💬 NLP

Comparative Analysis of AutoML and BiLSTM Models for Cyberbullying Detection on Indonesian Instagram Comments

This study evaluates and compares machine learning and deep learning models for detecting cyberbullying in Indonesian Instagram comments, finding that while BiLSTM with attention mechanisms achieves the highest overall performance, Logistic Regression remains a competitive and resource-efficient alternative.

Original authors: Raihana Adelia Putri, Aisyah Musfirah, Anggi Puspita Ningrum, Luluk Muthoharoh, Ardika Satria, Martin Clinton Tosima Manullang

Published 2026-04-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Raihana Adelia Putri, Aisyah Musfirah, Anggi Puspita Ningrum, Luluk Muthoharoh, Ardika Satria, Martin Clinton Tosima Manullang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine Instagram as a massive, bustling digital town square. While it's a place for sharing photos and connecting, it has a dark side: cyberbullying. This is when people use the internet to harass, threaten, or humiliate others. In Indonesia, this is a serious problem affecting many young people.

This paper is like a detective report trying to figure out the best way to build a "digital security guard" that can automatically spot these mean comments in the Indonesian language before they hurt anyone.

Here is the story of their investigation, broken down simply:

1. The Suspects (The Data)

The researchers gathered a "wanted list" of 650 Instagram comments.

The Source: They came from the comment sections of famous Indonesian artists and influencers.
The Mix: The list was perfectly balanced: 325 comments were mean (Bullying) and 325 were nice (Non-Bullying).
The Challenge: These weren't formal essays. They were messy, short, and full of slang, abbreviations, and emojis—just like real teenagers talking.

2. The Cleaning Crew (Preprocessing)

Before the computers could read these comments, they had to be cleaned up. Imagine trying to read a book written in a language where people constantly misspell words or stretch them out (like "baaaanget" instead of "banget").

The researchers built a special cleaning machine with six steps:

Case Folding: Turning everything to lowercase (so "Hello" and "hello" are the same).
Cleaning: Removing hashtags, links, and @mentions.
Slang Normalization: Fixing the slang (turning "bgt" back into "banget").
Stopword Removal: Throwing away common words like "and" or "the" that don't help identify bullying.
Stemming: Chopping words down to their root (turning "running" into "run").
Tokenization: Cutting the sentence into individual word pieces.

Without this step, the computers would be confused by the messy internet slang.

3. The Contestants (The Models)

The researchers pitted two different types of "detectives" against each other to see who could catch the bullies best.

Team A: The Classic Detectives (Machine Learning)

These are the old-school, reliable methods. They look at the words and count how often specific "bad" words appear.

Naive Bayes: A quick guesser based on probability.
Logistic Regression: A steady calculator that draws a line between good and bad.
SVM (Support Vector Machine): A sharp-eyed classifier that tries to find the perfect boundary between the two groups.
Tool: They used a method called TF-IDF, which is like a highlighter that marks words that are unique and important to the sentence.

Team B: The Deep Thinkers (Deep Learning)

These are the advanced AI detectives that try to understand the context and the flow of the sentence, not just the words.

Bi-LSTM: A model that reads the sentence from left-to-right AND right-to-left at the same time, like reading a book while also remembering what you just read.
Bi-LSTM + Attention: The same model, but with a "spotlight" (Attention mechanism). This spotlight tells the model to pay extra attention to the most emotional or important words in the sentence.

4. The Results: Who Won?

The researchers ran a race to see which model could correctly identify the bullying comments.

The Classic Winner: Among the old-school detectives, Logistic Regression was the champion. It got it right about 85.25% of the time. It was fast, efficient, and didn't need a supercomputer to run.
The Deep Learning Winner: The Bi-LSTM with the "Spotlight" (Attention) took the overall crown. It got it right about 84.62% of the time.
- Wait, isn't that lower? Actually, it's very close! The "Spotlight" helped the model understand the nuance of the bullying better than the others, even if the raw percentage was slightly lower than the best classic model.
- The standard Bi-LSTM (without the spotlight) did worse (78.46%), proving that the "spotlight" was crucial.

5. The Verdict

The paper concludes with a few key takeaways:

Deep Learning is the "Smartest": The model with the "spotlight" (Bi-LSTM + Attention) is the best at understanding the complex, messy context of Indonesian slang and emotional attacks.
Machine Learning is the "Most Efficient": If you don't have a powerful computer or need something that runs very fast, the classic Logistic Regression is a very strong, practical choice. It performed almost as well as the complex AI but is much lighter.
Cleaning is Key: The study emphasizes that if you don't clean the slang and fix the typos first, even the smartest AI will fail.

The Limitations (The Fine Print)

The authors are honest about the limits of their study:

Small Sample: They only used 650 comments. It's like judging a whole country's taste in music based on a survey of 650 people. A bigger dataset would make the results more reliable.
Specific Source: All comments came from famous artists' pages. Bullying might look different on a regular person's page.
Simple Labels: They only labeled comments as "Bullying" or "Not Bullying." They didn't break it down further (e.g., "Body Shaming" vs. "Hate Speech").

