-
Notifications
You must be signed in to change notification settings - Fork 1
/
Remark.txt
40 lines (35 loc) · 1.7 KB
/
Remark.txt
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
36
37
38
39
Remark:
Date: Mar/5/2019
1. Initialized model prediect 87%(300) as class2(means error 1).
Init the model parameter using standard normal distribution.
Meeting: Mar/7/2019
1. predicted anomaly error rollback
2. Init -> normal; (clustering methods)
3. Plot F1, ...
Result. 1. result shows in result_back_predict_negative_error_300.eps
xls in incrementalAD/Experiments/compare_Mar_5.ods.
It will increase and drop times, then predict all to postive.
2. not cluster, another easy way.
result shown as result_back_predict_negative_newinit.eps
3.
Meeting: Mar/19/2019
0. notice the batch size.
1. plot 2 class results
2. chose class plot, not metrics plot.
3. ROC curve.
Result:
0. result folder is the result of test_MCL21LS_compare_Back.m (not newinit, 100 batchsize)
1. plot 2-class results
whether only 2 classes or statistic 2 postive error negative error?
two classes:
still follow the negative rollback manner? same reason.
if need to prone to normal? (as distribution is not extreme imbalance)
N: /L1LS/test_L1LS_Back_negative.m Prone_normal=false;
The wave highly related to batch size, but the genreal trend is the same.
Y: /L1LS/test_L1LS_Back_negative.m Prone_normal=true;
The result is similar to N.
result and figure: excute /L1LS/test_L1LS_Back_negative.m
postive error negative error:
Don't know how to calculate TP,TN,FP,FN
3. ROC is not for distance based algorithms, especially for possibility based algorithmsbut we can still use it as a generalization.[min:step:max]
We save the ROC curve as gif, and show the AUC value in the code.