Learning from Points and Patches: A Pre-trained Hybrid Attention Transformer for Time Series Analysis

这个问题是要把给定的纯文本新闻正文排版成语义化的HTML片段，需要遵循特定的规则。首先看正文内容，是关于TimeHAT框架的学术性描述，没有明显的分节标题，所以不需要用

或

。接下来看每个自然段落，正文里的内容是由多个句子组成的，但需要确认是否是自然段落。原文中从开头到“effective time-series pre-training requires strategies tailored to temporal data.”是第一段，然后从“We propose TimeHAT...”到最后是第二段吗？仔细看原文，原文是一整段吗？不，原文的结构是：先讲背景和挑战，然后提出TimeHAT的两个问题，分别解释what to learn和how to learn，接着是实验评估，最后是代码链接。但原文的文本呈现是一个连续的段落吗？看用户提供的正文，其实是一整段，但需要拆分成自然段落吗？不对，规则说每个自然段落用

包裹，所以要看原文的自然段落划分。用户提供的正文里，所有内容都是连续的，没有换行，但实际上可能需要根据语义拆分？不，规则要求保留原文所有信息，不得增删改内容，所以原文的段落结构是怎样的？用户提供的正文是一段，但仔细看内容，其实是多个句子组成，但作为一个自然段落？不对，再看用户给的正文，原文是： “Time-series data generated by Web services and connected systems support applications such as user activity analysis, service monitoring, and resource management. However, learning transferable representations from heterogeneous time series remains challenging due to diverse temporal patterns, variable sequence lengths, and limited high-quality annotations. Despite the success of pre-training in natural language processing and computer vision, effective time-series pre-training requires strategies tailored to temporal data. We propose TimeHAT (Hybrid Attention Transformer), a general-purpose pre-training framework for time-series analysis that addresses two complementary questions: what to learn and how to learn. For what to learn, we introduce a Unified Point-Patch Masking Strategy (UPPMS) that integrates point-level and patch-level supervision within a unified framework for contextual reconstruction and causal extrapolation, enabling the model to capture both fine-grained temporal variations and broader temporal patterns. For how to learn, we develop a Hybrid Attention Encoder (HAE) that combines prefix-constrained bidirectional attention in initial layers with directional hybrid attention in deeper layers, aiming to preserve effective contextual interactions while mitigating rank degeneration. We evaluate TimeHAT across real-world datasets and four downstream tasks: classification, long-term forecasting, imputation, and anomaly detection. Experiments cover fine-tuning, few-shot, and zero-shot transfer settings, demonstrating competitive performance across diverse time-series analysis tasks. Code is available at https://github.com/TangPeiwang/TimeHAT.” 这是一个完整的段落，但需要拆分成自然段落吗？规则说“每个自然段落用

包裹”，所以如果原文是一个自然段落，就用一个

？但看内容，其实是多个部分：背景挑战、提出框架、两个问题的解释、实验评估、代码链接。但原文是连续的，没有换行，所以应该作为一个自然段落？不对，再看规则里的示例，比如长文用

/

划分小节，但原文没有小节标题，所以不需要。然后看是否有列表、引用、代码，这里没有列表、引用、代码，所以主要是

包裹，重要句子用。 接下来找重要句子，比如提出的框架名称“TimeHAT (Hybrid Attention Transformer)”，两个策略“Unified Point-Patch Masking Strategy (UPPMS)”和“Hybrid Attention Encoder (HAE)”，这些是重要的，应该用标注。还有关键的实验结果部分？比如“demonstrating competitive performance across diverse time-series analysis tasks”可能也是重要的？ 然后处理每个句子，确保保留所有信息。比如第一句到“tailored to temporal data.”是背景，然后用

包裹，接着从“We propose TimeHAT...”到结尾是第二段？不，原文是连续的，所以应该拆分成多个

吗？或者原文是一个自然段落，所以用一个

？不对，再仔细看规则，规则说“每个自然段落用

包裹”，所以需要根据原文的自然段落划分。用户提供的正文是一段，但实际在排版时，可能需要根据语义拆分吗？不，规则要求保留原文所有信息，不得增删改内容，所以原文的段落结构是怎样的？用户提供的正文是一段，所以用一个