In short: To catch cyberbullies in Indonesian Instagram comments, you need a good cleaning crew first. Then, you can choose between a fast, reliable classic detective (Logistic Regression) or a highly intelligent, context-aware AI with a spotlight (Bi-LSTM + Attention), depending on how much computing power you have.

1. Problem Statement

Cyberbullying on social media platforms, particularly Instagram, poses a severe threat to the mental health of Indonesian youth. While automated detection systems exist, most research focuses on Twitter/X data. Instagram comments present unique challenges due to their informal nature, heavy use of slang, emojis, character elongation, and mixed languages. Furthermore, there is a lack of comparative studies evaluating the trade-offs between traditional Machine Learning (ML) approaches and Deep Learning (DL) architectures specifically for Indonesian-language Instagram content.

2. Methodology

Dataset

Source: 650 Instagram comments scraped from accounts of Indonesian artists and influencers.
Structure: Balanced binary classification dataset with 325 "Bullying" and 325 "Non-Bullying" labels.
Characteristics: Informal Indonesian text with an average length of 63 characters. The dataset is clean (no duplicates/missing values) and covers diverse linguistic contexts.

Preprocessing Pipeline

A specialized pipeline was developed to handle Indonesian linguistic nuances using specific libraries (indoNLP, nlp-id, PySastrawi):

Case Folding: Conversion to lowercase.
Cleaning: Removal of mentions, hashtags, URLs, and non-alphabetic characters.
Slang Normalization: Replacing slang and character elongations (e.g., "bgt" $\rightarrow$ "banget").
Stopword Removal: Eliminating non-discriminative common words.
Stemming: Reducing words to root forms based on Indonesian morphology.
Tokenization: Splitting text into tokens.

Modeling Approaches

The study compared two distinct modeling paths:

Machine Learning (ML) Path (AutoML):
- Tools: PyCaret framework with Grid Search for hyperparameter tuning.
- Feature Extraction: TF-IDF (Term Frequency-Inverse Document Frequency).
- Algorithms Tested: Naive Bayes, Logistic Regression, and Support Vector Machine (SVM).
- Validation: 5-Fold Cross-Validation.
Deep Learning (DL) Path:
- Tools: PyTorch.
- Architectures:
  1. Bi-LSTM: Bidirectional Long Short-Term Memory to capture sequential context.
  2. Bi-LSTM + Bahdanau Attention: Adds an attention mechanism to weight significant keywords.
- Configuration: Batch size 32, Embedding size 128, Adam optimizer, Cross-Entropy Loss, Early Stopping (patience=3).
- Validation: Static 80/10/10 split (Train/Val/Test) with random seed 42.

3. Key Contributions

Comparative Analysis: A direct comparison of three classical ML algorithms against two DL architectures specifically for Indonesian Instagram comments, filling a gap in literature that often focuses on Twitter.
Domain-Specific Preprocessing: Demonstration of the critical impact of a tailored NLP pipeline (slang normalization and morphological stemming) on model performance for informal Indonesian text.
Practical Framework: Identification of the optimal model for different deployment scenarios (high accuracy vs. resource constraints), providing a foundation for automated content moderation systems in Indonesia.

4. Experimental Results

Machine Learning Performance:

Logistic Regression emerged as the best-performing ML model.
- Accuracy: 85.25%
- F1-Score: 85.22%
SVM followed with 82.61% accuracy.
Naive Bayes performed the lowest at 77.99% accuracy.

Deep Learning Performance:

Bi-LSTM + Attention achieved the highest overall performance.
- Accuracy: 84.62%
- F1-Macro: 84.58%
Standard Bi-LSTM lagged behind with 78.46% accuracy, highlighting the significant boost (approx. 6%) provided by the attention mechanism.

Comparative Insight:

Accuracy: Logistic Regression (ML) slightly outperformed Bi-LSTM+Attention (DL) by 0.63 percentage points.
Contextual Understanding: While ML models were highly competitive, the DL models demonstrated superior capability in capturing complex sequential dependencies and emotional context, evidenced by the significant performance jump when adding the Attention mechanism.
Efficiency: ML models offered faster training times and lower computational requirements, making them viable for resource-constrained environments.

5. Significance and Conclusion

Model Selection: For high-accuracy requirements where computational resources are available, Bi-LSTM with Attention is the recommended approach. However, for lightweight, rapid deployment, Logistic Regression offers comparable performance with significantly lower overhead.
Preprocessing Importance: The study confirms that domain-specific preprocessing (slang normalization and stemming) is not optional but essential for achieving high performance on informal Indonesian social media data.
Limitations & Future Work: The study is limited by a relatively small dataset (650 samples) and a binary label scheme that does not capture nuanced categories (e.g., body shaming vs. hate speech). Future research should utilize larger datasets and explore Transformer-based models (e.g., BERT variants) to further enhance detection capabilities.

Conclusion: The paper establishes that while Deep Learning models with attention mechanisms are theoretically superior for contextual understanding, traditional Machine Learning models remain highly effective and competitive for cyberbullying detection in Indonesian social media, provided that rigorous, language-specific preprocessing is applied.