Tax et al. 2017 — Predictive Business Process Monitoring with LSTM Neural Networks

Seminal LSTM-PPM paper; published at CAiSE 2017. Often cited as the reference point for neural PPM.

Contribution

A single LSTM architecture handles multiple prediction tasks via different output heads:
1. Next activity + its timestamp.
2. Case continuation (suffix) — recursive next-activity prediction.
3. Remaining cycle time — derived from continuation.
Empirical comparison against tailor-made baselines (transition systems, non-parametric regression, stochastic Petri nets) on four real-life logs. LSTM wins consistently.
Addresses the "tailor-made and non-generalisable" critique of prior PPM techniques — accuracy of classical methods is highly dataset- and prefix-length-sensitive.

Input: one-hot activity + relative-time features (time-since-previous-event, time-in-case) per step.
Single or stacked LSTM layer.
Output heads: softmax (next activity) + linear (timestamp).
Greedy recursive decoding for suffix generation.

Outperforms Evermann et al. 2017 (sources/2017-evermann-deep-learning-runtime) for next-activity.
Outperforms van der Aalst et al. annotated-transition-system for remaining time.
Outperforms Polato et al. for the same task.

Predecessor: Evermann et al. 2017 — first application of LSTM to PPM (single task only).
Successors: Camargo et al. 2019 (data-aware LSTM), Navarin et al. 2017 (data-aware remaining time), ProcessTransformer 2021 (sources/2021-bukhsh-processtransformer).

sources/2023-riess-temporal-loss-remaining-cycle-time — Riess uses the Tax et al. LSTM-prefix-log approach as the baseline architecture for his temporal-loss experiments.
sources/2024-riess-synbps-simulation-framework — Tax et al.'s qualitative finding about repeated-activity sequences degrading LSTM performance is the motivating example for SynBPS's controlled-variable simulation framework.