1. Jure Žbontar
  2. jrs2012

Commits

Jure Žbontar  committed 623de43

Rename.

  • Participants
  • Parent commits 56320e6
  • Branches default

Comments (0)

Files changed (2)

File jrs2012/jrs2012.tex

View file
+\documentclass[14pt]{beamer}
+
+\title{Team ULjubljana's Solution to the JRS 2012 Data Mining Competition}
+\author{
+    Jure Zbontar \and Marinka Zitnik \and Miha Zidar \and \\
+    Gregor Majcen \and Matic Potocnik \and Blaz Zupan
+}
+
+\institute{
+    University of Ljubljana,
+    Faculty of Computer and Information Science
+}
+
+\date{Joint Rough Set Symposium, 2012}
+
+\begin{document}
+\frame{\titlepage}
+
+\begin{frame}
+\frametitle{Overview}
+\begin{itemize}
+\item Train a set of diverse models.
+\item Combine the predictions with stacking.
+\item Decide on the threshold values.
+\end{itemize}
+\end{frame}
+
+\begin{frame}
+\frametitle{Logistic Regression}
+\begin{columns}
+
+\column{.7\textwidth}
+
+\begin{itemize}
+\item Train a L2 regularized logistic regression model on each label separately.
+\item Obtain an F-score of 0.51555.
+\end{itemize}
+
+\column{.3\textwidth}
+\includegraphics[width=\textwidth]{img/f-scores5.pdf}
+
+\end{columns}
+\end{frame}
+
+
+\begin{frame}
+\frametitle{F-score Logistic Regression}
+
+\begin{itemize}
+
+\item Try to maximize F-score directly.
+\item The hypothesis of logistic regression is
+
+$$ h_\theta(x^{(i)}) = \frac{1}{1 + e^{-\theta^T x^{(i)}}} $$
+
+\item We used
+
+$$
+h_\Theta(x^{(i)}) = 
+\left[
+\begin{matrix}
+
+\frac{1}{1 + e^{-\Theta_1^T x^{(i)}}} \\
+\frac{1}{1 + e^{-\Theta_2^T x^{(i)}}} \\
+\vdots \\
+\frac{1}{1 + e^{-\Theta_{83}^T x^{(i)}}}
+
+ \end{matrix}
+\right]
+$$
+
+\end{itemize}
+\end{frame}
+
+
+
+\begin{frame}
+\frametitle{F-score Logistic Regression}
+
+$$
+\mbox{fscore}_i
+= 2 \frac{\alert<4>{|\mbox{\{true labels\}}_i \cap \mbox{\{predicted labels\}}_i|}
+}{\alert<3>{|\mbox{\{predicted labels\}}_i|} + \alert<2>{|\mbox{\{true labels\}}_i|}}
+$$
+
+\begin{itemize}
+\item Trying to optimize fscore directly is hard.
+\item We optimize a smooth approximation instead.
+\end{itemize}
+\
+$$
+\mbox{fscore\_approx}_i = 2
+\frac
+    {\alert<4>{\sum_{l=1}^{83} h_{\Theta_l}(x^{(i)}) y_l^{(i)}}}
+    {\alert<3>{\sum_{l=1}^{83} h_{\Theta_l}(x^{(i)})} + 
+    \alert<2>{\sum_{l=1}^{83} y_l^{(i)}} }
+$$
+
+\end{frame}
+
+
+\begin{frame}
+\frametitle{F-score Logistic Regression}
+\begin{columns}
+
+\column{.7\textwidth}
+
+\begin{itemize}
+\item The F-score obtained with this method was 0.51819.
+\end{itemize}
+
+\column{.3\textwidth}
+\includegraphics[width=\textwidth]{img/f-scores4.pdf}
+
+\end{columns}
+\end{frame}
+
+
+\begin{frame}
+\frametitle{Logistic Regression / Neural Networks}
+\begin{columns}
+
+\column{.7\textwidth}
+\begin{itemize}
+\item Use 5-fold CV and logistic regression to obtain predictions for
+all training cases.
+
+\item Feed the predictions into a multilayer perceptron.
+\item The F-score obtained was 0.52487.
+\end{itemize}
+
+\column{.3\textwidth}
+\includegraphics[width=\textwidth]{img/f-scores3.pdf}
+
+\end{columns}
+\end{frame}
+
+
+\begin{frame}
+\frametitle{Stacking}
+\begin{columns}
+
+\column{.7\textwidth}
+
+\begin{itemize}
+\item Use 5-fold CV to obtain predictions for every base learner.
+\item Train a multilayer perceptron on every label separately.
+\item Stacking improved the F-score to 0.53378.
+\end{itemize}
+
+\column{.3\textwidth}
+\includegraphics[width=\textwidth]{img/f-scores2.pdf}
+
+\end{columns}
+\end{frame}
+
+
+\begin{frame}
+\frametitle{Thresholding}
+\begin{columns}
+
+\column{.7\textwidth}
+\begin{itemize}
+\item Select the best threshold for each label.
+\item Use a greedy approach.
+\item The F-score improved to our final score of 0.53579.
+\end{itemize}
+
+\column{.3\textwidth}
+\includegraphics[width=\textwidth]{img/f-scores1.pdf}
+
+\end{columns}
+\end{frame}
+
+
+\end{document}

