Automatic Domain Splitting (ADS)

• Automatic Domain Splitting (ADS) is a method for automatically partitioning problem spaces into adaptive domains based on criteria such as nonlinearity, residual structure, or orthogonality.
• It spans diverse applications including unsupervised domain adaptation, self-supervised domain mining for online advertising, anisotropic submesh decomposition for PDEs, and nonlinear uncertainty mapping.
• ADS techniques enhance model performance and computational efficiency by isolating heterogeneous components and enabling parallel processing, while reducing reliance on hand-crafted domain definitions.

Automatic Domain Splitting (ADS) denotes a class of procedures in which a model or solver partitions a problem into domains, subspaces, or subsets automatically rather than through fixed hand-crafted boundaries. In the cited literature, the same underlying idea appears in several technically distinct forms: feature-space decomposition into shared and private components for unsupervised domain adaptation, self-supervised domain mining for online advertising, anisotropic submesh decomposition for PDEs, and adaptive splitting of uncertainty sets for nonlinear transformations (Bousmalis et al., 2016, Sun et al., 2024, Trotti, 2023, Losacco et al., 2023). This suggests that ADS is better understood as a methodological pattern—automatic partitioning driven by optimization, residual structure, or a nonlinearity criterion—than as a single canonical algorithm.

1. Semantic range of the term “domain”

In this literature, “domain” is not restricted to a physical computational region. It can denote a source or target data distribution in domain adaptation, a latent fine-grained traffic segment inferred from advertising data, an anisotropic submesh in a discretized PDE, or a subset of an uncertainty set under nonlinear transformation (Bousmalis et al., 2016, Sun et al., 2024, Trotti, 2023, Losacco et al., 2023). The mechanism of “splitting” therefore also varies: orthogonal representation partitioning, vector-quantized routing, mesh-based decomposition, or recursive subdivision based on a nonlinearity index.

The distinction matters because methods with superficially similar names solve different problems. In Domain Separation Networks, the split is between domain-private and domain-shared latent subspaces. In Adaptive$^2$, the system first learns domains adaptively using a domain mining module by self-supervision and then uses a shared&specific network. In ASDSM, the split is imposed on a uniformly discretized PDE through highly anisotropic submeshes and hole-filling. In LOADS, splitting is triggered whenever imposed linearity constraints are not met. A plausible implication is that ADS should be characterized by its criterion for partitioning rather than by any one application area.

2. Representation-space ADS in unsupervised domain adaptation

Domain Separation Networks formulate automatic domain splitting as an explicit decomposition of image representations into two subspaces: one component which is private to each domain and one which is shared across domains. The architecture contains a shared encoder $E_c(x)$, private encoders $E_p^s(x)$ and $E_p^t(x)$, a shared decoder $D(h)$, and a task classifier $G(h_c)$. For an input $x$, the shared features are $h_c = E_c(x)$, the private features are $h_p = E_p(x)$, the reconstruction is $\hat{x} = D(h_c + h_p)$, and the predicted label is $E_c(x)$0 (Bousmalis et al., 2016).

The split is automatic in the strict sense used by the paper: the model learns what is shared versus private parameters through joint optimization and do not require manual partitioning. The objective is a weighted sum

$E_c(x)$1

where $E_c(x)$2 is defined on the labeled source domain, $E_c(x)$3 is applied to both domains using a scale-invariant mean squared error, $E_c(x)$4 encourages orthogonality between shared and private representations through

$E_c(x)$5

and $E_c(x)$6 aligns shared features across domains either through a Domain-Adversarial (DANN) loss with a gradient reversal layer or through an MMD loss with an RBF kernel.

The operational logic is tightly coupled across losses. The difference loss is the key mechanism preventing leakage of domain-unique information into the shared space. The similarity loss pulls together shared features from different domains. The reconstruction loss ensures that the private and shared components jointly retain enough information to reconstruct the input. The task loss constrains the shared component to remain task-relevant because only the shared representation is exposed to the classifier. The paper reports that DSNs consistently outperform previous methods like DANN, MMD, and CORAL across multiple domain adaptation benchmarks, and that the difference loss is critical in ablation. A further consequence of the decomposition is interpretability: reconstructions from only shared features or only private features expose what the model treats as content versus style; in MNIST $E_c(x)$7 MNIST-M, the shared reconstructions capture digit shape while the private reconstructions capture color/background.

