search
HomeTechnology peripheralsAIGraph-aware contrastive learning improves multivariate time series classification effects

This paper in AAAI 2024 was jointly published by the Singapore Agency for Science, Technology and Research (A*STAR) and Nanyang Technological University, Singapore. It proposed a method of using graph-aware contrastive learning to improve multivariate time series classification. Experimental results show that this method has achieved remarkable results in improving the performance of time series classification.

Graph-aware contrastive learning improves multivariate time series classification effectsPicture

Paper title: Graph-Aware Contrasting for Multivariate Time-Series Classification

Download address: https://arxiv. org/pdf/2309.05202.pdf

Open source code: https://github.com/Frank-Wang-oss/TS-GAC

1. Overall introduction

Author Based on existing contrastive learning methods, a method called graph-aware contrast (TS-GAC) is proposed, which aims to solve the spatial consistency problem of multi-sensors in MTS data. TS-GAC consists of two main components: graph enhancement and graph comparison. Graph enhancement improves spatial consistency through node and edge enhancement to maintain sensor stability and relevance. The graph comparison introduces multi-window time comparison to maintain time consistency. Through extensive experimental verification, this method achieves optimal performance on various MTS classification tasks. The findings highlight the importance of considering spatial consistency in contrastive learning of MTS data and provide a comprehensive solution that significantly improves classification performance. This research is of great significance for further improving the effect of contrastive learning and provides a powerful tool for processing MTS data.

Graph-aware contrastive learning improves multivariate time series classification effectsPicture

2. Model structure

The method proposed in this article mainly includes two parts: graph enhancement and graph comparison.

In order to effectively enhance MTS data, we introduce node and edge enhancement to generate weak views and strong views. Node enhancement includes frequency domain and time domain enhancement to fully enhance graph nodes. First, we apply frequency domain enhancement to enhance the nodes, and then segment the enhanced samples into multiple windows according to the dynamic local patterns in the MTS data (as shown in Figure 2). In each window, we use node temporal augmentation and perform feature extraction on the window through a one-dimensional convolutional neural network. Subsequently, we construct a graph for each window and further enhance the graph through edge enhancement. Finally, we use a graph neural network-based encoder for graph processing and feature learning.

Graph-aware contrastive learning improves multivariate time series classification effectsPicture

Graph comparison: including node-level comparison and graph-level comparison to achieve spatial consistency. Node-level comparison ensures the robustness of node features by pulling corresponding sensors closer in different views and pushing different sensors farther away in different views. Graph-level comparison further ensures the robustness of global features by comparing samples in different views.

The goal of this architecture is to achieve spatial consistency in contrastive learning, and specific enhancement and contrast techniques are provided for MTS classification. By first applying node enhancement, then utilizing temporal enhancement within nodes, and finally edge enhancement processed by GNN, this method is able to generate weak and strong views with different spatial and temporal properties for each sample. The innovation of this method is that it not only considers temporal consistency, but also enhances spatial consistency through graph structure, providing a new perspective for in-depth analysis and processing of MTS data.

3. Graph enhancement module

In view of the characteristics of MTS data, that is, as data collected by multiple sensors, the author proposes two main methods of node and edge enhancement:

Node enhancement: divided into frequency domain enhancement and time domain enhancement. Frequency domain enhancement converts the signal of each sensor to the frequency domain, enhances the extracted frequency features, and then converts the enhanced frequency features back to the time domain to obtain an enhanced signal. Specifically, the discrete wavelet transform is used to decompose the signal through high-pass and low-pass filters to represent macro-trends and micro-trends within the signal. Time domain enhancement takes into account the dynamic characteristics of MTS data by dividing each MTS sample into multiple windows and performing time domain enhancement in each window.

Edge enhancement: aims to enhance the correlation between sensors, that is, the edges in the constructed graph. The nodes (sensors) and edges (correlations between sensors) are first defined through a graph building process. Then, the correlation between sensors is effectively enhanced through the edge enhancement method. In this step, considering that strong correlation is more important than weak correlation in the feature propagation of GNN, when performing edge enhancement, the strongest s correlations are retained to ensure the stability of topological information, and The remaining correlations are replaced with random values ​​to enhance the edge connections.

Through these enhancement strategies, the author aims to generate weak views and strong views so that the subsequent contrastive learning process can learn robust sensor features and inter-sensor relationships. The design of these enhancement strategies takes into account the multi-source and dynamic nature of MTS data, and enhances the capabilities of CL by providing data views from different angles, so that more robust and generalized representations can be learned.

4. Graph comparison module

The paper proposes a graph-aware comparison method, which specially designs node and edge enhancement and graph comparison strategies to enhance the spatial consistency of MTS data. It mainly includes three levels of comparison: multi-window time comparison, node-level comparison, and graph-level comparison.

Multi-Window Temporal Contrasting (MWTC): This method ensures the temporal consistency of each sensor at the sensor level and maintains the robustness of temporal dependence within MTS data through predictive coding. sex. MWTC maintains the robustness of temporal patterns by summarizing past window information in one view and comparing it to future windows in another view.

