{ "candidate_id": "ATT-0086", "reviewer": "C", "rubric": "PAPER-RUBRIC-1.0", "frozen": true, "frozen_at": "2026-11-08T02:54:10.868178+00:00", "identity": { "candidate_id": "ATT-0086", "title": "CrossLinear: Plug-and-Play Cross-Correlation Embedding for Time Series Forecasting with Exogenous Variables", "authors": [ "Zhou, Pengfei", "Liu, Yunlong", "Liang, Junli", "Song, Qi", "Li, Xiangyang" ], "arxiv_id": "2505.23116", "arxiv_version": "v1", "first_arxiv_date": "2025/05/29", "latest_arxiv_date": "2025/05/29", "source_urls": [ "https://doi.org/10.1145/3711896.3736899" ], "publication_status": "KDD 2025, DOI 10.1145/3711896.3736899 verified", "full_text_path": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0086/ATT-0086_paper.pdf", "metadata_source_path": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0086/ATT-0086_arxiv.html", "full_text_sha256": "0235fd4cd7a51ae05b313532abdc7d0fe4dc86ac23bc29c1105f5e3cccb814f8", "review_state": "ready_for_five_new_independent_reviews", "completed_new_reviews": 0, "source_versions_not_asserted_identical": true }, "source_sha256_verified": "0235fd4cd7a51ae05b313532abdc7d0fe4dc86ac23bc29c1105f5e3cccb814f8", "reading_started_at": "2026-11-08T02:51:14Z", "reading_finished_at": "2026-11-08T02:54:10.868178+00:00", "reading_time_note": "以UTC记录开始和完成时刻", "read_coverage": { "pdf_pages": 13, "scope": "正文、参考文献及全部随附附录的逐页文本；关键图式按需要视觉复核", "visual_pages": [ 9, 12, 13 ] }, "legacy_read_first": true, "legacy_full_reviews_missing": true, "other_new_reviews_read": false, "experiments_run": false, "external_predecessor_fulltext_checked": false, "resource_notes": "本次未执行代码、训练或推理；未将仓库入口存在等同资源完整或可运行。", "claims": [ { "claim_id": "CL01", "claim": "CrossLinear在外生变量预测上具有竞争力", "judgment": "多数据支持，非所有范式或数据最优", "pages": "6–7" }, { "claim_id": "CL02", "claim": "Sum更好反映内外生变量重要性", "judgment": "等价卷积参数化下尚不能区分语义与优化路径", "pages": "7,11–13" }, { "claim_id": "CL03", "claim": "插件带来一致性能改善", "judgment": "部分成立，有明确退化；DLinear Traffic汇总改善暂停引用", "pages": "8,12" } ], "risks": [ { "severity": "重要问题", "certainty": "已确认", "description": "表11 DLinear Traffic原分行平均为0.302/0.3845而非0.323/0.404，加插件0.308/0.39125；按分行数实际退化，暂停该汇总改善主张", "pages": "8,12" }, { "severity": "重要问题", "certainty": "已确认", "description": "无随机重复及区间，很多领先差异很小", "pages": "6–8,11–12" }, { "severity": "重要问题", "certainty": "未交代", "description": "全目标遮挡与RevIN统计量计算顺序不明，不能把其直接解释为无任何目标历史信息", "pages": "4,8" }, { "severity": "重要问题", "certainty": "已确认", "description": "权重热图并非直接物理依赖的独立验证，残差和卷积可重参数化", "pages": "9,11–13" } ], "scores": [ { "dimension_id": "A1", "dimension_name": "问题重要性", "score": 85, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "目标预测与辅助变量融合广泛且重要，轻量方法有现实价值。", "evidence_pages": "1–3" }, { "dimension_id": "A2", "dimension_name": "原创性", "score": 60, "status": "tentative", "confidence": "中", "judgment_range": "50–70", "reason": "主要为卷积残差嵌入与线性预测的组合，未独立核对全部近邻前作。", "evidence_pages": "2–4" }, { "dimension_id": "A3", "dimension_name": "实质推进程度", "score": 70, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "两范式多数据及跨骨干收益，仍有任务和插件反例。", "evidence_pages": "6–8,12" }, { "dimension_id": "A4", "dimension_name": "知识与解释增量", "score": 60, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "重参数化解释有价值，直接物理依赖及语义隔离证据不足。", "evidence_pages": "7,9,11–13" }, { "dimension_id": "B1", "dimension_name": "问题定义与评价有效性", "score": 75, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "历史外生信息集与两类任务清楚，目标选择和缺失协议仍有限。", "evidence_pages": "3–5,8" }, { "dimension_id": "B2", "dimension_name": "技术与论证正确性", "score": 75, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "主体运算和卷积合并基本正确，参数热图与依赖语义不能直接等同。", "evidence_pages": "4,9,11–13" }, { "dimension_id": "B3", "dimension_name": "数据与评测流程完整性", "score": 60, "status": "tentative", "confidence": "中", "judgment_range": "50–70", "reason": "数据分割和配置范围可见，遮挡归一化顺序、目标身份与选择程序不完整。", "evidence_pages": "5,8,11" }, { "dimension_id": "B4", "dimension_name": "比较对象与比较公平性", "score": 65, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "TimeXer等直接对手充分，但固定十轮非同收敛预算，缺简单外生回归。", "evidence_pages": "5,11" }, { "dimension_id": "C1", "dimension_name": "核心主张的证据充分度", "score": 65, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "部分预测和插件有效，但DLinear Traffic汇总有矛盾且普遍改善不成立。", "evidence_pages": "6–8,12" }, { "dimension_id": "C2", "dimension_name": "机制归因与竞争解释区分度", "score": 60, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "多种融合对照有价值，等价参数化和容量优化混杂未分离。", "evidence_pages": "7–9,11–13" }, { "dimension_id": "C3", "dimension_name": "结果稳定性与不确定性处理", "score": 30, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "未报告随机重复次数、种子或误差区间，千分级优势稳定性未知。", "evidence_pages": "6–8,11–12" }, { "dimension_id": "C4", "dimension_name": "适用范围与边界验证", "score": 70, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "多数据、范式、遮挡及超参较丰富，缺失协议和真实漂移边界有限。", "evidence_pages": "6–9,13" }, { "dimension_id": "D1", "dimension_name": "过程透明度", "score": 65, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "公式、数据数目和搜索范围明确，精确配置和汇总一致性有缺口。", "evidence_pages": "4–5,11–12" }, { "dimension_id": "D2", "dimension_name": "研究资源可获得性", "score": 75, "status": "tentative", "confidence": "中", "judgment_range": "65–85", "reason": "GitHub及Zenodo DOI有利资源保存，但本次未验证内容和执行。", "evidence_pages": "1" }, { "dimension_id": "D3", "dimension_name": "主张克制与报告一致性", "score": 45, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "一致改善、近乎免费和物理依赖结论过强，DLinear均值错误改变收益方向。", "evidence_pages": "1–2,8–9,12" }, { "dimension_id": "D4", "dimension_name": "表达清晰度", "score": 75, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "整体结构与全表清晰，个别轴维度叙述及平均标签不准确。", "evidence_pages": "3–4,6–13" }, { "dimension_id": "E1", "dimension_name": "复杂度合理性", "score": 80, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "层数少、残差和共享头职责明确，复杂结构负担低。", "evidence_pages": "3–5" }, { "dimension_id": "E2", "dimension_name": "资源效率与成本收益", "score": 60, "status": "tentative", "confidence": "中", "judgment_range": "50–70", "reason": "单设置训练时间有优势，多对多N平方因素及插件实际增量成本未验证。", "evidence_pages": "9" }, { "dimension_id": "E3", "dimension_name": "可迁移与可复用程度", "score": 85, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "五骨干插件和简单接口支持较高复用价值，不能保证每次获益。", "evidence_pages": "8,12–13" }, { "dimension_id": "E4", "dimension_name": "后续研究启发性", "score": 80, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "残差重参数化的优化作用、失效条件和变量选择问题具体且可迁移。", "evidence_pages": "7–9,11–13" }, { "dimension_id": "F1", "dimension_name": "与当前问题的相关性", "score": null, "status": "unknown", "confidence": "低", "judgment_range": null, "reason": "未提供具体当前研究目标、知识背景、资源和决策，不以通用价值替代个人评价。", "evidence_pages": "不适用页码：用户背景缺失" }, { "dimension_id": "F2", "dimension_name": "相对于已有知识的学习收益", "score": null, "status": "unknown", "confidence": "低", "judgment_range": null, "reason": "未提供具体当前研究目标、知识背景、资源和决策，不以通用价值替代个人评价。", "evidence_pages": "不适用页码：用户背景缺失" }, { "dimension_id": "F3", "dimension_name": "当前使用可行性", "score": null, "status": "unknown", "confidence": "低", "judgment_range": null, "reason": "未提供具体当前研究目标、知识背景、资源和决策，不以通用价值替代个人评价。", "evidence_pages": "不适用页码：用户背景缺失" }, { "dimension_id": "F4", "dimension_name": "当前决策信息价值", "score": null, "status": "unknown", "confidence": "低", "judgment_range": null, "reason": "未提供具体当前研究目标、知识背景、资源和决策，不以通用价值替代个人评价。", "evidence_pages": "不适用页码：用户背景缺失" } ], "body_zh": "## 阅读判断与独立详评\n\nCrossLinear 是强调简单变量融合的预测方法论文，也是一项插件式结构研究。对于想在通道独立骨干前加入低成本外生信息的人，建议定向查阅其方法与重参数化讨论；关于恢复直接物理依赖、插件一致改善和近乎零开销的表述则应保留。本轮先读历史摘录，再读二〇二五年五月第一版全部十三页及附录 A 至 E，会议论文集没有另行独立阅读。