3. Self-supervised domain mining and routing

Adaptive$E_c(x)$8 generalizes the notion of automatic domain splitting from a binary source/target setting to fine-grained latent domain discovery in online advertising. Its core claim is that fine-grained “domain” patterns exist that are difficult to hand-craft, so the framework first learns domains adaptively using a domain mining module by self-supervision and then employs a shared&specific network to model shared and conflicting information. The Domain Mining Module uses VQ-VAE. Raw user and item features are embedded and concatenated, an FFN produces an interaction representation $E_c(x)$9, an encoder produces $E_p^s(x)$0, and vector quantization assigns the sample to the nearest codebook element:

$E_p^s(x)$1

The label $E_p^s(x)$2 is the domain index for the input sample, and the VQ-VAE decoder reconstructs the input representation under a loss combining reconstruction, quantization, and batch regularization (Sun et al., 2024).

This domain assignment is obtained automatically as the closest codebook vector for each sample, with no need for manual labels or expert curation. During co-training, the VQ-VAE operates on a gradient-stopped copy of embeddings so as not to destabilize the main network. The mined domain indicator becomes a routing signal for the Adaptive Domain Modeling Module. That module contains a shared network

$E_p^s(x)$3

and a collection of domain-specific networks. Routing can be hard,

$E_p^s(x)$4

or soft through similarity-weighted aggregation over all codebook vectors. The outputs are fused as

$E_p^s(x)$5

and the overall loss is

$E_p^s(x)$6

The empirical significance lies in the treatment of domain definition itself as a learned object. On Avazu and Criteo, Adaptive$E_p^s(x)$7 consistently outperformed baselines including standard MLP, MMoE, PLE, ADIN, and PEPNet. The paper further reports that traditional domain adaptation methods with hand-crafted domains perform no better than single-domain models under fair FLOPS conditions, and that hand-crafted domain splits sometimes performed no better or even worse than single-domain (MLP) models. In online deployment in the live streaming scenario of Kuaishou Advertising System, the model was deployed with 10% traffic for 15 days and obtained +2.376% revenue and +2.522% conversions, while the variant using hand-crafted domains failed to outperform the online baseline. The broader implication is that automatic domain identification can be a first-order modeling issue rather than a preprocessing convenience.

4. ADS for PDE discretizations: anisotropic submeshes, skeletons, and residual updates

In numerical PDE solving, automatic domain splitting appears in the Anisotropic Submeshes Domain Splitting Method (ASDSM), a two-level algorithm for the $E_p^s(x)$8-dimensional advection-diffusion equation. The method solves the discretized problem over highly anisotropic submeshes, merges those solutions into an initial approximation of the fine solution, fills the remaining holes by local solves, and then iteratively refines the result using the residual. In 2D, the construction uses a fine mesh $E_p^s(x)$9, anisotropic submeshes $E_p^t(x)$0 and $E_p^t(x)$1, a coarse mesh $E_p^t(x)$2, and hole regions $E_p^t(x)$3, with node-matching conditions

$E_p^t(x)$4

The main discretized system is written as $E_p^t(x)$5, with analogous systems on anisotropic and coarse meshes (Trotti, 2023).

The algorithm has four steps. First, it computes initial guesses on the anisotropic submeshes and on the coarse mesh, producing $E_p^t(x)$6, $E_p^t(x)$7, and $E_p^t(x)$8. Second, it merges them into a global approximation $E_p^t(x)$9; Richardson extrapolation is used at intersection points when three solutions are available. Third, for each hole, the merged solution supplies Dirichlet BCs, and all hole problems are solved independently, which is fully parallelizable. Fourth, it computes the residual

$D(h)$0

projects residuals onto the anisotropic meshes, constructs an approximate error, finds a scaling $D(h)$1 minimizing the residual norm, and updates through

$D(h)$2

The method’s stated theoretical justification is threefold. By working on highly anisotropic meshes, each submesh problem is essentially a $D(h)$3-dimensional problem, reducing computational intensity. After the initial guess, the residual is localized: it is zero in the filled holes and non-zero only on the skeleton. And all hole solves are mutually independent, maximizing parallelization. The paper also states a storage advantage of order $D(h)$4 rather than $D(h)$5 in 2D. Empirically, ASDSM achieves comparable accuracy with fewer degrees of freedom than standard sparse grid or uniform grid direct/iterative solvers, offers potentially much lower memory and computational cost per iteration, and exposes massive parallelism in the fill-holes step. Its main limitation is stagnation in further residual reduction when the coarse structure cannot resolve the solution finely enough or when the residual becomes irregular at coarse points; artificial residual smoothing can help overcome some stagnation.

