Towards Diverse Perspective Learning with Selection over Multiple Temporal Poolings (AAAI24)

• This is the author code implements "Towards Diverse Perspective Learning with Selection over Multiple Temporal Poolings," a paper accepted at AAAI 2024.
• It builds upon the official code of DTP github and softDTW github based on PyTorch.
• For further details, please refer to the original DTP and softDTW papers.

Overview

In Time Series Classification (TSC), temporal pooling methods that consider sequential information have been proposed. However, we found that each temporal pooling has a distinct mechanism, and can perform better or worse depending on time series data. We term this fixed pooling mechanism a single perspective of temporal poolings. In this paper, we propose a novel temporal pooling method with diverse perspective learning: Selection over Multiple Temporal Poolings (SoM-TP).

• We investigate data dependency arising from distinct perspectives of existing temporal poolings.
• We propose SoM-TP, a new temporal pooling method that fully utilizes the diverse temporal pooling mechanisms through an MCL-inspired selection ensemble.
• We employ an attention mechanism to enable a non-iterative ensemble in a single classifier.
• We define DPLN and perspective loss as a regularizer to promote diverse pooling selection.

Running the codes

STEP 1. Download the benchmark datsets for time series classification

• The datasets can be downloaded form the UCR/UEA repository.
• Create a directory named "data" and store downloaded datasets within it.

STEP 2. Train the CNN classifier with various temporal poolings including SoM-TP

For traditional temporal poolings,

python main.py --model=ConvPool --pool=DTP

and for SoM-TP

python main.py --model=SoMTP

STEP 3. Run LRP (Layer-wise Relevance Propagation: XAI input attribution method)

For traditional temporal poolings,

python LRP.py --model=ConvPool --pool=DTP

and for SoM-TP

python LRP.py --model=SoMTP

SoM-TP performance

1. Comparison with traditional temporal poolings
2. Comparison with advanced TSC methods
3. SoM-TP dynamic selection
4. LRP comparison

Citation

@article{Seong_Kim_Choi_2024,
title={Towards Diverse Perspective Learning with Selection over Multiple Temporal Poolings}, volume={38},
url={https://ojs.aaai.org/index.php/AAAI/article/view/28743}, DOI={10.1609/aaai.v38i8.28743},
number={8},
journal={Proceedings of the AAAI Conference on Artificial Intelligence},
author={Seong, Jihyeon and Kim, Jungmin and Choi, Jaesik}, year={2024}, month={Mar.}, pages={8948-8956} }