What's it doing?
An article describing how this recommender system was created was published in Significance, the magazine of the Royal Statistical Society and American Statistical Associationhttps://rss.onlinelibrary.wiley.com/doi/10.1111/j.1740-9713.2019.01317.x
Imagine you have just finished playing a boardgame that you really enjoyed and asked those around you if they could recommend another game. Recommendations are likely to be more reliable if they come from someone who also enjoy the game you have just played, and they have a broad knowledge of other board games. If you could ask thousands of people who have played the game and collate their responses, taking account of whether they did or did not like the game, it should be possible to devise a useful recommendation algorithm. This application does this by utilising the ratings given by users of the BoardGameGeek website.
The method recommends those games that provide the biggest ratings contrast between those users who enjoyed the original game and those who either did not enjoy it or have not rated it. This contrast is scaled to lie approximately on the 0 to 10 range and shown as the User Score. The User Score is also adjusted to remove an unwanted bias that is evident between games published in a similar time period. Clearly the more users who have rated both games, the more reliable will be the User Score. A Bayesian method has been used to combine information available prior to anyone playing the game with the information contained in the ratings data. This prior information comes from the attributes of each game, such as the designer, mechanics and playing time. The Game Score, scaled to be comparable to the User Score, quantifies the prior information. The Bayesian method combines these scores, weighting them according to the relative precisions of the User and Game Scores, to give the Like Score. It is this value that is used to order the recommendations.
The Like Scores are also used to provide a Strength rating, with 3 stars implying a strong recommendation. Clearly it is not possible for the algorithm to know which aspects of the game particularly appealed to you, so there can be no guarantee that any of the recommendations will be good for you. All that can be inferred is that the games recommended are liked significantly more by those who liked the original game than those who did not like it or have not rated it. Also, games are only recommended if those who liked the original game gave it a good rating (>7) on average.
As well as computing the Like Score the Bayesian method also computes the precision of this score. A Surety rating is provided that represents this precision. A Surety of 3-stars implies high precision and so one can be confident in the reliability of the Like Score. Typically, low Surety ratings will be associated with those games that have had relatively few user ratings.
Glossary