Paper deep dive
MOSAIC: Multi-Domain Orthogonal Session Adaptive Intent Capture for Prescient Recommendations
Abderaouf Bahi, Mourad Boughaba, Ibtissem Gasmi, Warda Deghmane, Amel Ourici
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 4/14/2026, 2:07:30 AM
Summary
MOSAIC is a multi-domain sequential recommendation framework that uses a triple-encoder architecture to decompose user preferences into three orthogonal components: domain-specific, domain-common, and cross-sequence-exclusive. It employs adversarial training, domain masking, and a dynamic token-level gating mechanism to adaptively integrate these preferences for improved recommendation accuracy and interpretability.
Entities (5)
Relation Signals (3)
MOSAIC → uses → Triple-encoder architecture
confidence 98% · Our approach employs a triple-encoder architecture
MOSAIC → addresses → Session-based recommendation
confidence 95% · MOSAIC, a Multi-Domain Orthogonal Session Adaptive Intent Capture framework for prescient recommendations.
MOSAIC → implements → Dynamic gating mechanism
confidence 95% · Additionally, a dynamic gating mechanism modulates the relative contribution of each component
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Capturing user intent across heterogeneous behavioral domains stands as a fundamental challenge in session-based recommender systems. Yet, existing multi-domain approaches frequently fail to isolate the distinct contribution of cross-domain interactions from those arising within individual domains, limiting their ability to build rich and transferable user representations. In this work, we propose MOSAIC, a Multi-Domain Orthogonal Session Adaptive Intent Capture framework that explicitly factorizes user preferences into three orthogonal components: domain-specific, domain-common, and cross-sequence-exclusive representations. Our approach employs a triple-encoder architecture, where each encoder is dedicated to one preference type, enforced through domain masking objectives and adversarial training via a gradient reversal layer. Representational alignment and mutual independence constraints are jointly optimized to ensure clean preference separation. Additionally, a dynamic gating mechanism modulates the relative contribution of each component at every timestep, yielding a unified and temporally adaptive session-level user representation. We conduct extensive experiments on two large-scale real-world benchmarks spanning multiple domains and interaction types. The ablation study validates that each component domain-specific encoding, domain-common modeling, cross-sequence representation, and dynamic gating contributes meaningfully to the overall performance. Experimental results demonstrate that MOSAIC consistently outperforms state-of-the-art baselines in recommendation accuracy, while simultaneously providing interpretable insights into the interplay between domain-specific and cross-domain preference signals. These findings highlight the potential of orthogonal preference decomposition as a principled strategy for next-generation multi-domain recommender systems.
Tags
Links
- Source: https://arxiv.org/abs/2604.10147v1
- Canonical: https://arxiv.org/abs/2604.10147v1
Trouble viewing inline? Open PDF directly →
Full Text
73,554 characters extracted from source content.
Expand or collapse full text
MOSAIC: Multi-Domain Orthogonal Session Adaptive Intent Capture for Prescient Recommendations Abderaouf Bahi1§ , Mourad Boughaba2, Ibtissem Gasmi1 , Warda Deghmane1 , and Amel Ourici3 1Computer Science and Applied Mathematics Laboratory (LIMA), Faculty of Science and Technology, Chadli Bendjedid University, P.O. Box 73, El Tarf 36000, Algeria 2Electro-Mechanical Systems Laboratory (LSEM), Faculty of Technology, Badji Mokhtar University, P.O. Box 12, Annaba 23000, Algeria 3Mathematical Modeling and Numerical Simulation Laboratory (LAM2SIN), Faculty of Technology, Badji Mokhtar University, P.O. Box 12, Annaba 23000, Algeria §Corresponding author: Abderaouf Bahi (a.bahi@univ-eltarf.dz) Abstract Capturing user intent across heterogeneous behavioral domains stands as a fundamental challenge in session-based recommender systems. Yet, existing multi-domain approaches frequently fail to isolate the distinct contribution of cross-domain interactions from those arising within individual domains, limiting their ability to build rich and transferable user representations. In this work, we propose MOSAIC, a Multi-Domain Orthogonal Session Adaptive Intent Capture framework that explicitly factorizes user preferences into three orthogonal components: domain-specific, domain-common, and cross-sequence-exclusive representations. Our approach employs a triple-encoder architecture, where each encoder is dedicated to one preference type, enforced through domain masking objectives and adversarial training via a gradient reversal layer. Representational alignment and mutual independence constraints are jointly optimized to ensure clean preference separation. Additionally, a dynamic gating mechanism modulates the relative contribution of each component at every timestep, yielding a unified and temporally adaptive session-level user representation. We conduct extensive experiments on two large-scale real-world benchmarks spanning multiple domains and interaction types. The ablation study validates that each component—domain-specific encoding, domain-common modeling, cross-sequence representation, and dynamic gating—contributes meaningfully to the overall performance. Experimental results demonstrate that MOSAIC consistently outperforms state-of-the-art baselines in recommendation accuracy, while simultaneously providing interpretable insights into the interplay between domain-specific and cross-domain preference signals. These findings highlight the potential of orthogonal preference decomposition as a principled strategy for next-generation multi-domain recommender systems. 