Games with a Purpose for Understanding Latent Factors

Matrix Factorization is a commonly used technique in model-based Collaborative Filtering. Unfortunately, produced recommendations are often hard to comprehend by users due to the model’s statistical nature. However, some evidence exist that latent factors of such models actually contain semantic meaning. With this research line we strive to harness the motivational power of computer games in order to gain insights into semantic characteristics of latent factor models produced by Matrix Factorization.

LittleMissFits: The gameplay of Little Miss Fits consists of finding a mismatch in a set of five movies. In the background four of these belong to one dimension of an underlying factorization, whereas one movie belongs to another dimension. This movie is the mismatch to be found. By analyzing success rates of players we are able to determine whether the respective dimensions were perceived as distinctive. We also test dimensions of models that are trained with different parameters in order to evaluate if the parameterization has an influence on comprehensibility of the resulting model.

MuchoMatcho: By leveraging the crowd this Game with a Purpose allows to elicit descriptions for the dimensions of a Matrix Factorization model, this way revealing the otherwise hidden semantics of the underlying latent factors. The game follows the output-agreement method and shows representative items for each latent factor to randomly matched pairs of users. For each factor consecutively, the two players then type in terms that they think describe best the commonalities of the respective items until a match is found. By analyzing the terms collected, meaningful descriptions for each factor can be derived as a by-product of gameplay. In this manner, the online game contributes to a better understanding of latent factors in Recommender Systems.