<![CDATA[Tourism recommendation systems have become increasingly important in supporting intelligent travel planning and improving tourist experiences through personalized destination suggestions. However, most conventional recommendation approaches fail to effectively model sequential tourist behavior and contextual travel activities. This study proposes a predictive tourism destination recommendation system using a dual-input Long Short-Term Memory (LSTM) architecture capable of simultaneously learning destination sequences and tourist activity patterns. The proposed framework aims to generate contextual and adaptive destination recommendations based on tourist travel histories in Rokan Hulu Regency, Indonesia. Due to the limited availability of real-world sequential tourism datasets, a realistic synthetic dataset was constructed by considering destination categories, tourist typologies, geographical distance, popularity scores, visit duration, and activity diversity. The preprocessing stage involved tokenization, sequence padding, and tensor transformation to prepare the data for deep learning–based sequential modeling. The proposed dual-input LSTM model was trained using destination sequences and activity sequences as parallel inputs and evaluated using Top-K Accuracy metrics. Experimental results demonstrate that the proposed model achieved a Top-1 Accuracy of 39.72%, a Top-3 Accuracy of 76.21%, and a Top-5 Accuracy of 92.40%. Comparative evaluation also shows that the proposed architecture outperformed both a Random Forest classifier and a single-input LSTM model across all evaluation metrics. The findings indicate that integrating tourist activity information significantly improves contextual recommendation quality and predictive performance. Overall, this research contributes to the development of intelligent sequence-aware tourism recommendation systems and demonstrates the effectiveness of dual-input deep learning architectures for modeling contextual tourist travel behavior.]]>
