A cross-region transfer learning method for classification of community service cases with small datasets

Abstract

The precise classification of community service cases is the most fundamental aspect of intelligent community service systems. However, data imbalance makes it challenging to achieve the desired level of precise classification. Existing transfer learning methods use open Internet knowledge for identifying case features and mining potential feature relations. However, community service cases have the characteristics such as short text length and non-public content, which restrict the transfer learning modes. In this paper, a cross-region transfer learning method is proposed to solve the classification problem of cases with small datasets in data imbalance situation while considering the perspective of regional cooperation. First, an ontology modeling method is applied to standardize the case features, reducing the effect of semantic ambiguity on transfer results. Secondly, to improve the effectiveness of source domain classification, this paper utilizes an extended marginal fisher analysis where the distance is measured by the inner product between data. Next, the mapping matrix from target domain to source domain is learned through domain adaptation. Finally, the method is verified based on the empirical data from Lanzhou and Beidaihe in China. Experimental results on classifiers show the proposed method helps regions to improve case classification rates significantly through knowledge complementation. The proposed approach can be followed to build case-based community service systems of reasonable accuracy and limited sample sizes.