1 Introduction Session-based recommendation, which aims to anticipate a user’s next interaction by mining preference signals from historical behavioral sequences, has become an indispensable component of modern online platforms such as e-commerce, streaming services, and content discovery systems [1, 2, 3]. Nevertheless, conventional single-domain models are inherently constrained by data sparsity [4], which amplifies cold-start difficulties and skews preference learning toward the dominant patterns of a single domain [5, 6]. To overcome these limitations, multi-domain sequential recommendation has emerged as a compelling paradigm that harnesses user behavioral signals across heterogeneous domains to enrich preference modeling and improve recommendation quality [7, 8, 9]. Contemporary multi-domain approaches exploit both individual-domain sequences and cross-domain sequences that consolidate interactions spanning several domains [10, 11, 12, 13]. Individual domain sequences encode domain-specific behavioral patterns, whereas cross-domain sequences can amplify latent preference signals that remain imperceptible within a single domain [14, 15]. For instance, a user’s inclination toward the romance genre may appear too weak to detect in the Book domain alone, yet becomes prominently discernible once reinforced by correlated signals from the Movie domain within the cross-sequence [10]. Simultaneously, cross-domain sequences may carry redundant cues that duplicate patterns already captured by individual-domain sequences, while other signals remain exclusively discoverable through cross-domain co-occurrence [16, 17]. Consequently, effectively separating overlapping information from cross-sequences while preserving domain-distinctive signals is paramount for constructing more informative and transferable user representations [18, 19]. Beyond redundancy, cross-domain sequences can surface entirely novel behavioral patterns absent from any individual domain in isolation [7]. A user who first reads a literary classic and subsequently watches its film adaptation exemplifies a cross-domain behavioral trajectory exclusively uncovered through the joint sequence [11, 20]. However, the relevance of such cross-domain signals is not static — it fluctuates dynamically across timesteps [16, 21]. A romance preference inferred from a cross-domain session may carry little predictive value when forecasting the next interaction in a fantasy-oriented or action-driven context [22]. This temporal variability underscores the necessity of an adaptive mechanism capable of modulating the relative influence of cross-domain and domain-specific signals at each prediction step [21, 23]. To harness these diverse behavioral signals, prior works have leveraged graph neural networks [24, 25], Transformer-based encoders [1, 14, 26], and gating mechanisms [21] to integrate cross-domain information into user representations. More recent contributions have further incorporated cross-sequences into their architectures [16, 11], refining domain representations through shared-attention and alignment strategies [17, 27]. These advances collectively reflect the growing recognition that cross-domain sequences constitute a rich and transferable source of user preference signals that single-domain models fundamentally cannot exploit [8, 5]. Despite this progress, existing multi-domain sequential recommendation models face two persistent shortcomings: (1) insufficient disentanglement of overlapping information between cross-domain and individual-domain sequences [18, 19, 28], and (2) limited capacity to adaptively regulate the degree of cross-domain influence at each timestep [16, 21]. When redundant cues are not properly separated, models risk capturing duplicate rather than complementary features [29, 30]. Equally, cross-domain signals that prove informative in one behavioral context may become irrelevant or even counterproductive in another, degrading overall recommendation quality [5, 23]. To address these challenges, we propose MOSAIC, a Multi-Domain Orthogonal Session Adaptive Intent Capture framework for prescient recommendations. MOSAIC introduces a triple-encoder architecture that explicitly disentangles and dynamically integrates user preferences into three orthogonal components. The specific encoder captures domain-exclusive embeddings from individual-domain sequences, enforcing their independence from shared and cross-sequence information. The cross encoder learns cross-sequence representations that jointly model cross-sequence-exclusive signals and enhanced domain-common preferences through an alignment objective. During recommendation, a Transformer encoder processes the user’s most recent interactions, whose outputs interact with precomputed specific and cross preferences via a token-level cross-attention gating mechanism. This gating adaptively modulates the contribution of each preference type at every timestep, and the resulting