Nolle et al. 2019 — BINet: Multi-perspective Anomaly Classification

Extension of Nolle et al.'s 2018 BINet with three architectural variants, automated threshold heuristics, and a rule-based classifier over BINet outputs. Anomaly detection, not strictly prediction, but uses the same RNN-based trace modelling as neural PPM.

Contribution

Multi-perspective RNN — jointly models control-flow and data attributes.
Attribute-level detection — the anomaly is localised to the specific attribute value, not just the case.
Runtime deployability — can train during execution, adapts to concept drift.
Five assumptions it operates under: no domain knowledge, no clean dataset, no reference model, no labels, no manual threshold — i.e., truly unsupervised.

Significance

Bridges PPM and anomaly/conformance-adjacent work. Relevant for APM's self-modification (specifically the "detect misbehaviour" requirement in adaptation triggers) and the benchmark-contamination challenge (C3) in sources/2026-calvanese-agentic-bpm-manifesto.

Connections

Concepts: concepts/predictive-process-monitoring · concepts/self-modification Authors: entities/timo-nolle