Bukhsh, Saeed, Dijkman 2021 — ProcessTransformer

Landmark paper bringing Transformer architectures to PPM. Arguably the most cited Transformer-PPM reference.

Contribution

Transformer encoder stack applied directly to event sequences (no recurrent component).
Learned activity embeddings replace one-hot encoding.
Self-attention captures long-range dependencies.
Evaluated on nine public logs; ~80%+ next-activity accuracy on average.

Limitations (noted by authors)

Remaining-time / event-time performance is "competitive" but not clearly superior to LSTM baselines — contrasted with the clear next-activity win.
Positional encoding is standard sinusoidal; process-specific positional biases remain unexplored.

Significance

Opened the Transformer-PPM era (2021+), extended by MTLFormer (sources/2023-wang-mtlformer-multitask-transformer) and countless subsequent variants. Cited as baseline in nearly every post-2021 PPM paper.

Connections

Concepts: concepts/transformer-ppm · concepts/next-activity-prediction · concepts/remaining-time-prediction Related sources: sources/2017-tax-lstm-process-prediction (predecessor) · sources/2023-wang-mtlformer-multitask-transformer (successor)