papers/hanzo-ai-commerce/sections/06-lessons-learned.tex

\section{Lessons Learned}


\subsection{Data Quality Trumps Model Complexity}

Early iterations focused on sophisticated model architectures. We found greater improvements from:
\begin{itemize}
\item Fixing data pipeline bugs.
\item Adding missing features (device type had been dropped).
\item Correcting label noise (returns flagged as purchases).
\end{itemize}

Simple models on clean data outperformed complex models on noisy data.

\subsection{Feedback Loops Require Careful Design}

Recommendations create feedback loops: recommended items get more views, which makes them appear more popular, which increases recommendation probability. We address this through:
\begin{itemize}
\item Exploration-exploitation balancing.
\item Position debiasing in training.
\item Diversity constraints in serving.
\end{itemize}

\subsection{Explain ability Builds Trust}

Merchants hesitated to trust AI pricing and fraud decisions. Adding explanations (``Price increased due to low inventory and high demand'') increased adoption from 34\% to 78\% of eligible merchants.