In this study, we propose a novel deep learning method for cross-domain recommendations that effectively combines attention mechanisms, autoencoders, and multitask learning. Our approach leverages multiple datasets from diverse domains and incorporates domain-specific encoders, a shared self-attention mechanism, and a multilayer perceptron (MLP) to capture both intra-domain and inter-domain relationships. By jointly modeling these interactions, we improve recommendation accuracy across domains. Experimental results using the MovieLens dataset demonstrate that our proposed cross-domain recommendation system outperforms traditional approaches including matrix factorization, standard MLPs, and self-attention-based baselines.