RonTon05/3phase_MTL_Residual_TypeFusion

Multi-Task Trust & Type Classifier

This model is a fine-tuned Multi-Task learning model based on the Hugging Face transformers library, designed to simultaneously evaluate content trust (Binary Classification) and classify content types (10-class Classification).

Evaluation Results

📊 TASK 1 — TRUST REPORT (Binary Classification)

• Accuracy: 81.23%
• Macro F1: 45.14%

Classification Report:

Class	Precision	Recall	F1-Score	Support
0 (Trustworthy)	0.8267	0.9788	0.8964	2735
1 (Untrustworthy)	0.0333	0.0036	0.0064	563
Macro Avg	0.4300	0.4912	0.4514	3298
Weighted Avg	0.6913	0.8123	0.7444	3298

Confusion Matrix:

Actual \ Predicted	Predicted 0	Predicted 1
Actual 0	2677	58
Actual 1	561	2

---

📊 TASK 2 — TYPE REPORT (10-class Classification)

• Accuracy: 47.91%
• Macro F1: 41.55%

Classification Report:

Class	Precision	Recall	F1-Score	Support
0	0.3955	0.4291	0.4117	247
1	0.6250	0.6250	0.6250	32
2	0.5097	0.6270	0.5623	126
3	0.4878	0.6863	0.5703	204
4	0.2000	0.2529	0.2234	87
5	0.4869	0.3635	0.4163	718
6	0.5667	0.6859	0.6206	1041
7	0.1451	0.3415	0.2036	82
8	0.3571	0.0990	0.1550	101
9	0.4662	0.3030	0.3673	660
Macro Avg	0.4240	0.4413	0.4155	3298
Weighted Avg	0.4833	0.4791	0.4694	3298

---

Technical Specifications

Model Architecture and Objective

The model leverages a shared encoder backbone (e.g., PhoBERT) with custom classification heads designed for Multi-Task Learning (MTL):

1. Trust Head: Binary classification output layer.
2. Type Head: 10-class classification output layer.

Bias, Risks, and Limitations

⚠️ Imbalance Notice: Task 1 (Trust) shows extreme class imbalance (Class 0 has 2,735 samples while Class 1 only has 563). As reflected in the confusion matrix, the model struggles significantly to recall Class 1 (Recall: 0.0036). Special sampling techniques or loss re-weighting (e.g., Class Weights) are highly recommended for future iterations to improve performance on minority classes.