---
title: "Attention Mechanism in Predictive Business Process Monitoring"
type: source
tags: [ppm, attention, lstm, deep-learning]
authors: [Jalayer Abdulrahman; Kahani Mohsen; Beheshti Amin; Pourmasoumi Asef; Motahari-Nezhad Hamid Reza]
year: 2020
venue: "IEEE EDOC 2020: 181–..."
kind: paper
raw_path: "raw/Predictive process monitoring/Attention 2020.pdf"
created: 2026-04-13
updated: 2026-04-13
key_claims:
  - LSTM hidden states suffer from mis-information and accuracy degradation when used to predict the next event.
  - Adding an attention mechanism over all hidden states (NMT-inspired) yields more accurate predictions.
  - Attention also offers interpretability of individual activities' contribution to the prediction.
---

# Jalayer et al. 2020 — Attention Mechanism in PPM

Applies **attention** (the NLP / Neural Machine Translation innovation) to LSTM-based [[concepts/predictive-process-monitoring|PPM]]. Represents the **transition from pure LSTM to attention-augmented architectures** that preceded full Transformer adoption ([[sources/2021-bukhsh-processtransformer]]).

## Contribution
- Augments the standard LSTM-PPM pipeline with an attention layer over all hidden states.
- Argues plain LSTM relies too heavily on the single last hidden state as context.
- Attention weights additionally act as a built-in interpretability signal.

## Connections
**Concepts:** [[concepts/lstm-ppm]] · [[concepts/transformer-ppm]] · [[concepts/next-activity-prediction]]