03983ntm a2200337 i 4500 MY-KuUP 20260805094658.0 t||||fr|||| 000 0 ta 260805t20262026my a|||fr|||| 000 0 eng d THE0010798 (Local) Hardback UMPSA eng UMPSA rda FK .B46 2026 r Thesis Bhowmik Abhijit, author. EDUAE-LSTM : Bhowmik Abhijit a model for detecting fake reviews in the education using an educational attention-enhanced lstm and random forest technique / Kuantan, Pahang : UMPSA, 2025 © 2025 xiii, 129 pages : illustrations ; 30 cm. + 1 CD-ROM. rdacontent text rdamedia unmediated rdacarrier volume rda text file PDF Faculty of Computing Thesis (Doctor of Philosophy) -- Universiti Malaysia Pahang - 2026 Includes bibliographical references The increasing reliance on online student review systems in higher education has intensified the need to ensure the authenticity and reliability of such feedback. Educational reviews differ significantly from product and service reviews due to multi-aspect sentiments, domain-specific terminology, limited metadata availability, and strict privacy regulations. These unique characteristics make existing fake review detection methods from other domains unsuitable for academic settings, where rating–comment inconsistencies, subjective bias, and strategically manipulated evaluations pose challenges to fair and data-driven teacher assessment. This research aims to address these challenges by developing a hybrid analytical framework specifically designed for the educational domain by fulfilling three objectives namely, i) Feature identification, ii) Model development and iii) Evaluation of the developed model. The proposed approach integrates an Education Attention-Enhanced Long Short-Term Memory model (EduAE-LSTM) for aspect-based sentiment extraction with a Random Forest classifier for fake review detection. A large-scale teacher evaluation dataset from the American International University-Bangladesh (AIUB), containing over 2.2 million entries, was utilized alongside benchmark datasets such as SemEval-2014 Task 4, MOOC Reviews, and Coursera Course Reviews. Linguistic, contextual, and behavioral features, including polarity, subjectivity, rating–comment consistency, and review length were operationalized to enable accurate and interpretable classification. Experimental results demonstrate that the EduAE-LSTM surpasses standard CNN, GRU, and LSTM baselines, achieving an F1-score of 0.86 and ROC-AUC of 0.95. When integrated with the Random Forest classifier, the hybrid model achieved 96.9% accuracy, 0.9999 precision, 0.9620 recall, and a 0.9806 F1-score, outperforming all comparative machine learning models. Domain validation using expert judgments yielded a 96% agreement rate and no significant difference in predictions (p > 0.05). Cross-domain evaluations further confirmed the robustness of the framework, with F1-scores ranging from 0.94 to 0.99 across benchmark datasets. This research contributes both theoretically and practically by advancing NLP methodologies for education-specific text mining, proposing a validated hybrid deep+ensemble architecture for fake review detection, and offering an interpretable, data-driven tool to strengthen transparency and trust in academic evaluation systems. The societal impact aligns with SDG 4 (Quality Education) by enhancing the credibility of educational assessments, and SDG 16 (Peace, Justice, and Strong Institutions) through promoting ethical, trustworthy, and data-driven institutional decision-making. Faculty of Computing Dissertations Universities and colleges Dissertations Theses Dissertations lcc THESIS 0 0 lcc 0 -2 NFIC 20000 20000 2026-08-05 0 FK .B46 2026 r Thesis T000005598 2026-08-05 09:32:23 1 2026-08-05 THESIS 0 0 lcc 0 -2 NFIC 20000 20000 2026-08-05 0 CD14038 T000005599 2026-08-05 09:32:47 1 2026-08-05 ACCOM 105393 105399