One promising avenue towards increasing player entertainment for individual game players is to tailor player experience in real-time via automatic game content generation. Modeling the relationship between game content and player preferences or affective states is an important step towards this type of game personalization. In this paper we analyse the relationship between level design parameters of platform games and player experience. We introduce a method to extract the most useful information about game content from short game sessions by investigating the size of game session that yields the highest accuracy in predicting players' preferences, and by defining the smallest game session size for which the model can still predict reported emotion with acceptable accuracy. Neuroevolutionary preference learning is used to approximate the function from game content to reported emotional preferences. The experiments are based on a modified version of the classic Super Mario Bros game. We investigate two types of features extracted from game levels; statistical level design parameters and extracted frequent sequences of level elements. Results indicate that decreasing the size of the feature window lowers prediction accuracy, and that the models built on selected features derived from the whole set of extracted features (combining the two types of features) outperforms other models constructed on partial information about game content.