Thanks for your valuable comments!

W1. Comparison with [1]

Thanks for pointing it out, and we will include a discussion and comparison of [1] in the revised version. Time Series Editing (TSE) proposed in our work is essentially different to [1], though these two works share some similarities in conditional generation.

Specifically, the term "constraint" mentioned in [1] is actually similar to the term "condition" or "attribute" used in our work and [2]. The task of [1] can be formulated as $\mathbf{x}^{tgt}=\Phi(\mathbf{a}^{tgt})$, where $\mathbf{a}^{tgt}$ is the target constraint, $\mathbf{x}^{tgt}$ is the desired output and $\Phi$ denotes the function of generative model to be learned. However, our proposed TSE is formulated as $\mathbf{x}^{tgt}=\Phi(\mathbf{x}^{src}, \mathbf{a}^{src}|\mathbf{a}^{tgt})$, which generates $\mathbf{x}^{tgt}$ based on the input source sample $(\mathbf{x}^{src}, \mathbf{a}^{src})$ and the target condition $\mathbf{a}^{tgt}$. Since [1] and [2] are both "conditional generation" studies, we compare with the more recent work [2] in our current version. We will discuss and compare with [1] in the revised verison.

W2. Averaged results

Thanks for your feedback.

Below we provide the experimental results of the average performance of editing different attributes for reference, hoping to help you better understand the performance of different methods. We also provide the results for all 6 edit-preserve attribute groups in the pdf of the general response.

In the following tabels we present the avergaed results of editing different attribute combinations on the Synthetic and Air datasets. The Synthetic dataset has 3 attribtues: trend type, trend direction and season cycle. The Air dataset has 2 attributes: city and season. The results show that our proposed method TEdit outperforms the baseline on all the evaluation metrics.

• Table 1. Average performance of editing different attributes on Synthetic.

• Table 2. Average performance of editing different attributes on Air.

W3. Multi-resolution Thank you for sharing these works, it will actually enhance the discussion of related works of our paper.

As mentioned by your comments, one work is almost concurrent with our paper and these works primarily focus on the forecasting task, aiming to make more accurate predictions. Our work, however, focuses on the editing task, which requires to generate the corresponding time series regarding the given editing attributes as condition. The reason we propose the multi-resolution approach is that different attributes impact the time series at different scales. For example, the trend type has a global impact, while the season number focuses on more local details. In our work, we consider both the multi-resolution aspects of the time series and attributes, and leverage multi-resolution mechanism on both modeling and generation, which is more reflective of reality. Thank you again for mentioning these related works. We will discuss them in the revised version of our paper.

[1] Coletta et al. On the constrained time-series generation problem. NeurIPS'2023

[2] Narasimhan, Sai Shankar, et al. Time weaver: A conditional time series generation model. ICML'2024

Thanks for your constructive feedback!

W1. Edited attribute set

Thanks for your comments.

In the tables below, we show the average experimental results of editing different types and numbers of attributes on the Synthetic dataset. It can be seen that our TEdit outperforms baseline method on all metrics. The subscripts "edit" and "prsv" refer to the edited and preserved attributes. For detailed results of each combination, please refer to the PDF in the General Response. We will add more results in the revised version.

• Table 1: averaged results of different edit-preserve attribute groups on Synthetic.

W2. Standard errors

For each setting, we run the experiment three times, and calculate the mean and standard deviation (STD). These results can be seen in the PDF file in the General Response. From the results, the improvement of our method against the baseline is significantly larger than the STD values in almost all the settings, which illustrates that our method's improvements are significant. Moreover, the STD values of our method is much smaller than the baseline, showing that our method has better stability.

W3. RaTS

Briefly speaking, we follow [1] and leverage an external model TAP (Time series - Attribute Pretraining) similar to CLIP [2] to estimate this probability. The details of TAP are presented in Appendix G.

$p(a_{k}^{tgt}|x)$ is calculated in the following steps.

• Use TAP to encode time series $\mathbf{x}$ and the $N_k$ possible values, e.g., linear and exponential, of the $k$-th attribute $a_k$, e.g., trend type, into embeddings $\mathbf{h}_{x}$ , and $\mathbf{h}\_{a_k}^{n}$, where $n\in[1,..., N_k]$.
• Obtain $p(a_k^n|x)$ by applying a softmax over the cosine similarity between $\mathbf{h}_{x}$ and $\mathbf{h}\_{a_k}^{n}$.

W4. MSE for real data

MSE is indeed a straightforward metric for measuring the quality of generated samples. However, MSE can be used only when source and target samples share all the attributes, including hidden attributes such as noise, except for the edited attributes. We can guarantee such a 1-1 pair for synthetic data. However, in real-world datasets, there are many unseen attributes and interference which make the time series vary a lot even with the same attribute combination. For example, air quality is not only influenced by city and season, but also other factors such as weather, wind and the industrialization level. We may not say that the edited time series by modifying city attribute from "Boston" to "New York" is exactly the same as the specific data sample with city of New York in the dataset.