File jrs2012/jzb.tex

-\documentclass[14pt]{beamer}
-
-\title{Team ULjubljana's Solution to the JRS 2012 Data Mining Competition}
-\author{
-    Jure Zbontar \and Marinka Zitnik \and Miha Zidar \and \\
-    Gregor Majcen \and Matic Potocnik \and Blaz Zupan
-}
-
-\institute{
-    University of Ljubljana,
-    Faculty of Computer and Information Science
-}
-
-\date{Joint Rough Set Symposium, 2012}
-
-\begin{document}
-\frame{\titlepage}
-
-\begin{frame}
-\frametitle{Overview}
-\begin{itemize}
-\item Train a set of diverse models.
-\item Combine the predictions with stacking.
-\item Decide on the threshold values.
-\end{itemize}
-\end{frame}
-
-\begin{frame}
-\frametitle{Logistic Regression}
-\begin{columns}
-
-\column{.7\textwidth}
-
-\begin{itemize}
-\item Train a L2 regularized logistic regression model on each label separately.
-\item Obtain an F-score of 0.51555.
-\end{itemize}
-
-\column{.3\textwidth}
-\includegraphics[width=\textwidth]{img/f-scores5.pdf}
-
-\end{columns}
-\end{frame}
-
-
-\begin{frame}
-\frametitle{F-score Logistic Regression}
-
-\begin{itemize}
-
-\item Try to maximize F-score directly.
-\item The hypothesis of logistic regression is
-
-$$ h_\theta(x^{(i)}) = \frac{1}{1 + e^{-\theta^T x^{(i)}}} $$
-
-\item We used
-
-$$
-h_\Theta(x^{(i)}) = 
-\left[
-\begin{matrix}
-
-\frac{1}{1 + e^{-\Theta_1^T x^{(i)}}} \\
-\frac{1}{1 + e^{-\Theta_2^T x^{(i)}}} \\
-\vdots \\
-\frac{1}{1 + e^{-\Theta_{83}^T x^{(i)}}}
-
- \end{matrix}
-\right]
-$$
-
-\end{itemize}
-\end{frame}
-
-
-
-\begin{frame}
-\frametitle{F-score Logistic Regression}
-
-$$
-\mbox{fscore}_i
-= 2 \frac{\alert<4>{|\mbox{\{true labels\}}_i \cap \mbox{\{predicted labels\}}_i|}
-}{\alert<3>{|\mbox{\{predicted labels\}}_i|} + \alert<2>{|\mbox{\{true labels\}}_i|}}
-$$
-
-\begin{itemize}
-\item Trying to optimize fscore directly is hard.
-\item We optimize a smooth approximation instead.
-\end{itemize}
-\
-$$
-\mbox{fscore\_approx}_i = 2
-\frac
-    {\alert<4>{\sum_{l=1}^{83} h_{\Theta_l}(x^{(i)}) y_l^{(i)}}}
-    {\alert<3>{\sum_{l=1}^{83} h_{\Theta_l}(x^{(i)})} + 
-    \alert<2>{\sum_{l=1}^{83} y_l^{(i)}} }
-$$
-
-\end{frame}
-
-
-\begin{frame}
-\frametitle{F-score Logistic Regression}
-\begin{columns}
-
-\column{.7\textwidth}
-
-\begin{itemize}
-\item The F-score obtained with this method was 0.51819.
-\end{itemize}
-
-\column{.3\textwidth}
-\includegraphics[width=\textwidth]{img/f-scores4.pdf}
-
-\end{columns}
-\end{frame}
-
-
-\begin{frame}
-\frametitle{Logistic Regression / Neural Networks}
-\begin{columns}
-
-\column{.7\textwidth}
-\begin{itemize}
-\item Use 5-fold CV and logistic regression to obtain predictions for
-all training cases.
-
-\item Feed the predictions into a multilayer perceptron.
-\item The F-score obtained was 0.52487.
-\end{itemize}
-
-\column{.3\textwidth}
-\includegraphics[width=\textwidth]{img/f-scores3.pdf}
-
-\end{columns}
-\end{frame}
-
-
-\begin{frame}
-\frametitle{Stacking}
-\begin{columns}
-
-\column{.7\textwidth}
-
-\begin{itemize}
-\item Use 5-fold CV to obtain predictions for every base learner.
-\item Train a multilayer perceptron on every label separately.
-\item Stacking improved the F-score to 0.53378.
-\end{itemize}
-
-\column{.3\textwidth}
-\includegraphics[width=\textwidth]{img/f-scores2.pdf}
-
-\end{columns}
-\end{frame}
-
-
-\begin{frame}
-\frametitle{Thresholding}
-\begin{columns}
-
-\column{.7\textwidth}
-\begin{itemize}
-\item Select the best threshold for each label.
-\item Use a greedy approach.
-\item The F-score improved to our final score of 0.53579.
-\end{itemize}
-
-\column{.3\textwidth}
-\includegraphics[width=\textwidth]{img/f-scores1.pdf}
-
-\end{columns}
-\end{frame}
-
-
-\end{document}