Summary: Sequence Analysis of Game Data

Aggregating the data

Games are played and experienced sequentially, with players taking actions and the games calculating responses to those actions. When we aggregate behavioral telemetry data, we lose the sequential information and context of the action. In many situations, such loss is not a problem, but there are other instances where studying the sequence of player actions can be useful or even necessary. In this chapter, we introduced the idea of looking at sequences of player behavior and covered some of the basic methods for analyzing sequential data.

Exploring the data in R

Specifically, we discussed different methods to explore sequence data using the TraMineR and arulesSequences packages in R. We also explored the use of the frequent sequence pattern mining algorithm, SPADE, on game data. In addition, we discussed clustering techniques—particularly OM—and showed how to apply them using the TraMineR package.

Usage of TraMineR

Using TraMineR to examine frequent sequences allows for at-a-glance analysis of game data and facilitates the identification of high-level trends. However, it doesn’t provide any quantitative measures of similarity between sequences.

Usage of SPADE

Different techniques present various pros and cons. Using SPADE to mine frequent subsequence patterns allows for a clear and granular observation of common behavior patterns, but it loses the big picture due to its focus on subsequences without taking continuity into account. Without continuity, we lose some aspects of temporal sequences that may be important for the question being examined.

Why do we do clustering?

Clustering combines sequences into clusters based on their similarity to one another. Clustering is always dependent on the actual distance measure.

How are similarity measures formed?

There are various ways to derive distance measures for sequences, but these can be biased and require human input for encoding domain and game knowledge. Through our experience, we often found it helpful to develop our own similarity or distance measures based on what we believe are important differences between states.

The main discussion of the chapter

While in this chapter we mostly focused on sequence action analysis, upcoming chapters will introduce more advanced techniques that allow us to construct probabilistic graphical methods and add more information about the game state rather than simply the sequences of actions that occurred.