摘要:为了缓解神经网络的"黑盒子"机制引起的算法可解释性低的问题，基于使用证据推理算法的置信规则库推理方法（以下简称RIMER）提出了一个规则推理网络模型.该模型通过RIMER中的置信规则和推理机制提高网络的可解释性.首先证明了基于证据推理的推理函数是可偏导的，保证了算法的可行性；然后，给出了规则推理网络的网络框架和学习算法，利用RIMER中的推理过程作为规则推理网络的前馈过程，以保证网络的可解释性；使用梯度下降法调整规则库中的参数以建立更合理的置信规则库，为了降低学习复杂度，提出了"伪梯度"的概念；最后，通过分类对比实验，分析了所提算法在精确度和可解释性上的优势.实验结果表明，当训练数据集规模较小时，规则推理网络的表现良好，当训练数据规模扩大时，规则推理网络也能达到令人满意的结果.



Abstract:The black-box working mechanism of artificial neutral network brings the confusion of interpretability. Therefore, a rule inference network is proposed based on rule-base inference methodology using the evidential reasoning approach (RIMER). It is interpretable by the rules and the inference engine in RIMER. In the present work, the partial derivatives of inference functions are proved as the theoretical fundamental of the proposed model. The framework and the learning algorithm of rule inference network for classification are presented. The feed forward of rule inference network using the inference process in RIMER contributes for the interpretability. Meanwhile, parameters in belief rule base such as attribute weights, rule weights and belief degree of consequents are trained by gradient descent as in neural network for belief rule base establishment. Moreover, the gradient is simplified by proposing the "pseudo gradient" to reduce the learning complex during the training process. The experimental results indicate the advantages of proposed rule inference network on both interpretability and learning capability. It shows that the rule inference network performs well when the scale of the training dataset is small, and when the training data scale increases, it also achieves comforting results.



PDF全文下载地址:

http://jos.org.cn/jos/article/pdf/5920