LACE2

[amritbhanu]

Experiment:

• Modified the Cliff and morph code to adjust binary values. They work for continuous values
• Sharing from UCI, so training available from that and tested on 6 companies. They are 'AOL', 'Facebook', 'PayPal', 'Google','Apple', 'Yahoo'.
• 0.3 cliff, 0.2 alpha, 0.5 beta.

Baseline:

• Keeping 0.3 for UCI training without morphing and compared against those 6 companies.

Results:

• DELTA improvement of LACE over baseline. Positive the better for all measures except false alarm in which case negative the better.

AOL

Apple

Facebook

Google

Paypal

Yahoo

[timm]

so if we use naive bayes, pdf never gets worse and other measures improve, a little

and if we other learners the get LARGE reductions in pf and recall?

[amritbhanu]

Not all learners for large reductions in pf and recall. We only see that in Dec trees and rbf svm.

Prof. here we can use tuning to find the cliff percentage, alpha, beta, about how much to share and get the best performance.

Amritanshu Agrawal PhD Student CS @ NC State aagrawa8@ncsu.edu http://amritag.wixsite.com/amrit

On Thu, May 4, 2017 at 5:51 PM, Tim Menzies notifications@github.com wrote:

so if we use naive bayes, pdf never gets worse and other measures improve, a little

and if we other learners the get LARGE reductions in pf and recall?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/ai-se/LAS-Phishing/issues/18#issuecomment-299319461, or mute the thread https://github.com/notifications/unsubscribe-auth/ADTGCEp0JiATcnyTWM2aDq_P4m860FbCks5r2khhgaJpZM4NRPkz .

[timm]

need a monthly report

[amritbhanu]

report: https://github.com/ai-se/LAS-Phishing/blob/master/reports/may17.md

[amritbhanu]

Experiment:

• Continuing from the previous comment,
• Modified the Cliff and morph code to adjust binary values. They work for continuous values. Now the regions for 0 is randomly picked between (0 - 0.5) and for 1, it is picked from (0.5 - 1)
• Sharing from UCI, so training available from that and tested on 6 companies. They are 'AOL', 'Facebook', 'PayPal', 'Google','Apple', 'Yahoo'.
• 0.3 cliff

Baseline:

• Keeping 0.3 for UCI training without morphing and compared against those 6 companies.

Results:

• Lace rbf kernel svm performs better in most cases
• Surprising in some cases LACE performed better in few datasets with few learners. With others, it didnt change much. So LACE either performed better or stayed constant.

AOL

Apple

Facebook

Google

Paypal

Yahoo