{
"paper_id": "ATT-0097",
"paper_type": "多尺度MLP预测方法及附带理论论证",
"reading_recommendation": "定向读多尺度混合与自适应预测头，暂停采信定理证明；经验收益需按数据集与跨度分层。",
"detailed_review_zh": "## 总体判断与历史线索复核\nAMD的实用想法是以平均下采样提取多尺度信息、由粗到细残差混合，再用时间/通道交互和自适应多个预测器组合输出。它比单一线性头更灵活，能在若干基准上得到较好精度，且提供较丰富的消融与重复实验。阅读价值主要在模块组织和成本折中，理论部分目前不能作为采用方法的依据。旧摘要曾指出附录证明和部分结果问题，本次先读取该待核线索，再独立检查原文公式、表格和图像，确认关键证明错误，但没有用旧总分59.2作为新分数锚点。旧评全文、旧评分文件并未实际提供。当前全文为2025年4月15日arXiv v2，不能假定与首次2024年6月稿或会议版完全相同。\n\n## 方法贡献与机制\nMDM由平均池化生成多尺度序列，以可学习映射逐层回传到细尺度；DDI对patch执行时间递推和通道映射，用系数控制通道信息强度；AMS使用带噪声的选择器，为多个预测器生成随预测时点变化的权重。本文所谓TopK不是只激活少数专家，而是对前若干项和其余项作不同变换，再softmax得到全部非零的稠密权重，这点应与稀疏MoE区别。TimeMixer、多尺度分解、MLP混合及线性专家组合都是相关近邻，真正增量是组合方式与动态加权策略，原创性不宜按“首次多尺度”评估。\n\n各专家共享经混合后的输入，并非被硬性约束只预测某一独立尺度，所以从门控图看出TP5或TP16权重较高，并不足以证明对应真实可识别的时间尺度。对通道关联的缩放是合理工程选择，但两张输入热图也不足以证明跨通道学习一般会破坏分布。自注意力在不带位置时对置换等变，并非所有带位置模型都丢失顺序；随机打乱AMD中的表征导致变差，也不能直接推出它比所有注意力机制更懂时间顺序。应把这些说法当作解释性假设，而不是完成了机制验证。\n\n## 理论主张暂停采信\n原图附录第10页中，式(13)的分母a−b在式(14)变成d(a−b)，外面的1/d却保持不变，确实少了倍数d。用f(j)=j、d=2即可直接检查：下采样均值f1(a)=2a−1/2，相邻粗尺度索引斜率为2，而原f的Lipschitz常数是1；原推导“仍不超过K”不成立。可以用重新缩放坐标或改为dK修补这一局部结论，但必须明确修改，不能维持当前等式。\n\n第二处不只是少一个常数。式(21)到式(22)把两列不同权重的差异丢掉，仅用函数值差乘最大权重作界。取常数输入gn=1、K=0，同时两列权重不同，则前式可以非零，后式却为零，构成直接反例。固定矩阵不等于其各列相同，也不能自动保留关于时间索引的原Lipschitz常数。之后式(25)把L+t写成P+1+t也缺条件，线性映射构造混用未来点、维度和索引，没有解决从有限历史定义到未来延拓的问题。即使“存在某个有界误差线性预测器”在额外有限域假设下很弱甚至容易成立，本文展示的具体证明和界仍不成立。我将其列为仅针对理论支撑的致命问题，暂停该主张，不据此否定训练模型的所有数值实验。\n\n## 实验和比较\n主表纳入TimeMixer、PatchTST、iTransformer、DLinear等较直接基线，作者说重新运行并在多个输入长度中取最好，较只选一个不适合基线的窗口有进步。四个预测跨度和多种长短期数据，提供了相当完整的误差覆盖。但是“最佳”输入长度是否只按验证集选择、各模型超参数预算是否相当、实际每行采用何配置，全文未闭合。不能直接说存在测试集调参，应该将选择规则作为待补的公平性条件。短期PEMS比较中缺少完整的任务专用空间交通模型族，也不包含主长期对照TimeMixer，故优势的适用范围是所列模型，而非交通预测全领域。\n\n结果并非持续SOTA。ETTh2的720步AMD MSE为0.402，TimeMixer为0.342，约差17.5%；Exchange-720上AMD0.750也差于DLinear0.643；Traffic的多个短跨度亦不领先。相反ETTm1较短跨度、ETTm2部分跨度和Solar长跨度的结果较好，短期PEMS的MSE也有明显竞争力。MAE并不总同向，例如PEMS03和04的AMD MAE均0.198，分别高于TimesNet0.192和0.195。应用时应按损失和尺度选择，不能把一列最佳数量等同全任务最佳。\n\n## 消融、稳定性与说明一致性\nWeather和ECL上有去MDM、去DDI、去AMS、均匀权重、随机顺序、稀疏专家以及损失项消融，能支持多个组件在这两个条件下有用。表7给出六组数据、四个跨度的五种子均值和小标准差，说明自身训练随机性较小，是实证优势；仍缺对应基线和所有消融的配对方差，也不覆盖ETTh2这种明显弱项。数据选择和模型容量是机制解释的竞争因素，删除整个AMS与单一线性头的比较不等于完全等参数对照。\n\n另一个需要澄清的关键规格在表6：许多数据集标α=0，而若它确指式(10)的α，前k项变为−1、其他项变为0，后续softmax反而降低前k项权重，与“突出主导模式”的解释相反。也可能表头误把通道缩放β写成α，全文不足以判定，不能擅自替作者选择实现。这会影响可重复性和门控机制解释。正文说全实验8个预测器，但解释图展示TP16，可能来自另一个敏感性配置，需标注而非断言造假。\n\n## 成本、复用与最终建议\n图5在Weather、输入512预测96、同批大小下给训练时间和GPU内存，AMD八专家约17ms/iter、1349MB，相比PatchTST195ms和7048MB较有吸引力。但DLinear约7ms、999MB更轻，因而“优于所有MLP效率”不准确；不能把单条件训练时间推到所有通道数、所有预测长度或部署推理。文中说插件保持相同计算资源，却未附逐插件成本表，且表4有个别恶化，如DLinear在ECL-192的MSE从0.152到0.159，不能写全行改善。\n\n数据划分、训练轮数、学习率和部分模型维度给得较细，代码入口明确。仍有α/β、下采样符号、初始化及门控配置的不一致，本次未访问实际代码快照或运行实验。最值得带走的是按输入模式动态组合预测器的工程方向，以及通道交互强度应有条件地控制；理论可用性、尺度可解释性及均预算比较都需要补证。四个F维度缺个人背景，维持未知。",
"reading_record": {
"legacy": "先读旧摘要，逐条独立复核式13–14、21–22与ETTh2/720；未读旧全文或其他新评",
"full_text": "PDF1–14全部正文、参考文献、证明、伪代码、设置、额外实验和局限阅读",
