CC-Select: Diverse Selection Methods

• CC-Select is an umbrella term for selection methods that address uncertainty and combinatorial challenges across diverse domains.
• In display advertising, it aligns with CECS, using an encoder–decoder architecture to efficiently select multi-element ad creatives and boost performance metrics like CTR.
• In selective classification and conformal selection, CC-Select frameworks deploy calibrated, confidence-aware mechanisms to control risk and enhance prediction reliability.

“CC-Select” is not a single standardized designation in the arXiv literature. The label is used directly for calibrated selective classification, is mapped in practice to CECS for multi-element creative selection in display advertising, appears as a conceptual shorthand around contrastive feature selection and confidence-aware selective classification, and is discussed in relation to conformal selection methods such as cfBH and NCCS (Fisch et al., 2022, Zhang et al., 2023, Weinberger et al., 2023, Wu et al., 2024, Choi, 21 Jun 2026). This suggests a shared motif of selection under uncertainty or combinatorial structure, but the underlying objectives, statistical assumptions, and algorithmic mechanisms differ.

1. Terminological scope and disambiguation

Across the cited literature, “CC-Select” functions as a context-dependent label rather than a unique algorithmic name. One paper uses it explicitly for calibrated selective classification; another states that the query naturally maps to “Creative Combination Selection” or “Cross-Element Combinatorial Selection”; another uses it as a conceptual description for contrastive feature selection; and a conformal-selection paper discusses “CC-Select / cfBH” as an existing family against which NCCS is positioned (Fisch et al., 2022, Zhang et al., 2023, Weinberger et al., 2023, Choi, 21 Jun 2026). A related selective-classification paper further states that there is no explicit method named “CC-Select” in that work, but that a generic term such as “Confidence-aware Contrastive Selective” would describe its family (Wu et al., 2024).

Usage in the literature	Canonical name in the paper	Domain
Creative combination selection	CECS	Display advertising
Calibrated selective classification	CC-Select	Selective classification
Contrastive feature selection	CFS	Contrastive analysis
Confidence-aware contrastive selective classification	CCL-SC	Selective classification
Null-calibrated conformal selection related to CC-Select/cfBH	NCCS	Conformal selection

This terminological spread creates a recurring misconception: that “CC-Select” refers to a single transferable framework. The literature instead supports a narrower reading in each domain. In display advertising it denotes an encoder–decoder selector over creative elements; in selective classification it denotes abstention with calibrated uncertainty; in conformal selection it refers to prediction-oriented score-based selection procedures that NCCS reinterprets through target-membership scoring.

2. Cross-element combinatorial selection in display advertising

In display advertising, “CC-Select” maps to CECS, the “Cross-Element Combinatorial Selection” framework for multi-element ad creatives (Zhang et al., 2023). The task is to choose, for each request $(u,C)$, a creative combination $A$ that maximizes

$P(A, label=1 \mid C, u; \Theta),$

where $A$ is an ad creative combination, $C$ is the set of candidate elements, $u$ is the user, and $label=1$ denotes positive user feedback. The combinatorial difficulty is explicit: with $M$ creative element types and average candidate count $N$, enumeration scales as $N^M$. The paper argues that mainstream industrial practice reduces this to $A$0 by selecting each element independently, but thereby ignores cross-element interactions.

CECS addresses that gap with an encoder–decoder architecture. Its encoder, Cross-Element Interaction (CEI), computes a central representation for each category by mean pooling,

$A$1

followed by $A$2. Each candidate embedding is then updated through attention over category prototypes, yielding an interaction-aware representation $A$3. Its decoder, Cascade-Element Selection (CES), recasts creative combination as a sequential selection problem over categories. A GRU propagates selection state,

$A$4

and a pointer-style mechanism produces a distribution over candidates in the current category. The stated effect is to avoid explicit enumeration of $A$5 combinations and instead estimate creative elements sequentially in $A$6 steps.

Training uses per-category cross-entropy together with uncertainty weighting,

$A$7

with $A$8 a learnable scalar for category $A$9. Evaluation is specialized to combinatorial creative selection rather than conventional ranking. The paper defines Hit Ratio and Precision with type weights $P(A, label=1 \mid C, u; \Theta),$0, $P(A, label=1 \mid C, u; \Theta),$1, and $P(A, label=1 \mid C, u; \Theta),$2 for banner, main title, and subtitle.

