CRI.Z: Efficient Relative-Importance Measure

• CRI.Z is a relative-importance measure that uses reduced SVD to assign explained variance directly, aiding variable ranking.
• It eliminates the complex reallocation step from CRI, reducing computational burden in high-dimensional contexts (p > n).
• Empirical studies show CRI.Z improves support recovery and prediction accuracy in correlated predictor scenarios compared to traditional methods.

CRI.Z is a computationally efficient relative-importance measure for variable ranking and filter-based selection, introduced as a variant of Comprehensive Relative Importance (CRI) in “Variable Selection Using Relative Importance Rankings” (Chang et al., 13 Sep 2025). It is formulated for standardized linear-model settings, including $p>n$, and uses a reduced singular value decomposition (SVD) of the predictor matrix to assign each predictor the explained variance captured in an orthogonal basis. In the paper’s formulation, CRI.Z is a parameter-free generalization of the CAR score to arbitrary $X$, with substantially lower computational burden than General Dominance (GD) and lower memory cost than full CRI (Chang et al., 13 Sep 2025).

1. Conceptual position within relative-importance analysis

CRI.Z arises from the paper’s attempt to connect variable selection with relative importance (RI) analysis. The stated motivation is that RI measures incorporate both direct and combined effects of predictors, thereby addressing a key limitation of marginal correlation, which ignores dependencies among predictors. Within this framework, GD, CRI, and CRI.Z are used not merely for post-hoc explanation but for predictor ranking before model creation (Chang et al., 13 Sep 2025).

The paper situates CRI.Z alongside two other RI measures. GD measures the average incremental contribution of a predictor to model fit, typically $R^2$, across all possible submodels. CRI generalizes Relative Weights (RW) to arbitrary $X$, including $p>n$, by transforming correlated predictors to an orthogonal basis, allocating explained variance there, and reallocating it back to the original predictors. CRI.Z is derived from CRI by replacing the reallocation term with the identity, so that each predictor is assigned the explained variance captured in the orthogonal basis directly. This suggests that CRI.Z preserves the core orthogonalization logic of CRI while removing the computationally expensive reallocation step (Chang et al., 13 Sep 2025).

The resulting measure is positioned as the preferred RI score when $p$ is large. The paper explicitly contrasts this with GD, which is computationally prohibitive beyond small $p$, and with CRI, whose $O(p^2)$ reallocation matrix can be heavy in large problems. In that sense, CRI.Z is not presented as a separate modeling paradigm, but as the efficient ranking component within an RI-based variable-selection pipeline (Chang et al., 13 Sep 2025).

2. Mathematical construction

The common preprocessing step is to center $y$ and each column $x_j$ of $X$0 to zero mean and scale them to unit $X$1 norm. Let the reduced SVD of the standardized predictor matrix be

$X$2

with rank $X$3, $X$4, $X$5, and $X$6 (Chang et al., 13 Sep 2025).

CRI is defined through two components. The vector

$X$7

contains the correlations, or loadings, of $X$8 in the orthogonalized latent coordinates, while

$X$9

encodes how orthogonal components map back to original predictors. The CRI score vector is

$R^2$0

where $R^2$1 denotes the Hadamard product (Chang et al., 13 Sep 2025).

CRI.Z removes the reallocation step entirely. Its definition is

$R^2$2

so that the predictor-specific score is

$R^2$3

The paper describes this as replacing the CRI reallocation term with the identity. A plausible implication is that CRI.Z can be read as the squared correlation of each original predictor with the orthogonalized representation of the response, without the subsequent redistribution imposed by CRI (Chang et al., 13 Sep 2025).

In the full-column-rank case with $R^2$4, CRI.Z reduces to Johnson’s minimal transformation and to the CAR score. The corresponding orthogonalization is

$R^2$5

with scores

$R^2$6

This special case anchors CRI.Z in the earlier relative-weights and CAR literature while extending the construction to arbitrary $R^2$7 through SVD (Chang et al., 13 Sep 2025).

3. Algorithmic form and computational profile

The CRI.Z algorithm is deliberately minimal. After standardization, one computes the reduced SVD $R^2$8, then forms

$R^2$9

The vector $X$0 is returned as the CRI.Z score vector (Chang et al., 13 Sep 2025).

The computational cost is dominated by the reduced SVD, with complexity $X$1. The subsequent multiplications require only $X$2 time and $X$3 memory. Because CRI.Z does not construct a $X$4 reallocation matrix, it avoids the $X$5 time and space costs that appear in CRI. It also avoids the exponential $X$6 model-fitting burden of GD, which requires enumeration of all submodels and is described as feasible only for small $X$7 (Chang et al., 13 Sep 2025).

This computational profile is central to the method’s intended use. The paper explicitly recommends CRI.Z when $X$8 is large and especially when $X$9, where reduced-SVD computation remains feasible because $p>n$0. The method is also described as parameter-free, in contrast to CAR’s shrinkage-based workaround for arbitrary $p>n$1, because CRI.Z requires no covariance shrinkage or tuning (Chang et al., 13 Sep 2025).

The implementation guidance is correspondingly simple. The repository implements CRI and CRI.Z in R, while GD and CAR are obtained from third-party packages, specifically relaimpo and care. The paper also notes that CRI.Z can be implemented directly with base R SVD (Chang et al., 13 Sep 2025).

