Summary
The paper presents the TIMBA model for multivariate time series imputation (MTSI). This work builds on the idea of using diffusion models to learn the latent distribution of time series data, enabling the generation of complete sequences from incomplete data. It also replaces the time-oriented Transformer model with a bidirectional Mamba model containing S6 blocks. A GNN-based model is used to represent the time series data, which serves as conditional variables for the diffusion models and the inputs for the Conditional Feature Extraction Module (CFEM). The Noise Estimation Module (NEM) is used to model the final noise. The paper claims that this approach enhances the spatiotemporal representation for the imputation task.
Overall, the paper highlights various techniques widely used in the time-series and machine learning communities. Although it attempts to combine these techniques to solve the imputation task, it is difficult to understand the utility of each component and the relationships between them. I believe a few components could be combined to perform imputation more effectively. Therefore, this study proposes an over-designed architecture that may lack sufficient rationale to justify each component.
Strengths
1. It explores some interesting technical combinations of existing models, such as bidirectional Mamba and S6, which could be beneficial for the imputation task as it requires bidirectional context.
2. The GNN model is used to capture spatial relationships between variables, which could complement temporal dependencies.
3. It examines the sensitivity to various levels of missing values in the imputation tasks.
Weaknesses
1. This paper introduces a variety of existing techniques, combining them into a new model, TIMBA. However, it is primarily a combination of existing techniques and offer little in terms of new insights or techniques.
2. The technical contribution is limited. For instance, the second and the third contribution are benchmark experiments and ablation studies, which are generally not considered as contribution.
3. Although this paper focuses on diffusion and bidirectional Mamba, many components appear unnecessary. In diffusion-based time series models, a typical setup includes only a backbone model and a diffusion model. For example, the foundational work TimeGrad uses RNN as the backbone model [1]. In the imputation task, the foundational work, CSDI, employs a single-layer transformer encoder as the backbone and demonstrates that bidirectional models generally outperform unidirectional ones. Therefore, this study would likely only need bidirectional Mamba as the backbone for imputation.
[1] Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting.
[2] CSDI: Conditional Score-based Diffusion Models for Probabilistic Time Series Imputation
4. Althought this paper proposes ablation studies as a contribution, however, the benefits of various blocks are not known. For example, the study believes that the bidirectional Mamba with S6 is a better design, while we do not know whether the S6 can improve performance. A complete ablation studies should be conducted for all involved components, helping justify the necessity of each component. For example, for your proposed Mamba architecture, it should include the ablation on Mamba with and without S6, unidirectional vs bidirectional.
5. Many blocks are not discusses and do not have clear rationale. For example, the Conditional Feature Extraction Module (CFEM) and Noise Estimation Modules (NEM) are not discussed sufficiently. It does not show the motivation behind these blocks. We also do not know why these models should be combined with other blocks. The authors may address the following questions, but are not limited to these: what is the specific purpose of the CFEM in this architecture? How does the NEM contribute to the overall performance of the model?
6. I find the use of GNN-based representation for time series confusing, as its necessity is unclear.Could you elaborate on why GNN-based representation is necessary for this task, and what specific advantages it provides over alternative approaches?
7. Although it is positioned as a diffusion model for imputation tasks, it only compares with two baselines in this category, CSDI and PriSTI. Most of the baselines are from traditional ML methods, which may not be suitable for comparison.