Element Selection: Methods & Applications

• Element selection is a family of constrained decision problems that choose candidates from finite sets across domains, ranging from creative advertising to materials discovery.
• Methodologies leverage combinatorial optimization, local search, approximate ranking, and semantic parsing to address unique feedback types and correctness criteria.
• Applications of element selection yield practical improvements such as higher click-through rates, energy efficiency, and faster computations in diverse systems.

“Element selection” is not a single technical notion but a family of selection problems in which the “elements” may be creative components in advertising, coordinates in a vector, GUI or web objects, reflecting elements in RIS/STARS hardware, bitvector positions in rank/select data structures, chemical elements in a phase field, or even program choices parameterized by future loss. Across these literatures, the common core is a constrained decision problem over a finite candidate space, but the mathematical object being selected, the admissible feedback, and the notion of correctness vary sharply by domain (Zhang et al., 2023, Ono, 14 Feb 2026, Marchini et al., 2019, Vasylenko et al., 2022, Plotkin et al., 4 Apr 2025).

Domain	What is selected	Typical objective
Display advertising	One creative element per type	Maximize click probability
Dimension reduction	A subset of coordinates	Minimize regression or reconstruction MSE
Data structures and order statistics	An index, predecessor, or approximate rank element	Support rank/select or approximate order selection
GUI and web systems	Objects, pages, or DOM elements	Match user intent or a prior locator
Wireless systems	Active reflecting/transmitting elements	Improve secrecy or energy efficiency
Materials discovery	Sets of chemical elements	Rank phase fields by property and novelty
Programming semantics	A choice from alternatives	Minimize or maximize future loss

1. Meanings and problem formulations

In display advertising, element selection is formulated as choosing exactly one candidate from each creative type to form a complete ad creative $A=(e_1,\ldots,e_k)$, with the objective

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

where $u$ is the user and $C$ is the set of candidate creative elements (Zhang et al., 2023). The central difficulty is combinatorial: if type $i$ has $N_i$ candidates, the total number of possible creatives is $\prod_i N_i$, or $N^k$ in the uniform case.

In multiplication-free dimension reduction, element selection means choosing a subset of coordinates $$I=\{\sigma(1),\dots,\sigma(K)\}\subset\{1,\dots,N\}$$ and keeping only those entries of $\mathbf{x}$, so that runtime reduction requires no multiplications: $\max_{\Theta}\; P(A, label = 1 \mid C, u; \Theta),$0 The subset is scored by the minimum mean-squared error of linear regression from selected elements to a target vector, yielding the objective

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

or, in block form, $\max_{\Theta}\; P(A, label = 1 \mid C, u; \Theta),$2 (Ono, 14 Feb 2026).

In GUI research, element selection denotes the foundational operation of selecting objects before dragging, copying, deleting, grouping, or formatting them. The main complications are not algebraic but interactional: small or thin targets, heterogeneous object types, hidden or off-screen objects, overlapping layers, and ordering semantics in multiple-object selection (Mishra, 2014). In web testing, the same phrase refers to localizing a DOM or visual web element corresponding to an older oracle element on a changed page (Nass et al., 2023).

In wireless communications, the phrase typically denotes subset activation of reflecting elements. In IRS secrecy analysis, element subset selection activates only $\max_{\Theta}\; P(A, label = 1 \mid C, u; \Theta),$3 reflecting elements, chosen as the top-$\max_{\Theta}\; P(A, label = 1 \mid C, u; \Theta),$4 magnitudes of $\max_{\Theta}\; P(A, label = 1 \mid C, u; \Theta),$5, to minimize secrecy outage probability under outdated CSI (Li et al., 2022). In STARS-aided ISAC, element selection is encoded by binary on-off variables $\max_{\Theta}\; P(A, label = 1 \mid C, u; \Theta),$6, with $\max_{\Theta}\; P(A, label = 1 \mid C, u; \Theta),$7 turning an element off and $\max_{\Theta}\; P(A, label = 1 \mid C, u; \Theta),$8 making it controllable (Shen et al., 22 Jul 2025). In active RIS-assisted massive MIMO, the number of active elements is not represented by an explicit integer variable in the main optimization, but emerges from continuous amplitude control $\max_{\Theta}\; P(A, label = 1 \mid C, u; \Theta),$9 under a global RIS amplification-power budget (Jr et al., 2024).

In materials discovery, the “elements” are chemical elements, and selection occurs at the phase-field level rather than the composition level. A phase field is the set of constituent chemical elements that can form one or more crystalline phases, independent of stoichiometry or structure, and the task is to decide which element sets should be explored because they are likely to yield stable phases with high target properties (Vasylenko et al., 2022).

In programming-language semantics, the selection monad defines element selection abstractly as a function

$u$0

meaning that a selection function chooses an element of $u$1 based on a loss function $u$2 (Plotkin et al., 4 Apr 2025). This is the most general formulation in the set of sources: an element is whatever the language or model takes as a candidate choice.

2. Combinatorial and representational selection

The advertising paper “Cross-Element Combinatorial Selection for Multi-Element Creative in Display Advertising” treats selection as a structured combinatorial optimization problem and proposes CECS, a Cross-Element Combinatorial Selection framework composed of a cross-element interaction encoder and a cascade-element selection decoder (Zhang et al., 2023). The encoder builds interaction-aware representations by attending from each candidate element to type-level central vectors, while the decoder turns creative construction into a sequential cascade over element types using a GRU state and pointer-style attention. This replaces independent per-type CTR scoring with conditional selection across types while preserving practical $u$3 complexity. On offline metrics, the best multi-CTR baseline, AutoInt + MMoE, achieved $u$4 and $u$5, whereas CECS achieved $u$6 and $u$7. In deployment, compared with an MMoE-based Multi-CTR baseline, CECS produced a $u$8 CTR lift and a $u$9 GMV lift (Zhang et al., 2023).

The dimension-reduction paper frames element selection as subset selection under a regression-fidelity criterion and introduces a swap-based local search accelerated by the matrix inversion lemma (Ono, 14 Feb 2026). A swap replaces one selected coordinate with one unselected coordinate, and the change in objective $C$0 is evaluated by a rank-2 update of $C$1 and a rank-1 update of $C$2, avoiding full recomputation of $C$3. On MNIST with $C$4 and $C$5, the normalized reconstruction loss was $C$6 for the proposed element selection and $C$7 for PCA; the paper explicitly notes that PCA is better in reconstruction error, but element selection is multiplication-free at runtime (Ono, 14 Feb 2026). The accelerated local search reduced runtime from about $C$8 seconds in a naive implementation to about $C$9 seconds, and to about $i$0 seconds with additional vectorization, corresponding to a $i$1 speed-up (Ono, 14 Feb 2026).

The N-CMTF paper uses element selection in a coordinate-descent sense. “Columnwise Element Selection for Computationally Efficient Nonnegative Coupled Matrix Tensor Factorization” proposes Cut-CD, which computes an importance score for each scalar coordinate in a column, normalizes the scores within the column, and updates only those elements whose normalized importance exceeds a column-wise cut-off equal to the mean normalized importance (Balasubramaniam et al., 2020). The scalar update is

$i$2

and the importance estimate is

$i$3

The paper argues that, unlike rowwise greedy CD, the updates and importance scores in column $i$4 depend only on the gradients of that column. Its reported complexity for one factor update is $i$5 for Cut-CD versus $i$6 for GCD, and the tabled speedup grows linearly with rank, reaching $i$7 at $i$8 for $i$9 (Balasubramaniam et al., 2020).

Taken together, these works show three distinct meanings of “selection” inside optimization itself: selecting one candidate per slot, selecting a subset of coordinates, and selecting which coordinates are worth updating. This suggests that the term often names a computational bottleneck rather than a single mathematical primitive.

3. Rank, predecessor, and approximate order selection

A second major usage treats selection as an order-statistic operation. “Compact Fenwick trees for dynamic ranking and selection” studies dynamic rank and select on sequences and bitvectors, where select is implemented as predecessor search on prefix sums: $N_i$0 The paper develops compact Fenwick variants and a dynamic bitvector whose updates, ranking, and selection all run in logarithmic time with only a few percent space overhead (Marchini et al., 2019). For $N_i$1, one reported configuration uses about $N_i$2 bits per bit, compared with $N_i$3 bits per bit for Prezza’s dynamic bitvector in the same setting, while also improving practical performance for select through level-order layout and byte compression (Marchini et al., 2019).

“Fast Approximate Rank Determination and Selection with Group Testing” studies total orders augmented with one-sided group tests $N_i$4 or $N_i$5 (Liyanage et al., 16 Jul 2025). Exact minimum or maximum can then be found by a Las Vegas algorithm with $N_i$6 expected queries. For approximate rank determination, the paper gives a Monte Carlo algorithm with query complexity

$N_i$7

returning $N_i$8 such that $N_i$9 with probability at least $\prod_i N_i$0. For approximate selection, it gives a Monte Carlo algorithm with expected query complexity

$\prod_i N_i$1

which outputs an element with probability at least $\prod_i N_i$2, and conditioned on output, the element has the desired approximate rank with probability $\prod_i N_i$3 (Liyanage et al., 16 Jul 2025).

“Multiparty Selection” moves the same order-statistic problem into communication complexity (Chen et al., 2020). For exact selection, it gives a deterministic protocol in the $\prod_i N_i$4-party broadcast model with $\prod_i N_i$5 communication to find the median of the multiset union. For approximate selection, it finds an $\prod_i N_i$6-mediocre element in $\prod_i N_i$7 communication under disjointness and density assumptions, and in the two-party case obtains $\prod_i N_i$8 communication under stronger assumptions (Chen et al., 2020). This is a different resource regime: the complexity measure is broadcast bits, not arithmetic operations or comparisons.

“Robust Max Selection” studies adversarially corrupted comparisons rather than noisy but consistent comparisons (Dang et al., 2024). In this model, the uncorrupted elements are totally ordered, but queries involving corrupted elements may have arbitrary output and may induce cycles. The paper proves that any correct algorithm must output a set of size at least $\prod_i N_i$9 containing the uncorrupted maximum, and that this bound is tight. Restricted to output size exactly $N^k$0, deterministic algorithms require $N^k$1 queries, while a randomized two-stage algorithm achieves $N^k$2 queries with high probability (Dang et al., 2024).

A common misconception is that “selection” in algorithms always means returning a single winner. In robust max selection, that is impossible in general: the model itself forces set-valued output (Dang et al., 2024). A related misconception is that select is always a stand-alone search primitive; the Fenwick-tree paper shows that in succinct data structures it is often implemented as predecessor search on prefix sums (Marchini et al., 2019).

4. Interface and web element selection

In interface research, selection is an interaction mechanism rather than an optimization variable. “Inventions on selecting GUI elements” surveys selection difficulties documented in patents: precision problems for very small or thin objects, the tediousness of selecting many objects one by one, semantic ambiguity in heterogeneous multi-object selections, ordering and grouping logic, off-screen objects, overlapping layers, and the inefficiency of a single pointing device (Mishra, 2014). The cited inventions propose several recurring patterns: time-space object containment, list-box-based indirect selection of notebook pages, explicit “Select All” and “Refresh All” buttons scoped to windows or wells, bimanual interaction with two pointing devices, and tolerance expansion through enlarged invisible hit zones for small objects (Mishra, 2014).

The web-testing paper treats element selection as localization of a changed web element in a new DOM or visual layout (Nass et al., 2023). Its baseline, VON Similo, merges sufficiently overlapping DOM nodes into a visual web element and ranks candidates using weighted similarity over properties such as Tag, Visible Text, Class, Id, Name, HRef, Location, Area, Shape, Alt, Is Button, XPath, and Neighbor Text. VON Similo LLM then takes the top 10 candidates from that ranking and asks GPT-4 to choose the most similar one from structured JSON input. On a dataset of 804 web element pairs from 48 real-world web applications, the detailed results table reports that VON Similo localized 734 elements and failed on 70, while VON Similo LLM localized 764 and failed on 40, corresponding to $N^k$3 versus $N^k$4 located and a $N^k$5 reduction in failures (Nass et al., 2023). The abstract, however, reports reducing failed localizations from 70 to 39, a 44 percent reduction (Nass et al., 2023). The same source reports average localization times of $N^k$6 ms for VON Similo and $N^k$7 ms for VON Similo LLM, with a total GPT-4 API cost of $N^k$8 for 804 prompts (Nass et al., 2023).

These interface papers treat semantic and contextual interpretation as integral to selection. In the GUI-patent survey, context appears as time, hierarchy, or z-order (Mishra, 2014). In web localization, it appears as neighboring text, layout, and large-language-model reasoning over semantically related labels such as “Health & Beauty” and “Beauty, Health & Hair” (Nass et al., 2023).

5. Wireless and electromagnetic element subset selection

In wireless systems, element selection typically denotes subset activation under channel uncertainty or circuit-power constraints. “IRS-Assistance with Outdated CSI: Element subset selection for secrecy performance enhancement” studies a BS-side IRS with $N^k$9 reflecting elements and outdated CSI of the IRS-user channels (Li et al., 2022). The proposed ESS method sorts IRS elements by descending $$I=\{\sigma(1),\dots,\sigma(K)\}\subset\{1,\dots,N\}$$0, activates the top $$I=\{\sigma(1),\dots,\sigma(K)\}\subset\{1,\dots,N\}$$1, and turns off the remaining $$I=\{\sigma(1),\dots,\sigma(K)\}\subset\{1,\dots,N\}$$2 elements. The paper’s central claim is that minimal secrecy outage probability can be achieved using a subset of the IRS, and that the optimal number of selected reflecting elements can be effectively found by closed-form expressions (Li et al., 2022). Under outdated CSI, activating all elements is not always best, because Bob’s gain degrades while Eve’s SNR grows linearly with the number of active elements; this produces a non-monotonic SOP-versus-$$I=\{\sigma(1),\dots,\sigma(K)\}\subset\{1,\dots,N\}$$3 curve and an optimal $$I=\{\sigma(1),\dots,\sigma(K)\}\subset\{1,\dots,N\}$$4 (Li et al., 2022).

“Joint Active and Passive Beamforming for Energy-Efficient STARS with Quantization and Element Selection in ISAC Systems” embeds element selection into a mixed-integer EE maximization problem for STARS-aided ISAC (Shen et al., 22 Jul 2025). Each STARS element has a binary state $$I=\{\sigma(1),\dots,\sigma(K)\}\subset\{1,\dots,N\}$$5, where $$I=\{\sigma(1),\dots,\sigma(K)\}\subset\{1,\dots,N\}$$6 switches the element off and $$I=\{\sigma(1),\dots,\sigma(K)\}\subset\{1,\dots,N\}$$7 makes it active. The paper jointly optimizes BS beamforming, STARS amplitudes and phase shifts, quantization levels, and element selection via alternating optimization, using Dinkelbach’s transform, SCA, PDD, PCCP, heuristic search for quantization, and integer relaxation for the on-off variables (Shen et al., 22 Jul 2025). In the reported simulations, optimized element selection improves EE by $$I=\{\sigma(1),\dots,\sigma(K)\}\subset\{1,\dots,N\}$$8 over a fixed 75%-on baseline, and the coupled STARS architecture often has the best EE because its hardware requires fewer PIN diodes per element (Shen et al., 22 Jul 2025).

“Energy-Efficient Active Element Selection in RIS-aided Massive MIMO Systems” considers active RIS-assisted massive MIMO with a global amplification budget rather than explicit binary variables (Jr et al., 2024). The paper’s central design question is how many RIS elements should be active to maximize EE and outperform an entirely passive RIS. Its reported results state that the proposed algorithm can provide up to $$I=\{\sigma(1),\dots,\sigma(K)\}\subset\{1,\dots,N\}$$9 higher EE than the entirely passive RIS, that the active RIS can operate with less than half of the reflecting elements for the entirely passive RIS, and that active RIS should be equipped with a reasonable number of reflecting elements above $\mathbf{x}$0 to achieve complete utilization of the available amplification power (Jr et al., 2024).

A recurrent misconception in this literature is that activating more reflective elements is always better. The IRS secrecy paper explicitly shows that, under outdated CSI, using all elements can worsen secrecy outage (Li et al., 2022). The STARS and active-RIS EE papers show the analogous point for circuit power: more active elements increase hardware power consumption and may reduce energy efficiency unless activation is co-designed with beamforming and quantization (Shen et al., 22 Jul 2025, Jr et al., 2024).

6. Materials discovery and semantic formulations

At the opposite end of abstraction, “Element selection for functional materials discovery by integrated machine learning of elemental contributions to properties” elevates the problem from selecting features or coordinates to selecting sets of chemical elements (Vasylenko et al., 2022). PhaseSelect represents a material by its phase field, learns element embeddings from an element–environment co-occurrence matrix, applies multi-head attention over the elements in a phase field, and then uses supervised heads for classification and regression together with an unsupervised ranking autoencoder for novelty (Vasylenko et al., 2022). For a phase field $\mathbf{x}$1, PhaseSelect outputs a merit probability $\mathbf{x}$2, a predicted maximum property value $\mathbf{x}$3, and a novelty or synthetic-accessibility score based on reconstruction error $\mathbf{x}$4 (Vasylenko et al., 2022). On 5-fold cross-validation, the reported classification results include approximately $\mathbf{x}$5 accuracy and ROC-AUC $\mathbf{x}$6 for superconductors at the 10 K threshold, $\mathbf{x}$7 accuracy and ROC-AUC $\mathbf{x}$8 for magnets at the 300 K threshold, and $\mathbf{x}$9 accuracy and ROC-AUC $\max_{\Theta}\; P(A, label = 1 \mid C, u; \Theta),$00 for bandgap prediction at the 4.5 eV threshold (Vasylenko et al., 2022). The paper’s conceptual claim is that selection at the phase-field level reduces combinatorial complexity and aligns with the early-stage experimental question “which elements to combine?” (Vasylenko et al., 2022).

The programming-languages paper “Handling the Selection Monad” provides the most abstract account of element selection (Plotkin et al., 4 Apr 2025). In its basic form,

$\max_{\Theta}\; P(A, label = 1 \mid C, u; \Theta),$01

so a selection function chooses an element of $\max_{\Theta}\; P(A, label = 1 \mid C, u; \Theta),$02 based on a loss function $\max_{\Theta}\; P(A, label = 1 \mid C, u; \Theta),$03. The paper extends algebraic effect handlers with choice continuations that expose future losses to the handler, allowing a programmer-supplied operation to inspect the possible losses of alternative choices before deciding how to resume the computation (Plotkin et al., 4 Apr 2025). It gives an operational semantics for the higher-order language $\max_{\Theta}\; P(A, label = 1 \mid C, u; \Theta),$04, proves progress and type soundness, proves termination for a subset satisfying a mild hierarchical constraint on operation types, provides a selection-monad denotational semantics for that subset, and proves soundness and adequacy results (Plotkin et al., 4 Apr 2025). Its Haskell implementation demonstrates element selection as greedy maximization over finite sets, gradient descent over parameters, hyperparameter tuning, minimax, and Nash-equilibrium-style choice (Plotkin et al., 4 Apr 2025).

These two papers make explicit that “element” may denote either a physical chemical constituent or an abstract branch of a computation. The shared structure is loss-aware choice over a candidate set, but the surrounding semantics differ radically.

7. Recurring themes, limitations, and clarifications

Taken together, these sources suggest three recurring axes along which element-selection problems differ. The first is what counts as an element: a coordinate, a UI object, a reflecting surface component, a chemical symbol, or an abstract result alternative (Ono, 14 Feb 2026, Mishra, 2014, Li et al., 2022, Vasylenko et al., 2022, Plotkin et al., 4 Apr 2025). The second is what feedback is available: exact losses, surrogate importance scores, prefix-sum statistics, noisy or adversarial comparisons, or semantic context from neighboring layout and text (Balasubramaniam et al., 2020, Marchini et al., 2019, Dang et al., 2024, Nass et al., 2023). The third is what correctness means: optimality, approximation, inclusion of a target in a candidate set, or satisfaction of system-level constraints such as secrecy outage, rate, and energy efficiency (Liyanage et al., 16 Jul 2025, Dang et al., 2024, Li et al., 2022, Shen et al., 22 Jul 2025).

Several limitations recur. Local search and relaxation methods may terminate at local optima or depend on initialization, as stated explicitly for swap-based dimension reduction and for mixed-integer wireless designs (Ono, 14 Feb 2026, Shen et al., 22 Jul 2025). Dataset bias and the absence of explicit negative examples complicate phase-field selection in materials discovery (Vasylenko et al., 2022). Proprietary-model cost, latency, token limits, and security considerations constrain LLM-based web element localization (Nass et al., 2023). Perfect CSI assumptions limit the direct transfer of several wireless subset-selection results to realistic deployments (Li et al., 2022, Shen et al., 22 Jul 2025, Jr et al., 2024). In robust comparison settings, even the output type changes: the correct object is a set, not a point estimate (Dang et al., 2024).

A final clarification is that “element selection” should not be conflated with exhaustive optimal search. CECS is designed precisely to avoid enumerating all creative combinations while still modeling interactions (Zhang et al., 2023). Group-testing selection and multiparty selection both accept approximation to reduce query or communication complexity (Liyanage et al., 16 Jul 2025, Chen et al., 2020). PhaseSelect intentionally selects at the phase-field level rather than the composition level (Vasylenko et al., 2022). In the selection monad literature, programmer-defined choice may deliberately replace globally optimal argmin or argmax with heuristic optimization routines (Plotkin et al., 4 Apr 2025). The term therefore names a broad research pattern: structured choice under constraints, where the structure of the candidate set, the feedback channel, and the admissible error jointly determine the algorithmic form.