AAI_2025_Capstone_Chronicles_Combined

‭ResolveAI‬

‭The experimental results indicate that the binary classification model, based on a‬ ‭bidirectional LSTM architecture, learned meaningful patterns from the training data, as‬ ‭evidenced by a steady reduction in training loss from 0.6934 in epoch 1 to 0.1217 by epoch 14,‬ ‭and an increase in training accuracy from about 50% to over 96%. However, the validation‬ ‭metrics reveal a different trend. Initially, validation loss decreased from 0.6936 in epoch 1 to‬ ‭0.6539 in epoch 4, and validation accuracy improved modestly from 50% to around 62%.‬ ‭Beyond epoch 4, while training performance continued to improve, validation loss began‬ ‭increasing, reaching as high as 1.6170 by epoch 13, and validation accuracy plateaued around‬ ‭65%–67%. This divergence indicates that the model began to overfit the training data, learning‬ ‭details that do not generalize well to unseen examples.‬ ‭Regarding performance metrics, the model achieved a final F1 Score of approximately‬ ‭0.64, which, combined with accuracy, precision, and recall values observed during training,‬ ‭suggests moderate overall performance. Although the model successfully learned the task, the‬ ‭overfitting issue raises concerns about its robustness in a real-world setting. In practice,‬

‭20‬

68