Feature selection is considered to be a key preprocessing step in machine learning and pattern recognition. Feature evaluation is one of the key issues for constructing a feature selection algorithm. In this work, we propose a new concept of neighborhood margin and neighborhood soft margin to measure the minimal distance between different classes. We use the criterion of neighborhood soft margin to evaluate the quality of candidate features and construct a forward greedy algorithm for feature selection. We conduct this technique on eight classification learning tasks. Compared with the raw data and other three feature selection algorithms, the proposed technique is effective in most of the cases.