{ "candidate_id": "ATT-0097", "reviewer": "C", "rubric": "PAPER-RUBRIC-1.0", "frozen": true, "frozen_at": "2026-11-08T02:43:44.092367+00:00", "identity": { "candidate_id": "ATT-0097", "title": "Adaptive Multi-Scale Decomposition Framework for Time Series Forecasting", "authors": [ "Hu, Yifan", "Liu, Peiyuan", "Zhu, Peng", "Cheng, Dawei", "Dai, Tao" ], "arxiv_id": "2406.03751", "arxiv_version": "v2", "first_arxiv_date": "2024/06/06", "latest_arxiv_date": "2025/04/15", "source_urls": [ "https://ojs.aaai.org/index.php/AAAI/article/view/33908" ], "publication_status": "AAAI 2025, DOI 10.1609/aaai.v39i16.33908 verified", "full_text_path": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0097/ATT-0097_paper.pdf", "metadata_source_path": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0097/ATT-0097_arxiv.html", "full_text_sha256": "02aad9b5b876e3b53d6c34114d6e658eb557898d57f20402baf512b3e264b05e", "review_state": "ready_for_five_new_independent_reviews", "completed_new_reviews": 0, "source_versions_not_asserted_identical": true }, "source_sha256_verified": "02aad9b5b876e3b53d6c34114d6e658eb557898d57f20402baf512b3e264b05e", "reading_started_at": "2026-11-08T02:41:51Z", "reading_finished_at": "2026-11-08T02:43:44.092367+00:00", "reading_time_note": "以UTC记录开始和完成时刻", "read_coverage": { "pdf_pages": 14, "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": [ 4, 7, 10 ] }, "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": "AMD提升多尺度预测", "judgment": "部分支持，最近基线存在反例", "pages": "5–7" }, { "claim_id": "CL02", "claim": "理论保证多尺度线性预测可行性", "judgment": "当前证明存在直接错误，暂停作为可靠结论使用", "pages": "9–10" }, { "claim_id": "CL03", "claim": "动态权重及模块有复用价值", "judgment": "局部消融和插件支持，但尺度可解释性不足", "pages": "6–7" } ], "risks": [ { "severity": "关键问题", "certainty": "已确认", "description": "定理证明同常数保持及不同列权重上界错误；暂停CL02当前证明", "pages": "10" }, { "severity": "重要问题", "certainty": "文稿已确认、实现未知", "description": "式十alpha=0反向压低主导项，与表六和叙述不一致", "pages": "4,12" }, { "severity": "重要问题", "certainty": "已确认", "description": "一致SOTA受ETTh2等反例限制", "pages": "5" } ], "scores": [ { "dimension_id": "A1", "dimension_name": "问题重要性", "score": 80, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "多尺度与通道信息取舍是预测中的重要难点。", "evidence_pages": "1–3" }, { "dimension_id": "A2", "dimension_name": "原创性", "score": 60, "status": "tentative", "confidence": "中", "judgment_range": "50–70", "reason": "相对TimeMixer及线性专家混合属于有用组合，前作全文核对有限。", "evidence_pages": "2–4" }, { "dimension_id": "A3", "dimension_name": "实质推进程度", "score": 65, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "有若干任务提升及插件证据，但最近基线并非稳定全面胜出。", "evidence_pages": "5–7" }, { "dimension_id": "A4", "dimension_name": "知识与解释增量", "score": 50, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "消融给出局部认识，尺度解释与理论支撑不足。", "evidence_pages": "6–7,9–10" }, { "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": "长短预测与MSE/MAE明确，泛化及可解释性缺更直接测量。", "evidence_pages": "3,5–7" }, { "dimension_id": "B2", "dimension_name": "技术与论证正确性", "score": 35, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "理论有两处可构造反例的错误，门控公式与零alpha配置也矛盾。", "evidence_pages": "4,9–10,12" }, { "dimension_id": "B3", "dimension_name": "数据与评测流程完整性", "score": 65, "status": "tentative", "confidence": "中", "judgment_range": "55–75", "reason": "划分数量齐全，最佳输入选择与预处理拟合范围未完整说明。", "evidence_pages": "5,11–12" }, { "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": "搜索基线输入长度较合理，但调参预算、短期最近对手及效率参照仍有缺口。", "evidence_pages": "5–7,12" }, { "dimension_id": "C1", "dimension_name": "核心主张的证据充分度", "score": 60, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "经验性能部分支持，理论可行性证明和一致SOTA表述不能成立。", "evidence_pages": "5–7,10" }, { "dimension_id": "C2", "dimension_name": "机制归因与竞争解释区分度", "score": 65, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "平均权重与随机序列对照有意义，尺度可辨识和替代集成解释未区分。", "evidence_pages": "6–7" }, { "dimension_id": "C3", "dimension_name": "结果稳定性与不确定性处理", "score": 70, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "六数据五种子波动明确，但对手及消融缺配对统计。", "evidence_pages": "12,14" }, { "dimension_id": "C4", "dimension_name": "适用范围与边界验证", "score": 65, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "多任务多尺度参数覆盖，缺噪声与有限历史边界的系统证据。", "evidence_pages": "5–7,12–14" }, { "dimension_id": "D1", "dimension_name": "过程透明度", "score": 65, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "多数训练配置和算法公开，门控参数标签、公式形状仍有歧义。", "evidence_pages": "4,11–12" }, { "dimension_id": "D2", "dimension_name": "研究资源可获得性", "score": 65, "status": "tentative", "confidence": "中", "judgment_range": "55–75", "reason": "代码入口明确，未核验依赖版本、快照和运行。", "evidence_pages": "1,11–12" }, { "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": "一致最优、通道解释、理论担保和资源不变措辞超过支持范围。", "evidence_pages": "1–2,5–7,10" }, { "dimension_id": "D4", "dimension_name": "表达清晰度", "score": 70, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "模块图与消融易读，公式错误妨碍可靠理解。", "evidence_pages": "3–7,10" }, { "dimension_id": "E1", "dimension_name": "复杂度合理性", "score": 65, "status": "rated", "confidence": "中", "judgment_range": null, "reason": "模块各有实证动机，但密集专家成本与简单集成对照仍需界定。", "evidence_pages": "4,6–7" }, { "dimension_id": "E2", "dimension_name": "资源效率与成本收益", "score": 65, "status": "tentative", "confidence": "中", "judgment_range": "55–75", "reason": "单数据固定配置训练收益不错，不是所有轻量模型更快，缺广泛成本。", "evidence_pages": "7" }, { "dimension_id": "E3", "dimension_name": "可迁移与可复用程度", "score": 75, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "MDM和AMS可分离，已有两种基线插件结果。", "evidence_pages": "7" }, { "dimension_id": "E4", "dimension_name": "后续研究启发性", "score": 70, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "尺度语义与通用条件集成的区分是具体机制问题。", "evidence_pages": "4,6–7" }, { "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\nAMD 是方法论文，辅以理论论证和效率实验。建议定向查阅其多尺度混合及自适应预测器设计，但附录定理的现有证明暂停作为可靠结论使用。旧条目给出了证明风险与若干实验反例，本轮先将其作为待核线索，再独立核对十四页全文；旧完整评语仍未提供，历史总分不沿用，也不计入新评。本文只评价二〇二五年四月第二版，保留它与会议版本可能不同的边界。\n\n方法从平均池化获得多种时间尺度，由粗到细做残差混合，再通过按片段的时间交互和可缩放的通道交互，生成多个预测器的输入。门控网络根据输入分配随预测位置变化的权重，密集组合各预测器。相对 TimeMixer 的多尺度建模和已有线性专家混合，主要增量是把尺度混合、双依赖处理和动态聚合组合成完整系统。这些部件有明确动机，但不能把每个专家自动解释为一个真实物理尺度；训练目标没有给出专家与可辨识尺度的一一对应约束。原文所谓自相关建模也应由具体操作理解，不能仅由命名推导出统计自相关被准确恢复。\n\n实证方面，长期预测覆盖 ETT、天气、电力、交通、汇率和太阳能等常用数据，短期部分还有四个 PEMS 集。作者说明重跑基线并搜索六种输入长度，这比固定一个不适合所有基线的长度更合理。长预测结果确有若干竞争优势，例如 ETTh1 中长预测和部分 ETTm、Solar 条目。但一致最先进的表述不成立：ETTh2 的七百二十步 MSE 为零点四零二，TimeMixer 为零点三四二；Weather 最长预测也落后，Exchange 最长预测不如 DLinear。PEMS 的 MSE 改善较明显，但部分 MAE 仍由其他方法领先，短期表还缺最接近的 TimeMixer 及专用交通图模型，因而结论应限定于所比较方法和指标。\n\n消融比单纯去模块更丰富，包括随机顺序、平均权重、稀疏专家、通道强度以及负载均衡项。随机顺序和平均聚合使结果下降，支持时间结构和自适应权重确实有作用；模块移植到 DLinear、MTS-Mixers 的结果也提供了可复用证据。但实验主要集中 Weather 和 ECL，不能直接证明所有数据都具有同样机制。平均权重对照比删除全部专家更接近控制参数量，值得肯定；另一方面，重排会同时改变输入分布，不能仅据此推导自注意力必然丢失序列信息。通常带位置编码的 Transformer 也不能笼统称为顺序不可知。通道分支越强越差说明特定配置会引入干扰，不足以证明学到的关系就是错误相关。\n\n理论部分的风险已独立确认。式十三到十四把分母从 a-b 改成 d(a-b)，却保留外面的四分之一式平均因子的一般形式一除以 d，没有补回相应倍数。一个直接反例是原序列 f(t)=t，Lipschitz 常数为一，下采样率 d=2 后 f1(a)=2a-0.5，相对于新索引的常数为二而非一。这否定了证明中的同常数保持步骤，并非仅是排版美观问题。式二十一到二十二进一步将不同位置的列权重混在同一个最大值里，遗漏权重差乘函数值的项。取常数非零序列，原常数可以为零，而不同列权重仍会生成不恒定输出，足以说明该上界推导没有成立。\n\n后半证明也没有补救这些问题：式二十五直接令 g(L+t)=g(P+1+t)，未给出相应长度或周期条件；任意可学习线性变换之后的周期性与平滑常数没有控制，有限历史表示的未来扩展也未清楚定义。定理文字仅说存在有限预测误差界，本身很弱，不能推出较小误差、训练可找到该模型或多尺度优于单尺度。这里暂停的是当前证明及其用于支持机制的结论，不是声称所有有界预测命题都不可能成立，更不把理论失效等同于已观测实验结果不存在。\n\n另一个新发现的公式和配置矛盾在门控。式十对低于阈值的项用 αlog(1+u)，对高于阈值的项用 αexp(u)-1；表六多数数据的 α 为零。按文稿代入，低项得到零、高项得到负一，后续 softmax 反而给予高项更小权重，与强调主导模式的解释相反。该式已视觉复核，但尚未查看代码，因此可能是公式括号或参数标签错误，不能断言实际实现必然如此。当前材料需要把算法陈述与配置一致性视为未解决问题。\n\n复现方面，附录提供了数据划分数量、各数据学习率、批量、轮数、隐藏维度、专家数量和仓库入口。五个随机种子的均值与波动覆盖六个数据集，这是实质优点，不能误写为完全没有重复实验。局限在于主要对手和关键消融没有对应成对区间，输入长度与其他超参数以何种验证程序选最佳仍未充分交代。部分差距可以稳定判断，千分级差距则不宜过度排序。图五实测 Weather、固定批量、输入五百一十二和预测九十六时的训练耗时及显存：八专家 AMD 约十七毫秒、一千三百四十九 MB，明显快于较大的 Transformer，但不如七毫秒的 DLinear 便宜，且没有 TimeMixer 的同图对比。不能据此声称比所有 MLP 都高效或插件增加后资源绝对不变。\n\n可带走的是将粗细尺度残差混合与按位置的专家权重分开建模，二者可独立借鉴。更深的研究问题是专家权重是否反映真实尺度变化，还是只是一种通用条件集成；若没有可辨识尺度仍能同样有效，解释应转向后者。优先阅读第三至四页的模块定义，以及第十页证明与第十四页重复结果。个人适配四维因缺少用户具体目标与资源保留未知。" }