Code vulnerabilities pose risks to software security. This paper combines data-driven (deep learning) and rulebased methods to enhance vulnerability detection, leveraging unsupervised learning to prune the hybrid model. A neural network encodes source code into vectors preserving semantic syntactic properties, which are classified using a Random Forest ensemble. A novel pruning algorithm, utilizing clustering techniques, removes low-impact trees to prevent overfitting. Evaluation on Java and C datasets shows that the proposed approach outperforms state-of-the-art rule-based, neural network, and vulnerability detection methods in accurately classifying methodlevel vulnerabilities