Ikibazo gihuza ibibazo · Hagati Urwego
Ikibazo cya Lasso na Elastic Net Regression Ikibazo
Ongera usuzume uburyo ibihano L1 na L2 bigira ingaruka kuri coefficient, kuki gupima no gutandukanya ibintu, nuburyo abahanuzi bafitanye isano bigira ingaruka kumahitamo.
Ikibazo 1 Bya 8
A lasso fit assigns an exactly zero coefficient to a feature. What does that indicate within this fitted model?
Reba buri kibazo8 ibibazo
Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.
- A lasso fit assigns an exactly zero coefficient to a feature. What does that indicate within this fitted model?
- The feature is excluded from its linear prediction
- The feature has no association in every reasonable model
- The feature was removed during data preprocessing
- The feature has no variation in the source population
- Which penalty combination defines elastic net in the guide?
- L1 penalty with alpha denoting only the L1 share
- L1 plus L2
- L1 penalty only, with no L2 component
- L2 penalty with L1 used only during feature scaling
- Two highly correlated predictors alternate as the one retained by lasso across resamples. What does this show?
- Standardization guarantees stable selection whenever predictors are correlated
- Any change in selected features proves the test set leaked into cross-validation
- Their selection can be unstable despite similar predictions
- Lasso is guaranteed to assign identical coefficients across resamples
- Why should scaling be learned inside each cross-validation training fold?
- Fit the scaler once on all rows, then split for cross-validation
- Fit scaling parameters on validation rows only
- Scale after selecting the final penalty using test data
- To keep validation-fold information out of preprocessing
- In the common mixing-ratio convention, which endpoint corresponds to pure lasso?
- The endpoint with all L2 penalty
- The endpoint with all L1 penalty
- The endpoint with no penalty
- The setting that removes the intercept
- A dataset has predictors measured in dollars and in tiny fractions. Why can scaling matter for penalized regression?
- Feature scaling cannot affect penalty behavior when units change
- Rescaling the target is the only preprocessing relevant to coefficients
- The penalty acts on coefficient magnitudes, which depend on feature units
- Scaling after fitting leaves the selected coefficients unchanged
- When should a final test set be used if penalty settings are chosen by cross-validation?
- Only if training error is zero
- Repeatedly to choose each penalty value
- Before fitting the preprocessing pipeline
- After model selection, for a final evaluation
- Which statement about a zero lasso coefficient is appropriately cautious?
- It is a result of this penalized fit, not universal proof of no relationship
- It proves this predictor has no causal effect
- It will remain zero across populations and penalty settings
- It establishes that no nonlinear relationship exists