In this paper we removed two key weaknesses of Playtracer,
a powerful visual datamining system for games with states
and state transitions, by introducing the notion of state fea-
tures. These features collapse states without losing salient
information and make Playtracer output signicantly easier
to understand. We show how this method allows us to use
Playtracer to analyze previously untraceable data from the
educational game Refraction, as well as two previously un-
traceable games with continuous state spaces, Hello Worlds
and Foldit.
Designers have full control over what they view as features,
and the use of dierent sets of features will result in out-
put that only analyzes the facets of the game those features
capture. Based on our experiences, we suggest three gen-
eral principles for the selection and use of features. First,
high-level features tend to collapse the state space more ag-
gressively, which is generally desirable for complex games.
Second, states should share the same value for a feature if
they are conceptually the same to prevent the appearance
of multiple states which are in fact identical. Finally, each
additional feature causes more fragmentation of the state
space, so only the minimum number of features should be
used. Keeping these principles in mind when choosing fea-
tures makes it much more likely that Playtracer's output will
be useful for analyzing how players interact with a game.