Artificial Intelligence

Abstract

Interpretability is a key requirement for deploying neural networks in regulated domains. We compare LIME, SHAP, Grad-CAM, and integrated gradients on three classification tasks and assess faithfulness, stability, and computational cost. Results show that SHAP provides the best faithfulness-stability trade-off at moderate overhead.

Abstract

Class imbalance in tabular datasets degrades classifier performance in high-stakes applications. We adapt conditional GAN architectures (CTGAN, TVAE) for synthetic minority-class generation and evaluate on six benchmark datasets, achieving a 6.8 percentage-point F1 improvement over SMOTE on average.

Abstract

We model a four-way urban intersection as a Markov decision process and train a DQN agent to minimise average vehicle waiting time. In simulation, the RL policy reduces waiting time by 31% compared to a fixed-cycle baseline and adapts to demand fluctuations without re-training.