A topic that is currently gathering a lot of interest in the games research community is how player behavior in-game relates to the subjective experience of playing a game. The implications are obvious: if we can analyze the behavior of a player and find out whether or not the player is having a good experience playing the game, and if not why, there is a lot of revenue that we can prevent loosing from players quitting. However, there are some challenges towards this.
Play style and progression are two important keywords in game analytics, as all methods for optimizing design and monetizations strategies revolve around them. When building games, we design with a specific operational space for the players in mind, so finding out how the players operate in that space is important when evaluating the effectiveness of a design. Similarly, how the patterns in behavior – the play styles – change over time or any other dimension of progression is important when evaluating design, engagement and monetization strategies. These types of analysis are thus important across both main perspectives we have on users: as players and as customers.
Play style and progression analysis are closely interlinked in practice, but formally a bit different. On one hand, play style analysis seeks to identify patterns in how people play a game. Play styles can be identified across the entire game, or just a section of it. Progression analysis on the other hand focuses on evaluating how people progress through a game. In practice, these are often combined to explore how play styles vary across a game.
These kinds of analysis can be performed using a great variety of techniques, from basic aggregation of key behavioral features to complex machine learning approaches.
Fundamentally though, progression analysis has four components:
- Feature selection: choosing the right behavioral features (manually or automated). This is not always easy!
- Dimensionality reduction: Reducing the number of input features (optimization)
Various approaches are possible – isomap projection, principal components, clustering, classification – a large number of techniques are available, all with specific strengths and weaknesses. Data normalization is a key step in most machine learning-driven processes, given the varied nature of most input features. There is a strong human element in most unsupervised learning methods, for example in determining the number of groups/clusters/classes.
- Progression mapping: mapping reduced/compound features to one or more progression dimensions.
Various approaches are again possible, but a fundamental plot of progression features vs. progression dimension can work wonders to bring the data into a format in which they can be readily analyzed.
A challenge here is selecting the right dimension or dimensions of progression. Progression in games can be temporal (playtime), spatial, content-based, story related, skill related, or any combination in between.
Progression can also be measured in different ways, for example through character wealth, level, kills, mushrooms collected … Often multiple measures need to be combined to develop a comprehensive evaluation.
- Visualization (and implementation): developing stakeholder-flexible interactive visualizations of results, implementing the results (and starting the analysis cycle over)
Finding the Patterns
In a recent study published at the Computational Intelligence in Games 2013 Conference, we presented a pretty flexible but also simple method which combines cluster analysis – a technique for reducing the dimensionality of complex datasets – and progression analysis, in order to investigate if and how players changed their behavior as they played Tomb Raider: Underworld (2008, Crystal Dynamics/Square Enix).
The case we are describing here uses the levels of Tomb Raider: Underworld as the dimension of progression, and focuses on behavioral features that relate to the core gameplay, e.g. jumping, shooting, etc., but in principle the method can be applied to any progression dimension and any behavioral measures.
The virtual trails of 62,000 players were analyzed using a specific type of cluster analysis called Archetype Analysis, which finds extreme behavioral profiles in the dataset and then compares all other players to these. The technique also allows any single player to be described using combinations of the Archetypes. An Archetype could for example be a play style that is characterized by being extremely aggressive, focusing on completing the game as fast as possible, and highly skilled at handling enemies, but very bad at solving the puzzles in the game.
Archetype analysis provided the means for finding groups of players who behaved similarly across each of the 7 levels in the Tomb Raider: Underworld game (there is also a prologue level). This allowed us to evaluate how the players of the game grouped into different play styles across the game – and there was a lot of variation.
Play styles across game levels
The analysis revealed that the majority of players follow one of four highly varied but also more or less consistent play styles. In addition, a handful of play styles that only appeared in some levels were identified. Each of these corresponds to distinct ways of playing Tomb Raider: Underworld.
The most frequent play style focused on exploring the environment and collecting as many of the numerous rewards and treasures in the game as possible. Another focused on completing the game as fast as possible, dying often along the way. One of the most infrequent play styles, comprising of just a few percent of the players, was the hard-core play style, which focused on completing the game fast, dying infrequently, making a general mockery of the enemy AI, and even locating much of the treasure and discovering many of the secrets hidden in the game along the way.
There was however much variation in how players adopted the different play styles. The number of players in each play style group varies immensely throughout the game. In some Tomb Raider levels, one play style might dominate, while in subsequent levels it might contain only a fraction of the players. The analysis basically shows that people vary their playing behavior throughout a game – either by choice or because the design of different levels encourages or forces specific behaviors. In the background of these results is the decreasing number of players that remain in the game, with only a portion completing the game, as is typical in single-player games – this may be why we see fewer unique play styles towards later levels.
The combination of play style and progression analysis provides an overview of the entire game, and forms the basis for evaluating the overall patterns. The next step is to perform drill-down analyses on specific clusters and see how they fare in detail for specific levels – or even segments of levels. This is where we for example identify problems blocking progression for specific groups of players. We will cover this process in a future post.
If you would like to read more about progression and play style analysis, the reading list recently published here on the blog provides a good starting point.
About the author
Anders Drachen is a veteran Data Scientist and Lead Game Analyst at Game Analytics. His work in the game industry as well as in data and game science is focused on game analytics, business intelligence for games, game data mining, game user experience, industry economics, business development and game user research. He is one of the most published experts worldwide on the topic of game analytics, user research, game data mining, and user profiling, having authored more than 60 research publications on game analytics, user testing, and business intelligence in game development, and co-edited the first book on the topic.