The reported dataset contains 7.4 million samples, 5,494,110 users, and 157,611 shops, with 90% training and 10% testing. Under candidate truncation to length 5 per type, CECS reaches $P(A, label=1 \mid C, u; \Theta),$3 and $P(A, label=1 \mid C, u; \Theta),$4, exceeding the strongest Multi-CTR baseline, AutoInt + MMoE, at $P(A, label=1 \mid C, u; \Theta),$5 and $P(A, label=1 \mid C, u; \Theta),$6. In online A/B tests against an MMoE-based Multi-CTR baseline, CECS produces a $P(A, label=1 \mid C, u; \Theta),$7 CTR lift and a $P(A, label=1 \mid C, u; \Theta),$8 GMV lift. The paper explicitly notes that there is no module called “CC-Select”; rather, the entire CECS framework is a creative combination selection algorithm. In this usage, CC-Select is therefore best understood as a practical name for CECS rather than as a separate method.

3. Calibrated selective classification

In selective classification, CC-Select is the method introduced in “Calibrated Selective Classification” (Fisch et al., 2022). The setting augments a base classifier $P(A, label=1 \mid C, u; \Theta),$9 with a selector $A$0, so that the system predicts when $A$1 and abstains when $A$2. The central claim is not merely that abstention should improve effective accuracy, but that the accepted predictions should also carry well-calibrated probabilities. This property is formalized as selective calibration: $A$3 for all confidence levels $A$4 in the accepted region.

The paper defines coverage $A$5, selective risk

$A$6

and selective calibration errors such as S-BCE and S-TCE. Its methodological contribution is a separate selector network trained to minimize a kernel-based surrogate of selective calibration error, S-MMCE, rather than to rank examples only by confidence. The selector is a 3-layer feed-forward ReLU MLP with hidden size 64 and sigmoid output. Its input is a meta-feature vector derived from the base model: top probability, predicted class index, full probability vector, and density or outlier features computed from the last hidden representation, including KDE score, Isolation Forest score, one-class SVM score, and average distance to $A$7 nearest neighbors.

The training objective relaxes selective calibration through S-MMCE-based surrogates and decouples coverage control from selector learning. A soft selector $A$8 is trained with a regularized loss that discourages collapse to reject-all behavior, and a threshold $A$9 is chosen on an unlabeled tuning set to enforce target coverage. For robustness under distribution shift, the paper uses a training strategy inspired by group distributionally robust optimization: perturbations generate shifted datasets, selective calibration error is evaluated on these groups, and the selector is optimized on the worst groups.

Experiments are reported on CIFAR-10-C, ImageNet-C, and lung cancer risk assessment from NLST to MGH. Across coverage levels, S-MMCE-trained selectors achieve lower selective calibration error AUC than the full model, confidence-based selection, and outlier-based baselines. On CIFAR-10-C, the reductions in S-TCE$C$0 AUC versus confidence-based selection are summarized as “~18–25% or more,” and the per-corruption breakdown shows improvements across all corruption types. On the lung cancer task, the selector yields substantial reductions in selective calibration error and Brier score AUC while avoiding the simple behavior of rejecting all high-risk cases. In this usage, CC-Select denotes a calibrated abstention mechanism layered on top of a fixed probabilistic predictor.

4. Contrastive selective-classification and contrastive feature-selection usages

A later selective-classification line does not name its method CC-Select, but explicitly positions it as a confidence-aware contrastive selective framework of that kind (Wu et al., 2024). “Confidence-aware Contrastive Learning for Selective Classification” introduces CCL-SC, where the confidence function is Softmax Response,

$C$1

and the theoretical motivation is a generalization bound that depends on intra-class feature variance $C$2. This shifts the design focus from output heads to the feature layer $C$3. The method defines positives as samples correctly predicted as the anchor’s class and negatives as samples whose true label is not the anchor’s class but that are incorrectly predicted as that class. A momentum encoder and class-conditioned queues support a confidence-aware supervised contrastive loss $C$4, weighted by $C$5, and training uses

$C$6

The reported result is significantly lower selective risk than state-of-the-art methods across almost all coverage degrees on CIFAR-10, CIFAR-100, CelebA, and ImageNet, with further gains when combined with SAT+EM.

A different use appears in contrastive analysis, where “contrastive feature selection methods like CC-Select” are described as methods for finding features that carry signal in a target dataset that is absent or attenuated in a background dataset, although the actual method introduced there is CFS rather than CC-Select (Weinberger et al., 2023). CFS works with two unlabeled datasets, target $C$7 and background $C$8, and learns a background proxy $C$9 from $u$0 using an autoencoder. It then freezes $u$1 and learns a gated feature subset on $u$2 via

$u$3

Its information-theoretic analysis shows

$u$4

supporting a two-stage separation between background variation $u$5 and salient variation $u$6. Empirically, CFS-Pretrained and CFS-SG consistently outperform CAE, DUFS, STG, and LassoNet on Grassy MNIST and four biomedical datasets.

