Many changes to results and conclusions and such

content/experiments/experiments.tex

@@ -287,75 +287,8 @@ MAE                & 0.810{\tiny{$\pm$0.001}} \\
 Focal              & 0.798{\tiny{$\pm$0.009}} \\
 BCE-W              & 0.788{\tiny{$\pm$0.008}} \\
 Asym.              & 0.729{\tiny{$\pm$0.007}} \\
-LibAUC             & --- \\
+LibAUC             & 0.835{\tiny{$\pm$0.002}} \\
 \bottomrule
 \end{tabular}
 \end{table}
 
-
-
-\section{Ablation Study}\label{sec:experiments:ablation}
-\addcontentsline{tocheb}{section}{\protect\numberline{\secnumforhebrewtoc}{ניתוח אבלציה}}
-
-% Isolate the contribution of ROLL's individual design choices to determine which
-% components drive its performance and noise robustness.
-
-\subsection{Effect of KDE Smoothing}\label{sec:experiments:ablation:kde}
-
-% Compare ROLL-AoC against a hard-threshold variant (no KDE; threshold estimated
-% by sorting scores and selecting the rank corresponding to FPR $= \alpha$).
-% Expected finding: without smooth KDE the gradient signal collapses and training
-% is unstable, confirming that differentiability is not just a theoretical convenience.
-
-\subsection{Effect of Kernel Choice}\label{sec:experiments:ablation:kernel}
-
-% Compare Gaussian vs.\ sigmoid kernel in ROLL-AoC on representative datasets.
-% The sigmoid kernel admits closed-form gradients (see \Cref{appendix:kde-sigmoid});
-% this subsection assesses whether kernel choice materially affects accuracy or
-% whether performance is robust to kernel selection.
-
-\subsection{ROLL-AoC vs.\ ROLL-TPR}\label{sec:experiments:ablation:variants}
-
-% Systematically compare the two ROLL variants across all benchmarks.  Discuss when
-% targeting a single operating point (ROLL-TPR) helps (e.g.\ Credit Card Fraud, where
-% a strict FPR budget is natural) vs.\ when whole-curve optimization (ROLL-AoC) is
-% preferable (heterogeneous or unconstrained evaluation).
-
-
-\section{Sensitivity Analysis}\label{sec:experiments:sensitivity}
-\addcontentsline{tocheb}{section}{\protect\numberline{\secnumforhebrewtoc}{ניתוח רגישות}}
-
-% Characterize ROLL's robustness to its key hyperparameters so practitioners can
-% apply it without exhaustive tuning.
-
-\subsection{Bandwidth Sensitivity}\label{sec:experiments:sensitivity:bandwidth}
-
-% Plot AUC vs.\ log-bandwidth $h$ for ROLL-AoC across a representative subset of
-% datasets (at least one from KEEL, CIFAR-10N, and large-scale).  Report the range
-% of $h$ over which performance stays within 1\% of its peak --- this defines a
-% practical safe tuning zone.  Show that Silverman's rule falls within this zone,
-% justifying it as a default.
-
-\subsection{Operating-Point Sensitivity (ROLL-TPR)}\label{sec:experiments:sensitivity:alpha}
-
-% For ROLL-TPR, sweep the target FPR budget $\alpha$ and measure TPR achieved at
-% each setting.  Validate that the trained model's operating point tracks the
-% specified $\alpha$, demonstrating that ROLL-TPR reliably targets the operating
-% point it was trained for and generalises to nearby values.
-
-
-\section{Summary}\label{sec:experiments:summary}
-\addcontentsline{tocheb}{section}{\protect\numberline{\secnumforhebrewtoc}{סיכום}}
-
-% TODO: write 2--3 paragraphs synthesising findings.
-% (1) Noise robustness is the primary result: ROLL-AoC is best at 20\% KEEL noise
-%     and best at all three CIFAR-10N noise levels. The story is consistent across
-%     very different settings (small tabular vs. image).
-% (2) Large-dataset picture: strong on Credit Card Fraud and Home Credit; clear
-%     failure on Bank Marketing (0.511 vs 0.746 for GCE) -- flag as open question.
-% (3) ROLL-TPR is inconsistent: best on Credit Card Fraud (0.974), poor elsewhere.
-%     Its value is as a targeted tool when a specific FPR budget is fixed.
-% Bridge: Discussion chapter \Cref{chapter:discussion_and_conclusions} addresses
-%     the Bank Marketing failure and the mechanism behind noise robustness.
-
-TODO write this section.