"visual_checks": "p10证明逐式核对，p7效率图，p12参数表，p13敏感性图",
"code_executed": false,
"data_versions": "2406.03751v2；无外部代码/数据快照核验"
},
"claims": [
{
"claim_id": "CL01",
"claim": "多尺度混合和动态专家提高部分时序任务精度",
"claim_type": "经验性能",
"conditions": "所列数据和选择配置",
"support_status": "支持但不是持续全任务SOTA",
"location": "表1–4"
},
{
"claim_id": "CL02",
"claim": "定理1所示推导证明多尺度线性预测的误差界",
"claim_type": "理论",
"conditions": "原文Lipschitz与任意固定混合矩阵条件",
"support_status": "暂停；关键等式与不等式有反例",
"location": "附录式13–16、21–32"
},
{
"claim_id": "CL03",
"claim": "专家门控辨认真实主导尺度",
"claim_type": "机制解释",
"conditions": "需独立尺度真值与唯一性证据",
"support_status": "仅可视化和性能关联，未证可识别",
"location": "图3、式7–10"
}
],
"risks": [
{
"risk_id": "R01",
"severity": "致命问题",
"certainty": "原图及解析反例已确认，仅限理论支撑",
"description": "式13→14遗漏d，式21→22丢失权重列差；当前定理证明链不成立",
"location": "p10式13–24",
"question": "是否能在明确坐标、矩阵范数/列变化及未来延拓条件下给出新证明？"
},
{
"risk_id": "R02",
"severity": "重要问题",
"certainty": "规格歧义已确认，实际代码未知",
"description": "表6许多α=0与式10高权重选择解释相反，可能α/β误记",
"location": "p4式10、p12表6",
"question": "表6参数实际对应哪个公式，完整门控设置是什么？"
},
{
"risk_id": "R03",
"severity": "重要问题",
"certainty": "评估选择过程未完整说明",
"description": "最佳输入长度的验证选择程序及各模型调参预算未闭合",
"location": "表1说明、实验设置、附录实现",
"question": "能否逐行给验证准则与最终配置？"
},
{
"risk_id": "R04",
"severity": "重要问题",
"certainty": "表格直接支持边界",
"description": "ETTh2长跨度和部分MAE明显落后，插件并非全行改善",
"location": "表1、2、4",
"question": "哪些数据特征可预判这种取舍？"
}
],
"rubric_version": "PAPER-RUBRIC-1.0",
"reviewer_id": "A",
"paper_title": "Adaptive Multi-Scale Decomposition Framework for Time Series Forecasting",
"paper_version": "2406.03751v2",
"first_public_date": "2024/06/06",
"version_date": "2025/04/15",
"assessment_date": "2026-11-08",
"analysis_mode": "standard",
"independence": {
"other_new_reviews_read": false,
"legacy_read_before_fulltext": true,
"legacy_full_reviews_available": false,
"legacy_records": [
{
"source_record_id": "S5-097",
"source_id": "S5",
"source_library_file_id": "libfile_849759d66a808191813dbe9bd1605bb4",
"source_filename": "已粘贴的文本(5).txt",
"source_line": 29,
"source_number": 97,
"title": "Adaptive Multi-Scale Decomposition Framework for Time Series Forecasting",
"source_url": "https://ojs.aaai.org/index.php/AAAI/article/view/33908",
"raw_line": "| 13 | [P097 Adaptive Multi-Scale Decomposition Framework for Time Series Forecasting](https://ojs.aaai.org/index.php/AAAI/article/view/33908) | 59.2 | 100% | 35% | 55.5–63.5 | 74 | AAAI 2025 会议论文；补充理论与结果为 arXiv:2406.03751v2（2025-04-15） | **重要问题**：arXiv v2 附录 A 式 (13)→(14) 把分母由 a-b 变为 d(a-b) 时遗漏 d；式 (21)→(22) 的不等式也忽略不同列权重。；
**重要问题**：表 1 的 ETTh2/720 中 AMD MSE 0.402，TimeMixer 0.342；长期预测表现取决于数据与预测长度。；
**重要问题**：基线各模型调参预算和最佳输入长度选择程序未在已检查文字中完整交代；影响相对效果的可比性。 | [评审](论文/P097/评审.md) · [评分数据](论文/P097/评审.json) · 无单独复核文件 |",
"raw_cells": [
"13",
"[P097 Adaptive Multi-Scale Decomposition Framework for Time Series Forecasting](https://ojs.aaai.org/index.php/AAAI/article/view/33908)",
