A new LogitBoost algorithm for multiclass unbalanced data classification

Jie Song, Xiaoling Lu, Miao Liu, Xizhi Wu

Abstract: LogitBoost algorithm is an extension of Adaboost algorithm. It replaces the exponential loss of Adaboost algorithm to conditional Bernoulli likelihood loss. LogitBoost-J algorithm further extends the LogitBoost to multiclass situation. But like LogitBoost algorithm and Adaboost algorithm, LogitBoost-J algorithm is not suitable for unbalanced data classification. This paper proposes a new LogitBoost algorithm for multiclass unbalanced data classification. The experiment on practical data shows that this new algorithm performs better than LogitBoost-J algorithm and is competitive to BABoost algorithm.