Technical inquiry regarding Latent TSCI implementation #1
Description
Dear Dan Waxman,
I noticed you are the primary contributor to the latent TSCI modules on GitHub, I thought it best to consult you directly regarding some implementation nuances. I find the framework you’ve implemented remarkably elegant and technically inspiring.
Specifically, I encountered a consistency issue when using the Coupled_Double_pendulum_sample function. I observed that:
*Sampling Consistency: Unless multiple_sample_rate is set to 1 and the random indices are generated outside the data-generation loop, getpath.py struggles to run successfully due to mismatched Trajectory IDs across reconstructed datasets.
*TSCI Computation: My understanding is that the TSCI score should be averaged across trajectories with matching IDs. However, I found that enforcing ID consistency (by taking intersections or filling missing IDs) significantly degrades performance. The screenshot below illustrates the results I obtained using concatenated trajectories.
*Model Performance: Despite tuning the GRU-ODE hyperparameters and achieving a training loss around 0.1, the absolute TSCI scores (Pearson/Cosine) in my current setup remain lower than those of CCM, although TSCI shows a more robust capability in identifying the correct causal direction.
I am now concerned that, in an effort to save storage space, I may have reduced the size of the generated dataset and the reconstructed files, which could have hindered the model’s ability to learn properly. As I am unsure whether this configuration is viable, I would greatly appreciate your guidance.
Given your deep involvement in the implementation, would you be willing to share the specific source code or the exact hyperparameter configurations used for the Section 3.2 experiments? Seeing the full training pipeline for the latent variable models would be immensely helpful for my research at USTC.
I assure you that any shared materials will be used strictly for academic purposes, and your work will be fully and properly cited in any future publications.
Thank you very much for your time and for your significant contributions to the community. I look forward to the possibility of learning from your implementation.