Currently, social networks has been used for its users, and exploited by mechanisms of quality measurement systems and recommendation of products and services. The Recommendation Systems (SR) have used the data of the social networks and in parallel they have applied the sentiment and affective analysis in such data. However, there is still a concern in increasing the accuracy of the sentiment and affective analysis. This article introduces an SR, which extracts the texts of the users of the social networks and suggests musical styles based on the sentiment analysis by lexical approach and based on the affective analysis through the machine learning. The Convolutional Neural Network algorithm used for the emotion classification of hapiness, sadness, anger, fear, disgust and surprise presented a precision higher than the found in related works. Classification results of the F-Measure were of 0.98 e 0.96 for the emotion of sadness and anger, respectively. In addition, SR was assessed by means of subjective tests and the experimental results show that 97% of users approved the SR proposal.