Rama-Maneiro, Vidal, Lama 2020 — Deep Learning PPM Review & Benchmark

Hub survey/benchmark paper for deep-learning PPM. Combines a systematic literature review (SLR; selection criteria, architecture analysis) with an empirical benchmark of 10 concrete approaches on 12 public logs.

Why this matters

Provides a coherent taxonomy of DL-PPM methods: by prediction target (next-activity / remaining-time / outcome), architecture (LSTM / GRU / CNN / attention / ensemble), and encoding strategy.
The benchmark is one of the primary references for choosing a PPM approach.
Uses the same public logs (BPI Challenges, Helpdesk, Sepsis) as the community standard — results are directly comparable.

Bridges

Together with sources/2016-teinemaa-outcome-ppm-review (outcome-PPM review) and sources/2019-verenich-survey-ppm (remaining-time survey), this is part of the survey triad a PPM researcher should start from.

Cited by

sources/2023-riess-temporal-loss-remaining-cycle-time — Riess cites Rama-Maneiro et al. as the canonical DL-PPM benchmark reference when motivating the need for a third evaluation axis (temporal consistency) beyond accuracy and earliness.
sources/2023-riess-phd-thesis-ppm — used in the thesis background as the primary DL-PPM survey.

External-validity caveat

The 12-log benchmark is the community standard but constitutes a closed sample — cross-log transfer is not evaluated, so benchmark-leading results do not guarantee performance on a specific organisation's process. The external-validity critique Cartwright & Hardie apply to policy RCTs ("it worked there, but will it work here?"; see sources/2023-anjum-rocca-phi403-lecture-11-is-more-data-better and sources/2023-anjum-rocca-phi403-lecture-19-what-rcts-do-not-show) applies directly to PPM benchmarking.

Connections

Concepts: concepts/predictive-process-monitoring · concepts/lstm-ppm · concepts/transformer-ppm · concepts/trace-encoding · concepts/rct-limitations