\relax 
\citation{Boehm1981}
\citation{kemerer87}
\citation{Boehm2000,Jorgensen2007,Shepperd2007,Mendes2007}
\citation{Auer2006}
\citation{Menzies2006}
\citation{baker07}
\citation{Pendharkar2005}
\citation{Mendes2008a}
\citation{Li2009}
\@writefile{toc}{\contentsline {section}{\numberline {1}Introduction}{1}}
\newlabel{sec:introduction}{{1}{1}}
\citation{Kadoda2000,Mendes2003,Li2009}
\citation{Briand1999,jeffery2001}
\citation{Shepperd1997,Mendes2003}
\citation{Mendes2002,Mendes2003}
\citation{Scott1992}
\citation{Wand1994}
\citation{Palpanas2003,John1995,Frank03}
\citation{Shepperd2007}
\citation{Jorgensen2004}
\citation{Boehm2000}
\citation{Jorgensen2009}
\@writefile{toc}{\contentsline {section}{\numberline {2}Background}{3}}
\newlabel{sec:background}{{2}{3}}
\@writefile{toc}{\contentsline {subsection}{\numberline {2.1}Software Effort Estimation}{3}}
\newlabel{subsec:software-effort-estimation}{{2.1}{3}}
\citation{Boehm1981}
\citation{Menzies2006,Kadoda2001}
\citation{Boehm1981,Kitchenham2009}
\citation{Myrtveit}
\citation{Shepperd2001}
\citation{Myrtveit}
\citation{Myrtveit}
\citation{Myrtveit}
\citation{Shepperd1996}
\citation{Mendes2003}
\citation{Li2009}
\@writefile{toc}{\contentsline {subsection}{\numberline {2.2}ABE}{4}}
\newlabel{subsec:ABE}{{2.2}{4}}
\citation{Mendes2002a}
\citation{Briand1999}
\citation{Myrtveit1999}
\citation{Angelis2000}
\citation{Mendes2002a}
\citation{Angelis2000}
\citation{Mendes2003}
\newlabel{equ:euclid}{{1}{5}}
\newlabel{equ:irwm}{{2}{5}}
\citation{Wand1994}
\citation{Wand1994,Scott1992}
\citation{Wand1994}
\citation{Scheid2004}
\citation{Palpanas2003}
\citation{Frank03}
\citation{John1995}
\@writefile{lof}{\contentsline {figure}{\numberline {1}{\ignorespaces We see a Gaussian kernel density estimate built on individual data points. Each point is in the center of a kernel and its effect is distributed to its neighborhood. The sum of all kernels make up the final Gaussian kernel density estimate.}}{6}}
\newlabel{fig:pointwise-kernel}{{1}{6}}
\@writefile{toc}{\contentsline {subsection}{\numberline {2.3}Kernel Density Estimation}{6}}
\newlabel{subsec:kernel-estimation}{{2.3}{6}}
\citation{Mendes2002,Mendes2003}
\citation{Scheid2004}
\citation{Scott1992,Cressie1993}
\citation{Scheid2004,Wand1994}
\citation{John1995}
\newlabel{equ:general-kernel}{{3}{7}}
\newlabel{equ:general-kernel-b}{{4}{7}}
\newlabel{equ:general-kernel-g}{{5}{7}}
\@writefile{lof}{\contentsline {figure}{\numberline {2}{\ignorespaces The formulas for different kernels used in this study. In formulas $\rho = {\begingroup x - X_i\endgroup \over h}$. Note that IRWM kernel has different characteristics and its calculation details were provided in Section 2.2\hbox {}.}}{7}}
\newlabel{fig:kernel-formulas}{{2}{7}}
\citation{Scheid2004}
\citation{Scott1992,Cressie1993}
\citation{Mendes2002,Mendes2003}
\newlabel{fig:uniform-kernel}{{3(a)}{8}}
\newlabel{sub@fig:uniform-kernel}{{(a)}{8}}
\newlabel{fig:triangular-kernel}{{3(b)}{8}}
\newlabel{sub@fig:triangular-kernel}{{(b)}{8}}
\newlabel{fig:epanechnikov-kernel}{{3(c)}{8}}
\newlabel{sub@fig:epanechnikov-kernel}{{(c)}{8}}
\newlabel{fig:gaussian-kernel}{{3(d)}{8}}
\newlabel{sub@fig:gaussian-kernel}{{(d)}{8}}
\newlabel{fig:irwm-kernel}{{3(e)}{8}}
\newlabel{sub@fig:irwm-kernel}{{(e)}{8}}
