We propose a method for jointly learning multiple related matrices, and show that, by sharing information between the two matrices, such an approach allows us to improve predictive performances for items where one of the matrices contains very sparse, or no, information. While the above justification has focused on recommender systems, the approach described is applicable to any two datasets that relate to a common set of items and can be represented in matrix form. Examples of such problems could include image data where each image is associated with a set of words (for example captioned or tagged images); sets of scientific papers that can be represented either using a bag-of-words representation or in terms of their citation links to and from other papers; corpora of documents that exist in two languages.