铅笔与橡皮

外部exp模型以及平均值的梯度积分算法

 662  2 分钟  

review

Average cascading & average sliding windows

last week I proposed a new method to extern the text and delete text: Calulate the average of the contribution of sentences rather than sum of all tokenziers.

Result

It will take a long time for test all data so I just test half test (1000 data)

Here are results:

Method Test Data Number of cycle for explain Original model exp model Improved accuracy
Cascading 4142 10 79.50% 79.48% -0.02%
Sliding windows 4142 10 79.50% 79.54% +0.047
low frequency 1245 2 83.22% 81.56% -1.66%
AVG frequency - 2 79.50% - -
AVG cascading 4142 10 79.50% 79.40% -0.10%
AVG Sliding windows - 10 79.50% - -
Remove individual tokenizers - 10 79.50% - -

Conclusion

According to the current results(Cascading, Sliding windows, low frequency, AVG cascading) Not very helpful for model performance.

The remaining few methods are not expected to improve much

Extern Seletion model

Through the test records of the model, we can obtain a large number of test records, including: changes in probability and changes in f1. Among them, the label of the data we manually add the highest f1 value index:

1
2
3
4
5
6
7
8

"""f1_score = 
[[1,1,1,1,1,1,1,1,1,1],
[0.8,0.8,0.8,1,1,1,0.5,0.5,0.5],
[0,0,0,0,0,0.5,1,1,1,1,1]]
..."""
for i in f1_score:
   label.append(f1_score.index(max(f1_score)))

Training data : Probability changes in the interpretation process

Label : index of highest f1 score

model type : Multiclass Model (10 categories, since the number of explanations is 10, predict the index with the highest f1 value based on the probability change)

MODEL

1
2
3
4
5

net = torch.nn.Sequential(
   torch.nn.Linear(2, 500),
   torch.nn.ReLU(),
   torch.nn.Linear(500, 11),
)

Unbalanced Data

20220111213059

initial state, accuracy = 0.95%

20220111213219

initial state, accuracy = 92.7592%

Although the accuracy is high, the test has no effect because 99% of the labels in the data are 0

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35

print(f1)
$[[1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
 [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
 [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
 [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
 [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
 [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
 [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
 [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
 [0.6666666666666666,
  0.6666666666666666,
  0.6666666666666666,
  0.6666666666666666,
  0.6666666666666666,
  0.6666666666666666,
  0.6666666666666666,
  0.6666666666666666,
  0.6666666666666666,
  0.6666666666666666,
  0.6666666666666666],
 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
 [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
 [0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5],
 [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
 [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
 [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
 [0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5],
 [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
 [0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5],
 [0.2222222222222222, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5]...

Balanced Data

20220112112417

initial state, accuracy = 0.95%

20220112112402

data : 300, accuracy : 0.95->27.8%

Possible reasons for this situation

1.Not enough data (only 300data)

solution : test more dataset and collect more data

2.The exp data random data without features, and the model cannot find useful features.

solution : We need to prove that explainable AI is a regular technique or use some mathematical method to prove that the data is not random data

20220112112505