The document proposes using support vector machines (SVMs) with tree kernels to classify and rank moves in the game of mahjong based on a player's hand. Tree kernels allow SVMs to operate on tree structures representing mahjong hands without requiring feature engineering. An experiment applies this method to positions from expert mahjong games, achieving 57% accuracy in ranking expert moves first. The method shows promise for classifying positions that linear features cannot and could be improved by incorporating additional information from the game state.