Order-Preserving Pattern Mining Enhances Structure-Aware Time Series Forecasting

Guangzhao Chai, Jing Liu, Youxi Wu, Yan Li

Keywords: Time Series, Frequent Order-Preserving Patterns, Loss Function, Machine Learning, Data Mining

TL;DR: To the best of our knowledge, This is the first study that embeds intrinsic pattern structures into the learning process, underscoring the potential of structural priors in improving the stability and accuracy of deep forecasting models.

Abstract:

Traditional time series forecasting models tend to focus on numerical fitting, mak-
ing it difficult to explicitly model and leverage the relative ordering patterns in-
herent in time series. This often results in suboptimal predictions when dealing
with data segments that exhibit clear pattern regularities. To address this gap, this
paper introduces Order-Preserving Patterns (OPPs) into time series forecasting for
the first time and proposes a novel model that explicitly incorporates prior pattern
knowledge by leveraging frequent OPPs as explicit priors. The proposed model
utilizes a convolutional neural network to perform feature dimensionality reduc-
tion on high-dimensional labeled time series, extracting one-dimensional repre-
sentations suitable for pattern mining. It then applies a sliding window and sup-
port counting strategy to discover frequent OPPs. An OPP matching mechanism
is proposed to distinguish between OPP and non-OPP training samples. Addition-
ally, a pattern constrained loss function is designed to guide the predicted values
toward consistency with the prior pattern logic. This constraint is imposed from
three perspectives—right boundary, left boundary, and intermediate positions—to
ensure order alignment with the tail elements of the OPPs. Experimental results
show that under the 'Perturbation Boundary' window sizes across ten real-world
and public benchmark datasets, the proposed OPPCL model consistently achieves
substantially lower MSE compared with state-of-the-art methods. In particular,
it yields at least 31.45% and 37.30% reductions on the SWaT and Electricity
datasets, respectively. The improvement becomes more pronounced when the
window size exceeds the 'Perturbation Boundary'. Code is available at this repos-
itory: https://anonymous.4open.science/r/OPPCL-B070/.

Primary Area: learning on time series and dynamical systems

Submission Number: 2124