In the cloud based service provisioning industry, one of the main challenges that providers face involves keeping existing tenants engaged while attracting new ones. To address this, providers need to gain insights about customer satisfaction. In that regards, support ticket data, understood as the main way of communication between both parties, can be mined to obtain an estimation of customer satisfaction by means of the polarity of the sentiment extracted from the report descriptions. To that end, in this work we propose a model which can learn a feature representation for sentiment polarity changes, from the sequence of tickets emitted by a given customer during the period associated with the service subscription term. Then, that resulting feature representation, combined with other handcrafted features related to contract and ticket data, is passed to a classifier which estimates the likelihood of service subscription renewal by the customer. Experiment results using real data from a service provider shows that learned representation of sentiment polarity changes from support ticket data in combination with other handcrafted features improves the accuracy in predicting subscription renewals. Moreover, our architecture is flexible enough incorporate and integrate several feature representations and give more expressive power to the prediction.