🎲 Board games recommender system

See the app deployed here.

Team

• Matěj Bukáček
• Petr Janík
• Jakub Kraus
• Michal Salášek

Consultation 1 - April 4th 10:00

• problem formulation, clarification of the purpose, for whom, why

  • recommendation of board games to users who like to play board games and want to try new ones

• planned usage

  • website visitor rates 10 games and gets paged recommendations sorted from the most relevant
  • on each page there will be 5 recommendations from each algorithm randomly interleaved
  • for each recommendation there will be a button to reveal the algorithm used

• hypothetical business model

  • referral links / eshop

• specific data with some basic analysis (value distributions etc.)

  • dataset:
    • relevant columns:
      • games
        • ImagePath - Image http:// path
      • user_ratings
        • BGGId - BoardGameGeek game ID
        • Rating - Raw rating given by user
        • Username - User giving rating
  • data preprocessing:
    • removed NaNs from usernames - 63 ratings removed
    • there were multiple ratings of the same game by some users - duplicates removed, only the latest rating kept ( ratings don't have timestamps, so we assumed file user_ratings.csv is ordered chronologically and kept only the last occurence of Username - BGGId pair in the file) -> 32 687 ratings removed (0.17%)
    • removed users with less than 10 ratings
    • removed games with less than 10 ratings
  • analysis (after data preprocessing):
    • 18 340 221 user ratings
    • 224 556 users
    • 21 919 games, 157 mechanics, 217 themes, 10 subcategories (only 10 030 games (45.76%) belong to some subcategory)
    • density of user ratings matrix: 0.37%
    • average number of ratings per user: 81.67
    • median number of ratings per user: 39
    • maximum number of ratings per user: 6478
    • average rating: 7.10 (range [0, 10])
    • 
    • 
    • 
    • 

• specific proposal for algorithms that you want to implement

  • memory based CF (Pearson correlation coefficient, item-based for easier interpretation of results) - Jakub Kraus
  • model based CF (gradient descent) - Michal Salášek
  • content based (TF-IDF) - Matěj Bukáček
  • possibly another content based

• implementation of some naive baseline

  • for CF (prediction is one number):
    • global average rating
  • for content based (prediction is a list of games):
    • top-rated games (by available Rank:boardgame property)
    • most rated games
    • random games

• basic idea of evaluation approach

  • because we are doing a prototype project, we will focus more on the manual evaluation and less on the technical evaluation
  • technical (only for CF):
    • data will be split into 80% and 20%
      • at the very end, once we choose the best model, we will use the 20% of data for final evaluation
      • 80% will be used for training
        • it will be split into 80% and 20% using 5-fold cross-validation
          • for models that require validation set (to stop the training) this 80 % will be split again into 80% and 20%
    • used metric will be RMSE
  • by real users’ happiness:
    • we will give our system to real users and collect their feedback
  • by archetype users:
    • we will create several archetypes of users and evaluate our system on them (by seeing if the recommendations are what we would expect)

Consultation 2 - April 25th 10:00

• functional demo of several recommendation algorithms (at least some basic version)
• preliminary subjective evaluation (does it seem to do something meaningful?)
• specific examples of "interesting" recommendations (both good and poor)
• detailed plan for evaluation
• specific problems encountered, questions