\@writefile{lof}{\contentsline {figure}{\numberline {3}{\ignorespaces General shapes of the kernel types used in this research. Note that IRWM is not symmetric unlike other kernels, that is due to the fact IRWM has a different weighting procedure than other kernels (see Equation 2\hbox {}).}}{8}}
\@writefile{lof}{\contentsline {subfigure}{\numberline{(a)}{\ignorespaces {Uniform Kernel}}}{8}}
\@writefile{lof}{\contentsline {subfigure}{\numberline{(b)}{\ignorespaces {Triangular Kernel}}}{8}}
\@writefile{lof}{\contentsline {subfigure}{\numberline{(c)}{\ignorespaces {Epanechnikov Kernel}}}{8}}
\@writefile{lof}{\contentsline {subfigure}{\numberline{(d)}{\ignorespaces {Gaussian Kernel}}}{8}}
\@writefile{lof}{\contentsline {subfigure}{\numberline{(e)}{\ignorespaces {IRWM Kernel for \textit {k}=10}}}{8}}
\newlabel{fig:kernels}{{3}{8}}
\newlabel{fig:bandwidth-1}{{4(a)}{9}}
\newlabel{sub@fig:bandwidth-1}{{(a)}{9}}
\newlabel{fig:bandwidth-05}{{4(b)}{9}}
\newlabel{sub@fig:bandwidth-05}{{(b)}{9}}
\newlabel{fig:bandwidth-02}{{4(c)}{9}}
\newlabel{sub@fig:bandwidth-02}{{(c)}{9}}
\newlabel{fig:bandwidth-005}{{4(d)}{9}}
\newlabel{sub@fig:bandwidth-005}{{(d)}{9}}
\@writefile{lof}{\contentsline {figure}{\numberline {4}{\ignorespaces We see the effect of bandwidth on kernel density estimation. From Figure 4(a)\hbox {} to Figure 4(d)\hbox {}, the bandwidth gets smaller and smaller and we observe a transition from over-smoothing to under-smoothing.}}{9}}
\@writefile{lof}{\contentsline {subfigure}{\numberline{(a)}{\ignorespaces {h=0.1}}}{9}}
\@writefile{lof}{\contentsline {subfigure}{\numberline{(b)}{\ignorespaces {h=0.05}}}{9}}
\@writefile{lof}{\contentsline {subfigure}{\numberline{(c)}{\ignorespaces {h=0.02}}}{9}}
\@writefile{lof}{\contentsline {subfigure}{\numberline{(d)}{\ignorespaces {h=0.005}}}{9}}
\newlabel{fig:choice-of-bandwidth}{{4}{9}}
\@writefile{toc}{\contentsline {section}{\numberline {3}Methodology}{9}}
\newlabel{sec:methodology}{{3}{9}}
\@writefile{toc}{\contentsline {subsection}{\numberline {3.1}Weighting Method}{9}}
\newlabel{subsec:weighting-method}{{3.1}{9}}
\newlabel{equ:wkde}{{6}{9}}
\citation{Boehm1981}
\citation{Desharnais1989}
\citation{Alpaydin2004}
\newlabel{equ:weight}{{7}{10}}
\newlabel{equ:effort-update}{{8}{10}}
\@writefile{toc}{\contentsline {subsubsection}{\numberline {3.1.1}Uniform vs. Non-Uniform Weighting}{10}}
\citation{Kitchenham2007}
\newlabel{fig:weighting-abe}{{3.1.1}{11}}
\@writefile{lof}{\contentsline {figure}{\numberline {5}{\ignorespaces In the case of ABE0 all instances are given equal probability values, hence equal weights. However, uniform kernel prefers some instances over the others: Only a certain portion of the instances are given equal weights.}}{11}}
\newlabel{fig:weight-vs-non-weight}{{5}{11}}
\@writefile{toc}{\contentsline {subsection}{\numberline {3.2}Data}{11}}
\@writefile{lof}{\contentsline {figure}{\numberline {6}{\ignorespaces We used 237 projects coming from 3 datasets. Datasets have different characteristics in terms of the number of attributes as well as the measures of these attributes.}}{11}}
\newlabel{fig:datasets}{{6}{11}}
\citation{Kadoda2000}
\citation{Molokken-Ostvold2004,Briand1999,Foss}
\citation{Stensrud}
\citation{Stensrud2002}
\citation{Mendes2003}
\@writefile{toc}{\contentsline {subsection}{\numberline {3.3}Experiments}{12}}
\@writefile{toc}{\contentsline {subsection}{\numberline {3.4}Performance Criteria}{12}}
\newlabel{EquationMRE}{{9}{12}}
