David vs. Goliath in Next Activity Prediction: Argmax vs. LSTM, Transformer, and LLM
A new paper on arXiv benchmarks the performance of various models for Next Activity Prediction (NAP), a key component of predictive process monitoring. The study compares large language models (LLMs), Transformers, LSTMs, and a simple argmax baseline across seven real-life event logs. Surprisingly, the results indicate that pretraining and model size do not consistently improve performance, and the basic argmax baseline often matches or rivals that of billion-parameter LLMs. AI