This paper presents a novel feature selection method for classification of high dimensional data, such as those produced by microarrays. It includes a partial supervision to smoothly favor the selection of some dimensions (genes) on a new dataset to be classified. The dimensions to be favored are previously selected from similar datasets in large microarray databases, hence performing inductive transfer learning at the feature level. This technique relies on a feature selection method embedded within a regularized linear model estimation. A practical approximation of this technique reduces to linear SVM learning with iterative input rescaling. The scaling factors depend on the selected dimensions from the related datasets. The final selection may depart from those whenever necessary to optimize the classification objective. Experiments on several microarray datasets show that the proposed method both improves the selected gene lists stability, with respect to sampling variation, as well as the classification performances.