Node-level Contrasting (NC): NC learns robust sensor-level features by comparing sensors in different views within each MTS sample. This involves maximizing the similarity between corresponding sensors in two views while minimizing the similarity between different sensors in those views.

Graph-level Contrasting (GC): GC promotes robust global-level feature learning by comparing samples within each training batch. This strategy works by maximizing the similarity between corresponding samples in two views while minimizing the similarity between different samples in those views.

These contrastive learning strategies work together to enhance representation learning of MTS data through graph structures, thereby improving classification accuracy. The article also highlights the importance of temporal comparison to maintain temporal consistency for each sensor, as well as the role of graph comparison in learning robust features at the sensor and global levels. By combining node-level and graph-level comparisons, this method can effectively learn complex spatial and temporal patterns in MTS data and achieve significant improvement in MTS classification performance.

5. Experimental results

In the experimental part, the paper compares the performance on ten public multivariate time series data sets and compares it with the existing state-of-the-art methods. . These datasets include Human Activity Recognition (HAR), ISRUC sleep stage classification, and sub-datasets in the UEA dataset such as finger movements, spoken Arabic numerals, etc. For fair comparison, all methods use the same encoder. Experimental results show that TS-GAC achieved the best performance on eight of the data sets, especially on the HAR and ISRUC data sets. Compared with other methods, the accuracy increased by 1.44% and 3.13% respectively.

Graph-aware contrastive learning improves multivariate time series classification effectsPicture

At the same time, the author also visualized the model features. The visualization results enabled TS-GAC to extract more discernible sensor-level features. . At the same time, compared with other methods, TS-GAC can obtain more consistent sensor-level features for data from different viewing angles.

Graph-aware contrastive learning improves multivariate time series classification effectsPicture

The authors also conducted ablation studies to evaluate the impact of the designed enhancement and contrast techniques on model performance. The ablation study tested different variants, including those that removed node enhancement, removed edge enhancement, removed graph-level contrast, removed node-level contrast, and removed multi-window temporal contrast. The results show that graph enhancement and graph contrast techniques are extremely effective in improving the spatial consistency of MTS data, and the full TS-GAC shows better performance than any variant that reduces contrast loss.

Graph-aware contrastive learning improves multivariate time series classification effectsPicture

In addition, the author also analyzed the sensitivity of the model, including the impact and retention of hyperparameters (such as λMWTC, λGC, λNC) Effect of number of edges. These analyzes further confirm the effectiveness and robustness of the proposed method.

Overall, the experimental results highlight the ability of TS-GAC to achieve optimal performance on multiple MTS classification tasks, proving that the proposed graph enhancement and graph comparison techniques improve the spatial consistency of the model on MTS data. Importance and validity of sexual aspects.

The above is the detailed content of Graph-aware contrastive learning improves multivariate time series classification effects. For more information, please follow other related articles on the PHP Chinese website!

Statement
This article is reproduced at:51CTO.COM. If there is any infringement, please contact admin@php.cn delete
如何使用C#编写时间序列预测算法如何使用C#编写时间序列预测算法Sep 19, 2023 pm 02:33 PM

如何使用C#编写时间序列预测算法时间序列预测是一种通过分析过去的数据来预测未来数据趋势的方法。它在很多领域,如金融、销售和天气预报中有广泛的应用。在本文中,我们将介绍如何使用C#编写时间序列预测算法,并附上具体的代码示例。数据准备在进行时间序列预测之前,首先需要准备好数据。一般来说,时间序列数据应该具有足够的长度,并且是按照时间顺序排列的。你可以从数据库或者

如何使用XGBoost和InluxDB进行时间序列预测如何使用XGBoost和InluxDB进行时间序列预测Apr 04, 2023 pm 12:40 PM

XGBoost是一个流行的开源机器学习库,可用于解决各种预测问题。人们需要了解如何使用它与InfluxDB进行时间序列预测。 译者 | 李睿审校 | 孙淑娟XGBoost是一个开源的机器学习库,它实现了优化的分布式梯度增强算法。XGBoost使用并行处理实现快速性能,很好地处理缺失值,在小型数据集上执行良好,并防止过拟合。所有这些优点使XGBoost成为回归问题(例如预测)的一种流行解决方案。预测是各种业务目标的关键任务,例如预测分析、预测维护、产品规划、预算等。许多预测或预测问题都涉及到时间序

用于时间序列概率预测的分位数回归用于时间序列概率预测的分位数回归May 07, 2024 pm 05:04 PM

不要改变原内容的意思,微调内容,重写内容,不要续写。“分位数回归满足这一需求,提供具有量化机会的预测区间。它是一种统计技术,用于模拟预测变量与响应变量之间的关系,特别是当响应变量的条件分布命令人感兴趣时。与传统的回归方法不同,分位数回归侧重于估计响应变量变量的条件量值,而不是条件均值。”图(A):分位数回归分位数回归概念分位数回归是估计⼀组回归变量X与被解释变量Y的分位数之间线性关系的建模⽅法。现有的回归模型实际上是研究被解释变量与解释变量之间关系的一种方法。他们关注解释变量与被解释变量之间的关

