Commit Guru 150

[Suvodeep90]

Commit Guru 150

[timm]

Methods: 'ns', 'nd', 'nf', 'entropy', 'la', 'ld', 'lt', 'ndev', 'age',
'nuc', 'exp', 'rexp', 'sexp','fix'
FX: commitguru: 13 attributes

[timm]

FSS

CFS(https://github.com/ai-se/bellwether_community/blob/master/src/CFS.py)
paper: https://www.cs.waikato.ac.nz/ml/publications/1997/Hall-LSmith97.pdf
paper: https://www.cs.waikato.ac.nz/~mhall/HallHolmesTKDE.pdf
thesis: https://www.cs.waikato.ac.nz/~mhall/thesis.pdf
parameter:

temporal selection: https://arxiv.org/pdf/1803.05067.pdf

[timm]

LEARNER: logistic regression

• [ ] FFT *8, FFT * 16 : might be able to avoid FSS
• [ ] Dodge

[timm]

Hyper parameter optimizer

• DE on SMOTE
• DE on LogReg
• nothing on FSS

suggestions:

• don't search M+N space, search M then N

[timm]

Success criteria

• recall
• false alarm
• IFA: section5 : https://xin-xia.github.io/publication/icsme173.pdf
• PCI@20%: probably not
• POpt20 (can't do cause of not LOCS).
• popt20 surrogate
• precision is bad (http://menzies.us/pdf/07precision.pdf)

[timm]

DATA

150 projects
rows:

• chart of rows per project spited on rows
• chart of ratio of defectives

[timm]

what's the PEEKING mechanism?

None. no zzz

Just the activity in the commit

[timm]

LABELLING
using keyword labelling

not yet active learning yet

[timm]

TRAIN-TEST rig

• Leave one out * 150
  - everything is inside

suggestions:

• move somethings out to a one time pre-preprocessr
• don't do all, do a. tournament small groups that grow

[timm]

RELATED WORK

has anyone used this data to get PD>66 nd PF < 33 before?

1. Predicting crashing releases of mobile applications - uses some metrics collected from commit guru along with other code related metrics. (recall- between .5 to .7, prec - ~0.2)

2. Just-In-Time Bug Prediction in Mobile Applications: The Domain Matters! - uses commit guru metrics with FSS - (58(p),25(R),34(F1))

3. Software Maintenance at Commit-Time - (90.75(P) 37.15(R) 52.72(F1))