5. ADS for nonlinear uncertainty mapping

A distinct and mathematically explicit form of automatic domain splitting is the low-order ADS approach, LOADS, for mapping an initial uncertainty set through a nonlinear transformation. LOADS uses second-order Taylor expansions to approximate the transformation, a nonlinearity index (NLI) to detect when local linearity assumptions are violated, and an automatic splitting rule that subdivides the domain whenever the NLI exceeds a user-chosen threshold. For a nonlinear transformation

$D(h)$6

the state is parameterized by

$D(h)$7

and the second-order Taylor expansion induces a Jacobian decomposition into nominal and varying parts. The NLI is defined as the ratio between the Frobenius norm of polynomial upper bounds on Jacobian variation and the nominal Jacobian norm:

$D(h)$8

with

$D(h)$9

The domain is treated as quasi-linear if $G(h_c)$0 and is split otherwise (Losacco et al., 2023).

The splitting direction is selected by a directional NLI,

$G(h_c)$1

When splitting along direction $G(h_c)$2, the domain is divided into three subsets, and the splitting size per direction is limited by $G(h_c)$3 to prevent over-refinement. Because repeated splitting can produce rapid growth in the number of domains, LOADS couples the refinement rule with a merging algorithm. At scheduled merging breaks, sibling sets generated by previous splits are recursively recombined if the NLI of their union falls below the same threshold.

The method extends naturally to Gaussian mixtures for probability density propagation. Each subdomain is associated with a Gaussian kernel; splitting uses a library-based update of weights, means, and covariances,

$G(h_c)$4

The paper emphasizes the cost-accuracy tradeoff controlled by the NLI threshold, the maximum splits per direction, the uncertainty domain size, and the merging-break frequency. Lower thresholds increase accuracy but also computational cost; overly restrictive split limits can degrade accuracy; overly small uncertainty domains underrepresent tails; more frequent merging limits domain explosion but may reduce adaptivity. With suitable parameter choices, the method achieves accuracy comparable to high-fidelity Monte Carlo propagation with orders-of-magnitude lower computational cost.

6. Cross-cutting principles, misconceptions, and adjacent terminology

Across these works, ADS is not synonymous with finer partitioning in itself. The cited methods all embed splitting inside a larger loop consisting of a criterion, a local model, and a recomposition mechanism: orthogonality plus reconstruction in DSN, vector quantization plus shared/specific routing in Adaptive$G(h_c)$5, residual projection plus hole filling in ASDSM, and NLI-based subdivision plus merging in LOADS (Bousmalis et al., 2016, Sun et al., 2024, Trotti, 2023, Losacco et al., 2023). This suggests that the central technical question is not merely how to split, but how to guarantee that the split interacts constructively with supervision, reconstruction, residual correction, or uncertainty propagation.

A common misconception is that “domain splitting” necessarily refers to spatial decomposition of a PDE or a geometric region. The literature shows a broader spectrum: domain-private and domain-shared latent subspaces in representation learning, automatically mined latent domains in advertising systems, anisotropic computational submeshes in PDE solvers, and recursively subdivided uncertainty sets in nonlinear mappings. Another misconception is that manually chosen domains are intrinsically adequate. Adaptive$G(h_c)$6 directly contests this by reporting that hand-crafted domains may perform no better than single-domain models under fair FLOPS conditions. Similarly, LOADS and ASDSM both show that uncontrolled refinement is not automatically beneficial: LOADS introduces explicit merging to limit growth of domains in time, and ASDSM reports stagnation when the coarse structure or residual irregularity becomes limiting.

The acronym “ADS” also has an adjacent but distinct use in recommender systems: Adaptive Domain Scaling. That model is not an automatic domain splitting method in the sense of ASDSM or LOADS. Instead, it enhances personalization capability in target-aware sequence modeling across multiple domains through Personalized Sequence Representation Generation (PSRG) and Personalized Candidate Representation Generation (PCRG), dynamically learning both the user behavioral sequence item representation and the candidate target item representation under different domains (Chai et al., 8 Feb 2025). Its reported results include offline gains on Taobao and two billion-scaled industrial datasets, online gains of +0.52% CPM and +1.00% ADVV on Douyin Ads, +0.79% GMV per user, +0.36% Orders per user, and +0.89% GMV per Mille on Douyin E-commerce, with training time increasing only slightly (+3.6%) and online latency not significantly changed. The terminological overlap is therefore substantial, but the methodological object is different: Adaptive Domain Scaling adapts representations under multiple domains, whereas automatic domain splitting partitions the problem structure itself.

From an encyclopedia perspective, the unifying interpretation of ADS is operational rather than disciplinary. In all cases, the method attempts to place complexity where heterogeneity is detected: in private subspaces when domains differ, in mined routing indices when domain identity is latent, in anisotropic submeshes when dimensionality can be reduced, or in recursively refined uncertainty subsets when Jacobian variation becomes too large. That common pattern explains why the term has proved useful across representation learning, scientific computing, and uncertainty quantification, even though the mathematical machinery differs substantially from one field to another.