A New Anytime Classifier Basing on AAPE: Anytime Averaged Probabilistic with Weight Estimator (AAPWE)

In the before researches, most researchers often address on the classifying problem how to get higher accuracy and more efficiency. Then they have proposed some algorithms for this intention [W.S. Yerazunis, 2004; L. Androutsopoulos et al., 2000; G.L. Webb et al., 2005]. Mostly, these algorithms need very strong resource to calculate the predicted probability of each class label and to classify instance belong to the class label with the highest predicted probability. They need run a longer time and use more CPU resources to get a satisfactory result till their calculation is finished. This method doesn´t suit the online application which has unknowing times. Then, it has been recent interest in framing the classification problem as an anytime classification algorithm [K. Ueno et al., 2006; Y. Yang et al., 2007]. The idea is that we must classify instance without knowing in advance how much time and resource we have available in online applications. The Anytime Averaged Probabilistic Estimator (AAPE) algorithm [Y. Yang et al., 2007] considers an interesting variation of the problem which to our knowledge has not been addressed before. It can be interrupted at anytime and gives an instance a best predicted class label with the highest averaged probability under variance time and computing resource. Then, we improve AAPE algorithm with weight for each super parent and propose the Anytime Averaged Probabilistic with Weight Estimator (AAPWE). We regard that every super parents have a weight denoting their influence for other attributes. The bigger weight one attribute have, the better effective the attribute has for classifying instance as the superparent. At last, we find that AAPWE has more effective than AAPE through experiment.