The classification task in dynamical environments is challenging. A reason for this is the change of their statistical properties over time. This characteristic is called concept drift and is one of the major topics in data mining. The objective of this thesis is to analyze, how accurate different systems are classifying in dynamical environments over a period of time. For this purpose, two different approaches are evaluated. One approach removes the features with the highest change in influence. The other is an ensemble based model which let experts vote between the outcomes. Although the models were not able to increase the accuracy after a long period of time, the results show that both models are able to achieve a higher accuracy than the baseline in particular cases. This outcome underlines that emotion detection in Twitter can be improved. New models or improvements to existing ones could be able to handle concept drift to achieve a higher accuracy.