Interviewed By James Freeman
Recently I came across a site called Board Game Finder. I am always curious about finding new games so i figured i would give it a shot. As I typed in games that i love it showed me a list of other games I might be interested in (many of which i already owned), this got me very curious about the inner workings of the website. After messing around with picking the top 5 games from some of my favorite game mechanics and categories to try and find new games to try i contacted the designers. They were kind enough to give me a few minutes of their time and break down some of the math and logic behind the site.
How did board game finder get started?
As most projects, it all started doing something unrelated. We downloaded the data from boardgamegeek.com for a small PhD class project that had nothing to do with recommendations. After that we had the idea to apply a recommendation system. We tried different recommendation methods and noticed that some worked well. Building a website was the natural next step to allow everyone to interact with the system. We met a few times to define the project in detail, and after that it all went smooth.
How long has the team been working on the project?
We started somewhere between late spring and early summer. We must say that all of us have our jobs, so Board Game Finder was a hobby project for which we worked in our free time: weekends and nights mainly. In many weeks we were busy and the progress was zero. Some other weeks were very productive. But overall, we have been working for a few months.
What was the biggest obstacle to overcome?
We found many big challenges. One of them was to build an application with an intuitive user interface, simple to use, and fast enough despite all the algorithmic processing behind the scenes. Another one was to develop an algorithm that works well for a very small number of likes, getting rid of the most popular games when needed. It was also difficult to assess the quality of the recommendations of different algorithms. From a logistic point of view, the biggest challenge was to decide on a name for the website. But the biggest overall challenge is still unsolved: finding good profile pictures for the team members. We had some great ideas for funny profile pictures but we had to resort to our standard pictures from our websites.
What is a matrix factorization approach?
In the context of recommender systems, matrix factorization is just the mathematical formalism behind the intuition that we all have. It basically states that each user has its own preferences, each game has its own attributes, and the rating that a user provides for a game is a function of that user's preferences and that game's attributes. In matrix factorization, these preferences and attributes are considered latent (unobserved), and our goal is to learn them from the observed ratings. In Board Game Finder, we use a model that combines matrix factorization with some other ideas from our recent research.
So in time as people rate more games future recommendations will become more accurate based on what people with similar taste have played?
Yes, eventually. Machine learning methods rely on data; in principle, the more data you collect, the better your predictions will be. This is why data is so precious nowadays, and why there are companies that are interested in selling or buying consumer data. That said, we must also keep in mind that the current database from boardgamegeek.com is fairly large, so this will be a slow process before the learned model shifts significantly. We also have some ideas to make the predictions more accurate by feeding meta-data into the model, but this is still ongoing work.
Has there been a game recommended to you specifically that you had never heard of that you now love?
Yes, definitely. Each of us has discovered something that sparked his interests. A few examples are T.I.M.E. Stories, Mysterium, Pandemonium, Mythos Tales, and StarCraft The Board Game. Some of us have also discovered games to give as gifts for our relatives or friends; two examples are Patchwork and Arboretum.
Weekly Blog Schedule.
Friday - Bi-Weekly
Saturdays - Bi-Weekly
A Gloomhaven Campaign
First Sunday Of The Month