Overview:

An interest in games lies in many of us—including myself—but only a few have found significance in the data behind those games. I grew up playing Clash Royale, and encountering a large dataset from the game’s top players sparked a deep curiosity in me. As described on the Clash Royale Wiki, the game is “a fast-paced brawler where you collect cards and duel players in real time. Destroy your opponent's Crown Towers, but be sure to defend your own.” Each match features two players, each with a deck of eight character cards that can be deployed using a finite elixir resource.

Dataset:

The dataset I’m using is publicly available on Kaggle: https://www.kaggle.com/datasets/s1m0n38/clash-royale-games/data

It consists of match statistics from top-ranked players spanning 15 seasons—from September 2022 to December 2023. Each row in the dataset represents a match and includes 23 columns, covering: Match Details: Date, game mode, and other contextual info Player Data (2 Players per Match): Unique IDs Trophy count before the match Number of crowns left (the target variable) 8 cards used in the deck (represented as unit IDs)

Objective:

The goal of this project is to predict the number of crowns left in a match using the average elixir cost of each player’s deck, employing supervised learning techniques.

Project Pipeline:

1. Data Setup & Environment Configuration (Appendix A)

• Set up Python virtual environments using Google Cloud Platform

• Configure Kaggle API access

• Install all required libraries and utilities

2. Exploratory Data Analysis (EDA) (Appendix B)

• Perform EDA on multiple CSV files containing match data

• Assess structure, identify missing values

• Analyze card (unit) usage distribution

3. External Data Integration & Cleaning (Appendix C)

• Map unit IDs to actual card names using an external JSON API

• Clean and preprocess data for further modeling

4. Machine Learning Pipeline with PySpark (Appendix D)

• Use PySpark to construct a Logistic Regression model

• Apply preprocessing steps including:

  • Feature encoding

  • Feature assembling

  • Training/testing data split

• Evaluate model performance using:

  • Accuracy

  • Precision

  • Recall

  • F1 Score