Game Reviews & Sales Analysis

Project Overview

This project explores the relationship between video game reviews, sentiment, and sales performance.
We combine text analysis (NLP) and machine learning models to answer:

• Do player sentiments in reviews predict game sales?
• How do genre and platform affect performance?
• Which machine learning models work best for sentiment prediction and sales prediction?

---

Dataset

We worked with two combined sources:

• Steam reviews dataset (dataset.csv) → game reviews and text. (Can be found here: https://www.kaggle.com/datasets/andrewmvd/steam-reviews )
• VG Sales dataset (vgsales.csv) → sales numbers, genre, platform, year. (Can be found here: https://www.kaggle.com/datasets/gregorut/videogamesales )

After preprocessing, we built a unified final_df with:

game	genre	platform	review_text	sentiment	sales	year	sentiment_score

---

Steps & Techniques

1. Data Cleaning

• Removed missing values and duplicates.
• Cleaned review text: lowercasing, punctuation removal, stopwords, lemmatization.
• Fixed formatting issues (spacing between words).
• Removed "Unknown" genres from analysis.

2. Exploratory Data Analysis (EDA)

• Distribution of genres and platforms.
• Sentiment distribution (positive, negative, neutral).
• Sales distribution by genre/platform.
• ANOVA + Regression checks to test statistical significance.

3. SQL Integration

• Data stored in SQLite database (gaming_reviews.db).
• Queried sales by genre/platform and sentiment distribution.

4. Hypothesis Testing

• H1: Games with positive sentiment have higher sales.
• H2: Genre & platform explain sales differences better than reviews alone.
• H3: Review sentiment predicts future success.

5. Regression Analysis

• Linear regression → Sentiment vs. Sales.
• Extended regression → Sales ~ sentiment + genre + platform.
• Result: genre/platform stronger predictors than sentiment alone.

6. Classification (Sentiment Prediction)

• Transformed reviews into TF-IDF features.
• Models tested: Logistic Regression, Random Forest, XGBoost.
• Best: Random Forest → Accuracy 91%.

7. Regression (Sales Prediction)

• Target: game sales (numeric).
• Features: review TF-IDF + sentiment score + genre/platform.
• Models tested:
  • Random Forest Regressor
  • Gradient Boosting
  • Ridge & Lasso Regression
• Best: Random Forest (R² = 0.46, RMSE ≈ 0.87).

8. Model Insights

• Sentiment classification → high accuracy.
• Sales regression → moderate predictive power.
• Genre & platform more influential than raw reviews.
• Missing real-world factors (marketing, hype, brand) likely limit predictive power.

---

Visualizations

• Top 10 game genres (bar plot).
• Sentiment distribution.
• ANOVA plots for sales by sentiment.
• Feature importance plots for sales models (words, genres, platforms).

---

Results Summary

• Sentiment classification worked very well.
• Sales regression was limited (R² ≈ 0.46).
• Business context (marketing, release timing, brand loyalty) needed to explain sales more fully.

---

Next Steps

• Incorporate external factors (marketing spend, social media hype).
• Try deep learning models (LSTMs, transformers) on larger compute resources.
• Build a dashboard (Streamlit) for interactive exploration.

---

Tech Stack

• Python (pandas, numpy, matplotlib, seaborn, scikit-learn, statsmodels)
• NLP (NLTK, scikit-learn TF-IDF)
• SQL (SQLite)
• ML Models: Logistic Regression, Random Forest, XGBoost, Ridge, Lasso, Gradient Boosting