\citation{Mendes2003}
\citation{Mendes2008a}
\citation{Conte1986}
\citation{Conte1986,Mendes2003}
\citation{Foss}
\newlabel{EquationMdRE}{{10}{13}}
\@writefile{lof}{\contentsline {figure}{\numberline {7}{\ignorespaces Pseudocode for Win-Tie-Loss Calculation Between Method \textit  {i} and \textit  {j}}}{13}}
\newlabel{FigureWinTieLossPseudocode}{{7}{13}}
\@writefile{toc}{\contentsline {section}{\numberline {4}Results}{13}}
\newlabel{sec:results}{{4}{13}}
\citation{supplementaryResults}
\@writefile{lof}{\contentsline {figure}{\numberline {8}{\ignorespaces Desharnais dataset $win,tie,loss$ statistics for ABE0 and N-ABE through Gaussian kernel. For each dataset we have $4$ of these tables (one for each kernel). In total it amounts to \textit  {19 Datasets $\times $ 4 tables = 76 tables}. It is infeasible to include all the tables in this paper, therefore an executive summary of $76$ tables is provided in Figure\nobreakspace  {}9\hbox {}. Furthermore, we provide all $76$ tables in excel format at http://goo.gl/qpQiD.}}{15}}
\newlabel{fig:win-tie-loss-desharnais}{{8}{15}}
\citation{Mendes2002,Mendes2003}
\@writefile{lof}{\contentsline {figure}{\numberline {9}{\ignorespaces Nine data sets comparing ABE0 to N-ABE. For every row in each cell, there are three symbols indicating the effect of N-ABE w.r.t. 3 different error measures. From left to right, the first symbol stands for N-ABE effect w.r.t. MdMRE, the second symbol w.r.t. MAR and the third one w.r.t. Pred(25). A ``$+$'' indicates that for majority of $k$ values (at least 3 out of 5 \textit  {k} values), N-ABE improved ABE0 in terms of $win-loss$ values. ``$-$'' indicates that N-ABE decreased the performance of ABE0 in the majority case. If the former conditions do not satisfy, then a ``$o$'' symbol is assigned. Note that dataset order here is the same as Figure\nobreakspace  {}6\hbox {}, yet the dataset names are abbreviated to 3 to 5 letters due to space constraints.}}{16}}
\newlabel{fig:summary-all1}{{9}{16}}
\@writefile{toc}{\contentsline {section}{\numberline {5}Discussion}{16}}
\newlabel{sec:discussion}{{5}{16}}
\citation{Palpanas2003,Frank03,John1995}
\@writefile{lof}{\contentsline {figure}{\numberline {10}{\ignorespaces Ten more data sets comparing ABE0 to N-ABE. Same format as Figure\nobreakspace  {}9\hbox {}.}}{17}}
\newlabel{fig:summary-all2}{{10}{17}}
\@writefile{lof}{\contentsline {figure}{\numberline {11}{\ignorespaces The comparison of ABE0 to N-ABE under IRWM kernel. Similar to Figure\nobreakspace  {}9\hbox {} three symbols indicate the effect of N-ABE w.r.t. 3 different error measures and ``$+$'' indicates that for majority of $k$ values N-ABE improved ABE0 in terms of $win-loss$ values. A ``$-$'' symbol indicates a decrease and a ``$o$'' symbol indicates neither decrease nor increase. Notice that subject to IRWM kernel, N-ABE fails to improve ABE0 w.r.t. 3 different performance measures.}}{18}}
\newlabel{fig:summary-irwm}{{11}{18}}
\@writefile{lof}{\contentsline {figure}{\numberline {12}{\ignorespaces An intuitive example. In a 3 analogy case, there is a $75\%$ change for a hypothetical test project between uniform and non-uniform weighting effort estimates.}}{18}}
\newlabel{fig:intuitive-example}{{12}{18}}
\citation{Alpaydin2004}
\citation{Milic2004}
\citation{Robson2002}
\@writefile{toc}{\contentsline {section}{\numberline {6}Threats to Validity}{19}}
