Mitigasi Catastrophic Forgetting pada Sequential Multi-Source Domain Adaptation Untuk Analisis Sentimen Menggunakan Metode Experience Replay

SANAD, SYARIF (2026) Mitigasi Catastrophic Forgetting pada Sequential Multi-Source Domain Adaptation Untuk Analisis Sentimen Menggunakan Metode Experience Replay. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221257-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Analisis sentimen lintas domain menghadapi tantangan domain shift yang semakin kompleks ketika data domain baru tersedia secara bertahap, bukan serentak. Metode Weighting Scheme-based Unsupervised Domain Adaptation (WS-UDA) yang ada saat ini mengasumsikan ketersediaan semua data secara simultan, sehingga tidak dapat diterapkan langsung pada skenario sekuensial tanpa mengalami catastrophic forgetting. Penelitian ini mengintegrasikan mekanisme Experience Replay ke dalam arsitektur WS-UDA yang diadaptasi untuk pelatihan sekuensial, mencakup penggantian discriminator multi-kelas menjadi discriminator biner dan penambahan memory buffer berbasis per-domain reservoir sampling. Eksperimen pada Amazon Review Dataset dan FDU-MTL Dataset menunjukkan bahwa metode usulan berhasil menurunkan forgetting rate sebesar 58,8% pada Amazon (4,90% → 2,02%) dan 30,2% pada FDU-MTL (1,125% → 0,786%) dibandingkan Naive Sequential, dengan buffer hanya membutuhkan 8,33% dan 3,12% dari total data masing-masing. Pada FDU-MTL, Average Source Accuracy metode usulan (87,92%) bahkan melampaui Oracle joint training (86,37%) dan paper WS-UDA asli (87,10%). Hasil ini membuktikan bahwa Experience Replay efektif memitigasi catastrophic forgetting pada Sequential Multi-Source Domain Adaptation sekaligus efisien secara komputasi.
========================================================================================================================================
Sentiment analysis across domains faces increasingly complex domain shift challenges when new domain data arrives incrementally rather than simultaneously. Existing Weighting Scheme-based Unsupervised Domain Adaptation (WS-UDA) methods assume simultaneous data availability, making them inapplicable to sequential scenarios without catastrophic forgetting. This study integrates an Experience Replay mechanism into a WS-UDA architecture adapted for sequential training, including replacing the multi-class discriminator with a binary discriminator and adding a per-domain reservoir sampling memory buffer. Experiments on the Amazon Review Dataset and FDU-MTL Dataset show that the proposed method reduces the forgetting rate by 58.8% on Amazon (4.90% → 2.02%) and 30.2% on FDU-MTL (1.125% → 0.786%) compared to Naive Sequential, with buffers requiring only 8.33% and 3.12% of the total data, respectively. On FDU-MTL, the Average Source Accuracy of the proposed method (87.92%) even surpasses Oracle joint training (86.37%) and the original WS-UDA paper (87.10%). These results demonstrate that Experience Replay effectively mitigates catastrophic forgetting in Sequential Multi-Source Domain Adaptation while remaining computationally efficient.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Analisis Sentimen, Catastrophic Forgetting, Experience Replay, Sequential Domain Adaptation, Weighting Scheme UDA, Sentiment Analysis, Catastrophic Forgetting, Experience Replay, Sequential Domain Adaptation, Weighting Scheme UDA
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering
Depositing User:	Syarif Sanad
Date Deposited:	30 Jul 2026 07:09
Last Modified:	30 Jul 2026 07:09
URI:	http://repository.its.ac.id/id/eprint/139843

Actions (login required)

View Item