旧完整评语缺失，历史分数和排序均不沿用。\n\n方法的核心是一层跨变量一维卷积，把目标变量和辅助变量压成一条长度不变的序列，再与目标历史通过可学习标量融合。随后进行分块线性投影，加位置嵌入，用展平后的线性头输出未来。输入明确只有历史外生变量，不涵盖预测期间已知的未来协变量，因此实用范围应按这个信息集理解。多对多扩展通过增加卷积输出通道，再共享后续时间模块完成。该设计简单、容易插入现有通道独立方法，确实具有工程复用价值，原创增量主要是面向特定预测目标的残差融合与轻量结构，而不是新定义了一种统计交叉相关估计量。\n\n关于机制，附录主动证明加权残差形式可以合并为一个卷积核，这一点值得肯定：将目标通道中心位置的单位选择核按权重加入普通卷积核，即可得到相同映射。因此，Sum 优于 Cross Only 可以说明某种参数化、初始化或训练路径更好，却不能说明后者表达不了内外生变量的重要性差异。一个自由卷积本来就可给各输入通道不同权重，并非“平等对待所有变量”。正文的内外生语义解释是可能的直观描述，现有对照没有把它与残差初始化、正则化、优化条件和容量控制区分开。\n\n整个归一化后骨干基本由线性或仿射操作组成，展平线性头已经区分各位置。固定位置嵌入在线性头下可表现为一个加性偏置，因此其消融增益不直接证明学到了新的位置关系。类似地，单层卷积的固定权重只是模型参数时间不变，并不保证从数据中隔离了真实不随时间变化的关系，也不保证排除了间接统计关联。把卷积权重求和后画成热图，并与几个天气常识相符，只能作为定性解释线索；权重受归一化、相关输入和后续投影共同影响，不能独立等同于相关系数或物理直接依赖。\n\n预测结果的覆盖值得肯定：七个常用长期数据和五个电价子集，同时报告多对一、多对多两种任务，并包含直接相关的 TimeXer 和 TiDE。多对一在 ECL、Traffic 等数据取得竞争力，多对多在 ETTm1 的优势也较明确。不过，主表许多差距很小，例如 ECL 多对一平均 MSE 为零点三二三对 TimeXer 的零点三二七；没有重复次数、种子、方差或区间，无法确认微小差距稳定。多对多 Traffic 的零点四八三明显不如 iTransformer 的零点四二八，ECL 和 ETTh2 的平均也没有领先 TimeXer，因此结论应限定为若干数据上的良好效果，而非普遍最优。\n\n插件实验比只展示自有骨干更有说服力，五种结构中很多平均误差下降，RLinear 在 Traffic 的收益尤其明显。然而，历史所提局部退化已独立核实：Traffic 九十六步 DLinear 的 MSE 从零点二六八变为零点三三八，ECL 的 DLinear 平均 MAE 也变差。进一步逐行核算发现更重要的汇总错误：表十一 Traffic 的 DLinear 原始四个 MSE 为零点二六八、零点三零二、零点二九八、零点三四零，算术均值应为零点三零二，而不是表中的零点三二三；原 MAE 四行均值为零点三八四五，而非零点四零四。加插件后四行均值分别为零点三零八和零点三九一二五，意味着按逐行数字两个平均指标都退化，并非汇总表呈现的改善。\n\n这项核算已对 PDF 图像复核，差异大于四舍五入能解释的范围。现阶段无法确定原始分行还是汇总数哪一个抄错，因此不替作者擅改真实实验结果；正确处理是暂停使用该组汇总改善作为证据。它也不否定 RLinear、SparseTSF 等其他组合已经展示的收益。ECL 的 DLinear 最长跨度 MAE 在主表与插件表还有零点四八七和零点四八一的差异，进一步说明汇总与分行记录的一致性需要谨慎对待。\n\n消融包含目标单独输入、卷积单独输入、拼接、加权和、位置嵌入，以及输入长度、核大小、隐藏宽度等敏感性。Endo Only 提供了有价值的“外生变量是否帮助”对照，但 ETTh1 的变化几乎可以忽略，说明并非所有任务都需要这些变量。拼接与加权和改变表示维度，仍混入复杂度因素；初始化分析比直接宣称语义隔离更贴近已有证据。缺失值测试覆盖目标或外生变量的随机遮挡与两种填充值，但没有与其他模型并行对照，也没有说明遮挡是在实例归一化之前还是之后。尤其目标全部遮挡时，若仍保留原目标历史的均值和尺度，便不能说全部历史信息都缺失；这里仅指出协议未交代，不能据此认定实际泄漏。\n\n数据与公平性方面，分割样本数、硬件、优化器、轮数和搜索范围都有提供。Weather 特殊异常值问题也被主动披露，维持已有协议是可以理解的，但不能把这种数值尺度下的微小误差直接推广成一般天气预测优势。多对一目标变量的具体选择没有为各数据清楚列出，目标选择是否具有代表性仍未知。统一十轮训练和遵循各官方配置，并不等于各基线充分收敛或获得相同调参预算。传统外生回归、简单多变量线性基线及电价领域的 NBEATSx 未进入比较，限制了“简单卷积为何必要”的结论强度。\n\n效率图在同批量 ECL 上显示较好的训练时间与误差折中，但没有插件加入前后完整的增量成本，也没有推理、显存和总调参开销。文中按历史长度写线性复杂度，在变量数、核宽、隐藏维度和预测长度固定时可以成立；却不能忽略多对多卷积核明确具有 N乘N乘三的规模。变量数增加时该部分有平方级参数与计算因素，不能从一维长度的复杂度直接推出任意高维任务几乎免费。资源入口同时提供 GitHub 与 Zenodo DOI，对版本保存有帮助，但本次未下载运行或验证归档内容。\n\n可带走的洞见是，将目标历史作为强残差基线，再让辅助变量提供可学习修正，是简洁且可组合的设计。真正待解释的是这种重参数化为何在有限训练预算下更有效，以及在哪些变量关系、目标选择和缺失机制下会失效。建议优先阅读第四页方法与第十一至十三页重参数化和完整插件表，避免只看汇总胜率。当前用户目标与资源未知，个人适配四维不评分。" }