The two Gamasutra features by Dmitry Nozhnin, head of analytics and monetization at Innova, are good examples of useful case examples from developers using analytics. Dmitry provides more than just an overview, digging into the specifics of features and models, which is a level of granularity beyond what is commonly published in the area (often due to confidentiality issues). While they are about a year old, I wanted to bring them to the attention of the community as they contain a lot of experiences from a team that went into data mining with very limited prior skills, and succeeded in predicting churn. The articles show that getting into the second-tier domain of Machine Learning is actually feasible even for people without a degree in Statistics, Data Mining and Computer Science.
In the first article, Dmitry covers the process Innova developed for predicting churn of new users. He describes how they ended up being able to with a high degree of precision predict when players churn, but having a harder time with defining why players churn. This is a common problem, and while we need more research on this, the explanation is possibly related to the fact that while human behavior is often seen following particular patterns, especially at large scales, the underlying motivations and reasons for each individual following that pattern, can vary substantially (Nick Lim from Sonamine touched on the “why” side of things recently). This highlights the benefit of integrating user-oriented analytics with game user research.
In the second article, Dmitry describes their methods for defining who veteran players are and predicting when they quit. One of the main conclusions is that defining the right features for analysis – this case the parameters for churn – is absolutely vital. This is an experience I definitely share.
One of the key conclusions Innova makes is that you get pretty good precision just by features that are derivates of playtime and activity – in their case days of activity and daily playtime (for the veteran players). They are not the only people to find this interesting pattern. Nick Lim mentioned it earlier, and the result has cropped up in e.g. Network Science and Computational Intelligence, as well as Web Analytics (some examples: , , ). I am working on a longer writeup of this particular pattern of player behavior.