Publications

Conference Papers

Data Augmentation for Industrial Multivariate Time Series via a Spatial and Frequency Domain Knowledge GAN

Published in 2022 IEEE International Symposium on Advanced Control of Industrial Processes (AdCONIP), 2022

The study utilized an AutoEncoder to reduce the correlation between variables, enabling the selection of representative variables for examining mutual information among those with a correlation value exceeding 0.7. Additionally, wavelet decomposition was employed to transform the data from the time domain to the frequency domain, facilitating the extraction of the variance component. For evaluation, the GAN-TEST and maximum mean discrepancy (MMD) metrics were used, demonstrating that the difference between the original and generated data in MMD was below 0.1, with an accuracy difference in GAN-TEST of less than 10%.

Recommended citation: J. C. Lin and F. Yang, "Data Augmentation for Industrial Multivariate Time Series via a Spatial and Frequency Domain Knowledge GAN," 2022 IEEE International Symposium on Advanced Control of Industrial Processes (AdCONIP), Vancouver, BC, Canada, 2022, pp. 244-249, doi: 10.1109/AdCONIP55568.2022.9894177.
Download Paper

Workshop Papers

LLM-augmented Preference Learning from Natural Language

Published in 2024 Economics and Computation workshop, 2024

The study focused on using large language models (LLMs) to tackle the scarcity of preference data. By generating preference data and testing various LLMs, the research identified optimal prompts to enhance understanding of preferences in texts. LLama2 was used to condense extensive text, emphasizing preference detection, while BERT’s output, guided by instructive sentences, improved classification accuracy. Techniques like masking and segment embedding were also employed to aid in entity comparison.

Recommended citation: Kang, I., Ruan, S., Ho, T., Lin, J. C., Mohsin, F., Seneviratne, O., & Xia, L. (2023). LLM-augmented Preference Learning from Natural Language. arXiv preprint arXiv:2310.08523.
Download Paper

Generating Election Data Using Deep Generative Models

Published in AI4SG workshop at AAAI-23, 2023

The study involved generating election data using deep generative models by utilizing strict order complete (SOC) election data, which contains comprehensive information about elections. The SOC election data was converted into images to account for the relationships between the rankings of alternatives. A conditional Deep Convolutional GAN (cDCGAN) was employed for learning, and an edge-trimming penalty was proposed to prevent overfitting or underfitting. Additionally, the quality of the generated data was assessed using an Auto-Encoder reconstruction error method, and a memorization metric was applied to evaluate the occurrence of overfitting

Recommended citation: Lin, Jui Chien, et al. "Generating Election Data Using Deep Generative Models."
Download Paper

JuiChien (Ryan) Lin