Sigit Purnomo

Sigit Purnomo

Senior Lecturer at Informatics Department, Universitas Atma Jaya Yogyakarta

When Classical Models Compete with Deep Learning: A Semantic Label Transformation for Peer Feedback Classification

1 minute read

This article was published in the 2026 International Conference on Information Technology and Computing (ICITCOM), published by IEEE.

Authors: Yohanes Sigit Purnomo W.P., Theresia Devi Indriasari

Language: English

Abstract:
Peer feedback often contains overlapping communicative intents, making multi-label classification challenging. This study proposes a semantic priority-based label transformation that converts multi-label Indonesian peer feedback into a single-label task by assigning a dominant label based on communicative importance. The final transformed dataset consisted of 723 instances across four categories: Appreciation, Problem, Suggestion, and Neutral. Models were evaluated under class imbalance using weighted F1-score, balanced accuracy, and Matthews Correlation Coefficient. Results show that classical machine learning and deep learning models exhibit different performance characteristics. ExtraTrees with word-level TF-IDF achieved stable and competitive performance across evaluation stages, while Transformer-based models obtained higher cross-validation performance but lower validation performance, indicating a generalization gap. RCNN_WE achieved the highest validation performance, but its lower cross-validation score suggests sensitivity to data partitioning rather than definitive evidence of stronger generalization. The priority-order ablation supports the proposed Problem > Suggestion > Appreciation > Neutral order because it achieved the best-balanced performance, despite an Appreciation-first order producing higher weighted F1-score. These findings highlight the importance of label design, class imbalance, and evaluation strategy in automated peer feedback classification.

Keywords: Peer feedback classification, Semantic label transformation, Multi-label to single-label classification, Machine learning, Deep learning, Text classification, Class imbalance

Conference: 2026 International Conference on Information Technology and Computing (ICITCOM)

Conference Date: August 5, 2026

Date Added to IEEE Xplore: October 8, 2026

Publisher: IEEE

Electronic ISBN: 979-8-3195-3459-0

Print on Demand ISBN: 979-8-3195-3460-6

DOI: 10.1109/ICITCOM70739.2026.11714417

Article URL: https://ieeexplore.ieee.org/document/11714417

How to Cite

If you extend or use this work, please cite the paper where it was introduced:

@inproceedings{Purnomo2026SemanticLabel,
    author = {Purnomo W.P., Yohanes Sigit and Indriasari, Theresia Devi},
    title = {When Classical Models Compete with Deep Learning: A Semantic Label Transformation for Peer Feedback Classification},
    booktitle = {2026 International Conference on Information Technology and Computing (ICITCOM)},
    year = {2026},
    publisher = {IEEE},
    isbn = {979-8-3195-3459-0},
    doi = {10.1109/ICITCOM70739.2026.11714417},
    url = {https://ieeexplore.ieee.org/document/11714417}
}

You may also enjoy

Mixed Approach Speech-to-Text Translation for Endangered Language

2 minute read

Published:

This study addresses the technological marginalization of endangered regional languages by evaluating speech-to-text translation for Dayak Ma’anyan, an extremely low-resource Austronesian language. It examines whether cascaded multilingual automatic speech recognition and machine translation models can provide effective Ma’anyan–Indonesian translation despite severe data scarcity.

Understanding Social-Cognitive-Norm Mechanisms Driving Disinformation Verification among Indonesian Young Adults on Social Media

2 minute read

Published:

This research aimed to develop an integrated theoretical model to explain the factors influencing verification behavior regarding social media disinformation among young adults in Indonesia. The model combined the Stimulus–Organism–Response (SOR) framework with the Norm Activation Model (NAM) and the Social Identity Theory (SIT) to examine the collective effects of social, cognitive, and moral processes in shaping responsible information behavior. An online cross-sectional survey was conducted with 746 respondents, who actively used social networking sites to obtain and share information. The results showed that verification behavior was primarily driven by information skepticism and personal norms, emphasizing the importance of critical thinking and moral obligation for responsible engagement. In the Organism stage, awareness of fake information, perceived deception, and critical consumption enhanced moral sensitivity and analytical reasoning. At the social level, factors such as collective memory, parasocial interaction, and status-seeking were reported to be significant identity-based stimuli in shaping cognitive and moral responses, with gender found to moderate these effects. Thematic analysis suggested that most young adults verified information through cross-checking and peer consultation, but were influenced by social validation. Theoretically, this research contributed to disinformation research by framing verification as a cognitive-normative process rather than a reactive behavior. Different initiatives were recommended by educational institutions, governmental bodies, and community organizations to strengthen moral reasoning, digital literacy, and civic responsibility in combating disinformation.

Automated Rubric-Based Classification of Student Peer Code Review Feedback

2 minute read

Published:

This study examines automated approaches for classifying student peer code review feedback in Bahasa Indonesia according to rubric-based code quality categories. The study used 2,281 student feedback items collected from peer code review activities in an introductory programming course. The dataset was annotated into seven labels and validated with strong inter-rater agreement. Three approaches were compared: classical machine learning, deep learning, and few-shot prompting using large language models. Random Forest with count vectorization produced the best performance with an F1-score of 0.9430, outperforming recurrent convolutional neural networks with FastText embeddings and few-shot prompting. The findings indicate that classical machine learning with token-based features can be an effective and interpretable baseline for supporting rubric alignment in peer code review tools, especially in computing education contexts using Bahasa Indonesia.

Leveraging Machine Learning in Student Peer Review: A Systematic Literature Review

2 minute read

Published:

Our study examines how machine learning techniques are integrated into student peer review processes, focusing on the challenges that motivate their adoption and the methods used to address them. Using Kitchenham’s systematic literature review framework, 328 articles were screened, and 25 empirical studies on machine learning applications in student peer review were selected. The findings show that machine learning is mainly used to manage large volumes of reviews, support automated grading, and improve feedback quality. Common techniques include classification, prediction, ranking, and clustering, which help improve the fairness, efficiency, and objectivity of peer review. This study provides a rigorous synthesis of machine learning adoption in student peer review and highlights its potential to enhance assessment accuracy, support learning outcomes, and guide future research and broader implementation in educational contexts.