strategies to explore the search space to find the optimal S/T block, and all of these methods are very time - consuming. For example, the gradient - based method needs to train a super - network representing the search space, which is much larger and more difficult to train than training a single S/T block alone.

3. **High training cost**: Even if a high - performance S/T block is found, existing methods still need to train this S/T block from scratch to handle unseen time - series prediction tasks, which usually takes several hours.

To solve the above problems, the paper proposes a fully - automated and efficient correlated time - series prediction framework (FACTS), aiming to complete the search and training process within a few minutes. The main contributions of FACTS include:

1. **Proposing an automatic pruning strategy** for generating a high - quality search space, which is suitable for unseen time - series prediction tasks.
2. **Proposing a zero - sample search strategy** that can find the optimal S/T block on any unseen time - series prediction task within a few minutes.
3. **Proposing a fast parameter adaptation strategy** that accelerates the training process on new tasks by introducing a learnable coefficient linear combination of pre - trained S/T block parameter weights, which can reduce the training time by up to 66%.

Through these innovations, FACTS can not only achieve state - of - the - art prediction accuracy, but also is far superior to existing manual and automated methods in terms of time and resource consumption.