The rapid growth of mobile applications and online e-commerce websites has made it easy to gather information to create an enormous quantity of training data to aid consumers in making decisions about what to purchase. On online shopping sites, helpful reviews analysis of user reviews can significantly increase users’ loyalty. People may significantly influence the market value of goods and customer confidence in e-commerce decisions by using ratings, and reviews. One major issue with users’ rating prediction models is that they ignore variations across users that fall inside the user’s preferences or reviews. In this paper, we develop a new balanced helpful recommendation model with quantifying users’ tendencies (BHRQUT) -based on personalized reviews and ratings to predict helpful reviews and improve recommendation accuracy by creating an auxiliary feature that is computed based on actual ratings and predicted ratings. Text sequence processing was acquired by experimental research on the influence of word vector embedding dimension and word frequency of review text, utilizing (NLP). These features were transformed into vectors based on the embedding layer to the balanced (CNN-BiLSTM) model. Experimental evaluations are performed on four review datasets from the 5-score Amazon domain and our model can significantly enhance the accuracy of helpful review text analysis by 97 percent. According to the experimental results when we compared with other deep recommendation approaches concerning multiple metrics and drew from the different experiments, the presented model can enhance the analyzability of user feedback by enhancing decision-making confidence without reducing accuracy.