"59.2",
"100%",
"35%",
"55.5–63.5",
"74",
"AAAI 2025 会议论文；补充理论与结果为 arXiv:2406.03751v2（2025-04-15）",
"**重要问题**：arXiv v2 附录 A 式 (13)→(14) 把分母由 a-b 变为 d(a-b) 时遗漏 d；式 (21)→(22) 的不等式也忽略不同列权重。；
**重要问题**：表 1 的 ETTh2/720 中 AMD MSE 0.402，TimeMixer 0.342；长期预测表现取决于数据与预测长度。；
**重要问题**：基线各模型调参预算和最佳输入长度选择程序未在已检查文字中完整交代；影响相对效果的可比性。",
"[评审](论文/P097/评审.md) · [评分数据](论文/P097/评审.json) · 无单独复核文件"
],
"normalized_title": "adaptivemultiscaledecompositionframeworkfortimeseriesforecasting",
"normalized_url": "ojs.aaai.org/index.php/aaai/article/view/33908",
"record_type": "legacy_scored",
"legacy_rank": 13,
"legacy_score": "59.2",
"legacy_coverage": "100%",
"legacy_tentative_share": "35%",
"legacy_range": "55.5–63.5",
"citation_verbatim": "74",
"legacy_review_version": "AAAI 2025 会议论文；补充理论与结果为 arXiv:2406.03751v2（2025-04-15）",
"legacy_risks_verbatim": "**重要问题**：arXiv v2 附录 A 式 (13)→(14) 把分母由 a-b 变为 d(a-b) 时遗漏 d；式 (21)→(22) 的不等式也忽略不同列权重。；
**重要问题**：表 1 的 ETTh2/720 中 AMD MSE 0.402，TimeMixer 0.342；长期预测表现取决于数据与预测长度。；
**重要问题**：基线各模型调参预算和最佳输入长度选择程序未在已检查文字中完整交代；影响相对效果的可比性。",
"legacy_material_links_verbatim": "[评审](论文/P097/评审.md) · [评分数据](论文/P097/评审.json) · 无单独复核文件",
"counts_as_independent_review": false
}
],
"legacy_counts_as_review": false
},
"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
},
"user_context": {
"research_question": null,
"prior_knowledge": null,
"constraints": null
},
"aggregation": {
"weights": null,
"weighted_score_observed": null,
"weighted_coverage": null,
"tentative_weight_share": null,
"notes": "保留24维原始评分，不预设总分；F组个人背景未知。"
},
"scores": [
{
"dimension_id": "A1",
"dimension_name": "问题重要性",
"score": 80,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "多尺度及通道依赖的成本效果问题重要",
"missing_information": null,
"source_ids": [
"S01"
],
"evidence_locations": [
"引言"
],
"claim_ids": []
},
{
"dimension_id": "A2",
"dimension_name": "原创性",
"score": 60,
"status": "tentative",
"confidence": "中",
"judgment_range": [
50,
70
],
"rationale": "多尺度混合与动态专家组合有增量，近邻设计很多",
"missing_information": "见理由及风险记录",
"source_ids": [
"S01"
],
"evidence_locations": [
"相关工作、方法"
],
"claim_ids": []
},
{
"dimension_id": "A3",
"dimension_name": "实质推进程度",
"score": 65,
"status": "tentative",
"confidence": "中",
"judgment_range": [
55,
75
],
"rationale": "部分数据有改进且训练较高效，普遍性有限",
"missing_information": "见理由及风险记录",
"source_ids": [
"S01"
],
"evidence_locations": [
"表1–4"
],
"claim_ids": []
},
{
"dimension_id": "A4",
"dimension_name": "知识与解释增量",
"score": 45,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "理论链失败，机制解释多为可视化假说",
"missing_information": null,
"source_ids": [
"S01"
],
"evidence_locations": [
"定理1、图3/4"
],
"claim_ids": []
},
{
"dimension_id": "B1",
"dimension_name": "问题定义与评价有效性",
"score": 75,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "多数据长短期目标清楚，短期基线范围有限",
"missing_information": null,
