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..
Adaptive Time Encoding for Irregular Multivariate Time-Series Classification
Authors: Sangho Lee, Kyeongseo Min, Youngdoo Son, Hyungrok Do
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our method achieves state-of-the-art performance with high computational efficiency in irregular multivariate time-series classification tasks. |
| Researcher Affiliation | Academia | 1School of Industrial and Systems Engineering, Gyeongsang National University 2Department of Industrial and Systems Engineering, Dongguk University-Seoul 3Department of Population Health, NYU Grossman School of Medicine |
| Pseudocode | Yes | In Algorithm A1, we present a pseudo-code of our method to describe its overall learning procedure. |
| Open Source Code | Yes | The code for reproducing our experimental results is available on Git Hub at https://github.com/shlee-labs/ATENet. |
| Open Datasets | Yes | P12 (Physio Net Mortality Prediction Challenge 2012) [7], P19 (Physio Net Sepsis Early Prediction Challenge 2019) [27], PAM (PAMAP2 Physical Activity Monitoring) [25] |
| Dataset Splits | Yes | All datasets were split into training (80%), validation (10%), and test (10%) sets. |
| Hardware Specification | Yes | All experiments were executed on a Py Torch platform using an Intel Core i9-10900X at 3.70 GHz CPU, 256 GB RAM, and an NVIDIA Ge Force RTX 4090 24 GB GPU. |
| Software Dependencies | No | All experiments were executed on a Py Torch platform using an Intel Core i9-10900X at 3.70 GHz CPU, 256 GB RAM, and an NVIDIA Ge Force RTX 4090 24 GB GPU. |
| Experiment Setup | Yes | In ATENet, the encoder 𝑓maps an irregular multivariate time series X = (𝒕, 𝒙) with 𝑉variables to 𝑉-dimensional time embeddings at 𝐾reference time points. The encoder 𝑓consists of multi-head attention with 𝐻embedding functions (attention heads), each deriving a time embedding vector of size 𝐿for each observed time point. In this study, both 𝐾and 𝐿were set to 128. For the P12-M, P12-L, P19, and PAM datasets, we set 𝐻to 2, 4, 2, and 1, respectively, while using learnable, learnable, sinusoidal, and learnable embedding functions in that order. The subsequent classifier 𝑔is constructed by a GRU followed by two fully connected layers, where the first layer includes batch normalization and a Ge LU activation function, and the second one uses a softmax as an activation function. The GRU dimension was set to 32, and the two fully connected layers had dimensions of 32 and 𝐶, respectively, where 𝐶is the number of classes. Additionally, the scaling parameter 𝜖in Eq. (3) and the smoothing parameter 𝜂in Eq. (12) of the main text were set to 128 and 0.1, respectively. For model training, we set the batch size 𝐵to 128 and ran the model for 20 epochs. For the P12-M dataset, we used the Adam optimizer with an initial learning rate of 0.0001 and assigned weights of 0.01 and 0.1 to the temporal and intervariable consistency regularization terms, 𝛼and 𝛽, respectively. In the P12-L dataset, we used the Adam optimizer with an initial learning rate of 0.001 and set both 𝛼 and 𝛽to 0.01. In the P19 dataset, we also employed the Adam optimizer, but with an initial learning rate of 0.01, and adjusted 𝛼and 𝛽to 0.01 and 1, respectively. For the PAM dataset, we maintained a learning rate of 0.01, while setting 𝛼to 0.1 and 𝛽to 0.01. |