Shapley 值：解释 LSTM 预测的维度问题

如何解决Shapley 值：解释 LSTM 预测的维度问题

我构建了一个 LSTM 来将时间序列的每个时间步分为三类之一。我的输入数据是维度（样本，时间步长 = 100，特征 = 62）。这工作得很好。

我想为此设置添加可解释性，并已转向 Shapley 值。但是尝试执行时出现错误：

explainer = shap.KernelExplainer(model = model.predict,data = XTrainS,link = "identity")

X_idx = 0

shap_value_single = explainer.shap_values(X = XTrainS[X_idx],nsamples = 100)


---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-113-1e443b4fff21> in <module>
      2 X_idx = 0
      3 
----> 4 shap_value_single = explainer.shap_values(X = XTrainS[X_idx],nsamples = 100)

~/miniconda3/lib/python3.8/site-packages/shap/explainers/_kernel.py in shap_values(self,X,**kwargs)
    181                 if self.keep_index:
    182                     data = convert_to_instance_with_index(data,column_name,index_value[i:i + 1],index_name)
--> 183                 explanations.append(self.explain(data,**kwargs))
    184 
    185             # vector-output

~/miniconda3/lib/python3.8/site-packages/shap/explainers/_kernel.py in explain(self,incoming_instance,**kwargs)
    202         # convert incoming input to a standardized iml object
    203         instance = convert_to_instance(incoming_instance)
--> 204         match_instance_to_data(instance,self.data)
    205 
    206         # find the feature groups we will test. If a feature does not change from its

~/miniconda3/lib/python3.8/site-packages/shap/utils/_legacy.py in match_instance_to_data(instance,data)
     85     if isinstance(data,DenseData):
     86         if instance.group_display_values is None:
---> 87             instance.group_display_values = [instance.x[0,group[0]] if len(group) == 1 else "" for group in data.groups]
     88         assert len(instance.group_display_values) == len(data.groups)
     89         instance.groups = data.groups

~/miniconda3/lib/python3.8/site-packages/shap/utils/_legacy.py in <listcomp>(.0)
     85     if isinstance(data,group[0]] if len(group) == 1 else "" for group in data.groups]
     88         assert len(instance.group_display_values) == len(data.groups)
     89         instance.groups = data.groups

IndexError: index 62 is out of bounds for axis 1 with size 62

模型摘要：

Model: "sequential_10"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
lstm_20 (LSTM)               (None,100,500)          1044000   
_________________________________________________________________
dropout_23 (Dropout)         (None,500)          0         
_________________________________________________________________
lstm_21 (LSTM)               (None,250)          751000    
_________________________________________________________________
dropout_24 (Dropout)         (None,250)          0         
_________________________________________________________________
dense_10 (Dense)             (None,3)            753       
=================================================================
Total params: 1,795,753
Trainable params: 1,753
Non-trainable params: 0

我真的很感激任何帮助或指导！