准确度为 1.0，同时训练损失和验证损失仍在下降

Question

我创建了一个 LSTM RNN，以便根据 gps 坐标预测某人是否在开车。这是数据示例（注意：x、y、z 是从 lat、lon 转换而来的 3d 坐标）：

                        x           y           z       trip_id,mode_cat,weekday,period_of_day
datetime            id                          
2011-08-27 06:13:01 20  0.650429    0.043524    0.758319    1   1   1   0
2011-08-27 06:13:02 20  0.650418    0.043487    0.758330    1   1   1   0
2011-08-27 06:13:03 20  0.650421    0.043490    0.758328    1   1   1   0
2011-08-27 06:13:04 20  0.650427    0.043506    0.758322    1   1   1   0
2011-08-27 06:13:05 20  0.650438    0.043516    0.758312    1   1   1   0

当我训练我的网络时，我的 training_loss 和 validation_loss 都下降了，但准确率在第一个 epoch 达到了 1.0。我确保我的训练和测试数据不一样。以下是我如何拆分训练和测试数据：

t_num_test = df["trip_id"].iloc[-1]*4//5
train_test_df = df.loc[df["trip_id"]<=t_num_test].copy(deep=True)
test_test_df = df.loc[df["trip_id"]>t_num_test].copy(deep=True)

features_train = train_test_df[["x","y","z","datetime","id","trip_id","mode_cat","weekday","period_of_day"]]
features_train.set_index(["datetime","id"],inplace=True)
dataset_train_x = features_train[["x","y","z","trip_id","weekday","period_of_day"]].values
dataset_train_y = features_train[["mode_cat"]].values

features_test = test_test_df[["x","y","z","datetime","id","trip_id","mode_cat","weekday","period_of_day"]]
features_test.set_index(["datetime","id"],inplace=True)
dataset_test_x = features_test[["x","y","z","trip_id","weekday","period_of_day"]].values
dataset_test_y = features_test[["mode_cat"]].values

下面是我建立网络的方式：

single_step_model = tf.keras.models.Sequential()
single_step_model.add(tf.keras.layers.LSTM(1,
                                           input_shape=x_train_single.shape[-2:]))
single_step_model.add(tf.keras.layers.Dropout(0.2))
single_step_model.add(tf.keras.layers.Dense(1, activation='sigmoid'))

single_step_model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), loss='binary_crossentropy',
                          metrics=['accuracy'])
.
.
.
single_step_history = single_step_model.fit(train_data_single, epochs=epochs,
                                            steps_per_epoch=evaluation_interval,
                                            validation_data=test_data_single,
                                            validation_steps=60)

And here is the graph displaying training_loss, validation_loss and accuracy

是什么导致了这个结果？如果重要的话，我将使用大约 500,000 个数据点和大约 8000 个唯一 trip_id.

请指教

编辑： # of Driving/Not Driving (Mode_cat: 1/0)

Answer 1

希望对您有所帮助！

我能想到的案例很少

您的数据集有偏差。它会导致大部分输入数据倾斜吗？检查其中 mode_cat 值的百分比。（都是全1，还是大部分全1？）
你的 X 值可以有一个 feature/column 是函数 y 是 x 值的函数（比如 y_val = m * x_col2 + x_col3 ?)
Accuracy 很好学，但尝试使用像 f1 score/confusion_matrix 这样的东西。

Link:

https://scikit-learn.org/stable/modules/generated/sklearn.metrics.f1_score.html#sklearn.metrics.f1_score

https://scikit-learn.org/stable/modules/generated/sklearn.metrics.confusion_matrix.html#sklearn.metrics.confusion_matrix

准确度为 1.0，同时训练损失和验证损失仍在下降

Accuracy of 1.0 while Training Loss and Validation Loss still decreasing

python

deep-learning

lstm

keras

tensorflow