This paper is concerned with a mean-variance portfolio optimization model with cardinality constraint for generating high-quality lists of recommendations. It is usually difficult to accurately estimate the rating covariance matrix required for mean-variance portfolio optimization because of a shortage of observed user ratings. To improve the accuracy of covariance matrix estimation, we apply shrinkage estimation methods that compute the weighted sum of the target and sample covariance matrices, and we propose two types of target matrices that work well for shrinkage estimation of the rating covariance matrix. Experimental results show that with appropriate parameter tuning, our method can improve the quality of recommendation lists produced by various collaborative filtering algorithms.