Feature Engineering and Data Preparation

Use it when a column carries extreme values. Robust scaling subtracts the median and divides by the interquartile range, statistics that a handful of extreme points barely move, so the bulk of the data still lands on a usable scale.

Interpolation produces fractional values on indicator columns, so a synthetic row comes out 0.5 on two different category flags. It represents no real category, and no such row could ever arrive at prediction time.

The weights changed the objective, so the fit targets the reweighted class mix rather than the real one. Balanced weights make both classes equally heavy, pulling the optimum toward 0.5. The scores still rank correctly but are no longer probabilities.

At 200 columns and 1,500 rows there are roughly seven rows per column, so some will look predictive by chance alone. Judge them by repeated cross-validation with the selection step refitted inside every fold, against a baseline without them.

Counts and sums are genuinely zero — nothing happened, and that is an observation. Means, maxima and shares are undefined because there is nothing to summarise, and recency has no value at all. Encode those with a sentinel plus a companion flag, never a silent zero.

Recursive feature elimination refits the model, ranks columns by its weights, drops the weakest, and repeats — hundreds of fits, times cross-validation folds. Cut cost by dropping a block per round, pre-screening with a cheap filter, or ranking with a faster model.

Forward stepwise selection evaluated hundreds of candidate subsets and reported the winner. The maximum of many noisy estimates is biased upward, so part of the winner's margin is luck — even when the selector saw only training rows.

Three rows cannot support a rate of 1.0. Smoothing shrinks the estimate toward the global rate, weighted by the row count, so some seller signal survives; a rare bucket discards it entirely. Prefer smoothing.

The fill comes from a statistic computed once on the training data and shipped with the model as a fixed parameter. Never recompute it from live traffic: the same request would then score differently depending on its neighbours.

Either it raises an error or it emits all zeros across that column's indicators. All zeros is the dangerous case: if a reference level was dropped, that pattern already means the reference level, so the unknown browser is silently scored as it.

Exponentiating is a convex map, so the average of log-scale predictions does not map back to the average price. Under symmetric log-scale errors it returns the conditional median; with normal log errors of standard deviation s, multiply by exp(s^2 / 2) to recover the mean.

SMOTE hurts when the classes overlap or the minority contains mislabelled points. Interpolating between two minority rows that sit on opposite sides of a majority cloud plants synthetic positives inside majority territory, pushing the boundary outward and costing precision.

Check the arithmetic before pruning: stored sparsely, 20,000 documents touching a hundred terms each is a few million values, not a billion. Cut with document-frequency thresholds at both ends, and let held-out folds decide how far to go.

First decide whether the dip is real or sampling noise in a thin band. If it is noise, merge that band with the neighbour it sits closest to in weight of evidence and re-cut under a monotone constraint, accepting a small loss of information value.

Decide by what is actually broken. If the model already ranks cases well and only the cut-off is wrong, move the operating point - it is free and reversible. Reweight when the fit itself ignores the rare class.

Decide on four axes: how much signal the identity carries, what the model family can consume, what state serving can refresh, and who must explain the feature. Start cheap and escalate only when a holdout says it pays.

Rarely for accuracy, often for everything else. Binning discards within-bin variation and imposes arbitrary cutpoints, so it usually costs predictive power. It earns its place when bins buy interpretability, stable reporting, or let a linear model express a non-monotone effect.

A variance-threshold filter drops every column whose spread across rows falls below a cutoff, removing constant and near-constant features. It ignores the target and depends on units, so it can discard a rare binary flag that predicts strongly.

When the cost of an error varies row by row, not just class by class. In insurance-claim triage, weighting each claim by the euro value at risk makes the model spend its capacity where the money is.

Hour 23 and hour 0 are one hour apart but 23 units apart as integers. Mapping the hour onto a circle with sin(2pih/24) and cos(2pih/24) makes them neighbours again, which matters to any distance- or magnitude-based model.

A raw count mixes how active an entity is overall with how it splits that activity and how long it was observed. Dividing by the entity's total gives composition, and dividing by days observed gives intensity, so the model can separate a heavy user from a lopsided one.

A hash function maps each domain to a bucket index modulo a fixed bucket count, say 2^20, and that bucket is the feature slot. Nothing is stored, so new domains need no special case. Colliding domains share one blended weight.

When the incomplete column is strongly predictable from the other columns. Both borrow that structure — kNN from similar rows, iterative imputation from a regression on the other features — where a median gives every gap the same answer.

Character n-grams win when the tokens themselves are unreliable: misspelled or run-together brand names, inconsistent punctuation, model codes. A typo changes only a few of a word's character n-grams, whereas it destroys the word unigram entirely, so overlap survives.

Usually no. Dropping a level exists to remove the exact redundancy between the full set of indicators and the intercept, which an L2 penalty already resolves on its own. Keeping every level and leaving the intercept unpenalised treats all levels symmetrically.

Min-max equalises each feature's range, not its spread. A heavy-tailed column whose maximum sits far above the bulk gets compressed near zero after scaling, so a well-spread bounded column ends up supplying almost all of the distance.

Rank normalisation replaces each value with its rank, rescaled to a uniform or Gaussian shape. It flattens any tail with no parametric assumption, but it keeps only the ordering: how far apart two values were is discarded.

Apply score = offset + factor times ln(odds), where factor = PDO / ln 2, so 20 / 0.693 is about 28.85, and offset = 600 minus factor times the log of the anchor odds. Every doubling of the odds then adds exactly 20 points.

Neither, until the definition is settled. A label whose meaning changed is a specification defect, not noise: choose one definition tied to the decision the model serves, write it down with edge cases, and rebuild the target consistently.