A recent trend within computational intelligence and games research is to investigate how to affect video game players' in-game experience by designing and/or modifying aspects of game content. Analysing the relationship between game content, player behaviour and self-reported affective states constitutes an important step towards understanding game experience and constructing effective game adaptation mechanisms. This papers reports on further refinement of a method to understand this relationship by analysing data collected from players, building models that predict player experience and analysing what features of game and player data predict player affect best. We analyse data from players playing 780 pairs of short game sessions of the platform game Super Mario Bros, investigate the impact of the session size and what part of the level that has the major affect on player experience. Several types of features are explored, including item frequencies and patterns extracted through frequent sequence mining.