Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Functional Complexity-adaptive Temporal Tensor Decomposition
Authors: Panqi Chen, Lei Cheng, Jianlong Li, Weichang Li, Weiqing Liu, Jiang Bian, Shikai Fang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experiments on both synthetic and real-world datasets, we demonstrate that CATTE not only reveals the underlying ranks of functional temporal tensors but also significantly outperforms existing methods in prediction performance and robustness against noise. |
| Researcher Affiliation | Collaboration | 1College of Information Science and Electronic Engineering, Zhejiang University 2Zhejiang Provincial Key Laboratory of Multi-Modal Communication Networks and Intelligent Information Processing 3Microsoft Research Asia |
| Pseudocode | Yes | Algorithm 1: Training process of CATTE |
| Open Source Code | Yes | The code is available at https://github.com/Ocean STARLab/CATTE. |
| Open Datasets | Yes | Datasets: We examined CATTE on three real-world benchmark datasets. (1) CA traffic, lane-blocked records in California. (https://smoosavi.org/datasets/lstw; (2) Server Room, temperature logs of Poznan Supercomputing and Networking Center. (https://zenodo.org/record/3610078#%23.Y8SYt3b MJGi); (3) SSF, sound speed field measurements in the pacific ocean... (https: //ncss.hycom.org/thredds/ncss/grid/GLBy0.08/expt_93.0/ts3z/dataset.html). |
| Dataset Splits | Yes | We followed [28, 19] to randomly draw 80% of observed entries for training and the rest for testing. |
| Hardware Specification | Yes | We compared per-iteration runtimes of CATTE and baselines on a system with an NVIDIA RTX 4070 GPU, Intel i9-13900H CPU, 32 GB RAM, and 1 TB SSD (Table 7, Appendix B.5). |
| Software Dependencies | No | The CATTE was implemented with Py Torch [35] and torchdiffeq library (https://github. com/rtqichen/torchdiffeq). ... The CATTE was trained using Adam [36] optimizer with the learning rate set as 5e 3. Explanation: While PyTorch and torchdiffeq are mentioned, specific version numbers for these libraries are not provided in the text. Adam is an optimizer rather than a software dependency. |
| Experiment Setup | Yes | For CATTE , we set the ODE state dimension J = 10 and the initial number of components of the factor trajectories R = 10. ... In Figure 2, we showed the predictive trajectories of entry value indexed in different coordinates. The dotted line represents the ground truth and the full line represents the the predictive mean learned by our model. The cross symbols represent the training points. The shaded region represents the predictive uncertainty region. Appendix B.1: The CATTE was implemented with Py Torch [35] and torchdiffeq library (https://github. com/rtqichen/torchdiffeq). We employed a single hidden-layer neural network (NN) to parameterize the encoder. Additionally, we used two NNs, each with two hidden layers, for derivative learning and for parameterizing the decoder, respectively. Each layer in all networks contains 100 neurons. We set the dimension of Fourier feature M = 32, the ODE state J = 5 and the initial number of components of the latent factor trajectories R = 5. The CATTE was trained using Adam [36] optimizer with the learning rate set as 5e 3. The hyperparamters {a0 r, b0 r}R r=1, c0, d0 and initial values of learnable parameters {αr, βr}R r=1, ρ, σ2, ι are set to 1e 6 (so that all the initial posterior means of {λr}R r=1 equal 1). We ran 2000 epochs, which is sufficient for convergence. |