Predicting LMS login frequency using transformer- and LLM-based time series models

초록

This study introduces two categories of time series models: Transformer-based models (Transformer, Reformer, Informer, and Autoformer) and LLM-based models (Promptcast, LLMTime, and Time-LLM). They have been actively researched as solutions to address the long-term dependency issues and computational costs arising from the lack of data parallelization in traditional time series models. These models are applied to predict login frequency on platforms used by universities for academic credit exchange, enabling a comparative analysis of their predictive performance. The results indicate that Transformer-based models generally achieve superior forecasting performance compared to LLM-based models; however, they also require significantly longer computational time. Specifically, the LLM-based models were implemented using GPT-3.5 and GPT-4o.

키워드