As part of my capstone project at Galvanize, I've created a recommender system to connect Steam users based on similarities in gaming preference. As an avid gamer myself, I would like to apply my knowledge of Data Science to a topic that I greatly enjoy.

Valve currently does not have a recommendation engine in place to for users to connect with one another.

• Collected information on 250,000 users, 12,500 games, and a total of 15,941,228 ratings

• Dimensionality Reduction: Removed games that saw an average playtime of less than 30 minutes across all users that own the game. The rationale for that is that games that have really low total playtime are probably not worth recommending. After filtering, I am left with 8484 games

• Game specific information including:

  • Game Price
  • Detailed Description of Game
  • Release Date (MMM YYYY)
  • Metacritic rating
  • Game Genres
  • Top 5 community rated tags (if available)

• As the utility matrix in question is sparse (~0.75% filled), a ranking matrix factorization with collaborative filtering were employed to predict ratings for unobserved data

• Users were recommended based on the degree of cosine similarity it shares with one another

• Two ratings model were created to help evaluate the performance of the recommender system:

  • Base rating model - ratings are directly proportional to the time the user spent playing a particular game
  • Random rating model - ratings are randomly generated and the proportions add up to 1

• To see if the base rating model performs better than random, a sample of 3,000 users were selected to score the two models

• Each of the two models generated a list of top 100 recommended users, and that list was compared with the actual friends the user has (Steam API has that function). The assumption here is that people who are friends share a degree of similarity in gaming preferences

• As illustrated by the table below, the base rating model outperforms the random model quite significantly. Please note that there are over 125 million Steam users altogether, so the low score observed should not be surprising

  Model	Score (Max 3000)
  Base Rating Model	59
  Random Rating Model	41

• The current model could only recommend users that are trained. I would like to extend recommendations to new & unseen users

• After matching two users, I would like to further suggest a multi-player game for them to play together

• Build an interactive web app