Applying machine learning to keyboard and mouse data is an important topic in human-computer interaction since gained knowledge from analyzing user interaction behaviour allows to improve system attributes such as interactivity and user experience. For this purpose, an expressive data representation is crucial for achieving meaningful predictive power. In contrast to previous works which mostly rely on handcrafted features, this work explores generalizable encodings in order to supply the machine learning model with less prefiltered inputs. Results on two datasets show that the proposed encodings can improve performance of interactive task recognition, since a time series representation, keeping track of mouse pointer coordinates and mouse button states in fixed time intervals, significantly outperformed the baseline of using handcrafted features in case of mouse data. Regarding keyboard data, applying a similar representation which tracks the key states also resulted in better predictive power than using manually extracted features. In addition, approaches based on techniques from natural language processing were competitive to the time series representation. This indicates that multiple encodings need to be considered when assessing how to encode keyboard data. Overall, our work shows that applications based on machine learning on keyboard and mouse data can benefit from selecting a less prefiltering encoding technique over handcrafted feature extraction.