The use of eye-tracking as a tool to provide cognitive context is rising in real-world systems. Though extensive research has been done on using machine learning and deep learning to classify sedentary activities using data captured by eye-trackers, there is a gap in analyzing the impact of the usage of different sedentary activities and feature extraction methods on the performance. In this paper, conventional machine learning algorithms are used to classify reading activities, captured by eye tracking devices. Multiple data pre processing methods and filters are used to extract fixations out of the raw data captured by the eye-trackers. Out of these fixations, a total of 16 features are used to classify activities. Using the optimal configurations found, a 0.99 user dependent and a 0.76 user independent score is obtained. All obtained results are compared to results obtained by peers performing similar research, who use either a different data set or deep learning instead of conventional machine learning. The papers using deep learning had a vastly strong performance for all user dependent evaluations, but performed poorer in the user independent evaluation. Overall, conventional machine learning performed better on user independent evaluation, where this paper obtained the best results for user independent evaluation, most likely due to the fact that Japanese reading material was used, which has very distinctive reading directions.