Sentiment analysis plays an essential role in identifying user preferences for recommender system. However, existing recommendation systems mainly focus on the relationships between items and the user’s purchasing behavior while ignoring the user’s sentiment information. To improve the accuracy of recommender system, in this paper, a multi-feature fusion recommendation model based on sentiment analysis (MFF-SA) is proposed. Firstly, rather than sentiment values, sentiment vectors are obtained by a new quantify sentiment method that based on natural language processing, which to gain more valuable insights about uses’ preferences. Secondly, to explicitly model the features interactions, a multi-head self-attention mechanism is introduced, which can capture multiple different aspects of the feature relationship simultaneously. Finally, experimental results on three datasets demonstrate that the proposed methods outperform the state-of-the-art methods in terms of effectiveness and feasibility with a 3.9% relative improvement.