Adaptive Knowledge Transfer Based on Locally Weighted Learning

Locally weighted learning (LWL), which is an effectual and flexible method for prediction problems, is widely used in many regression scenarios. The training data samples, referring to the history experience knowledge base, are required to help do regression by new queries. However, sometimes, the knowledge base tends to be helpless due to the lake of information, such as inadequate training data. In such cases, traditional locally weighted learning will be powerless due to less history or inappropriate experience if there are not an adaptive mechanism or other learning methods like knowledge transfer to assistant. In this paper, we propose an adaptive transfer learning mechanism to assistant LWL to do prediction. As there are many auxiliary training sets, we assign different optimal local models to take each training set as the learning basic, and combine those models into an integrated one adaptively to give the final prediction value by allocating weights for each model dynamically with the feedback prediction error. Importantly, this learning process is assigned for multi-domain knowledge bases transference and multi-locally-weighted-model integration. Moreover, we also give an analysis about how the selection of additional training domains affects the regression result. Experimental studies are based on climate data which contains the monthly average of global land air temperature from 1901 to 2002 on grids divided by 0.5 latitude and 0.5 longitude. Knowledge transfer is taken out from neighbor grids to a center. The results show that our mechanism performs much better than traditional LWL.