Summary: Data Abstraction

We’ve discussed different ways to reduce the dimensionality of game data.

Feature engineering

The process by which an expert develops a set of features or metrics based on their own knowledge of the game or domain. It is important to note that many industry groups use the metrics we discussed in the initial chapter. This is due to the amount of control one can have in developing such a set of metrics. Furthermore, such metrics provide parallels to many business metrics that other industries have developed. This makes it easier to use these metrics in reporting and general communication.

Feature extraction

The process where statistical models are used to develop new dimensions from the features in the raw dataset as a way to reduce dimensionality. This technique is popular; however, as we can see, it’s hard to integrate different measurement types. Often, the low-dimensional space isn’t as explainable, meaning that the new dimensions don’t have a clear meaning to the designer, thereby limiting the explainable power of the models that use them.

Feature selection

A set of techniques where a search is used to select the best features that represent the space. The technique is easy, but as with any search algorithm, it is often the case that it is difficult to converge on a global optimal. Thus, the different search methods used may provide different solutions. Further, the technique depends on the existence of a way to measure the quality of different variables or models produced, and this measure typically depends on the model produced and may be different based on what the designer or researcher is modeling.