Little Miss Fits
Can you solve her problem?

Welcome stranger!

Please tell me who you are.

About Little Miss Fits

Little Miss Fits is a Game with a Purpose, developed at University of Duisburg-Essen in order to investigate ways to make automatically learned models for recommending more transparent.

But what does that mean?!

Recommender systems are widely used today. We know them from websites such as Amazon, Spotify and Netflix. From an algorithmic point of view most modern methods for recommending analyze past user ratings statistically and derive predictions for future ratings. Thereby algorithms often rely on abstract models deduced from these ratings. Algorithms using such models have become very mature over the last decades. Yet, it typically remains opaque to users how and why items are recommended. While there are first attempts to explain recommendations to users (think of Amazon's popular "Users who bought ... also bought..." explanations), recommendations resulting from pre-trained models cannot be explained that easily.

One of the most popular model-based recommending algorithms is Matrix Factorization. You can imagine its underlying model as a huge multidimensional space in which items (e.g. movies) are placed according to how they have been rated. Each dimension of that space pertains to another characteristic implicitly expressed in user ratings. Thinking of movies, one dimension could denote whether a movie has many cuts. Unfortunately, dimensions are not necessarily that explicit and comprehensible. In fact, little is known about the presence of real-world semantics in models of Matrix Factorization.

With Little Miss Fits, we target at shedding light on semantics in these models. For each round of the game a set of movies is generated, containing 4 movies that are positioned high on one dimension of the underlying model. These movies are mixed with one movie that is low in this but high on another dimensions. If dimensions of the recommending model are understandable to humans, you, as a player, should be able to find the mismatching movie. By collecting many rounds played and analyzing the rate of successfully found mismatches, we can understand whether dimensions were perceived as distinctive and comprehensible. This could help us explaining recommendations better in the future.
Hence, with each time you play the game, you help us to make recommending algorithms more transparent, comprehensible and user friendly!

Thank you for playing Little Miss Fits! :)

For further information, please visit the projekt website.