Video Game Sales Prediction
Regression study predicting global game sales (R² > 0.7), comparing Linear, Lasso, and RBFN across a 1985–2016 chronological split.
- data
Problem
An SC1015 data-science mini-project: predict a game’s global sales from its attributes — domestic sales, rating, genre, platform, and user/critic scores — and, just as importantly, work out which model and which features actually do it best. Dataset: Video Games Sales and Rating (Kaggle).
Approach
- Cleaning: dropped nulls, removed outliers with the IQR rule, fixed dtypes, and target-encoded categoricals (Platform, Genre, Publisher, Developer, Rating) by their mean global sales.
- EDA: univariate (boxplots, histograms) and multivariate (correlation matrix, pairplots).
- A chronological split — train on 1985–2009, test on 2010–2016 — so the evaluation measures real generalization to a later era rather than random luck.
- Three models compared: Linear Regression (baseline), Lasso (L1 feature selection and overfitting control), and an RBFN (non-linear, K-means centres feeding a Ridge output).
Results
Lasso beat plain linear regression through L1 feature selection and better multicollinearity handling; the non-linear RBFN edged ahead overall with slightly lower MSE. Both cleared R² > 0.7.
What the data showed
Global sales track domestic sales most strongly — NA, EU, and other-region sales are the top numeric predictors — while Genre, Developer, and Platform carry the most signal among the categorical features.