📄 中文摘要
一种针对多元时间序列的与任务无关的发现层，无需预设线性、平稳性或下游目标，构建实体间的关系假设图。方法通过无监督序列到序列自编码器学习窗口级别的序列表示，随后将这些表示聚合成实体级别的嵌入。通过阈值化潜在相似度，诱导出稀疏相似性网络。该网络能够捕捉不同实体之间在行为模式上的潜在关联，即便这些模式在原始数据中不明显或以复杂非线性方式存在。整个过程不依赖任何标记数据，使其适用于数据标注成本高昂或先验知识缺乏的场景。通过这种方式，系统能够自动识别数据中隐藏的结构和相互作用，为后续的分析和建模提供基础。特别地，它能够处理时间序列中常见的非线性和非平稳特性，克服了传统方法对这些假设的依赖。所学习到的结构信息可以作为下游任务的输入，例如异常检测、聚类分析或因果推断，从而提升这些任务的性能。此外，方法通过对相似度进行阈值处理，能够生成稀疏且可解释的网络结构，有助于理解实体间的核心关系。
📄 English Summary
Latent Structural Similarity Networks for Unsupervised Discovery in Multivariate Time Series
A task-agnostic discovery layer for multivariate time series constructs a relational hypothesis graph over entities without assuming linearity, stationarity, or a downstream objective. The methodology employs an unsupervised sequence-to-sequence autoencoder to learn window-level sequence representations, subsequently aggregating these representations into entity-level embeddings. A sparse similarity network is then induced by thresholding a latent similarity measure. This network captures potential associations in behavioral patterns between different entities, even when these patterns are not explicit in the raw data or exist in complex, nonlinear forms. The entire process operates without reliance on any labeled data, making it suitable for scenarios where data annotation is costly or prior knowledge is scarce. Through this approach, the system automatically identifies hidden structures and interactions within the data, providing a foundation for subsequent analysis and modeling. Notably, it addresses the common nonlinear and non-stationary characteristics in time series, overcoming the dependence of traditional methods on such assumptions. The learned structural information can serve as input for downstream tasks, such as anomaly detection, clustering analysis, or causal inference, thereby enhancing their performance. Furthermore, by thresholding similarities, the method generates sparse and interpretable network structures, facilitating the understanding of core relationships between entities.