token-level outputs are aggregated into a unified session-level representation for next-item prediction. The main contributions of this work are summarized as follows: • We propose MOSAIC, a novel triple-encoder framework that explicitly disentangles user preferences into three orthogonal components: domain-specific, domain-common, and cross-sequence-exclusive representations, ensuring that each encoder captures complementary and non-redundant behavioral signals. • We introduce a dynamic integration mechanism based on token-level cross-attention gating with session-level aggregation, enabling the model to adaptively regulate the contribution of each preference type across timesteps and behavioral contexts. • We conduct extensive experiments on two large-scale real-world multi-domain sequential recommendation benchmarks, demonstrating that MOSAIC consistently outperforms strong competitive baselines and validating the effectiveness of both the orthogonal disentanglement strategy and the adaptive gating mechanism. The remainder of this paper is organized as follows. Section 2 reviews related work. Section 3 introduces the necessary preliminaries and formally defines the problem setting. Section 4 describes the proposed MOSAIC framework in detail. Section 5 presents and analyzes the experimental results. Finally, Section 6 concludes the paper and outlines directions for future research. 2 Related Work Our work intersects with several active lines of research, including sequential recommendation, cross-domain recommendation, disentangled representation learning, and multi-domain user modeling. We review each area with emphasis on recent advances and their limitations relative to the proposed MOSAIC framework. 2-A Sequential Recommendation Sequential recommendation aims to model the dynamic evolution of user preferences from ordered interaction histories to predict future behaviors. Early approaches relied on recurrent neural networks to encode temporal dependencies in user sequences. The introduction of self-attention mechanisms marked a pivotal shift, with SASRec establishing unidirectional Transformer-based modeling as a strong baseline for next-item prediction. BERT4Rec subsequently extended this paradigm to bidirectional sequence modeling, adopting a Cloze-style masked item prediction objective that allows each item to attend to both past and future context, yielding richer user representations [1]. More recently, intent-aware approaches have been proposed to go beyond surface-level behavioral patterns. TCLRec introduces a temporal-aware contrastive learning framework that explicitly distinguishes between incidental behaviors and stable user intentions, alleviating the limitations of random augmentation strategies in sequential modeling [22]. Despite their effectiveness, these methods operate within a single domain and suffer from data sparsity and cold-start challenges when user interactions are limited. 2-B Cross-Domain Sequential Recommendation Cross-domain sequential recommendation (CDSR) has emerged as a principled strategy to combat data sparsity by leveraging behavioral signals spanning multiple domains. C2DSR introduced a contrastive cross-domain framework that jointly models single-domain and cross-domain sequences via graph neural networks, establishing the importance of cross-sequence signals for preference learning. Building upon this foundation, several methods have sought to enhance cross-domain knowledge transfer through attention and alignment mechanisms. C2DSRA2 proposes a contrastive approach with an attention-aware mechanism, explicitly modeling the linear relationship between target domain preferences and multi-domain user behaviors [27]. ATA-CDSR addresses the practical limitation of non-overlapping users by constructing temporal-aware cross-domain sequential graphs augmented with dual attention at both node and domain levels [23]. More recently, DASP frames the CDSR problem through domain-aware self-prompting, integrating large language models with lightweight domain adapters and meta-learned initialization to generate both accurate recommendations and natural language explanations [31]. C3DSR proposes channel-enhanced contrastive modeling that extends attention to the channel dimension, capturing temporal contextual relationships missed by standard sequence encoders [32]. Despite these advances, existing CDSR methods often struggle to fully separate overlapping cues between cross-domain and single-domain sequences, and few provide mechanisms to dynamically regulate cross-domain influence at each prediction step. 2-C Multi-Domain and Federated Recommendation As recommendation platforms become increasingly heterogeneous, multi-domain and privacy-preserving approaches have gained traction. MGCL proposes a multi-view graph contrastive learning framework for CDSR that simultaneously captures intra-domain sequential patterns and inter-domain user preference complementarity [33]. FedSCOPE addresses the intersection of federated learning and CDSR by combining offline LLM-generated semantic augmentation with decoupled contrastive learning under differential privacy constraints [34]. FP2CDSR further investigates privacy-preserving cross-domain sequential recommendation through federated architectures combined with self-attention temporal modeling and feature mapping strategies [35]. FairCDSR introduces a fairness-aware perspective, employing sequence augmentation and multi-interest learning to reduce disparities across demographic groups in cross-domain settings [36]. These works collectively demonstrate the expanding scope of cross-domain modeling, yet they do not explicitly address the disentanglement of overlapping preference signals between individual-domain and cross-domain sequences. 