In response to the issues of rule redundancy and low activation rule consistency in the extended belief rule base(EBRB), this paper proposes an inference model based on a novel structural optimization framework under Relief algorithm framework and activation factor, which can be applied to classification and regression problems in machine learning. Specifically, based on the Relief algorithm, the model first assigns different weights to the extended belief rules by analyzing the relevance of historical data and its neighboring input and output to identify key rules, and achieves rule reduction by fusing with neighboring rules. Then, this paper introduces an activation factor in the process of calculating individual matching degrees and determines the value of the activation factor through offline optimization strategies to ensure the consistency and effectiveness of the activated rules. Finally, to verify the effectiveness and superiority of the proposed model, the comparison of the performance between the proposed method and some other types of EBRB models in the terms of regression and classification problems is conducted.