Next-Activity Prediction

A PPM task: given a prefix of an event trace, predict the next activity the case will execute.

Formulation

Input: prefix ⟨e₁, e₂, …, eₖ⟩ of a running case.
Output: activity label eₖ₊₁ ∈ A (multi-class classification over activity vocabulary).
Training data: every prefix-continuation pair extracted from completed cases.

Canonical approaches

Probabilistic / Markov — transition-matrix models over activity sequences.
Classical ML — features extracted via concepts/trace-encoding; classifier (random forest, decision tree, logistic regression).
RNN / LSTM — sequence models with activity embeddings; dominant approach 2016–2020.
Transformer / attention — rising since 2020–2021.
Graph Neural Nets (GNN) — treat the log as a graph; 2023 onward.

Evaluation

Accuracy / top-k accuracy — did the predicted activity match the true next?
F1 (weighted / macro) — handles class imbalance across activities.
Brier score — if probabilistic outputs are needed.

Full-suffix variant

Extend the prediction recursively to produce the whole remaining activity sequence. Accumulates error quickly; often evaluated with Damerau-Levenshtein distance between predicted and actual suffix.

concepts/predictive-process-monitoring · concepts/remaining-time-prediction · concepts/trace-encoding