今話題？の AROW (Adaptive Regularization Of Weight Vectors) の oll 向けの実装を見かけたので，Confidence-Weighted (CW) が Passive-Aggressive (PA)/LIBLINEAR に比べて全然ダメ (PA/LIBLINEAR: 91%, CW: 88%; Perceptron: 89%) だった例のタスク（学…