A novel discriminative model for online behavioral analysis with application to emotion state identification can extract discriminative characteristics from behavioral data and find the direction of optimal projection, leading to better utilization of behavioral information to produce more accurate recognition results.