\newlabel{sec:threats-to-validity}{{6}{19}}
\citation{Stensrud2002}
\citation{Premraj2007,Lum2008,Li2006}
\citation{Myrtveit,Foss}
\citation{Foss}
\citation{Palpanas2003,John1995}
\citation{Foss,Myrtveit,Jorgensen2007}
\citation{Foss}
\@writefile{toc}{\contentsline {section}{\numberline {7}Conclusions}{20}}
\newlabel{sec:conclusions}{{7}{20}}
\@writefile{toc}{\contentsline {subsection}{\numberline {7.1}Answers To Research Questions}{21}}
\bibstyle{abbrv}
\bibdata{myref}
\bibcite{supplementaryResults}{1}
\bibcite{Alpaydin2004}{2}
\bibcite{Angelis2000}{3}
\bibcite{Auer2006}{4}
\bibcite{baker07}{5}
\bibcite{Boehm2000}{6}
\bibcite{Boehm1981}{7}
\bibcite{Briand1999}{8}
\@writefile{toc}{\contentsline {section}{\numberline {8}Future Work}{22}}
\newlabel{sec:future-work}{{8}{22}}
\@writefile{toc}{\contentsline {section}{\numberline {9}Acknowledgements}{22}}
\newlabel{sec:acknowledgements}{{9}{22}}
\bibcite{Conte1986}{9}
\bibcite{Cressie1993}{10}
\bibcite{Desharnais1989}{11}
\bibcite{Foss}{12}
\bibcite{Frank03}{13}
\bibcite{jeffery2001}{14}
\bibcite{John1995}{15}
\bibcite{Jorgensen2004}{16}
\bibcite{Jorgensen2009}{17}
\bibcite{Jorgensen2007}{18}
\bibcite{Kadoda2000}{19}
\bibcite{kemerer87}{20}
\bibcite{Kitchenham2009}{21}
\bibcite{Mendes2007}{22}
\bibcite{Kitchenham2007}{23}
\bibcite{Li2006}{24}
\bibcite{Li2009}{25}
\bibcite{Lum2008}{26}
\bibcite{Mendes2002a}{27}
\bibcite{Mendes2008a}{28}
\bibcite{Mendes2002}{29}
\bibcite{Mendes2003}{30}
\bibcite{Menzies2006}{31}
\bibcite{Milic2004}{32}
\bibcite{Molokken-Ostvold2004}{33}
\bibcite{Myrtveit1999}{34}
\bibcite{Myrtveit}{35}
\bibcite{Palpanas2003}{36}
\bibcite{Pendharkar2005}{37}
\bibcite{Premraj2007}{38}
\bibcite{Robson2002}{39}
\bibcite{Scheid2004}{40}
\bibcite{Scott1992}{41}
\bibcite{Shepperd2007}{42}
\bibcite{Shepperd2001}{43}
\bibcite{Kadoda2001}{44}
\bibcite{Shepperd1997}{45}
\bibcite{Shepperd1996}{46}
\bibcite{Stensrud2002}{47}
\bibcite{Stensrud}{48}
\bibcite{Wand1994}{49}
\newlabel{fig:50-sample-points}{{13(a)}{25}}
\newlabel{sub@fig:50-sample-points}{{(a)}{25}}
\newlabel{fig:100-sample-points}{{13(b)}{25}}
\newlabel{sub@fig:100-sample-points}{{(b)}{25}}
\newlabel{fig:1000-sample-points}{{13(c)}{25}}
\newlabel{sub@fig:1000-sample-points}{{(c)}{25}}
\@writefile{lof}{\contentsline {figure}{\numberline {13}{\ignorespaces The effect of sample size and bandwidth on kernel density estimation. The choice of optimum bandwidth (\textit  {h} value) is important. However, even with the optimum bandwidth, one still needs enough number of points for successful kernel density estimation. In this figure the best $h$ value is $1$. In terms of sample size, the value of $50$ appears to be too small. As we increase the sample size to $100$, we get better fit between estimated and actual probability values. But for a very close fit, we need to go up to $1000$ sample points.}}{25}}
\@writefile{lof}{\contentsline {subfigure}{\numberline{(a)}{\ignorespaces {50 Sample Points: Note the bad fit due to low sample size.}}}{25}}
\@writefile{lof}{\contentsline {subfigure}{\numberline{(b)}{\ignorespaces {100 Sample Points: Note the better fit due to increased sample size.}}}{25}}
\@writefile{lof}{\contentsline {subfigure}{\numberline{(c)}{\ignorespaces {1000 Sample Points: Note the optimum fit due to high sample size.}}}{25}}
\newlabel{fig:sample-size-kernel-estimation}{{13}{25}}