"source_ids": [
"S01"
],
"evidence_locations": [
"实验、表1/2"
],
"claim_ids": []
},
{
"dimension_id": "B2",
"dimension_name": "技术与论证正确性",
"score": 30,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "关键证明可被直接反例否定；实际方法另保留经验判断",
"missing_information": null,
"source_ids": [
"S01"
],
"evidence_locations": [
"附录式13–24"
],
"claim_ids": []
},
{
"dimension_id": "B3",
"dimension_name": "数据与评测流程完整性",
"score": 65,
"status": "tentative",
"confidence": "中",
"judgment_range": [
50,
80
],
"rationale": "有划分和重复设置，数据/选择脚本未独立核验",
"missing_information": "见理由及风险记录",
"source_ids": [
"S01"
],
"evidence_locations": [
"表5/6"
],
"claim_ids": []
},
{
"dimension_id": "B4",
"dimension_name": "比较对象与比较公平性",
"score": 65,
"status": "tentative",
"confidence": "中",
"judgment_range": [
55,
75
],
"rationale": "声称重跑与多输入搜索，最佳选择程序不全",
"missing_information": "见理由及风险记录",
"source_ids": [
"S01"
],
"evidence_locations": [
"实验设置"
],
"claim_ids": []
},
{
"dimension_id": "C1",
"dimension_name": "核心主张的证据充分度",
"score": 60,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "实证支持条件性贡献，不支持理论界或持续最优",
"missing_information": null,
"source_ids": [
"S01"
],
"evidence_locations": [
"表1–4、定理1"
],
"claim_ids": []
},
{
"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,
"rationale": "消融较全但仅两集、容量和尺度可识别未充分分离",
"missing_information": null,
"source_ids": [
"S01"
],
"evidence_locations": [
"表3、图3"
],
"claim_ids": []
},
{
"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,
"rationale": "五种子六数据标准差较好，缺基线/消融配对方差",
"missing_information": null,
"source_ids": [
"S01"
],
"evidence_locations": [
"表7"
],
"claim_ids": []
},
{
"dimension_id": "C4",
"dimension_name": "适用范围与边界验证",
"score": 65,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "多领域跨度且限制讨论，失败机制未验证",
"missing_information": null,
"source_ids": [
"S01"
],
"evidence_locations": [
"表1、末尾局限"
],
"claim_ids": []
},
{
"dimension_id": "D1",
"dimension_name": "过程透明度",
"score": 60,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "配置较详细但关键α/β和理论符号歧义",
"missing_information": null,
"source_ids": [
"S01"
],
"evidence_locations": [
"表6、算法1"
],
"claim_ids": []
},
{
"dimension_id": "D2",
"dimension_name": "研究资源可获得性",
"score": 70,
"status": "tentative",
"confidence": "中",
"judgment_range": [
55,
85
],
"rationale": "提供代码入口，未核对应版本可运行性",
"missing_information": "见理由及风险记录",
"source_ids": [
"S01"
],
"evidence_locations": [
"p1链接"
],
"claim_ids": []
},
{
"dimension_id": "D3",
"dimension_name": "主张克制与报告一致性",
"score": 45,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "持续SOTA、全插件改善、普遍效率等超过表内证据",
"missing_information": null,
"source_ids": [
"S01"
],
"evidence_locations": [
"结论、表1/4、图5"
],
"claim_ids": []
},
{
"dimension_id": "D4",
"dimension_name": "表达清晰度",
"score": 65,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "模块清晰但符号、理论和解释性措辞需校正",
"missing_information": null,
"source_ids": [
"S01"
],
"evidence_locations": [
