12 docs tagged with "scikit-learn"

One decision left — how many rows the model looks at before each step. "Stochastic is faster" turns out to need a unit attached, because batch beat single-row descent on passes, on seconds and on final cost, single-row descent could not converge at all at a constant step size, and yet mini-batch 32 reached a loose target after reading 4% of one pass where batch read 50 full passes and still missed.

Turning categories into numbers without inventing an order — how much label encoding really costs, why it depends on the model, what cardinality does to memory, and the leak hiding inside target encoding.

Why MAE, MSE and RMSE can report three different numbers for the same predictions, why RMSE is never below MAE, what their ratio tells you about outliers, and why scikit-learn hands back a negative mean squared error.

Which models need scaling and which are provably indifferent, why one outlier ruins MinMaxScaler, what log transforms actually fix, and how large the scaling leak really is.

The loss function is how you tell a model what "good" means, and changing it changes the answer. Three losses gave three different slopes on identical data; absolute error ignored an outlier at 60, 200 and 10000 alike; counting mistakes turned out to be unminimisable, with 1998 of 2000 gradients exactly zero; and squared error on a sigmoid was 11014x weaker than cross-entropy exactly where the model was most wrong.

Why a value is missing decides what you may do about it — MCAR, MAR and MNAR, what mean imputation costs a distribution, when a smarter imputer pays, and the fit/transform contract that keeps it honest.

More than one feature — the normal equations solved by hand, why a coefficient changes when a correlated feature joins it, and the measurement showing Adjusted R² fails to penalise 100 columns of pure noise while cross-validation catches them cleanly.

What the IQR rule really flags, why outliers hide each other, the extreme point no single column can see, which treatment actually works, and the features a model cannot invent for itself.

Settles why R² can exceed 1 or go negative, why SST = SSR + SSE only holds for a least-squares fit, why adding a junk feature never lowers R² but does lower Adjusted R², and why scikit-learn's MAPE returns 0.0296 rather than 2.96%.

Fitting a line to one feature — the two coefficients derived by hand and checked against scikit-learn, the five error measures computed from the same residuals, and why a genuinely useful feature can look exactly like noise.

The five Python libraries that do the work, how they stack, scikit-learn's fit/predict/transform contract, and the seven-step pipeline every project follows.

Why a held-out set is mandatory, how much the score depends purely on random_state, and the four situations where a random split reports a number that is simply false.