For the real-world datasets, we follow the practices in computer vision domain [1], by utilizing external models to evaluate the quality of the edited samples. We acknowledge that evaluating generative models is challenging, especially for the newly proposed time series editing task. We will keep investigating better benchmarking solutions in this novel direction.

Q1. Correlated attributes

We also had similar concerns at the early stage of experiments. However, after statistically visualizing the data generated by bootstrapping, we found that the data generated by bootstrapping is a more complete complement to the original real data distribution, as shown in Fig. 5 in the paper. Training on both the real and the bootstrapped data makes the model fit the complete distribution and derive better performance, as illustrated by our experiments. Therefore, there is almost no overfitting outside the real distribution.

Q2. Distributional coverage

Thanks for your comments and sorry for the misleading description.

We use all the data, including both real and bootstrapped data, for T-SNE. Thus, the three subfigures of Fig. 5 share the same T-SNE model with the same projection, which means their data cohorts are the same. So it is reasonable to comapre the distribution coverage of the original real dataset, bootstrapped synthetic dataset, and the full dataset of the above two sets, in the three subfigures. We will further clarify this in the revised version.

Q3. Unseen attributes

Our understanding of the "unseen attribute values" is "unseen attribute combinations". In other words, the model has seen all $N_k$ values of each individual attribute $a_k$, but has not seen certain attribute combinations. This setting is the same as our synthetic datasets. There are 4 unseen attribute combinations in the test data. Therefore, our proposed TEdit could deal with the unseen attribute combinations. In the future, we plan to further explore this interesting setting for real-world datasets.

L1. Evaluating generated time series samples

Thanks for your insightful comments!

Assessing the reality of the generated data has always been a critical open challenge. This is extremely challenging even for images, where attributes such as color and size can be easily interpreted, not to mention time series, where the data is difficult to be interpreted. The classic work [3] asked humans to evaluate the reality of the generated images. We follow the practice of recent works [1] to use external models as evaluation. Nevertheless, assessing the reality can be an interesting future direction. We will further discuss it in the revised version.

[1] Li et al. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. NeurIPS'2023.

[2] Radford et al. Learning transferable visual models from natural language supervision. ICML'2021.

[3] Meng et al. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. ICLR'2022

Thanks for your appreciation and feedbacks! Below are the responses.

W1. Handling of diverse attribute types and their interactions.

Thanks for your comments.

The following tables 1~6 show the complete experimental results of editing different combination of attributes (trend type, trend direction and the number of season cycles), hoping to provide a better understanding of our method. Due to time limitation, in the following tables, we present the results of TEdit-TW and Time Weaver [1] on editing different sets of attributes on Synthetic dataset. From the tables, our TEdit-TW outperforms Time Weaver [1] on all attribute combinations. We also found that the difficulty of editing different attributes varies greatly. By comparing MSE, (1) editing season_cycle is the most difficult, editing trend_type is the easiest, and (2) editing multiple attributes is usually more difficult than a single one. For more detailed experimental results, please refer to the PDF file in the General Response.

• Table 1. Synthetic dataset: edit "trend type".

• Table 2. Synthetic dataset: edit "trend direction".

• Table 3. Synthetic dataset: edit "season cycle".

• Table 4. Synthetic dataset: edit "trend type" and "trend direction".

• Table 5. Synthetic dataset: edit "trend type" and "season cycle".

• Table 6. Synthetic dataset edit: "trend direction" and "season cycle".

W2. Reproducibility.

Thanks for your comments.

We will release our codes upon the acceptance of this paper, and we will figure out a better way to present the implementation. In the current version of our paper, we have listed the implementation details and training details in the appendix. Specifically, Appendix I provides the training configurations. Appendix H shows the architectures of our model. Appendix E and F present evaluation model and metrics. Moreover, Appendix C and D illustrate the optimization details and training algorithm. Appendix E discusses the data processing details. Appendix F explains the evaluations. We will refine the discussion of the above implementation details for better reproducibility.

W3. Writing suggestions.

Sorry for the inconvenience. We will make more rounds of proofreading and revise the details.

[1] Narasimhan et al. Time weaver: A conditional time series generation model. ICML'2024

Thanks for your appreciation and comments!

W1. Suggested new datasets.

Thanks for your suggestion. The datasets you mentioned are indeed related and useful, revealing a wider application space for our work.

