Feature Engineering in Game Data: Types, Missing Value and Outlier Handling

Feature engineering is a crucial step that determines the upper bound of model performance; it extracts useful information from raw data and applies domain knowledge to make features more effective for machine‑learning algorithms.

The article first distinguishes feature types: structured data (e.g., database tables) versus unstructured data (text, audio, images, video), and quantitative data (numeric measures such as player level or purchase count) versus qualitative data (categorical attributes like character gender or item class).

It then discusses missing‑value handling in game logs, noting common sources of dirty data and presenting several strategies: leaving missing values untouched (supported by some models), deleting incomplete records, filling with statistical values (mean, median, mode, group‑wise means), using model‑based predictions (e.g., regression, K‑Nearest Neighbors), and creating dummy variables to encode missingness.

For outlier handling, the article outlines detection techniques—including statistical analysis, the 3σ rule, model‑based detection, distance‑based methods, clustering, and box‑plot analysis—and suggests remedial actions such as deleting anomalous records, treating outliers as missing values, imputing with central tendencies, or simply ignoring them when appropriate.

Finally, it notes that beyond the discussed preprocessing steps, feature engineering also includes feature construction, transformation, and selection, which will be covered in future articles.