These two lines use “CC-Select” differently. CCL-SC is a feature-level method for reducing selective risk in abstaining classifiers, whereas CFS is an embedded feature selector for the contrastive-analysis setting. The shared vocabulary lies in “selection” guided by confidence or contrast, not in a common architecture.

5. Conformal selection, membership scores, and NCCS

In conformal selection, CC-Select appears as a reference point for prediction-oriented conformal selection methods such as cfBH, and NCCS is introduced as a target-membership-aware alternative (Choi, 21 Jun 2026). The problem is to select test candidates whose unknown responses fall in a target region $u$7 while controlling the false discovery rate. The paper argues that the natural score is the target-membership probability

$u$8

because any strictly monotone transform of $u$9 gives the Neyman–Pearson oracle ranking at a fixed null selection level.

NCCS learns a score $label=1$0 that approximates $label=1$1 by converting the problem into binary classification on

$label=1$2

Calibration then uses only confirmed non-target calibration examples,

$label=1$3

and defines

$label=1$4

Under score independence and null exchangeability, these are finite-sample valid null p-values. They can be combined with BY under arbitrary dependence or with BH under standard positive-dependence conditions.

The paper isolates when this reinterpretation matters. For mean-monotone targets, conventional prediction-oriented scores and membership scores induce essentially the same ranking. For interval-valued, variance-driven, multimodal, or multi-condition targets, the distinction becomes important because mean-based scores can be misaligned with selection power. In a mean-driven upper-tail experiment, cfBH-clip reports $label=1$5 and power $label=1$6, NCCS-BH reports $label=1$7 and power $label=1$8, and Oracle-BH reports $label=1$9 and power $M$0. In a variance-driven interval experiment, Mean-Null-BH collapses to $M$1 and power $M$2, whereas NCCS-BH gives $M$3 and power $M$4. In rare-target regimes, direct empirical-FDP thresholding can be anti-conservative, while NCCS trades power for finite-sample null validity. Within this literature, NCCS can therefore be read as a target-membership-aware reformulation of CC-Select-style conformal selection.

6. Distinct selector terminology in random constraint satisfaction

A separate 2026 line introduces “contested cluster selectors” ($M$5), not a method named CC-Select, but the terminology is relevant because it uses “selector” in yet another technical sense (Sheshadri, 14 Jun 2026). Here the objects are variables in random CSPs that are non-backbone, carry information about solution-cluster identity, and are repeatedly but unreliably forced by local propagation during backtracking search. The formal definition combines cluster informativeness,

$M$6

with local contestation, where a local propagation process emits a value for $M$7 with probability at least $M$8 but has conditional accuracy at most $M$9.

The empirical context is instrumented DPLL on random 3-SAT near the empirical satisfiability threshold, near-optimal random Vertex Cover, and random 3-XORSAT. A small number of high-contestedness variables accounts for a large fraction of observed backtracking cost. Pinning two or three high-contestedness variables to solution-consistent values reduces backtracking by 70–80% on the reference instances studied, and a static degree–polarity metric yields a simple $N$0 enumeration heuristic with a reported $N$1 speedup over baseline DPLL at $N$2. The XORSAT control sharpens the interpretation: Gaussian elimination exposes affine selector coordinates, whereas DPLL churn concentrates on pivot variables chosen in a poor coordinate system.

This framework addresses search hardness rather than prediction, ranking, or calibration. Its significance for the broader “CC-Select” vocabulary is mainly terminological. It shows that “selector” can denote cluster-label variables in random CSPs, and that the shared language of selection spans settings as different as advertising optimization, abstaining classification, conformal FDR control, contrastive feature selection, and backtracking complexity.

In the current arXiv record, “CC-Select” is therefore best treated as an overloaded research label. Its most explicit and stable uses are CECS in multi-element creative selection (Zhang et al., 2023), calibrated selective classification (Fisch et al., 2022), and conformal-selection procedures contrasted with NCCS (Choi, 21 Jun 2026). Related works extend the phrase toward feature-level selective classification (Wu et al., 2024), contrastive feature selection (Weinberger et al., 2023), and selector variables in random CSPs (Sheshadri, 14 Jun 2026). The term has continuity at the level of selection as an organizing principle, but not at the level of a single method, objective, or theoretical framework.