We have looked into these datasets, many of which are EEG time series with various attributes, such as subject and task information. For example, the SEED dataset used in [1] contains attributes such as EEG channels and subjects, which is similar to the Motor dataset (Sec. 4.1 and Appendix E) used in our paper. Potential applications of these datasets include completing time series of missing channels by editing time series of other channels. Recall that, time series editing is a relatively new task, and there is still much to explore. We plan to further investigate the evaluation benchmark of Time Series Editing (TSE) task and will use a broader range of datasets in future research.

W2. Dataset size, synthetic data details and baselines.

Thanks for your comments.

1. For the dataset size, we believe the size should not be considered "too small" since the datasets we used and many datasets used in recent studies, e.g., time series generation [2] and LLM based time series model [3], have approximately the same scale. In total, our Synthetic-1/Synthetic-2/Motor/Air have 10,660/13,680/12,800/6,850 samples. Many datasets used in [2] and [3] range between 10,000~30,000. Collecting larger datasets will be an important future work for our paper and the proposed TSE task, and we will keep gathering more datasets with larger scales.

2. For the synthetic dataset, due to space limitation in the main paper, its details are presented in Appendix E.1. Briefly speaking, the synthetic samples we constructed consist of four parts: trend, season, noise and bias.

• The trend has two attributes: trend types (linear, quadratic, exponential, logistic) and trend directions (up, down).
• The attribute of season is the number of cycles, with possible values of 0, 1, 2, 4.
• The noise is a combination of Gaussian noise and high-frequency noise.
• The bias of each sample is randomly sampled from a uniform distribution.

3. Regarding the baselines, we are exploring a brand new field, and to the best of our knowledge, this is the first work focusing on developing methods and evaluation benchmarks for TSE. There are only few related works, and adapting the models from other tasks to this one involves significant efforts. The baselines we selected are Time Weaver [2], the latest work of conditional time series generation, and CSDI [4], a classic time series imputation framework. Please note that the latest work [2] only compared with 2 traditional GAN-based methods due to the lack of proper baselines in conditional time series generation. We believe [2][4] effectively reflect the current state-of-the-art in related fields. However, we will consider adding more comparative analyses in the future. Thank you for the suggestion.

W3. Presentation suggestions.

Thank you for your suggestions and we will update the revised version accordingly for better readability.

1. About time series editing (TSE) and image editing.

1.1. Comparison of TSE and image editing.

Broadly speaking, TSE shares the same target with image editing: edit the given data according to the specified conditions. However, time series data is quite different from image data, and there are unique challenges for TSE.

• Time series has attributes such as recoding location of city, which are more difficult for model to interpret than the attributes such as color in image.
• Time series data has its own unique properties, e.g., variant seasonality and low signal-to-noise ratio.
• Moreover, the data of the time series corresponding to some specific conditions might be rare, making generation and editing quite challenging.

Note that, we propose to use multi-resolution modeling and generation with bootstrapping mechanism trying to resolve the problem of aforementioned problems.

1.2 Definition presented in Section 3.1

Though sharing similar definition of image editing task, in Section 3.1, we formally formulate TSE task mainly because editing is a relatively new concept for the time series research community. A formal problem formulation could help readers to easily understand the task and follow the paper.

2. The necessity of diffusion.

TSE is a kind of generation task, which requires powerful generative models. There are two considerations of using diffusion models. (1) Diffusion is the latest generation technique and some recent works [2] have demonstrated that diffusion models can significantly outperform traditional methods like Generative Adversarial Network (GAN). (2) To keep fair comparison to the latest conditional time series generation work [2], we leverage the backbone model and base generation algorithm as the same as the baseline works [2][4].

3. BYOL and SimCLR

The work BYOL you mentioned bootstraps the output of a neural network to serve as targets for a learning network to approximate. It only bootstraps on the learning targets and it is different from our method that utilizes the learned generative model to generate the samples of both input and targets for bootstrapping the learning of the target generative model. SimCLR adopts different data augmentation to build the target representation for contrastive learning, which does not leverage bootstrapping. We will add more discussions to other bootstrapping works to enhance the presentation of our work.

[1] Yi et al. Learning topology-agnostic eeg representations with geometry-aware modeling. NeurIPS'2023

[2] Narasimhan et al. Time weaver: A conditional time series generation model. ICML'2024

[2] Jin et al, Time-LLM: Time Series Forecasting by Reprogramming Large Language Models. ICLR'2024

[4] Tashiro et al. CSDI: Conditional score-based diffusion models for probabilistic time series imputation. NeurIPS'2021

Overall I like the work, especially the possibilities it brings to the field. I'd suggest the authors include more discussions during revision, especially regarding the applications that I suggested to encourage future follow-ups. The rebuttals also provided some other evidence for me to further advocate for the work and thus I will raise my score.