Cross-Click Operators: Interactions & Applications

• Cross-click operators are structured transformation mechanisms that construct and selectively weight multi-order feature interactions, boosting CTR prediction performance.
• They generalize across disciplines by unifying concepts from recommender systems, higher-dimensional algebra, and quantum photodetection through explicit operator formulations.
• Empirical studies demonstrate that selective operator design and cross-domain transfer techniques improve model interpretability, accuracy, and privacy in practical applications.

Cross-click operators are operator constructions that mediate “cross” interactions, but the term does not denote a single canonical object across the literature surveyed here. In click-through rate prediction, it denotes mechanisms that construct, weight, align, transfer, or fuse click-relevant interactions across feature orders, domains, time periods, and privacy boundaries; representative instances include selective-kernel weighting of explicit multi-order feature crosses, historical embedding transfer from natural content to advertisements, bidirectional cross-domain attention, lifelong attention pyramids over source-domain histories, and federated augmentation/alignment pipelines (Wang et al., 2024, Liu et al., 2024, Li et al., 2021, Hou et al., 2023, Qin et al., 21 Mar 2025). In a separate mathematical lineage, it denotes higher-dimensional generalizations of cross-product and curl operators, and in passive multi-basis photodetection it denotes POVM elements for simultaneous detector clicks in different measurement bases; a further operator-like interpretation appears in human–computer interaction, where click planning is modeled as a stochastic timing mechanism embedded in intermittent motor control (Lewintan, 2021, Ye, 5 Jul 2026, Park et al., 2018).

1. Terminological scope and common operator structure

Across the cited works, cross-click operators share an operator-theoretic role rather than a common codomain or application. In CTR research, the relevant operators act on sparse categorical features, domain-specific click logs, chronological sequences, or privatized gradients; in higher-dimensional analysis, they act on vectors, skew tensors, and differential fields; in photodetection, they are Fock-space POVM elements; and in click-planning theory, they act on relative kinematic state and temporal cues to produce a click-time distribution. This suggests a unifying perspective in which a cross-click operator is any structured transformation that encodes interactions not captured by a single local feature, a single domain, or a single event channel (Wang et al., 2024, Liu et al., 2024, Li et al., 2021, Qin et al., 21 Mar 2025, Lewintan, 2021, Ye, 5 Jul 2026).

Within recommender systems, several misconceptions are explicitly challenged. One is that cross-domain click transfer can be achieved by simply pooling source and target data; another is that all feature crosses should be treated equally once constructed. The surveyed CTR works instead emphasize explicit weighting of crosses, structured transfer pipelines, candidate-conditioned sequence filtering, and auxiliary alignment objectives. A related recurring assumption is identity alignment: multiple methods retrieve user or item embeddings by user_id and item_id, or rely on overlap users to learn mappings. This means their operators are not domain-invariant in the strict sense, but identity-linked and architecture-coupled.

A second common theme is selectivity. FiiNet suppresses “featureless” explicit crosses; E-CDCTR transfers historical embeddings and parameters but refreshes BN parameters; DASL uses dual attention rather than one-way source-to-target transfer; LCN separates representation alignment from sequence attention; FedCCTR-LM separates augmentation, disentanglement, and privacy perturbation. In each case, the operator is not a fixed algebraic product but a selective mechanism that determines which cross-interactions should dominate.

2. Explicit multi-order feature-cross operators for CTR prediction

In single-domain CTR prediction, the clearest explicit formulation is FiiNet, or Multiple Order Feature Interaction Importance Neural Networks, which proposes that recommendation quality depends not only on constructing feature crosses but on learning the importance of those crosses dynamically and in a fine-grained manner (Wang et al., 2024). The architecture has five stages: sparse input layer, embedding layer, SKNet layer, deep neural network layer, and output layer. The embedding output is

$E = [e_1, e_2, \cdots, e_i, \cdots, e_f],$

where $f$ is the number of feature fields, $e_i \in \mathbb{R}^K$, and $K$ is the embedding dimension.

The core cross-click operator appears in the SKNet layer through a Split–Fuse–Select procedure. In the split stage, FiiNet explicitly constructs second-order and higher-order feature crosses. The second-order branch uses

$\sum_{i=1}^{n-1}\sum_{j=i+1}^{n} w_{ij} x_i x_j,$

and the third-order branch uses

$$\sum_{i=1}^{n-2}\sum_{j=i+1}^{n-1}\sum_{k=j+1}^{n} w_{ijk} x_i x_j x_k.$$

The paper’s illustrated setting has a low-order/second-order branch $\widetilde{U}$ and a higher-order/third-order branch $\hat{U}$, while noting that the design is extensible to more branches. A practical implementation detail is that the incoming feature sequence is fixed, so feature combinations can be obtained by direct slicing.

In the fuse stage, the branch outputs are added elementwise,

$U = \widetilde{U} + \hat{U},$

then summarized by global mean pooling rather than max pooling: $$z_i = F_{gp}(e_i) = \frac{1}{k}\sum_{t=1}^{k} e_i^{(t)}.$$ Two fully connected layers then compute a gating vector,

$f$0

where $f$1 reduces dimension using reduction ratio $f$2 and $f$3 restores dimension. In the select stage, soft attention weights are assigned across branches, and the final selected representation is

$f$4

The paper interprets this as importance learning over explicit multi-order crossed representations at the channel/component level, with branch/order-wise soft selection.

After SKNet selection, the crossed representation is passed through an MLP,

$f$5

and combined with a linear term in the final CTR output,

$f$6

Training uses binary cross-entropy,

$f$7

with Adam, learning rate $f$8, weight decay $f$9, batch size $e_i \in \mathbb{R}^K$0, epochs $e_i \in \mathbb{R}^K$1, dropout $e_i \in \mathbb{R}^K$2, embedding dimension $e_i \in \mathbb{R}^K$3, Xavier initialization, seed $e_i \in \mathbb{R}^K$4, and PyTorch 1.11.0.

Empirically, FiiNet outperformed LR, FM, AFM, FFM, Wide&Deep, DeepFM, NFM, Deep&Cross, xDeepFM, FiBiNET, and AutoInt on two real datasets. On KuaiRec-big it achieved AUC $e_i \in \mathbb{R}^K$5 and Logloss $e_i \in \mathbb{R}^K$6; on Book-Crossing it achieved AUC $e_i \in \mathbb{R}^K$7 and Logloss $e_i \in \mathbb{R}^K$8. Ablations showed that the full model performed best, that removing the SKNet importance-learning module hurt performance, and that removing second-order crosses hurt more than removing higher-order crosses. The interpretability claim is correspondingly narrow and explicit: attention weights over crossed feature combinations can be extracted and compared before and after training, but the evidence is weight visualization and aggregate analysis rather than instance-level explanation.

3. Cross-domain transfer operators in CTR and sequential recommendation

Cross-domain CTR work generalizes the cross-click idea from feature order to source–target transfer. E-CDCTR formulates a two-domain environment $e_i \in \mathbb{R}^K$9 in which $K$0 is natural content and $K$1 is advertisement traffic, and proposes a tri-level asynchronous framework consisting of Tiny Pre-training Model (TPM), Complete Pre-training Model (CPM), and Advertisement CTR model (A-CTR) (Liu et al., 2024). TPM is updated monthly on half a year of natural data and stores the last three months of user and item embeddings; CPM is updated weekly on the last month’s natural data with the same full architecture and input feature space as A-CTR; A-CTR is updated daily, initialized from CPM except BN parameters, and fine-tuned only on advertisement data. Historical embeddings are concatenated as

$K$2

then aggregated by

$K$3

The transfer operator therefore has both feature-level and parameter-level components. Offline, Target-only reached $K$4 GAUC, KEEP reached $K$5, and E-CDCTR reached $K$6; online, E-CDCTR improved CTR by $K$7 and RPM by $K$8. The paper also reports a negative result: Source-only ($K$9) and Sample Merging ($\sum_{i=1}^{n-1}\sum_{j=i+1}^{n} w_{ij} x_i x_j,$0) were both worse than Target-only, showing that naive source usage can induce negative transfer.

DASL shifts from one-way transfer to symmetric bidirectional support between paired domains (Li et al., 2021). It uses overlap users as pivots, learns domain-specific user embeddings $\sum_{i=1}^{n-1}\sum_{j=i+1}^{n} w_{ij} x_i x_j,$1 and $\sum_{i=1}^{n-1}\sum_{j=i+1}^{n} w_{ij} x_i x_j,$2, and aligns them through an orthogonal map $\sum_{i=1}^{n-1}\sum_{j=i+1}^{n} w_{ij} x_i x_j,$3 satisfying

$\sum_{i=1}^{n-1}\sum_{j=i+1}^{n} w_{ij} x_i x_j,$4

with objectives of the form

$\sum_{i=1}^{n-1}\sum_{j=i+1}^{n} w_{ij} x_i x_j,$5

Sequential behavior is encoded by a GRU over the last 10 purchased or consumed items in timestamp order, and dual attention is defined by

$\sum_{i=1}^{n-1}\sum_{j=i+1}^{n} w_{ij} x_i x_j,$6

$\sum_{i=1}^{n-1}\sum_{j=i+1}^{n} w_{ij} x_i x_j,$7

This makes the relevance weighting in one domain depend on queries from both domains. Across Imhonet, Amazon, and Youku, DASL outperformed all baselines in both domains, with an average improvement over the second-best baseline of $\sum_{i=1}^{n-1}\sum_{j=i+1}^{n} w_{ij} x_i x_j,$8 AUC and $\sum_{i=1}^{n-1}\sum_{j=i+1}^{n} w_{ij} x_i x_j,$9 HR@10; at Alibaba-Youku it improved VV by $$\sum_{i=1}^{n-2}\sum_{j=i+1}^{n-1}\sum_{k=j+1}^{n} w_{ijk} x_i x_j x_k.$$0. The paper is explicit, however, that a full final CTR loss, exact fusion equation, optimizer choice, and orthogonality enforcement procedure are not provided.

LCN addresses a one-way but lifelong version of cross-domain CTR, in which the target prediction is a click in one domain while the most informative long sequence comes from a source domain with non-overlapping items (Hou et al., 2023). The target probability is formulated as

$$\sum_{i=1}^{n-2}\sum_{j=i+1}^{n-1}\sum_{k=j+1}^{n} w_{ijk} x_i x_j x_k.$$1

with standard binary cross-entropy for CTR. Its operator stack contains Cross Representation Production (CRP) and the LifeLong Attention Pyramid (LAP). CRP builds within-target, within-source, and cross-domain positive pairs from short-term sequences and optimizes contrastive losses based on cosine similarity, thereby aligning target and source item geometry. LAP then performs candidate-conditioned retrieval and refinement over the lifelong source sequence using Complete-Scope Attention, Median-Scope Attention, and Focused-Scope Attention. In the industrial setting, TWIN obtained AUC $$\sum_{i=1}^{n-2}\sum_{j=i+1}^{n-1}\sum_{k=j+1}^{n} w_{ijk} x_i x_j x_k.$$2, GAUC $$\sum_{i=1}^{n-2}\sum_{j=i+1}^{n-1}\sum_{k=j+1}^{n} w_{ijk} x_i x_j x_k.$$3, and Logloss $$\sum_{i=1}^{n-2}\sum_{j=i+1}^{n-1}\sum_{k=j+1}^{n} w_{ijk} x_i x_j x_k.$$4, whereas LCN-500-100 obtained AUC $$\sum_{i=1}^{n-2}\sum_{j=i+1}^{n-1}\sum_{k=j+1}^{n} w_{ijk} x_i x_j x_k.$$5, GAUC $$\sum_{i=1}^{n-2}\sum_{j=i+1}^{n-1}\sum_{k=j+1}^{n} w_{ijk} x_i x_j x_k.$$6, and Logloss $$\sum_{i=1}^{n-2}\sum_{j=i+1}^{n-1}\sum_{k=j+1}^{n} w_{ijk} x_i x_j x_k.$$7; online, LCN-200-50 improved CTR by $$\sum_{i=1}^{n-2}\sum_{j=i+1}^{n-1}\sum_{k=j+1}^{n} w_{ijk} x_i x_j x_k.$$8, stay time by $$\sum_{i=1}^{n-2}\sum_{j=i+1}^{n-1}\sum_{k=j+1}^{n} w_{ijk} x_i x_j x_k.$$9, and increased latency by only $\widetilde{U}$0 ms. The paper also shows that adding CRP improves SIM, ETA, SDIM, and TWIN, indicating that the representation-bridging operator is reusable beyond the full architecture.

4. Federated and privacy-preserving cross-domain click operators

FedCCTR-LM places cross-domain CTR prediction in a federated setting with overlapping users, two domains $\widetilde{U}$1 and $\widetilde{U}$2, and a merged chronological sequence $\widetilde{U}$3 in addition to domain-specific sequences $\widetilde{U}$4 and $\widetilde{U}$5 (Qin et al., 21 Mar 2025). The prediction targets are

$\widetilde{U}$6

and the global objective aggregates both domains over all users while keeping raw data local.

The framework has three modules. PrivAugNet uses LLM prompts to augment item features, synthesize user profiles, and expand interaction sequences with positive and negative interactions. IDST-CL uses three independent transformers $\widetilde{U}$7 over domain $\widetilde{U}$8, domain $\widetilde{U}$9, and mixed-domain $\hat{U}$0, with standard self-attention

$\hat{U}$1

It then enforces Intra-Domain Representation Alignment (IDRA) between original and augmented sequences and Cross-Domain Representation Disentanglement (CDRD) between $\hat{U}$2, $\hat{U}$3, and $\hat{U}$4. Domain-specific CTR heads use

$\hat{U}$5

The local objective is

$\hat{U}$6

AdaLDP supplies the privacy operator. Client gradients are clipped by

$\hat{U}$7

then perturbed as

$\hat{U}$8

with geometric decay

$\hat{U}$9

The server updates the global model via a FedAvg-style rule,

$U = \widetilde{U} + \hat{U},$0

On Amazon-derived Book–Movie and Food–Kitchen tasks, FedCCTR-LM achieved the best results on all datasets and metrics. The paper reports NDCG@10 gains over the best baseline of $U = \widetilde{U} + \hat{U},$1 for Book, $U = \widetilde{U} + \hat{U},$2 for Movie, $U = \widetilde{U} + \hat{U},$3 for Food, and $U = \widetilde{U} + \hat{U},$4 for Kitchen; corresponding MRR@10 gains are $U = \widetilde{U} + \hat{U},$5, $U = \widetilde{U} + \hat{U},$6, $U = \widetilde{U} + \hat{U},$7, and $U = \widetilde{U} + \hat{U},$8. Ablations show that removing PrivAugNet, IDRA, or CDRD degrades performance, while replacing AdaLDP with static LDP is worse than the adaptive schedule. The main caveat is also explicit: the formal privacy guarantee applies to gradient sharing, not clearly to the augmentation stage itself, and the experimental use of a Hugging Face API-hosted Llama-2 model raises unresolved deployment and privacy questions.

5. Cross-click operators as stochastic click-planning mechanisms

A distinct usage appears in human–computer interaction, where the Intermittent Click Planning model describes the process by which users plan and execute click actions during pointing and target tracking (Park et al., 2018). The model reformulates pointing success as a temporal decision problem. If $U = \widetilde{U} + \hat{U},$9 is the time at which the pointer first contacts the target and $$z_i = F_{gp}(e_i) = \frac{1}{k}\sum_{t=1}^{k} e_i^{(t)}.$$0 is the time at which it exits, then successful clicking requires that the click occur within that interval. The effective target-crossing duration is

$$z_i = F_{gp}(e_i) = \frac{1}{k}\sum_{t=1}^{k} e_i^{(t)}.$$1

where $$z_i = F_{gp}(e_i) = \frac{1}{k}\sum_{t=1}^{k} e_i^{(t)}.$$2 is the length of the trajectory segment intersecting the target and $$z_i = F_{gp}(e_i) = \frac{1}{k}\sum_{t=1}^{k} e_i^{(t)}.$$3 is relative pointer–target velocity.

The click operator is intermittent because planning is assumed to occur only during the last submovement before the click. The mean click time is modeled as an implicit aim point,

$$z_i = F_{gp}(e_i) = \frac{1}{k}\sum_{t=1}^{k} e_i^{(t)}.$$4

and actual click timing is Gaussian, with variance obtained by maximum-likelihood integration of two cues. The temporal cue obeys scalar timing,

$$z_i = F_{gp}(e_i) = \frac{1}{k}\sum_{t=1}^{k} e_i^{(t)}.$$5

where $$z_i = F_{gp}(e_i) = \frac{1}{k}\sum_{t=1}^{k} e_i^{(t)}.$$6 is the mean period of click repetition. The visual movement cue has uncertainty

$$z_i = F_{gp}(e_i) = \frac{1}{k}\sum_{t=1}^{k} e_i^{(t)}.$$7

where $$z_i = F_{gp}(e_i) = \frac{1}{k}\sum_{t=1}^{k} e_i^{(t)}.$$8 is cue viewing time, $$z_i = F_{gp}(e_i) = \frac{1}{k}\sum_{t=1}^{k} e_i^{(t)}.$$9 is visual encoding rate, and $f$00 is the lower bound on visual timing standard deviation. Cue integration yields

$f$01

The pointing error rate is then

$f$02

The empirical fit is strong in both stationary and moving-target settings. In Study 1, on stationary targets, the model achieved $f$03 and adjusted $f$04, compared with $f$05 for the Wobbrock baseline. In Study 2, on moving targets, it achieved adjusted $f$06, compared with $f$07 for the Huang baseline. The paper further reports that FPS gamers and non-gamers differed significantly only in $f$08, with gamers showing lower timing uncertainty. In this literature, a plausible interpretation is that the cross-click operator is a stateful stochastic trigger mechanism instantiated at final-submovement onset and parameterized by relative motion, cue-viewing time, and user-specific timing precision.

6. Higher-dimensional cross-product and curl operators

In higher-dimensional analysis, the relevant object is not a click model but a generalization of the cross product itself. The paper defines a bilinear, anti-commutative map

$f$09

denoted $f$10, whose coordinates are the antisymmetric minors of $f$11 and $f$12 (Lewintan, 2021). The decisive structural change is that the codomain is $f$13, not $f$14 except when $f$15. For fixed $f$16, the map $f$17 is represented by a unique matrix

$f$18

such that

$f$19

The operator is most naturally understood as antisymmetrization. If $f$20 vectorizes the independent skew entries and $f$21 is the inverse map, then

$f$22

equivalently

$f$23

This yields higher-dimensional versions of familiar $f$24-D identities. The Lagrange identity becomes

$f$25

so

$f$26

The vector triple-product analogue is

$f$27

and the generalized Room identity is

$f$28

The same matrix formalism induces higher-dimensional differential operators. For $f$29,

$f$30

Its formal partner is the new operator

$f$31

whose image is divergence-free: $f$32 The vector Laplacian decomposes as

$f$33

which replaces the classical $f$34-D formula $f$35. The same framework yields an explicit Helmholtz decomposition in all dimensions $f$36. A frequent misconception is that the higher-dimensional analogue should again be vector-valued; this paper argues that the natural object is instead bivector- or skew-matrix-valued antisymmetrization.

7. Cross-click operators in passive multi-basis photodetection and QKD

In quantum photodetection, cross-click operators are POVM elements for simultaneous clicks in detectors assigned to different measurement bases within a passive multi-basis analyzer (Ye, 5 Jul 2026). The setup consists of an input single-photon Hilbert space $f$37, a passive linear-optical network, threshold detectors $f$38, efficiencies $f$39, dark-count probabilities $f$40, and a partition of detectors into disjoint basis groups

$f$41

A cross click is the event that detectors belonging to different basis groups click in the same run.

The structural starting point is the silence operator. For any detector subset $f$42,

$f$43

with

$f$44

Because

$f$45

many-photon no-click events reduce to symmetric tensor powers of explicit single-photon contractions. Defining

$f$46

the cross-click operator is obtained by inclusion–exclusion: $f$47

The central spectral quantity is

$f$48

the minimum cross-click probability over all $f$49-photon states. The paper proves four main analytic results. First, monotonicity: $f$50 Second, two-sided exponential bounds and exact asymptotic rate: $f$51 hence

$f$52

Third, for ideal detectors and $f$53,

$f$54

Fourth, in the bipartite setting,

$f$55

These formulas matter because observed cross-click probabilities then bound the total weight of high-photon-number sectors. For example,

$f$56

The paper’s explicit analytic characterization replaces earlier finite-sector numerics in QKD and entanglement-verification analyses. In the balanced six-state polarization analyzer, where $f$57, the ideal-detector formula becomes

$f$58

Under efficiency mismatch, the decay rate is controlled by explicit single-photon spectral data, making the effect of detector imbalance analytically transparent.

In this quantum setting, therefore, a cross-click operator is neither a recommender-system interaction module nor a differential operator, but a measurable event operator on Fock space. The shared pattern with the other literatures is structural rather than semantic: the operator encodes interactions that become visible only when multiple channels, bases, photons, or histories are considered jointly rather than separately.