Study on Cost Prediction Modeling with SVM Based on Sample-Weighted

In the process of cost prediction modeling with support vector machine (SVM), the prediction accuracy is significantly impacted by the similarity between training samples and the predicted object. In traditional cost prediction modeling, the training data must be independent and identically distributed and every sample participating in training is treated equally. However, different samples owe the different contributions to the final prediction model in practice. Considering the characteristic information of the predicted object, the sample-weighted method based on the prediction error and sample similarity were proposed to reflect the contribution levels of samples to the model. Further, two kinds of combination strategies, such as the sum of weights and the produce of weights were proposed to compare the prediction performance. Experiments show that the prediction results can be effectively improved through sample-weighted. In both combination strategies, the prediction effect is better by the product of weights than by the sum of weights, which can be extended in practical applications.