LSTM / RNN for PPM

Recurrent neural network approaches — particularly LSTM (Long Short-Term Memory) and GRU — became the dominant paradigm in PPM after Tax et al. 2017.

Why RNNs for PPM

Naturally handle variable-length sequences (no padding-to-max needed at inference).
Learn activity embeddings jointly with the prediction task.
Capture long-range dependencies better than N-gram Markov models.
Easily extended to multi-task learning — joint next-activity + timestamp prediction.

Canonical architectures

Single-task LSTM (Tax et al. 2017 — sources/2017-tax-lstm-process-prediction)

Input: one-hot activity + relative time features per step.
Single LSTM layer → softmax output for next-activity OR regression head for remaining time.

Multi-task LSTM

Shared LSTM trunk, separate heads for next-activity + remaining-time.
Regularises via auxiliary task.

Data-aware LSTM (Navarin et al., Camargo, Pasquadibisceglie)

Event attributes (categorical → embeddings; numerical → normalized) concatenated to activity features before LSTM.
Outperforms control-flow-only models on remaining-time regression.

Stacked / bidirectional variants

Marginal gains; rarely justify the training cost on typical BPM logs.

Limitations

Sequential training — slow compared to Transformers.
Vanishing gradients on very long traces.
Hyperparameter-sensitive; hyperparameter optimisation (NSGA-II, genetic algorithms) non-trivially improves results (Di Francescomarino et al., Senderovich et al.).
Loss-function sensitivity — unweighted MAE is standard but does not optimise for earliness; temporally weighted L1 variants (exponential / power / moderate decay) can improve earliness at the cost of temporal consistency (sources/2023-riess-temporal-loss-remaining-cycle-time).

Superseded by

concepts/transformer-ppm (2021+) for long-range dependency modelling and training parallelism, though LSTM remains a competitive baseline.

concepts/predictive-process-monitoring · concepts/transformer-ppm · concepts/trace-encoding