i think Ghotra et al [17] used AUC effort vs recall not AUC(pd,pf). please check

. And found out that techniques like AdaBoost.NC had a better performance than the rest while others are planning to use SMOTE~\cite{gray20|

?? run tis into the last sentence "and they found that.."

here's a light weight description. mote that point3 has to be changed for numeric attributes

t

2.   DE scores each {\em pop}$_i$ according to various objective
   scores $o$. In the case of our goal models, the objectives are $o_1$ the sum of the cost
 of its decisions, $o_2$ the number of ignore edges, and the number of $o_3$ satisfied goals
 and $o_4$  softgoals.

 3. OPTIMIZE tries to each replace {\em pop}$_i$ with a mutant $q$
 built by extrapolating between three other members of population $a,b,c$.
 At probability $p_1$, for each decision $a_k \in a$, then
 $m_k= a_k \vee (p_1 < \mathit{rand}() \wedge( b_k \vee c_k))$.

 4. Each mutant $m$ is assessed by calling  $\text{SAMPLE}(\textit{model,prior=m})$;
 i.e. by seeing what can be achieved within a goal after first assuming
 that $\textit{prior}=m$.

 5.  To test if the mutant $m$ is preferred to {\em pop}$_i$, OPTIMIZE uses
  Zitler's continuous domination {\em cdom}
  predicate~\cite{Zitzler2004}. This predicate compares two sets of objectives
  from sets $x$ and $y$. In that comparison,
  $x$ is better than another $y$ if $x$  ``losses'' least.
  In the following, $``n''$ is the number of objectives and $w_j \in \{-1, 1\}$

shows if we seek to maximize $o_j$.
[
\begin{array}{rcl}
x \succ y & =& \textit{loss}(y,x) > \textit{loss}(x,y)\
\textit{loss}(x,y)& = &\sum_j^n -e^{\Delta(j,x,y,n)}/n\
\Delta(j,x,y,n) & = & w_j(o_{j,x} - o_{j,y})/n
\end{array}
]

5. OPTIMIZE repeatedly loops over the population, trying to replace items with mutants,
   until new better mutants stop being found.

6. Return the population.
   \\hline
   \end{tabular}
   \caption{Procedure OPTIMIZE: strives to find ``good'' priors which,
   when passes to SAMPLE, maximize the number of edges used
   while also minimizing cost, and
   maximizing satisfied hard goals and soft goals.
   OPTIMIZE is based on Storn's differential evolution optimizer~\protect\cite{storn1997differential}.
   OPTIMIZE is called by the RANK procedure of \fig{rank}.
   For the reader unfamiliar with the mutation technique of step 3 and the {\em cdom}
   scoring of step5, we note that these
   are standard practice in the search-based
   SE community\cite{Fu2016,krall2015gale}.
   }\label{fig:optimize}

• Change to histograms
• tune on x evaluate on x.
• auc1 and auc2 (loc,recall)

Weka: Total 20 attributes in each datasets. Datasets from top to bottom (high to low imbalance) . CFS attribute selection, and breadth first.

• ant : cbo, rfc, lcom, loc, cam, amc, max_cc
• redaktor - cbm, max_cc
• arc - cbo, rfc, ce, npm, cam
• ivy - wmc, cbo, rfc, ce, npm, loc, moa, amc
• prop - lcom, ce , loc, moa, max_cc
• tomcat - cbo, rfc, loc, moa , max_cc
• camel - cbo, lcom, ca, avg_cc
• jedit - rfc, moa

Most of these implementations are provided in Scikit-Learn~\cite{pedregosa2011scikit} and used in our code.

Results:

Accuracy

Recall

Precision

F_score

False Alarm

AUC(pf,pdf)

AUC(low, pd)

• original smote paper datasets and the measures
• include f2 score, use abs nos for oversampling and undersampling (50, 100, 200, 400)

Experiment:

• Goal - maximizing Fscore for each of 6 learners separately.
• Once the parameters are found, reporting all 5 evaluation measures.
• train, validation, and test sets.
• parameters tuned are:
  • m(20,50) and n(80,50) % of oversampling and undersampling respectively.
  • power of distance metric (r) (0.1 to 5)??
  • k=(2,20) exponential??
  • Didn't do the preprocessing part (exp = 0.3 to 3)

Results:

Accuracy

Recall

Precision

F_score

False Alarm

t then creates new instances using the selected instances and their neighbors

how?

Within Measure Assessment

AUC for SMOTE

AUC for SMOTUNED

Results by Tantithamthavorn et al [50] also suggested that
every dataset comes with different attributes. And also clas-
sification techniques often have configurable parameters that
control characteristics of these classifiers that they produce.
Now time has come to even think about hyperparameter opti-
mization of these techniques and come up with an automated
process [2], [16] to tune these parameters for every dataset.

• Doesn't include 3 datasets: prop, tomcat and jedit.
• AUC(pf,pd)

Conclusions:

• AUC the highest improvement.
• Recall good improvements
• Fscore = DT and RF most improvement
• Others either the same or worse.

Results:

Accuracy - Maximize

AUC - Maximize

Recall - Maximize

Precision - Maximize

F_score - Maximize

False Alarm - Minimise

By keeping change logs of the most recently or frequently
changed files are the most probable source of future de-
fects [11], [21], [7]. They

K metrics combined with OO (object-oriented) metrics perform better than all other metrics.

it leaves open issue like

• why we are using smote and not Pelayo's trial and error

more generally , lit reviews must respect and disrespect. respectfully present others work, then point out their fatal mistake and why this work in needeed

Hence,
our learning objective can be generally described
as ``obtaining a classifier that will provide high accuracy for the minority class without severely compromising the accuracy of the majority class''.

t is important to select how many synthetic examples to create ($m$) and how much undersampling ($m$) of majority class needs to be done. In this case number of oversampling and undersampling are the same.

• Doesn't include 3 datasets: prop, tomcat and jedit

Results:

Accuracy

AUC

Recall

Precision

F_score

False Alarm

increase width of fig 2,3,4,5. make full page wide (but dont increase font size)

SMOTE's super-sampling selects instances from the minority class and finds k'' nearest neighbors for each instance and then creates new instances using the selected instances and their neighbors until we have m'' numbers of minority class samples.