AirPods Pro 和 AirPods 3:如何更改力传感器控件AirPods Pro 和 AirPods 3:如何更改力传感器控件Apr 13, 2023 am 11:43 AM

AirPods Pro具有出色的功能,例如降噪和采用入耳式设计的透明模式。为了在这些功能和其他功能之间切换,AirPods Pro(和 AirPods 3)杆中的力传感器可提供轻松控制。继续阅读以了解如何更改 AirPods 控件,包括按压速度、按压持续时间以及左右力传感器的功能等等。AirPods gen 1 和 2使用耳塞上的双击进行控制。但AirPods Pro 和 AirPods 3使用内置在每个耳塞杆中的压力传感器。我们将介绍三种在降噪、透明度和关闭之间切换的方法,如何自定义力度传感器

时间序列预测+NLP大模型新作:为时序预测自动生成隐式Prompt时间序列预测+NLP大模型新作:为时序预测自动生成隐式PromptMar 18, 2024 am 09:20 AM

今天我想分享一个最新的研究工作,这项研究来自康涅狄格大学,提出了一种将时间序列数据与自然语言处理(NLP)大模型在隐空间上对齐的方法,以提高时间序列预测的效果。这一方法的关键在于利用隐空间提示(prompt)来增强时间序列预测的准确性。论文标题:S2IP-LLM:SemanticSpaceInformedPromptLearningwithLLMforTimeSeriesForecasting下载地址:https://arxiv.org/pdf/2403.05798v1.pdf1、问题背景大模型

国产传感器天花板!OV50K40传感器详解国产传感器天花板!OV50K40传感器详解Mar 20, 2024 am 09:56 AM

国产制造商豪威发布了新CMOS——OV50K40传感器。这款传感器是全球首个采用TheiaCel技术的智能手机图像传感器,其独特的设计使得单次曝光能够实现接近人眼级别的动态范围。OV50K40传感器拥有5000万像素的原生分辨率,并采用了1/1.3英寸的大底面积,每个像素的面积达到了1.2微米。此外,它还支持像素四合一输出,即将四个像素合并成一个,输出1250万像素的图像,从而带来更强大的感光能力。这款传感器的帧率支持120帧/秒和60帧/秒(HDR),可实现完整的5000万像素B

为什么有人认为安卓拍照可以吊打苹果?答案就这么直接为什么有人认为安卓拍照可以吊打苹果?答案就这么直接Mar 25, 2024 am 09:50 AM

在讨论安卓手机的拍照功能时,大多数用户对其给予了肯定,与苹果手机相比,用户普遍认为安卓手机的拍照表现更出色。这一观点并非毫无根据,实际原因也是显而易见的。高端安卓手机在硬件配置方面具有较大的竞争优势,尤其是摄像头传感器方面。许多高端安卓手机采用最新的、顶级的摄像头传感器,这些传感器在像素数、光圈大小和光学变焦能力等方面往往比同期发布的iPhone更为突出。这种优势使得安卓手机在拍摄照片和录制视频时能够提供更高质量的成像效果,满足用户对于摄影和摄像的需求。因此,硬件配置的竞争优势成为了安卓手机吸引

五个时间序列预测的深度学习模型对比总结五个时间序列预测的深度学习模型对比总结May 05, 2023 pm 05:16 PM

MakridakisM-Competitions系列(分别称为M4和M5)分别在2018年和2020年举办(M6也在今年举办了)。对于那些不了解的人来说,m系列得比赛可以被认为是时间序列生态系统的一种现有状态的总结,为当前得预测的理论和实践提供了经验和客观的证据。2018年M4的结果表明,纯粹的“ML”方法在很大程度上胜过传统的统计方法,这在当时是出乎意料的。在两年后的M5[1]中,最的高分是仅具有“ML”方法。并且所有前50名基本上都是基于ML的(大部分是树型模型)。这场比赛看到了LightG

See all articles

Hot AI Tools

Undresser.AI Undress

Undresser.AI Undress

AI-powered app for creating realistic nude photos

AI Clothes Remover

AI Clothes Remover

Online AI tool for removing clothes from photos.

Undress AI Tool

Undress AI Tool

Undress images for free

Clothoff.io

Clothoff.io

AI clothes remover

AI Hentai Generator

AI Hentai Generator

Generate AI Hentai for free.

Hot Article

R.E.P.O. Energy Crystals Explained and What They Do (Yellow Crystal)
2 weeks agoBy尊渡假赌尊渡假赌尊渡假赌
Repo: How To Revive Teammates
1 months agoBy尊渡假赌尊渡假赌尊渡假赌
Hello Kitty Island Adventure: How To Get Giant Seeds
1 months agoBy尊渡假赌尊渡假赌尊渡假赌

Hot Tools

Dreamweaver Mac version

Dreamweaver Mac version

Visual web development tools

Atom editor mac version download

Atom editor mac version download

The most popular open source editor

WebStorm Mac version

WebStorm Mac version

Useful JavaScript development tools

VSCode Windows 64-bit Download

VSCode Windows 64-bit Download

A free and powerful IDE editor launched by Microsoft

Notepad++7.3.1

Notepad++7.3.1

Easy-to-use and free code editor