---
title: "MTLFormer: Multi-Task Learning Guided Transformer Network for Business Process Prediction"
type: source
tags: [ppm, transformer, multi-task, deep-learning]
authors: [Wang Jiaojiao; Huang Jiawei; Ma Xiaoyu; Li Zhongjin; Wang Yaqi; Yu Dingguo]
year: 2023
venue: "IEEE Access 11: 76722–..."
kind: paper
raw_path: "raw/Predictive process monitoring/MTLFormer_Multi-Task_Learning_Guided_Transformer_Network_for_Business_Process_Prediction.pdf"
created: 2026-04-13
updated: 2026-04-13
key_claims:
  - Training separate prediction models per task (next activity, execution time, remaining time) leads to high compute cost and ignores inter-task relationships.
  - MTLFormer — Transformer with multi-task parallel training and shared representation — reduces compute and improves per-task accuracy.
  - Self-attention captures long-distance dependencies better than LSTMs on complex process behaviour.
  - Evaluated on four real-life logs.
---

# Wang et al. 2023 — MTLFormer: Multi-Task Transformer for PPM

**Multi-task Transformer** for [[concepts/predictive-process-monitoring|PPM]]. Predicts next activity, execution time of the next activity, and remaining time jointly via shared representations — explicitly leveraging relationships between tasks.

## Contribution
- Transformer architecture with **multi-task parallel training** and hard parameter sharing.
- Claimed improvements on both efficiency (single model, three predictions) and accuracy over single-task baselines.
- Evaluated on four benchmark event logs.

## Position
Extends the [[concepts/transformer-ppm|Transformer-PPM]] line (ProcessTransformer 2021) with a multi-task learning (MTL) scaffold — reflecting the broader NLP trend of consolidating tasks into unified models.

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