In this paper, an effective multi-concept classifier is proposed for video semantic concept detection. The core of the proposed classifier is a supervised classification approach called C-RSPM (Collateral Representative Subspace Projection Modeling) which is applied to a set of multimodal video features for knowledge discovery. It adaptively selects non-consecutive principal dimensions to form an accurate modeling of a representative subspace based on the statistical information analysis and thus achieves both promising classification accuracy and operational merits. Its effectiveness is demonstrated by the comparative experiment, as opposed to several well-known supervised classification approaches including SVM, Decision Trees, Neural Network, Multinomial Logistic Regression Model, and One Rule Classifier, on goal/corner event detection and sports/commercials concepts extraction from soccer videos and TRECVID news collections.