文档介绍：失败为成功之母
2
Asset Information
Record the internal device information, including …
IP address
Operation systemposed Sigmoid Aging is modified by sigmoid function into 0~1
V={vt1,vt2,…,vtk,..} : Count the times of error prediction (accuracy <1/2) before each round learning for each learner
vt’s initial value is 0,
a is the slope of sigmoid function
b is the move right parameter with different initial point
15
Sigmoid Function
Original Sigmoid function Ex: a=1
16
Sigmoid Aging - Forgetting Curve
Proposed Sigmoid Aging Function. Ex: a=2,b=4
17
Step1: Begin to Learn
Set the initial weight in the chunk and train the learner y1
Then we got the committee YT and the voting weight α1 of y1
{w1n}n=1~N
y1(x t1)
Committee t1
Xt1={..}
18
Step2: Got New Chunk Data
New incoming chunk Xt2={(X2n,W2n,z2n)}n=1~N
y1(x t1)
Xt2={..}
19
Step3: Count the Times of Error Prediction
New incoming chunk Xt2={(X2n,W2n,z2n)}n=1~N
y1(xt2)
Xt2 ={..}
If accuracy < ½ then vt=vt+1
20
Step4: Calculate the Example Weight
Calculate the example weight by previous mistakes
Set the initial weight W2n and pass the old committee Yt1 to get new weight W’2n
{w2n}n=1~N
y1(xt2)
Xt2={..}
{w’2n}n=1~N
21
Step5: Train the New Learner
Train the learner y2 by the Xt2={(X2n,W’2n,z2n)}n=1~N
Then we got the committee YTnew and the voting weight α2 of y2
Committee YT
{w’2n}n=1~N
y2(xt2)
New Committee YTnew
22
Using validation set recorded by the recent data in buffer.
Optima selection
leave the best performance committee as the winner from YT and YTnew
Step6: Validation Test
Original Committee:
New Committee:
23
Step 7: Repeat this Process till terminated
Repeat this process when we got the new chunk (go to Step2) till this process be terminated
24
Committee Decision Function
Decision function with aging
When a new alert instance is generated, the prediction class will be determined by the committee
25