The Web Conference (WWW)

\documentclass[sigconf,screen]{acmart}
\usepackage{graphicx}
\usepackage{amsmath,amssymb}
\usepackage{booktabs}

\acmConference[WWW '26]{The ACM Web Conference 2026}{April 2026}{City, Country}
\acmISBN{978-1-4503-XXXX-X/26/04}
\acmDOI{10.1145/XXXXXXX.XXXXXXX}
\setcopyright{acmlicensed}
\copyrightyear{2026}
\acmYear{2026}

\begin{document}

\title{SeaRank: Session-Aware Personalization\\for Cold-Start Web Search}

\author{First Last}
\affiliation{\institution{University of Example}\country{Country}}
\email{[email protected]}
\author{Jane Doe}
\affiliation{\institution{Example Research Labs}\country{Country}}
\email{[email protected]}
\renewcommand{\shortauthors}{Last and Doe}

\begin{abstract}
Personalization is a workhorse of modern search engines, but cold-start
users lack history to drive models. SeaRank exploits in-session
behavioral signals combined with a pretrained transformer to close the
gap. In live traffic experiments, SeaRank improves session-level
nDCG@10 by 6.2\% for new users without harming established users.
\end{abstract}

\keywords{web search, personalization, cold start, session models}
\maketitle

\section{Introduction}
New users are the hardest to serve well. Historical features are
missing by definition, yet the session signal---queries, clicks,
dwells---accumulates quickly.

\section{Related Work}
Session-based search, learning-to-rank, transformer-based rerankers.

\section{Method}
We train a cross-attention model over query, candidate document, and
session prefix. Pretraining uses public click logs; fine-tuning uses
small cold-start subsets.

\section{Experiments}
\begin{table}[t]
\centering
\begin{tabular}{lccc}
\toprule
Model & nDCG@10 & MRR & Sessions helped \\
\midrule
Baseline LTR      & 0.412 & 0.278 & --- \\
BERT reranker     & 0.443 & 0.298 & 11\% \\
\textbf{SeaRank}  & \textbf{0.478} & \textbf{0.322} & \textbf{23\%} \\
\bottomrule
\end{tabular}
\caption{Offline results on cold-start evaluation splits.}
\end{table}

\section{Live A/B Test}
Deployed to 5\% of new-user traffic, SeaRank improved session-level
nDCG@10 by 6.2\% with no regression for logged-in users.

\section{Conclusion}
Cold start is tractable when session signals are first-class citizens
in the ranking stack.

\bibliographystyle{ACM-Reference-Format}
\bibliography{refs}
\end{document}