This paper proposes DeepCoNN, a deep model that jointly learns user behaviors and item properties from review text to improve recommendation systems. The model consists of two parallel neural networks that are connected by a shared layer, allowing latent factors of users and items to interact. One network learns user behaviors from reviews written by users, while the other learns item properties from reviews written for items. The shared layer enables the interaction of these latent factors, similar to factorization machine techniques. Experimental results show that DeepCoNN significantly outperforms existing recommender systems on various datasets, including Yelp, Amazon, and Beer. It effectively alleviates the sparsity problem by leveraging review text, which contains rich information that traditional collaborative filtering methods ignore. DeepCoNN also uses pre-trained word embeddings to extract semantic information from reviews, improving rating prediction accuracy. The model is scalable and suitable for online learning scenarios. It outperforms state-of-the-art techniques in terms of prediction accuracy on all evaluated datasets. The experiments demonstrate that DeepCoNN achieves significant improvements in rating prediction, especially for users and items with fewer ratings. The model's ability to jointly model users and items based on review text makes it more effective than methods that rely solely on ratings. The results show that DeepCoNN can effectively alleviate the sparsity problem and improve the performance of recommendation systems.This paper proposes DeepCoNN, a deep model that jointly learns user behaviors and item properties from review text to improve recommendation systems. The model consists of two parallel neural networks that are connected by a shared layer, allowing latent factors of users and items to interact. One network learns user behaviors from reviews written by users, while the other learns item properties from reviews written for items. The shared layer enables the interaction of these latent factors, similar to factorization machine techniques. Experimental results show that DeepCoNN significantly outperforms existing recommender systems on various datasets, including Yelp, Amazon, and Beer. It effectively alleviates the sparsity problem by leveraging review text, which contains rich information that traditional collaborative filtering methods ignore. DeepCoNN also uses pre-trained word embeddings to extract semantic information from reviews, improving rating prediction accuracy. The model is scalable and suitable for online learning scenarios. It outperforms state-of-the-art techniques in terms of prediction accuracy on all evaluated datasets. The experiments demonstrate that DeepCoNN achieves significant improvements in rating prediction, especially for users and items with fewer ratings. The model's ability to jointly model users and items based on review text makes it more effective than methods that rely solely on ratings. The results show that DeepCoNN can effectively alleviate the sparsity problem and improve the performance of recommendation systems.