papers/hanzo-analytics-ml/sections/07-model-training-pipeline.tex

\section{Model Training Pipeline}


\subsection{Training Data Preparation}

\begin{algorithm}
\caption{Training Data Generation}
\begin{algorithmic}[1]
\Function{GenerateTrainingData}{$\mathcal{E}$, $horizon$}
    \State $samples \gets []$
    \For{$c \in customers(\mathcal{E})$}
        \For{$t \in observation\_points(c)$}
            \State $features \gets$ \Call{ComputeFeatures}{$c$, $t$}
            \State $label \gets$ \Call{HasPurchase}{$c$, $[t, t + horizon]$}
            \State $samples.append((features, label))$
        \EndFor
    \EndFor
    \State \Return $samples$
\EndFunction
\end{algorithmic}
\end{algorithm}

\subsection{Class Imbalance Handling}

Purchase events are rare (typically 2-5\% of sessions convert). We address imbalance through:

\begin{enumerate}
    \item SMOTE oversampling of minority class
    \item Class weights in loss function: $w_{pos} = \frac{n_{neg}}{n_{pos}}$
    \item Threshold optimization on validation set
\end{enumerate}

\subsection{Model Retraining}

Models are retrained on a weekly schedule:

\begin{itemize}
    \item Training data: Rolling 180-day window
    \item Validation: 14-day holdout
    \item A/B testing: 5\% traffic for new model evaluation
    \item Rollback criteria: >1\% AUC degradation
\end{itemize}