Abstract: The problem of recommending new items to users (often referred to as item cold-start recommendation) remains a challenge due to the absence of users' past preferences for these items. Item features from side information are typically leveraged to tackle the problem. Existing methods formulate regression methods, taking item features as input and user ratings as output. These methods are confronted with the issue of overfitting when item features are high-dimensional, which greatly impedes the recommendation experience. Availing of high-dimensional item features, in this work, we opt for feature selection to solve the problem of recommending top-N new items. Existing feature selection methods find a common set of features for all users, which fails to differentiate users' preferences over item features. To personalize feature selection, we propose to select item features discriminately for different users. We study the personalization of feature selection at the level of the user or user group. We fulfill the task by proposing two embedded feature selection models. The process of personalized feature selection filters out the dimensions that are irrelevant to recommendations or unappealing to users. Experimental results on real-life datasets with high-dimensional side information reveal that the proposed method is effective in singling out features that are crucial to top-N recommendation and hence improving performance.