Evaluation
Results
Training performance across 12 epochs and evaluation on the held-out test set.
Training Curve
Accuracy Over 12 Epochs
The model converges rapidly, reaching 99.79% training accuracy by the final epoch.
| Epoch | Training Accuracy | Loss | Progress |
|---|---|---|---|
| 1 | 92.50% | — | |
| 3 | ~96% | — | |
| 6 | ~98% | — | |
| 9 | ~99.3% | — | |
| 12 | 99.79% | 0.0060 |
Key Metrics
Model Performance Summary
Final Training Accuracy
99.79%
after 12 training epochs with Adam optimizer.
Final Training Loss
0.006
binary cross-entropy, indicating tight convergence.
Threshold Optimization
Precision-recall analysis identifies the optimal decision boundary for the imbalanced hate-speech class beyond the default 0.5.
Design Choices
Why This Architecture
| Component | Choice | Rationale |
|---|---|---|
| Word Embeddings | GloVe 50d | Captures semantic similarity without training from scratch on a small corpus. |
| Sequence Model | BiLSTM | Reads tweets both left-to-right and right-to-left, capturing full context. |
| Regularization | Dropout + BatchNorm | Reduces overfitting on the small (~25K training) dataset. |
| Threshold | PR-optimized | Class imbalance (7% hate speech) makes default 0.5 sub-optimal. |
| Loss Function | Binary Cross-Entropy | Standard for binary classification with sigmoid output. |