Fill missing values with the median, then min-max scale

• Compute the median only from the values that are present.
• After imputation, min and max are taken over the completed column. The median always lies between them, so they are the same as for the present values.

present = sorted(int(t) for t in toks if t != "NA")
m = len(present)
med = present[m // 2] if m % 2 else (present[m // 2 - 1] + present[m // 2]) / 2
col = [med if t == "NA" else int(t) for t in toks]
lo, hi = min(col), max(col)
scaled = [0.0 if hi == lo else (v - lo) / (hi - lo) for v in col]

Why the median? It is robust: one huge faulty reading moves the mean a lot but barely moves the median. Mean imputation suits roughly symmetric data. For categories, use the mode. More advanced options: KNN imputation, model-based imputation, or adding an "was missing" indicator column so the model can learn whether missingness itself carries information.

Data leakage. In a real ML pipeline, compute the median, min and max on the training split only, then apply those same numbers to the validation and test data. scikit-learn's SimpleImputer and MinMaxScaler inside a Pipeline do this: fit on train, transform on everything. Computing them on the full data set leaks information about the test set into training.

Min-max scaling vs standardisation

• Min-max maps values to [0, 1]. It is sensitive to outliers, because one extreme value squashes everything else together.
• Standardisation (z = (x − μ)/σ) centres the data at 0 with unit variance. It is preferred for many linear models, SVMs and PCA.
• Tree-based models (decision trees, random forests, gradient boosting) generally do not need either.

Guard the constant column: dividing by max − min = 0 would crash.