SHAP方法能用于多变量预测的DeepAR吗？

[bingshuangchalan]

SHAP方法不支持多变量预测，请问有什么解决方案？

代码如下：
import numpy as np
import pandas as pd
import paddle

from paddlets.datasets.tsdataset import TSDataset
from paddlets.models.forecasting import DeepARModel
from paddlets.transform.sklearn_transforms import StandardScaler
from paddlets.xai.post_hoc.shap_explainer import ShapExplainer

x = np.linspace(-np.pi, np.pi, 2000)

data = pd.DataFrame(
{
'time_col': pd.date_range('2022-01-01', periods=2000, freq='1h'),
'y1': np.sin(x) * 1 + np.random.randn(2000),
'y2': np.sin(x) * 2 + np.random.randn(2000),
'y3': np.sin(x) * 3 + np.random.randn(2000),
'y4': np.sin(x) * 4 + np.random.randn(2000),
'y5': np.sin(x) * 5 + np.random.randn(2000),
'y6': np.sin(x) * 6 + np.random.randn(2000),
'y7': np.sin(x) * 7 + np.random.randn(2000)}
)

ts = TSDataset.load_from_dataframe(
data,
time_col='time_col',
target_cols=['y1','y2','y3','y4','y5','y6','y7'],
freq='1h',
fillna_method='pre'
)

train, val = ts.split(0.8)
scaler = StandardScaler()
scaler.fit(train)
scaled = scaler.transform(ts)
train_scaled = scaler.transform(train)
val_scaled = scaler.transform(val)

deepar = DeepARModel(
in_chunk_len = 13,
out_chunk_len = 1,
max_epochs=100,
optimizer_params = dict(learning_rate=0.0002),
dropout = 0.15,
batch_size = 32, # 批量大小
patience=20,
num_samples = 101,
regression_mode="sampling",
output_mode="quantiles",
rnn_type_or_module = 'LSTM'
)
np.random.seed(2023)
paddle.seed(2023)

模型训练

deepar.fit(train_scaled, val_scaled)
se = ShapExplainer(deepar, train_scaled, background_sample_number=100, keep_index=True, use_paddleloader=False)
shap_value = se.explain(val_scaled, nsamples=100)

报错如下：

ValueError Traceback (most recent call last)
/tmp/ipykernel_35604/1646252647.py in
54 # 模型训练
55 deepar.fit(train_scaled, val_scaled)
---> 56 se = ShapExplainer(deepar, train_scaled, background_sample_number=100, keep_index=True, use_paddleloader=False)
57 shap_value = se.explain(val_scaled, nsamples=100)

/usr/local/lib/python3.7/dist-packages/paddlets/xai/post_hoc/shap_explainer.py in init(self, model, background_data, background_sample_number, shap_method, task_type, seed, use_paddleloader, **kwargs)
62 raise_if(shap_method not in ['kernel', 'deep'], 'Only support kernel shap and deep shap!')
63 raise_if_not(isinstance(background_data.freq, str), 'Only support timeindex data!')
---> 64 raise_if(len(background_data.get_target().columns) > 1, 'Only support univariate output!')
65
66 self._model = model

/usr/local/lib/python3.7/dist-packages/paddlets/logger/logger.py in raise_if(condition, message, logger)
154
155 """
--> 156 raise_if_not(not condition, message, logger)

/usr/local/lib/python3.7/dist-packages/paddlets/logger/logger.py in raise_if_not(condition, message, logger)
133 if not condition:
134 logger.error("ValueError: " + message)
--> 135 raise ValueError(message)
136
137

ValueError: Only support univariate output!

[Sunting78]

您好，抱歉，目前无法支持多变量。

[Sunting78]

长时间未回复，该issue已关闭，如仍有问题可以reopen或新开issue。