为了公平比较,10个数据集的抽样频率,回溯窗口,预测长度细节如表所示,日周期数据集是指Environ...

Tạo vào: 5 tháng 10, 2026

Trả lời bằng GPT-5.6 Thinking bởi Chat01

Câu hỏi

image.png

为了公平比较,10个数据集的抽样频率,回溯窗口,预测长度细节如表所示,日周期数据集是指Environment,回溯和预测窗口分别为L=96, H \in {48,96,192,336};周为周期的数据集包含Energy和Health (US),回溯和预测长度分别为L=24, H \in {12,24,36,48},月为周期的数据集包含Traffic, SocialGood, Security, Economy, Climate, Agriculture的回溯和预测眼界分别为L=8, H \in {6,8,10,12}.中文润色得更连贯,并翻译成顶会论文。以及类似参照物的表格

Suy nghĩ

Clarifying table details

I’m polishing the paragraph into conference-style English and matching the reference table’s structure. I’m also noticing the listed groups contain nine datasets, not ten, so the missing dataset should be checked before finalizing the table.

Clarified table details

Worked for 18s

Trả lời

可以。你这段更适合用“周期 / sampling periodicity”来组织,而不是反复使用“抽样频率、预测长度、预测眼界”等不同表述。顶会论文里建议统一使用 look-back window 和 forecasting horizon。

先提醒一个细节:你目前列出的数据集是 1 + 2 + 6 = 9 个,分别是 Environment、Energy、Health (US)、Traffic、SocialGood、Security、Economy、Climate、Agriculture。如果正文确实是 10 个数据集,还需要补充一个遗漏的数据集。

中文润色版

为保证不同数据集之间的公平比较,我们根据数据的周期特性统一设置回溯窗口(look-back window)和预测长度(forecasting horizon),具体配置如表 X 所示。对于日周期数据集 Environment,回溯窗口设置为 L=96L=96,预测长度设置为 H∈{48,96,192,336}H\in\{48,96,192,336\}。对于周周期数据集 Energy 和 Health (US),回溯窗口设置为 L=24L=24,预测长度设置为 H∈{12,24,36,48}H\in\{12,24,36,48\}。对于月周期数据集 Traffic、SocialGood、Security、Economy、Climate 和 Agriculture,回溯窗口设置为 L=8L=8,预测长度设置为 H∈{6,8,10,12}H\in\{6,8,10,12\}。这种设置使不同周期的数据集在与其时间尺度相匹配的历史上下文和预测范围下进行评估,从而保证实验比较的一致性与公平性。

如果想写得更像顶会实验设置部分,可以再压缩一点:

为保证公平比较,我们根据数据集的周期特性统一设置回溯窗口 LL 和预测长度 HH,具体配置如表 X 所示。对于日周期的 Environment,设置 L=96L=96 以及 H∈{48,96,192,336}H\in\{48,96,192,336\};对于周周期的 Energy 和 Health (US),设置 L=24L=24 以及 H∈{12,24,36,48}H\in\{12,24,36,48\};对于月周期的 Traffic、SocialGood、Security、Economy、Climate 和 Agriculture,设置 L=8L=8 以及 H∈{6,8,10,12}H\in\{6,8,10,12\}。

顶会论文英文版

Experimental Setup. To ensure a fair comparison across datasets, we configure the look-back window LL and forecasting horizon HH according to the underlying temporal periodicity of each dataset, as summarized in Table X. For the daily-periodic Environment dataset, we set L=96L=96 and evaluate forecasting horizons H∈{48,96,192,336}H\in\{48,96,192,336\}. For the weekly-periodic Energy and Health (US) datasets, we use L=24L=24 and H∈{12,24,36,48}H\in\{12,24,36,48\}. For the monthly-periodic Traffic, SocialGood, Security, Economy, Climate, and Agriculture datasets, we set L=8L=8 and H∈{6,8,10,12}H\in\{6,8,10,12\}. These period-specific configurations provide each dataset with an appropriate amount of historical context and forecasting range, thereby enabling consistent and fair evaluation across datasets with different temporal characteristics.

如果希望更简洁、更接近 NeurIPS/ICLR/ICML 常见写法,我更推荐这一版:

To ensure a fair comparison, we adopt period-specific look-back windows and forecasting horizons for all datasets, as summarized in Table X. Specifically, for the daily-periodic Environment dataset, we use L=96L=96 and H∈{48,96,192,336}H\in\{48,96,192,336\}. For the weekly-periodic Energy and Health (US) datasets, we set L=24L=24 and H∈{12,24,36,48}H\in\{12,24,36,48\}. For the monthly-periodic Traffic, SocialGood, Security, Economy, Climate, and Agriculture datasets, we use L=8L=8 and H∈{6,8,10,12}H\in\{6,8,10,12\}.

我更建议论文正文使用第二个英文版本,信息完整但没有过度解释。


类似你参考图风格的 LaTeX 表格

如果希望做成参考图那种“简洁、三线表、顶会风格”,可以直接写成:

latex
\begin{table}[t] \centering \caption{Dataset configurations under different temporal periodicities. $L$ denotes the look-back window and $H$ denotes the forecasting horizon.} \label{tab:dataset_setting} \setlength{\tabcolsep}{9pt} \renewcommand{\arraystretch}{1.08} \begin{tabular}{lccc} \toprule \textbf{Periodicity} & \textbf{Dataset} & \textbf{Look-back $L$} & \textbf{Forecasting Horizon $H$} \\ \midrule Daily & Environment & 96 & $\{48, 96, 192, 336\}$ \\ \midrule Weekly & Energy & 24 & $\{12, 24, 36, 48\}$ \\ & Health (US) & 24 & $\{12, 24, 36, 48\}$ \\ \midrule Monthly & Traffic & 8 & $\{6, 8, 10, 12\}$ \\ & SocialGood & 8 & $\{6, 8, 10, 12\}$ \\ & Security & 8 & $\{6, 8, 10, 12\}$ \\ & Economy & 8 & $\{6, 8, 10, 12\}$ \\ & Climate & 8 & $\{6, 8, 10, 12\}$ \\ & Agriculture & 8 & $\{6, 8, 10, 12\}$ \\ \bottomrule \end{tabular} \end{table}

如果想让表格更像你上传图片的形式,减少重复数字,更适合论文版面,我建议使用 \multirow:

latex
\begin{table}[t] \centering \caption{Forecasting configurations for datasets with different temporal periodicities. $L$ and $H$ denote the look-back window and forecasting horizon, respectively.} \label{tab:dataset_setting} \setlength{\tabcolsep}{8pt} \renewcommand{\arraystretch}{1.08} \begin{tabular}{llcc} \toprule \textbf{Periodicity} & \textbf{Datasets} & \textbf{$L$} & \textbf{$H$} \\ \midrule Daily & Environment & 96 & $\{48,96,192,336\}$ \\ Weekly & Energy, Health (US) & 24 & $\{12,24,36,48\}$ \\ Monthly & Traffic, SocialGood, Security, Economy, & 8 & $\{6,8,10,12\}$ \\ & Climate, Agriculture & & \\ \bottomrule \end{tabular} \end{table}

对于顶会论文,我会更推荐第二种紧凑表格:它不像第一种那样重复 LL 和 HH,占用空间更少,也和你给出的参考图那种“Setting + 多个实验配置”的排版风格更接近。

表题还可以进一步写成:

Table X: Dataset configurations for different temporal periodicities. LL and HH denote the look-back window and forecasting horizon, respectively.

如果是 NeurIPS / ICML / ICLR 风格,这套表述已经比较自然。

Chia sẻ Q&A này