"方法、附录"
],
"claim_ids": []
},
{
"dimension_id": "E1",
"dimension_name": "复杂度合理性",
"score": 65,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "多个模块均有消融依据， dense experts增加预算",
"missing_information": null,
"source_ids": [
"S01"
],
"evidence_locations": [
"表3、图5"
],
"claim_ids": []
},
{
"dimension_id": "E2",
"dimension_name": "资源效率与成本收益",
"score": 65,
"status": "tentative",
"confidence": "中",
"judgment_range": [
55,
75
],
"rationale": "单设置训练成本较好，但非所有MLP最优且无全面推理成本",
"missing_information": "见理由及风险记录",
"source_ids": [
"S01"
],
"evidence_locations": [
"图5"
],
"claim_ids": []
},
{
"dimension_id": "E3",
"dimension_name": "可迁移与可复用程度",
"score": 70,
"status": "tentative",
"confidence": "中",
"judgment_range": [
60,
80
],
"rationale": "可作插件，门控规范和额外成本待核",
"missing_information": "见理由及风险记录",
"source_ids": [
"S01"
],
"evidence_locations": [
"表4、表6"
],
"claim_ids": []
},
{
"dimension_id": "E4",
"dimension_name": "后续研究启发性",
"score": 70,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "动态专家与关联强度的条件性选择有研究价值",
"missing_information": null,
"source_ids": [
"S01"
],
"evidence_locations": [
"方法、局限"
],
"claim_ids": []
},
{
"dimension_id": "F1",
"dimension_name": "与当前问题的相关性",
"score": null,
"status": "unknown",
"confidence": null,
"judgment_range": null,
"rationale": "用户当前研究问题、已有知识、资源和决策条件未提供。",
"missing_information": "用户当前研究问题、已有知识、资源和决策条件未提供。",
"source_ids": [
"S01"
],
"evidence_locations": [],
"claim_ids": []
},
{
"dimension_id": "F2",
"dimension_name": "相对于已有知识的学习收益",
"score": null,
"status": "unknown",
"confidence": null,
"judgment_range": null,
"rationale": "用户当前研究问题、已有知识、资源和决策条件未提供。",
"missing_information": "用户当前研究问题、已有知识、资源和决策条件未提供。",
"source_ids": [
"S01"
],
"evidence_locations": [],
"claim_ids": []
},
{
"dimension_id": "F3",
"dimension_name": "当前使用可行性",
"score": null,
"status": "unknown",
"confidence": null,
"judgment_range": null,
"rationale": "用户当前研究问题、已有知识、资源和决策条件未提供。",
"missing_information": "用户当前研究问题、已有知识、资源和决策条件未提供。",
"source_ids": [
"S01"
],
"evidence_locations": [],
"claim_ids": []
},
{
"dimension_id": "F4",
"dimension_name": "当前决策信息价值",
"score": null,
"status": "unknown",
"confidence": null,
"judgment_range": null,
"rationale": "用户当前研究问题、已有知识、资源和决策条件未提供。",
"missing_information": "用户当前研究问题、已有知识、资源和决策条件未提供。",
"source_ids": [
"S01"
],
"evidence_locations": [],
"claim_ids": []
}
],
"sources": [
{
"source_id": "S01",
"source_type": "原论文全文及附录",
"location": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0097/ATT-0097_paper.pdf",
"version_or_date": "2406.03751v2",
"sha256": "02aad9b5b876e3b53d6c34114d6e658eb557898d57f20402baf512b3e264b05e",
"access_status": "全文可读；未运行代码或模型"
},
{
"source_id": "S02",
"source_type": "历史摘要，非完整旧评",
"location": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0097/legacy_reference_only.json",
"access_status": "先读；仅历史待核线索，不作新评票"
}
],
"review_chinese_char_count": 1796
}