Introduction

This is a data science and analytics project that uses historical Steam application data to build and evaluate machine learning models that predict app prices. This project aims to explore data preprocessing, feature engineering, model training and evaluation.

Data Source

The dataset is derived from publicly available Steam application metadata, containing records and over 70 features per game. Key variables include:

Pricing information (initial price, discounts)
Game attributes (genres, categories, release date)
User engagement metrics (reviews, playtime statistics)
Platform and language support
Developer and publisher information

Extensive preprocessing was required to address missing values, high-cardinality categorical variables, skewed distributions and outliers commonly observed in game pricing data.

Methods

Data Cleaning & Preprocessing

Removed invalid, duplicate and missing entries
Dropped unrelated or low-information features (e.g., identifiers, URLs, text fields not used for modeling)
Encoded categorical variables and scaled numerical features
Explored dimensionality reduction using PCA

Exploratory Data Analysis (EDA)

Examined price distributions and long-tail behavior
Analyzed correlations between pricing, reviews, and playtime
Identified outliers and nonlinear relationships

Modeling

Trained and evaluated multiple regression models, including:
- Random Forest
- CatBoost / XGBoost
Conducted hyperparameter tuning
Compared PCA vs non-PCA feature representations

Evaluation

Assessed model performance using RMSE and R²
Evaluated generalization using validation data
Interpreted feature importance to identify key price drivers

Steam Applications Price Prediction