COSCO:A Sharpness Aware Training Framework for Few-Shot Multivariate Time Series Classification

Abstract

Time series classification is an important task with widespread applications in many different domains. Recently, deep neural networks have achieved state-of-the-art performance in time series classification, but often require large datasets and expert-labeled supervision for effective model training. In data scarcity situations, deep learning models often experience a significant decline. In this paper, we propose a training framework named COSCO to mitigate the decline in performance from a loss function and optimization perspective for the few-shot multivariate time series classification problem, where only a few samples are available. Particularly, we propose to apply a prototypical loss function and the sharpness-aware minimization (SAM) technique to enhance the generalization ability for deep learning classifier in few-shot multivariate time series classification problem. Such optimization techniques can be arbitrarily used in any deep learning models. We demonstrate the effectiveness of our method with ResNet backbone on few-shot benchmark datasets formed by UCR classification data.

Environment

To run this project offer the option of using a Jupiter Notebook COSCO.ipynb for plug and play .

You are also able to run locally by cloning the repository and installing the dependencies.

  git clone https://github.com/JRB9/COSCO.git
  cd COSCO
  pip install -r requirements.txt

Datasets

For plug and play we have provided the Full Datasets, 1-shot and 10-shot versions in this repository in the Dataset folder. These datsets were gathered from UCR Time Series Classification Archive . The list of multivariate datasets are as follows:

ArticularyWordRecognition, BasicMotions, CharacterTrajectories, EigenWorms, Epilepsy, EthanolConcentration, FaceDetection, FingerMovements, HandMovementDirection, Heartbeat, JapaneseVowels, Libras, MotorImagery, NATOPS, PEMS-SF, PenDigits, RacketSports, SelfRegulationSCP1, SelfRegulationSCP2, SpokenArabicDigits, UWaveGestureLibrary. 

Reproducing

Using Jupiter Notebook

Run COSCO.ipynb in any Jupiter Notebook environment.

Running Locally

Executerun.py with the following configurable arguments:

python run.py

Argument Descriptions

• lr: Learning rate for the optimizer (e.g., 0.01).
• rho: Momentum parameter for optimization (e.g., 0.9).
• nEpoch: Number of epochs for training (e.g., 100).
• dataset: The name of the dataset to use (e.g., "BasicMotions").
• shot: Number of shots (support examples) per class (1 or 10).
• normalize: Whether to normalize the input data (True or False).
• model: The type of baseline model to use ("resnet" or "tapnet").
• sam: Whether to use Sharpness-Aware Minimization (SAM) (True or False).
• optimizer: The optimizer type for training ("sgd" or "adam").
• prototypical_loss: Whether to apply prototypical loss during training (True or False).
• prototypical_loss_type: The type of prototypical loss to use ("neg", "sim", "cos", "negexp").
• save_dir: Directory path to save the output (e.g., "/content/classification_data/").
• save_name: File name for saving results (e.g., "results.csv").

Example Usage

python run.py --dataset BasicMotions --model resnet --lr 0.001 --rho 0.9 --nEpoch 100 --shot 1 --normalize False --sam True --optimizer adam --prototypical_loss True --prototypical_loss_type neg --save_dir /content/classification_data/ --save_name results.csv

Citation

Submitted & accepted for publication in the CIKM '24 conference.

If you use this code or our methods in your research, please cite our paper:

@inproceedings{barreda2024cosco,
  title={COSCO: A Sharpness-Aware Training Framework for Few-shot Multivariate Time Series Classification},
  author={Barreda, Jesus and Gomez, Ashley and Puga, Ruben and Zhou, Kaixiong and Zhang, Li},
  booktitle={Proceedings of the 33rd ACM International Conference on Information and Knowledge Management},
  pages={3622--3626},
  year={2024}
}