Multi-Task Hybrid Deep Learning untuk Predictive Business Process Monitoring Berdasarkan Log Data dan Social Network

Timbulong, Yehezkiella Felicia Jeis (2026) Multi-Task Hybrid Deep Learning untuk Predictive Business Process Monitoring Berdasarkan Log Data dan Social Network. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221007-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (8MB) | Request a copy

Abstract

Predictive Business Process Monitoring (PBPM) dalam penelitian ini didefinisikan sebagai rumpun metode pemantauan proses bisnis prediktif yang memanfaatkan rekaman jejak kejadian historis (event log) pada sistem informasi untuk memproyeksikan kelanjutan langkah operasional dan alokasi kerja di masa depan secara proaktif. Kendala pada pendekatan PBPM existing saat ini adalah sifat pembelajarannya yang masih berfokus tunggal pada urutan aktivitas (activity-centric bias), sehingga belum melibatkan karakteristik social network, seperti pola serah terima tugas (handover of work) antar-sumber daya (resource) manusia dalam memprediksi aktivitas maupun pekerja selanjutnya. Untuk mengatasi batasan tersebut, penelitian ini mengusulkan metode Multi-Task Hybrid Deep Learning. Multi-Task pada penelitian ini adalah aktivitas dan social network. Hybrid Deep Learning pada penelitian ini menggabungkan dua metode ekstraksi fitur, yaitu Long Short-Term Memory (LSTM) untuk mengekstraksi fitur pola sekuensial dari event log, dan LSTM untuk mengekstraksi fitur topologi struktural dari social network. Hasil dari metode Multi-Task Hybrid Deep Learning adalah aktivitas selanjutnya (next-activity) dan sumber daya selanjutnya (next-resource). Evaluasi kinerja dilakukan secara head-to-head dengan membandingkan metode Multi-Task Hybrid Deep Learning terhadap Random Forest, Markov Model, dan Most Frequent menggunakan dataset BPI Challenge 2012 dan BPI Challenge 2019. Pada pengujian dataset BPI Challenge 2012, metode usulan mencapai rata-rata akurasi aktivitas sebesar 85,6% (meningkat 6,8% dari Random Forest, 23,9% dari Markov Model, dan 74,2% dari Most Frequent) dan rata-rata akurasi resource sebesar 74,9% (meningkat 38% dari Random Forest dan 81,8% dari Most Frequent). Selanjutnya, pada pengujian dataset BPI Challenge 2019, model menghasilkan akurasi aktivitas sebesar 74,3% (meningkat 4,6% dari Random Forest, 20,7% dari Markov Model, dan 59,3% dari Most Frequent) dan akurasi resource sebesar 38,7% (meningkat 21,9% dari Random Forest, 7% dari Markov Model, dan 69,1% dari Most Frequent).
=====================================================================================================================================
Predictive Business Process Monitoring (PBPM) is defined in this study as a set of methods for predictive business process monitoring that utilizes historical event logs in information systems to proactively project the continuation of operational steps and future work allocation. A limitation of current PBPM approaches is that their learning is still solely focused on the sequence of activities (activity-centric bias), and thus does not yet incorporate social network characteristics—such as patterns of work handover between human resources—in predicting subsequent activities or workers. To address these limitations, this study proposes a Multi-Task Hybrid Deep Learning method. In this study, “multi-task” refers to activities and social networks. The Hybrid Deep Learning approach combines two feature extraction methods: Long Short-Term Memory (LSTM) to extract sequential pattern features from event logs, and LSTM to extract structural topological features from social networks. The output of the Multi-Task Hybrid Deep Learning method is the next activity and the next resource. Performance evaluation was conducted in a head-to-head comparison by evaluating the Multi-Task Hybrid Deep Learning method against Random Forest, the Markov Model, and the Most Frequent method using the BPI Challenge 2012 and BPI Challenge 2019 datasets. In testing on the BPI Challenge 2012 dataset, the proposed method achieved an average activity accuracy of 85.6% (an improvement of 6.8% over Random Forest, 23.9% over the Markov Model, and 74.2% over the Most Frequent method) and an average resource accuracy of 74.9% (an improvement of 38% over Random Forest and 81.8% over the Most Frequent method). Furthermore, when tested on the BPI Challenge 2019 dataset, the model achieved an activity accuracy of 74.3% (an improvement of 4.6% over Random Forest, 20.7% over the Markov Model, and 59.3% over the Most Frequent method) and a resource accuracy of 38.7% (an improvement of 21.9% over Random Forest, 7% over the Markov Model, and 69.1% over the Most Frequent method).

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Deep Learning, Event Log, Multi-Task Learning, Predictive Business Process Monitoring, Social Network Analysis
Subjects:	H Social Sciences > HD Industries. Land use. Labor > HD58.87 Reengineering (Management) Business process T Technology > T Technology (General) > T57.74 Linear programming
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Yehezkiella Felicia Jeis Timbulong
Date Deposited:	01 Aug 2026 04:01
Last Modified:	01 Aug 2026 04:01
URI:	http://repository.its.ac.id/id/eprint/141606

Actions (login required)

View Item