Cross-Domain Affinity Learning

• Cross-domain affinity learning is a framework that quantifies and transfers similarity across heterogeneous domains, such as image-text and audio-visual data.
• It employs methods like domain adaptation, dual metric learning, and graph-based propagation to align representations with minimal overlap.
• This approach boosts performance in tasks like recommendation, visual matching, and drug discovery by leveraging dense source data for sparser targets.

Cross-domain affinity learning refers to a set of methodologies and principles aimed at quantifying, modeling, and transferring similarity or compatibility relationships—termed “affinities”—across two or more distinct domains. These domains may differ in input spaces, feature distributions, tasks, or modalities (e.g., audio-visual, image-text, or recommender system domains), and often lack direct overlap in samples or labels. Robust cross-domain affinity learning enables leveraging information from a well-annotated or densely observed domain to boost inference, matching, or recognition performance in a distinct, often sparser or unlabeled, target domain.

1. Formal Problem Scope and Key Notions

At its core, cross-domain affinity learning seeks mappings, embeddings, or similarity functions that are calibrated across two or more domains. The setting may take several forms:

• Matching/Similarity Learning: Given pairs $(x_A, x_B)$ from domains $A$ and $B$, learn a function $S(x_A, x_B)$ that assigns higher similarity to positive matches.
• Domain Adaptation/Transfer: Transfer item or user preferences, class structure, or metric information from a source domain to a target under covariate, conditional, or label shift.
• Multi-graph/Multi-view Learning: Couple correlated graphs, sequences, or sets by learning representations optimized for multiple domains.

The “affinity” is often defined via:

• Metric-based distances or Mahalanobis similarity (Faraki et al., 2021, Lin et al., 2016)
• Learned probabilistic or softmax-based classifier scores (Kanagawa et al., 2018)
• Graph-propagated similarities or random-walk transition matrices (Li et al., 2024)
• Pairwise or higher-order relationships defined over aligned or unaligned entities

A recurring challenge is the lack of direct overlap or correspondences—either at the sample, label, or feature level—necessitating methods that discover structural or semantic commonalities without relying on explicit pairing.

2. Methodologies for Cross-domain Affinity Construction

The literature distinguishes several architectural and algorithmic paradigms:

2.1 Domain Adaptation Networks

Domain adaptation architectures, such as Domain Separation Networks (DSN), decompose representations into shared (domain-invariant) and private components. These are trained with multi-part losses: supervised source-domain loss, unsupervised feature reconstruction, orthogonality/difference constraints, and adversarial domain confusion. This enables transfer of affinity from a labeled source to an unlabeled target without requiring user or item overlap (Kanagawa et al., 2018).

2.2 Metric Learning and Dual Alignment

Latent metric learning approaches introduce explicit cross-domain mappings—often orthogonal or bijective—between user or item embeddings. Dual learning frameworks iteratively align source and target via a latent orthogonal transform, enforcing bidirectional consistency and reducing the number of required anchor correspondences (overlap pairs) to enable scalable affinity transfer (Li et al., 2021).

2.3 Graph-based Affinity Propagation

Multi-graph neural networks construct item- or user-level affinity graphs for each domain, then share lower layers to capture domain-invariant structure and learn domain-specific adaptations in upper layers. Optimizations involving multiple-gradient descent allow for simultaneous multi-domain objective balancing, dynamically adapting shared representations to best fit the cross-domain affinities observed (Ouyang et al., 2019). Recent models leverage multi-intent (multi-channel) graph encoders and high-order random walks to disentangle fine-grained affinity signals and avoid negative transfer due to intent mismatch (Li et al., 2024).

2.4 Covariance- and Manifold-based Alignment

Methods based on Maximum Mean Discrepancy (MMD) or class-wise covariance matching facilitate intra-class or intra-manifold alignment between domains, supporting stratified transfer of affinity for activity recognition and other structured tasks (Wang et al., 2017, Tian et al., 2023). Discriminative graph self-learning dynamically adapts affinity matrices that couple source and pseudo-labeled target instances within a joint optimization, closely integrating label propagation and feature adaptation (Tian et al., 2023).

2.5 Cross-modal Metric and Attention Mechanisms

In multimodal settings (e.g., audio-visual, protein-ligand), cross-domain belief matching leverages hybrid encodings and attention-based fusion. For example, KEPLA integrates domain knowledge through a knowledge graph alignment term alongside attention-based fusion of fine-grained, cross-domain representations, enabling robust protein–ligand affinity predictions even under domain shift (Liu et al., 16 Jun 2025). CoPRA leverages cross-domain pretrained LLMs fused by a co-attention transformer, with bi-scope pretraining to amplify both global and local affinity understanding (Han et al., 2024).

