基于有意义扰动掩码的频谱预测解释方法

An Interpretation Method of Spectrum Prediction Based on Mask with Meaningful Perturbations

针对频谱序列预测问题中深度学习技术可解释性不足、现有方法解释效果不直观以及时间相关性难以体现等问题,提出了一种基于掩码方案的频谱预测解释方法。首先,生成与输入频谱数据同样大小的重要性掩码矩阵,通过显著图标注输入数据的重要性部分,获得对单一样本预测结果的可视化解释;其次,将解释问题转变为针对掩码的多目标优化问题,根据频谱数据的动态特性与相关性特点改进扰动方式,实现针对频谱预测问题的有意义扰动;最后,通过在优化目标中添加对时间步跳跃的惩罚项,体现了短的连续序列或者相邻时间步的时间相关性同样重要的先验知识。基于实测频谱数据的测试分析表明,所提的解释方法具有简洁直观和易于用户理解等特点。与基线方法相比,所标注的重要性部分凸显了中心频点和相邻频点的相关性。在性能恶化实验中,模型输出精度下降最多,平均绝对百分比误差(Mean Absolute Percentage Error,MAPE)指标平均分别比综合梯度方法、沙普利值采样和高斯扰动掩码方案高6.4%,26.2%和30.0%;在性能恢复实验中,模型输出精度改善最大,MAPE指标平均分别比前述三种对比方案低7.6%,32.2%和32.8%。

The deep learning based spectrum prediction approaches have the inherent insufficient interpretability feature,and existing interpretable methods can neither provide intuitive interpretations nor reflect the temporal correlation enough.For above problems,a spectrum prediction interpretability method based on masks is proposed.Firstly,by generating the salience matrix with the same shape as the spectrum inputs and noting the important part of the input data through saliency map,the visual interpretation of the prediction is obtained.Secondly,the interpretation problem is transformed into a multi-objective optimization problem for the masks.According to the dynamic characteristics and correlation in the spectrum data,the perturbation mode is improved and a meaningful perturbation method for the spectrum prediction is proposed.Finally,by adding the penalty term of timeslots’jump to optimization goals,the prior knowledge that the time correlation of the short continuous sequence or adjacent time steps is equally important is reflected.In the interpretation experiment on the real-collected spectrum data,it is shown that the proposed scheme is brief,intuitive and easy for users to understand.Compared with the baseline methods,the important part highlights the correlation between the central frequency and the adjacent frequency.The accuracy of the model decreases the most in the performance descending experiment.The mean absolute percentage error(MAPE)is 6.4%,26.2% and 30.0% higher than that of the Integrated Gradient,Shapley Value Sampling and the mask with Gaussian perturbations respectively.In the performance recovery experiment,the accuracy of the proposed model is shown to be recovered,and the MAPE is 7.6%,32.2% and 32.8% lower than that of the mentioned counterparts,respectively.