yearsj / imac Goto Github PK

• Clone the adapted MOA repository from here

• Build on windows: The easiest way is to use an IDE such as IntelliJ

• Build on linux:

  1. cd execmoa
  2. mkidr moa， mkdir moa/classifiers， mkdir moa/classifiers/trees
  3. javac -cp moa.jar HoeffdingTree.java
  4. mv *.class moa/classifiers/trees
  5. jar -uf moa.jar moa/classifiers/trees/*
  

  The code in execmoa comes from this GitHub

Synthetic data is all generated using the API proposed by MOA.

SEA: '-i 5000000 -f 10000 -s (generators.SEAGenerator -b)',

LED: '-i 1000000 -f 1000 -s (ConceptDriftStream -s (generators.LEDGeneratorDrift -d 1) -d (ConceptDriftStream -s (generators.LEDGeneratorDrift -d 3) -d (ConceptDriftStream -s (generators.LEDGeneratorDrift -d 5) -d (generators.LEDGeneratorDrift -d 7) -w 50000 -p 250000 ) -w 50000 -p 250000 ) -w 50000 -p 250000)'

AGR: '-i 10000000 -f 10000 -s (ConceptDriftStream -s (generators.AgrawalGenerator -f 1 -b) -d (ConceptDriftStream -s (generators.AgrawalGenerator -f 2 -b) -d (ConceptDriftStream -s (generators.AgrawalGenerator -b) -d (generators.AgrawalGenerator -f 4 -b) -w 500 -p 2500000) -w 500 -p 2500000 ) -w 500 -p 2500000)',

RBF: '-i 2000000 -f 10000 -s (generators.RandomRBFGenerator -c 15 -a 200 -n 100) ',

RTG: '-i 10000000 -f 10000 -s (generators.RandomTreeGenerator -c 25 -o 200 -u 200)',

Covertype. The forest covertype data set represents forest cover type for 30 x 30 meter cells obtained from the US Forest Service Region 2 Resource Information System (RIS) data.

KDD99. KDD99 dataset corresponds to a cyber-attack de- tection problem (i.e., attack or common access), an inherent streaming scenario since instances are sequentially presented as a time series.

MNIST8M. MNIST8M is the augmentation of original MNIST database by using pseudo- random deformations and translations.

We have added the IMAC method to the code provided by this github.

There are four algorithms:

• VFDT algorithm, which periodic executes split-attempts.
• OSM algorithm, which predicts the interval of split-attempts in VFDT by assuming that heuristic measure of the second-best attribute does not change and the the heuristic measure of best attribute increases as much as possible as the data arrives.
• CGD algorithm, which predicts the interval of split-attempts in VFDT by assuming that the difference of best attribute and second-best attribute does not change.
• IMAC algorithm, which determines the potential split timing in VFDT with incremental information.

trees/HoeffdingTree.java        // Clean implementation
trees/HoeffdingTreeMeasure.java // can evaluate split-delay and split-attempts

run.py // evaluate VFDT, OSM, CGD and IMAC
run_measure.py // evaluate split-delay and split-attempts of VFDT, OSM, CGD and IMAC