3. Representative Task Domains and Applications

Cross-domain affinity learning underpins a broad spectrum of tasks:

Task Category	Affinity Modeling Principle	Example Papers
Recommendation	User–item affinity, metric transfer	(Kanagawa et al., 2018, Li et al., 2021, Ahangama et al., 2019, Li et al., 2024, Ouyang et al., 2019)
Visual Matching	Paired similarity, affine metrics	(Lin et al., 2016, Faraki et al., 2021)
Drug Discovery	Protein–ligand affinities, cross-modal fusion	(Liu et al., 16 Jun 2025, Zhang et al., 23 Jan 2026)
Audio-visual Separation	Local/global cross-stream affinity	(Lee et al., 2021)
Activity Recognition	Intra-class manifold affinity	(Wang et al., 2017, Tian et al., 2023)
Multi-modal Biology	Cross-domain LMs for bio-affinity	(Han et al., 2024, Zhang et al., 23 Jan 2026)

In recommendation, cross-domain affinity learning enables leveraging dense source domains to improve sparse target performance for cold-start users or items—either through domain adaptation, dual metric learning, or VAE-based latent code transfer (Kanagawa et al., 2018, Li et al., 2021, Ahangama et al., 2019, Li et al., 2024). In drug discovery, explicit modeling of modular protein domains or knowledge graph alignment supports transfer of biochemical knowledge and motif recognition across highly heterogeneous molecular environments (Liu et al., 16 Jun 2025, Zhang et al., 23 Jan 2026).

4. Architectural and Optimization Considerations

Most cross-domain affinity learning methods employ hybrid architectures that mix domain-shared and domain-private pathways. Training objectives typically combine:

• Domain-supervised loss (e.g., cross-entropy, regression)
• Domain adaptation terms (e.g., adversarial confusion, difference losses)
• Metric or structure alignment (e.g., MMD, covariance, contrastive loss)
• Explicit affinity propagation or label smoothness over joint graphs

Optimization frequently uses alternating, multi-objective, or meta-learning regimes. For example, alternating updates of feature projections, affinity matrices, and label assignments arise in joint graph self-learning (Tian et al., 2023). Dual or iterative update loops are used for bidirectional affinity alignment (Li et al., 2021). Multiple-gradient-descent and distributionally regularized adapters enable scaling to many target domains while preserving generalization (Ouyang et al., 2019, Krishnan et al., 2020).

5. Evaluation Protocols and Empirical Findings

Empirical evaluation in cross-domain affinity learning typically involves:

• Transfer tasks between heterogeneous domains with user, item, or label distribution shift
• Metrics aligned with the downstream application: e.g., NDCG@K or recall@K for recommendation (Kanagawa et al., 2018, Li et al., 2024), RMSE or correlation for regression tasks (Liu et al., 16 Jun 2025, Zhang et al., 23 Jan 2026)
• Baselines spanning both classical (matrix factorization, CCA, TCA) and neural methods (deep NNs, VAEs, GNNs)
• Ablations on domain-adaptive components, overlap fractions, and hyperparameters

Key empirical insights include:

• Domain adaptation with strong confusion terms or explicit metric mapping is critical; removing these can degrade recall or generalization by up to 10% (Kanagawa et al., 2018).
• Dual metric learning and contrastive approaches can match or surpass state-of-the-art baselines with as few as 8–16 user overlaps (Li et al., 2021, Li et al., 2024).
• Multi-graph architectures with layer sharing and dynamic optimization outperform baselines for both link prediction and multi-domain recommendation (Ouyang et al., 2019).
• Cross-modal affinity models maintain separation performance and alignment even under significant asynchrony or jitter if both local and global regularization are applied (Lee et al., 2021).
• Hybrid knowledge-based alignment and data-driven fine-grained attention boosts transfer robustness and interpretability in biochemical affinity tasks (Liu et al., 16 Jun 2025).

6. Open Challenges and Frontiers

Persistent challenges include:

• Handling extreme domain misalignment or absence of explicit correspondences while avoiding negative transfer (Li et al., 2024).
• Modeling fine-grained, multi-intent or multi-modal affinities especially for overlapping but non-identical label/task sets.
• Developing interpretable, structure-aware affinity modules that respect domain-specific constraints, as in structural biology (Zhang et al., 23 Jan 2026).
• Generalizing graph- or manifold-based affinity propagation approaches to very large or dynamically evolving domain graphs.
• Automating meta-learning or adaptation in extreme one-to-many or many-to-many transfer scenarios, maintaining scalability and sample efficiency (Krishnan et al., 2020).

A plausible implication is that advances in representation disentanglement, structured meta-learning, and multi-modal graph approaches may further increase robustness and controllability in cross-domain affinity learning systems. Additionally, integration of domain knowledge (e.g., KGs in biology, taxonomies in recommender systems) into affinity construction modules is likely to remain a prominent direction, given the continuing improvements observed in recent hybrid models (Liu et al., 16 Jun 2025).

7. Representative Theoretical and Empirical Guarantees

While formal convergence guarantees are rare, certain geometric or Riemannian perspectives justify the use of (class or domain) covariance alignment as minimizing cross-domain discrepancy in structure-aware spaces (Faraki et al., 2021). Orthogonality constraints on domain mappings ensure invertibility and norm preservation, enabling unique determination of cross-domain matching with a provably small number of anchor correspondences (Li et al., 2021). Well-posed alternating optimization schemes on affinity matrices, embeddings, and label propagation variables guarantee practical stability and, in some cases, blockwise optimality (Tian et al., 2023). Empirical results generally confirm that these architectural and optimization principles yield reliable transfer under distribution shift, though performance is often upper-bounded by the strength of available domain-invariant signals.

---

In summary, cross-domain affinity learning is a foundational paradigm for a growing array of machine learning tasks, fusing advances in representation learning, metric geometry, graph theory, and domain adaptation. It is characterized by flexible architectural strategies, theoretically motivated regularization or alignment objectives, and empirically validated robustness across both classical and emerging multi-domain applications. Key practical and theoretical continued developments are expected in scalability, interpretability, disentanglement, and fine-grained cross-domain alignment.