11-09-2017, 12:09 PM
With the explosive emergence of vertical search domains, the application of the base classification model extends directly to different domains is no longer desirable because of domain differences, while building a unique ranking model for each domain is laborious for the labeling of data and time for training models. In this paper, we address these difficulties by proposing an algorithm based on regularization called SVM (RA-SVM), through which we can adapt an existing classification model to a new domain, so that the amount of tagged data and the cost of training is reduced while performance is guaranteed. Our algorithm only requires prediction of existing classification models, rather than their internal representations or auxiliary domain data. In addition, we assume that similar documents in the domain-specific feature space must have consistent classifications and add some constraints to control the margin and the RA-SVM slack variables adaptively. Finally, it is proposed to measure adaptability to quantitatively estimate whether an existing classification model can be adapted to a new domain. Experiments on Letor and two large-scale datasets traced from a commercial search engine demonstrate the applicability of the proposed classification adaptation algorithms and the measure of classification adaptability.
It can be understood in the following video:
It can be understood in the following video: