Domain Transformation to Graphs and GraphSAGE-Based Embedding for Performance Enhancement in Time-Series Classification

In this paper, we address the problem of improving time-series classification performance in graph environments. With the recent increase in graph analytics, many studies analyzing time-series within the graph domain have been introduced. SimTSC is a novel approach that transforms time-series data into graphs and applies them to node(i.e., time-series) classification; however, it incurs two major problems. Problem 1 is that the embedding model does not consider the relationships among nodes during the feature embedding process. Problem 2 is that in mini-batch data sampling, biased data can be selected depending on the variation in the number of time-series and/or the number of classes. To solve these two problems, in this paper we propose three classification models. First, to solve Problem 1, we propose SAGE-based Classification Model(SbCM), which performs node feature embedding using GraphSAGE. GraphSAGE is an effective embedding model that utilizes neighboring nodes’ information, enabling feature embedding by leveraging both the target node and its neighboring nodes. Second, to address Problem 2, we propose Quota-based Classification Model(QbCM), which constructs mini-batch data using quota sampling. Quota sampling chooses samples proportional to the population ratio, enabling mini-batch data to better resemble the original graph structure. Third, we propose SAGE-Quota integrated Classification Model(SQiCM), which integrates SbCM and QbCM to solve both Problems 1 and 2 at the same time. Extensive experimental results using UCR data demonstrate that our SbCM, QbCM, and SQiCM improve classification performance by 2.3 to 3.2 times compared to the existing SimTSC on large datasets with multiple classes.