Abstract:

The goal of a ranking learning algorithm is to obtain the best ranking function, which assigns each instance a score, and ranks instances according to their scores. The framework of a push ranking algorithm allows a certain range of ranking errors in the learning procedure. Let ε be the range of the errors, by using the symmetric εinsensitive exponential loss function and the symmetric εinsensitive logistic loss function to substitute the original loss function, two new classes of push ranking learning algorithms can be obtained. The experimental results show that the proposed new algorithms are effective.

Key words: ranking;bipartite ranking;push ranking;symmetric εinsensitive exponential loss function;symmetric εinsensitive logistic loss function