4. Use in ranking and filter-based selection

The intended workflow is to compute CRI.Z scores, rank predictors in descending order, and then select variables by a filter rule before fitting the final model. The paper describes two main selection strategies. The first is a top-$p>n$2 rule, in which $p>n$3 is chosen by validation or cross-validation. The second is thresholding, in which variables are retained when $p>n$4 exceeds a data-dependent threshold such as an elbow in the scree of scores (Chang et al., 13 Sep 2025).

The downstream model is then fit on the selected variables. Two variants are emphasized. “LS-RI” denotes least squares on the selected set, and “Ridge-RI” denotes ridge regression on the selected set. The paper recommends LS-RI for interpretability and Ridge-RI for predictive accuracy, particularly when predictors are correlated or $p>n$5 is high. It also recommends validation for choosing both $p>n$6 and the ridge penalty $p>n$7 (Chang et al., 13 Sep 2025).

The method is especially motivated by settings involving correlated clusters, suppressors, or weak predictors. The paper states that CRI.Z orthogonalizes $p>n$8 and scores variables by squared correlation in the orthogonal basis, which captures joint effects and mitigates suppression. As a result, variables with weak or zero marginal correlation can still receive large CRI.Z scores if they contribute in the orthogonalized space. The same logic is used to explain why RI-based methods can rank truly relevant variables more accurately than marginal correlation in the presence of multicollinearity (Chang et al., 13 Sep 2025).

The paper’s practical defaults are correspondingly conservative: always standardize inputs; use CRI.Z when $p>n$9 is large; choose the number of selected variables by validation; and prefer ridge when correlation is strong or $p$0 is large. This suggests that CRI.Z is best viewed as a ranking mechanism that is intentionally decoupled from the final predictive estimator (Chang et al., 13 Sep 2025).

5. Empirical behavior and comparative results

The empirical study is divided into ranking performance and predictive modeling. In the low-dimensional setting $p$1, all methods perform reasonably in an equicorrelated strong-signal example, though performance degrades as correlation increases. In the suppressor example, SIS fails, while GD and CRI slightly outperform CRI.Z; all RI measures remain robust. In the weak-predictor example, CRI.Z performs better than CRI under a weak signal, and SIS fails again (Chang et al., 13 Sep 2025).

In the high-dimensional “High-100” setting with $p$2, SIS performs poorly in the suppressor and weak-predictor examples. CRI.Z, together with CAR, consistently outperforms CRI, particularly under higher correlations. The paper also notes that none of the methods perfectly recover all true predictors in the hardest cases, but RI measures remain more reliable than SIS (Chang et al., 13 Sep 2025).

For predictive modeling, the paper reports that LS-RI variants, especially LS-CRI.Z, achieve superior support recovery measured by F1-score in clustered-predictor scenarios with moderate or high correlation. The comparison set includes best subset, forward stepwise, lasso, and relaxed lasso. At low signal-to-noise ratio, lasso and relaxed lasso may achieve lower relative test error (RTE) because of shrinkage, but Ridge-CRI.Z closes and often surpasses this gap, outperforming in both F1 and RTE when $p$3 is properly tuned (Chang et al., 13 Sep 2025).

In the high-dimensional experiments, LS-CRI.Z and LS-CAR emerge as the best RI variants and are described as robust and scalable. In clustered predictor scenarios, LS-RI variants achieve the highest F1-scores, while forward stepwise collapses under high correlation. Ridge-RI further improves RTE. The paper’s overall conclusion is that RI-based ranking, and CRI.Z in particular, is a powerful and competitive alternative to lasso-dominated selection pipelines, especially when predictor dependence is strong (Chang et al., 13 Sep 2025).

6. Theoretical interpretation, limitations, and scope of the label

The theoretical framing of CRI.Z is deliberately modest. GD is identified with the Shapley value for the cooperative game $p$4, inheriting fairness axioms and additivity. CRI and CRI.Z, by contrast, rely on orthogonalization rather than submodel averaging. The paper states that, under orthonormal columns and standardized $p$5, the sum of CRI.Z scores approximates the explained variance in the orthogonal basis, but it does not claim exact equality for arbitrary $p$6. It also states that no formal consistency guarantees are provided; the case for the method is empirical rather than asymptotic (Chang et al., 13 Sep 2025).

The limitations are correspondingly specific. GD is computationally infeasible for moderate or large $p$7. CRI can introduce extra variability in very high-dimensional or highly correlated settings because of the full reallocation step. Lasso and relaxed lasso may be preferable at extremely low signal-to-noise ratio, where variance reduction dominates. In settings with purely uncorrelated predictors and sparse strong signals, SIS and lasso can be competitive, while CRI.Z remains competitive but not necessarily superior (Chang et al., 13 Sep 2025).

The label “CRI.Z” also appears in other arXiv contexts with different meanings. In “SDR-CIR: Semantic Debias Retrieval Framework for Training-Free Zero-Shot Composed Image Retrieval,” the string occurs as a shorthand for zero-shot composed image retrieval, written as “CRI.Z; ZS-CIR,” rather than as a variable-selection score (Sun et al., 4 Feb 2026). In cosmic-ray transport literature, the notation can also be interpreted as the cosmic-ray ionization rate $p$8, as discussed in the software paper on criptic and in the molecular-cloud ionization work CRIME (Krumholz et al., 2022, Krause et al., 2015). Within statistics and machine learning, however, CRI.Z refers specifically to the SVD-based RI ranking measure proposed in (Chang et al., 13 Sep 2025).