？但看内容，其实是多个句子组成，但作为一个自然段落，所以用

包裹整个内容？但这样的话，内容太长，可能需要拆分？不对，再看用户给的示例，比如如果有多个自然段落，每个用

，但这里原文是一个自然段落，所以用一个

？ 不对，再仔细看原文的结构： 1. 背景和挑战：从开头到“effective time-series pre-training requires strategies tailored to temporal data.” 2. 提出框架和两个问题：从“We propose TimeHAT...”到“mitigating rank degeneration.” 3. 实验评估和结果：从“We evaluate TimeHAT...”到结尾。 这三个部分是自然段落吗？原文中是连续的，所以需要根据语义拆分成三个

吗？但规则说“保留原文所有信息，不得增删改内容”，所以拆分段落是否属于增删改？不，拆分段落是为了语义化，但原文的段落结构是怎样的？用户提供的正文是一段，所以应该用一个

？但看规则里的示例，比如长文用

/

，但这里没有，所以可能原文是一个自然段落，用一个

包裹，然后在其中用标注重要部分。 然后处理重要句子： - “TimeHAT (Hybrid Attention Transformer)”是框架名称，用 - “Unified Point-Patch Masking Strategy (UPPMS)”是策略名称，用 - “Hybrid Attention Encoder (HAE)”是编码器名称，用 - “demonstrating competitive performance across diverse time-series analysis tasks”是结果，用 然后处理句子中的链接，直接保留。 现在开始构建HTML： 首先，第一部分：“Time-series data generated by Web services and connected systems support applications such as user activity analysis, service monitoring, and resource management. However, learning transferable representations from heterogeneous time series remains challenging due to diverse temporal patterns, variable sequence lengths, and limited high-quality annotations. Despite the success of pre-training in natural language processing and computer vision, effective time-series pre-training requires strategies tailored to temporal data.” 这部分用

包裹。 然后第二部分：“We propose TimeHAT (Hybrid Attention Transformer), a general-purpose pre-training framework for time-series analysis that addresses two complementary questions: what to learn and how to learn. For what to learn, we introduce a Unified Point-Patch Masking Strategy (UPPMS) that integrates point-level and patch-level supervision within a unified framework for contextual reconstruction and causal extrapolation, enabling the model to capture both fine-grained temporal variations and broader temporal patterns. For how to learn, we develop a Hybrid Attention Encoder (HAE) that combines prefix-constrained bidirectional attention in initial layers with directional hybrid attention in deeper layers, aiming to preserve effective contextual interactions while mitigating rank degeneration.” 这部分用

包裹。 第三部分：“We evaluate TimeHAT across real-world datasets and four downstream tasks: classification, long-term forecasting, imputation, and anomaly detection. Experiments cover fine-tuning, few-shot, and zero-shot transfer settings, demonstrating competitive performance across diverse time-series analysis tasks. Code is available at https://github.com/TangPeiwang/TimeHAT.” 这部分用

包裹。 但原文是连续的，所以是否需要拆分成三个

？规则说“每个自然段落用

包裹”，所以如果原文是三个自然段落，就用三个

，但用户提供的正文是一段，所以可能需要根据语义拆分吗？或者原文的段落结构是怎样的？用户提供的正文是一段，所以应该用一个

？但这样的话，内容太长，不符合语义化。可能用户提供的正文在原始格式中是多个段落，但粘贴时变成了一段，所以需要根据语义拆分成多个

。 再看规则，规则说“保留原文所有信息，不得增删改内容”，所以拆分段落是否允许？其实，语义化排版需要根据内容的语义划分段落，所以即使原文是一段，也可以根据语义拆分成多个

，只要不改变内容。比如：

Time-series data generated by Web services and connected systems support applications such as user activity analysis, service monitoring, and resource management. However, learning transfer