https://www.kaggle.com/mlg-ulb/creditcardfraud#creditcard.csv

• To test adaboost with different thershold, run python3 adaboost.py -d creditcard.csv -n 10 -t 0.5
• To test adaboost with different upsample rate, run python3 adaboost.py -d creditcard.csv -s 1 -n 10
• To test fc_nn with different upsample rate, run python3 fc.py -d creditcard.csv -s 1 -n 10
• To test svm with different upsample rate, run python3 svm.py -d creditcard.csv -s 1 -n 10
• To test any model with just an upsample ratio but not a range just use "-n". For example, python3 svm.py -d creditcard.csv -n 10
• The upsample range will be [1, n]
• To automatically generate curve on three models, run python3 run_pipeline.py -d creditcard.csv -s 1 -n 10

• Add data parsing, tested on temp.csv (a snapshot of original file)
• Initialization of fully connected network with two dense layers

• All data points are detected as no-fraud, not acceptable
• Normalize data and delete one feature
• NN reaches a true positive of 65%
• Will start working on using SVM, Adaboost

• Initialization of Adaboost, havent used different threshold yet
• Rather than using only 100,000 data, models are trained on the whole dataset this time
• Both Adaboost and nn reach a true positive rate ranging from 70 to 75%
• No idea why we can reach this result, need further investigation

• Note that at this point dataset has not been upsampled yet - (meaning repopulate fraud cases to reach 4k cases). In other words, we still use the original data here
• Create SVM
• Figuring out the params to be used for SVM

• Figure out & clean up correct params for SVM
• Decided to run SVM twice - one with gammaa=1 and one without to see the difference in accuracy & runtime (Concluded that we need gamma=1)
• Investigated the built-in Roc curve functions for our use

• Add upsample function to manually increase weight on examples with true label
• Performance gets worse on adaboost with a ratio of 0.3
• Add roc curve

• Get rid of kfolds for SVM & still run single SVM multiple times to finish collecting results without upsampling

• Add another type of upsample, this time we only increase the number of true class by n times
• Fix the issue with upsample
• Adaboost reaches a recall rate of 0.85 when upsampling true class by 10 times
• NN reaches a recall rate of 0.798 when upsampling true class by 10 times
• Plan to enable adjusting threshold for Adaboost
• Plan to make roc curve on models with different upsample ratio

• Clean up the util.py to make some functions more understandable & Added comments to each method to make it more readable for future readers
• Update SVM to accommodate upsamples
• Will run SVM with 10X upsample to see the difference in performance
• will create a results section in README that contains the running results

• Revise comment style
• Modify parser and right now we can run adaboost with a threshold range
• Users can now run models with a range of upsample rate
• Start to produce final result and plan to delete intermediate code

• change the type for '--upsamplen' from 'float' to 'int'
• Modified SVM to accomodate new parsing features
• Added plot_recall_upsample_curve funciton
• Get all results from running SVM
• Finish ploting all models on the recall_upsample curve
• create generate_curves.py for creating recall_upsample curve
• all results updated. Ready to start working on presentation slides tomorrow

• Add roc curve & precision recall curves to SVM

• Get rid of hard code style file (generate_cruve.py)
• fcnn and adaboost are now able to automatically generate graph

• Run all files to get output graphs
• Mix last bit of stuff on all files

#running SVM with 1-10 upsample

#upsample=1, recall=0.78 [[85265 11] [ 36 131]]

#upsample=2, recall=0.82 [[85268 11] [ 56 255]]

#upsample=3, recall=0.81 [[85306 7] [ 81 344]]

#upsample=4, recall=0.84 [[85292 12] [ 95 486]]

#upsample=5, recall=0.82 [[85290 12] [ 134 597]]

#upsample=6, recall=0.81 [[85288 13] [ 166 714]]

#upsample=7, recall=0.82 [[85300 11] [ 186 831]]

#upsample=8, recall=0.85 [[85271 24] [ 174 1007]]

#upsample=9, recall=0.84 [[85329 23] [ 206 1065]]

#upsample=10, recall=0.84 - 1.53% [[85279 31] [ 233 1228]]

#=== #upsample=20, recall=0.87 - 3.17% [[85292 56] [ 383 2516]]

#positive case - 4.77% #upsample=30, recall=0.88 [[85293 58] [ 541 3831]]

#10% #upsample=60, recall=0.9 [[85224 103] [ 888 7936]]

Upsample range from 1 time to times: [[85293 13] [ 36 101]] [[85290 18] [ 56 226]] [[85287 33] [ 76 342]] [[85310 40] [ 106 429]] [[85321 36] [ 123 553]] [[85299 49] [ 118 715]] [[85294 44] [ 145 845]] [[85264 50] [ 193 969]] [[85326 41] [ 208 1048]] [[85259 63] [ 214 1235]]

Unsample param = 10: 0.4 [[19782 65531] [ 0 1458]] 0.405 [[19982 65331] [ 0 1458]] 0.410 [[20020 65293] [ 0 1458]] 0.415 [[20026 65287] [ 0 1458]] 0.420 [[20026 65287] [ 0 1458]] 0.425 [[20026 65287] [ 0 1458]] 0.430 [[20028 65285] [ 0 1458]] 0.435 [[20033 65280] [ 0 1458]] 0.440 [[20064 65249] [ 0 1458]] 0.445 [[20226 65087] [ 0 1458]] 0.450 [[20796 64517] [ 0 1458]] 0.455 [[22545 62768] [ 0 1458]] 0.460 [[26687 58626] [ 0 1458]] 0.465 [[35198 50115] [ 0 1458]] 0.470 [[48167 37146] [ 4 1454]] 0.475 [[63503 21810] [ 4 1454]] 0.480 [[76080 9233] [ 4 1454]] 0.485 [[82029 3284] [ 4 1454]] 0.490 [[84409 904] [ 30 1428]] 0.495 [[85091 222] [ 134 1324]] 0.500 [[85247 66] [ 225 1233]] 0.505 [[85286 27] [ 317 1141]]

Upsample range from 1 time to times: [array([[85273, 11], [ 55, 104]]), array([[85269, 23], [ 63, 235]]), array([[85279, 34], [ 68, 357]]), array([[85300, 24], [ 108, 453]]), array([[85300, 29], [ 123, 581]]), array([[85289, 26], [ 137, 729]]), array([[85292, 40], [ 136, 860]]), array([[85253, 36], [ 223, 964]]), array([[85353, 21], [ 238, 1011]]), array([[85267, 42], [ 209, 1253]])]