Critic Score Prediction
Predictive model and interactive dashboard for score estimation.
Overview
This project predicts critic scores from structured game metadata and visualizes model behavior in an interactive app.
It combines data preparation, feature engineering, model training, and explainability into one workflow.
Problem
Score outcomes are influenced by many interconnected factors, making manual estimation inconsistent.
Teams often need a quick baseline estimate before launch planning and marketing decisions.
- —Raw datasets include noisy and incomplete fields.
- —Feature interactions are non-linear and hard to reason about directly.
- —Predictions should be understandable, not just accurate.
Constraints
Data quality
Training data required careful cleaning to handle null values, outliers, and category imbalance.
Interpretability
Users needed confidence in predictions through visible feature impact and error context.
Interactive speed
Inference and chart updates had to remain responsive in a browser-based UI.
Key Decisions
Feature engineering pipeline
Built repeatable preprocessing and encoding steps to keep training and inference consistent.
Model comparison pass
Evaluated multiple regressors and selected the best trade-off between accuracy and stability.
Streamlit-first interface
Used Streamlit to rapidly ship an interface for prediction, diagnostics, and exploration.
Results
- —Shipped a working score prediction interface for quick experimentation.
- —Exposed feature influence to improve trust in model output.
- —Created a reusable ML app template for future predictive tools.
Takeaways
- —Data preparation quality has more impact than model swapping in many ML products.
- —Simple interpretability features dramatically improve user confidence.
- —Usable tooling turns models into products.