yazidsupriadi/mbert_gru_bot

IndoBERT-GRU for Bot Detection on Platform X

This model uses BERT-BASE-MULTILINGUAL-CASED to classify Twitter/X accounts as bots or humans based on textual content and numeric features.

Model Architecture

• Base Model: bert-base-multilingual-cased
• Classifier: GRU + 2-layer Feedforward (ReLU + Sigmoid)

Dataset

This dataset consists of Indonesian tweets with labels (0 for human, 1 for bot), combined with numeric features like favorite_count, retweet_count, reply_count, and quote_count.

Training Details

• Loss Function: Binary Cross Entropy (BCELoss)
• Optimizer: Adam (lr = 2e-5)
• Epochs: 10
• Max Token Length: 128
• Batch Size: 16

Training Log

Epoch	Train Loss	Val Accuracy	Precision	Recall	F1-score
1	0.9018	0.8632	0.8436	0.8928	0.8675
2	0.3000	0.8740	0.8526	0.9053	0.8781
3	0.2735	0.8890	0.8821	0.8988	0.8904
4	0.2590	0.8940	0.8920	0.8973	0.8946
5	0.2520	0.8958	0.9068	0.8829	0.8947
6	0.2409	0.8965	0.9163	0.8734	0.8943
7	0.2390	0.9020	0.9160	0.8858	0.9007
8	0.2317	0.9045	0.9068	0.9023	0.9045
9	0.2329	0.9052	0.9029	0.9088	0.9058
10	0.2288	0.9067	0.9101	0.9033	0.9067

Evaluation Results

Final Validation Accuracy: 0.5620 Final Precision: 0.5338 Final Recall: 1.0000 Final F1-score: 0.6960

Final ROC AUC: 0.9750

Confusion Matrix

ROC Curve

Classification Report

              precision    recall  f1-score   support

       Human       0.90      0.91      0.91      1994
         Bot       0.91      0.90      0.91      2006

    accuracy                           0.91      4000
   macro avg       0.91      0.91      0.91      4000
weighted avg       0.91      0.91      0.91      4000