A popular approach to recommender systems is to factor the ratingmatrix into two low rank matrices. Although this has proven to be agood way to predict ratings, the component scores in these matrix fac-torisation methods don’t serve as estimates of the underlying causesof the ratings. We propose to model the rating matrix using Multidi-mensional Item Response Theory (MIRT). We show that MIRT doesallow for a causal interpretation of the latent scores, and additionally,allows to control the exploration-exploitation trade off using comput-erized adaptive testing (CAT). We adapt MIRT to large proportionsof missing data using elastic net regularisation. Our model is vali-dated in parameter recovery simulations, and is applied to a binarizedversion of the movielens 1M dataset. Results show that parameterscan be recovered accurately for models with up to three dimensions,for various proportions of missing data. Performance on the movielensdataset with a three dimensional MIRT model–performing is on parwith matrix factorisation methods of up to 20 dimensions. In a secondsimulation study, we show how MIRT can be combined with CAT totackle the exploration-exploitation tradeoff, resulting in more accurateitem rankings. We conclude that MIRT can provide both more inter-pretability and more control over the exploration-exploitation tradeoff.We discuss directions for future work on MIRT based recommendation.