Predict player behaviour with the world's largest Alexa gameplay dataset

Featuring 6,800 hours of gameplay, including 380,000 decisions made by 14,000 players across 50,000 games

Spot patterns, analyse user behaviour and predict player performance with this real-world curated dataset prepared especially for Machine Learning enthusiasts. Perfect for beginners, with the depth to satisfy the curiosity of expert practitioners.

Updated on an ongoing basis, this dataset records the actions of users playing a simple number guessing game.

The data

Players were tasked with guessing an unknown number from 1-100. After each guess they were advised whether the target number was higher or lower. Data was collected from an Amazon Alexa skill developed by the dataset author.

Data files

Three files are provided. All reflect the same data, just cut in different ways for your convenience. You can work with whichever best suits your objectives.

• games.csv - details of all games (date, player, outcome). The best place to start to understand the data.
• guesses.csv - all players decisions , including statistical calculations to aid interpretation, analysis and model training.
• repeat-player-prediction.csv - details of each player's first game with a label indicating whether they later returned to play again or not.

Beginner Challenges

• Predict whether a first-time user will return
• Predict the number of guesses a user will need to guess the target number

Intermediate Challenges

• Predict if a user will continue or end a session
• Predict a user's next guess

Expert Challenges

• Predict when a regular player will next play

Data Inclusions/Exclusions

• Invalid guesses (those not understood by Alexa, or falling outside the range 1-100) were discarded.
• Repeated guesses and guesses that conflicted with previous hints were retained.
• Only completed games with no more than 15 guesses are included.
• Alexa user IDs have been anonymised to protect player privacy.