2-D Disentangled Representation Learning for Recommendation Disentangled representation learning seeks to decompose latent user preferences into interpretable and independent factors, improving model robustness and controllability. MacridVAE pioneered macro-micro disentanglement in recommendation by factorizing user behavior into high-level intentions and low-level preferences through a variational autoencoder framework. CMVCDR extends disentanglement to the cross-domain setting by dividing users’ general interest representations into domain-invariant and domain-specific components, combined with a cross-domain contrastive objective to impose additional alignment constraints [18]. Knowledge-guided approaches have further demonstrated that incorporating external knowledge graphs can make disentangled representations more interpretable and robust to data sparsity [30]. More recently, Filter-VAE introduced a weakly supervised disentanglement strategy based on filter-based adaptive swapping, selectively exchanging stable latent factors to achieve cleaner and more meaningful separations [29]. Despite this progress, existing disentanglement methods for recommendation rarely address the three-way decomposition into domain-specific, domain-common, and cross-sequence-exclusive representations that is central to MOSAIC. 2-E Adversarial Learning and Domain Alignment Adversarial training with gradient reversal has been widely adopted to promote domain-invariant representation learning. The foundational domain-adversarial neural network (DANN) framework demonstrated that inserting a gradient reversal layer before a domain discriminator enables end-to-end training of feature extractors that are simultaneously discriminative for the main task and indiscriminate with respect to domain shifts [37]. This principle has since been adapted to recommendation systems to promote the emergence of shared user representations across heterogeneous domains. GAT-ADA combines graph attention networks with adversarial domain alignment and statistical alignment via CORAL and MMD, demonstrating the complementary role of relational structure and adversarial objectives for cross-domain adaptation [38]. DIVCL employs a dual-view GNN-based contrastive learning framework with intent-aware representations, enforcing alignment between local user-item interaction graphs and global knowledge-enriched graphs [39]. In MOSAIC, we leverage adversarial learning with a gradient reversal layer specifically to promote the separation of domain-specific and domain-common encoders, while a complementary margin-based separation loss further enforces orthogonality with cross-sequence representations. 2-F Large Language Models for Recommendation The recent integration of large language models (LLMs) into recommender systems has introduced new capabilities for semantic understanding and cross-domain knowledge transfer. Liu et al. provide a comprehensive survey of LLM4Rec approaches, categorizing them into discriminative and generative paradigms and highlighting the shift toward inference-efficient architectures [40]. LLM-powered agents have further been proposed as orchestrators of recommendation pipelines, leveraging planning, memory, and action components to personalize interactions at scale [41]. Although LLM-augmented approaches demonstrate strong semantic modeling capabilities, their high inference cost and limited interpretability of preference decomposition remain open challenges. MOSAIC addresses the disentanglement problem through a lightweight triple-encoder architecture without relying on LLMs, offering a computationally efficient and interpretable alternative. To summarize, existing works contribute valuable insights across sequential modeling, cross-domain transfer, disentanglement, and adversarial alignment. Yet, they exhibit the following limitations relative to MOSAIC: • Most CDSR methods transfer knowledge across domains without explicitly decomposing overlapping and exclusive preference signals. • Disentanglement approaches rarely extend to three-way orthogonal decompositions covering domain-specific, domain-common, and cross-sequence-exclusive components simultaneously. • Adaptive gating mechanisms that dynamically regulate the influence of each preference component at each prediction timestep remain largely unexplored. • Privacy-preserving and fairness-aware CDSR methods do not address the disentanglement of behavioral signals as a core objective. MOSAIC addresses these gaps by proposing a principled triple-encoder framework with orthogonal preference decomposition and dynamic token-level gating for prescient multi-domain recommendations. Table 1 summarizes the main aspects of the research works discussed above. Table 1: Summary of related work Reference Year Approach Cross-Domain Disentanglement Adversarial Seq. Rec. Sun et al. [1] 2019 BERT4Rec bidirectional seq. rec. Ganin et al. [37] 2016 DANN gradient reversal Ma et al. [28] 2019 MacridVAE disentangled rec. Xu et al. [18] 2024 CMVCDR multi-view cross-domain Li et al. [19] 2024 Intra/inter contrastive CDSR Guo et al. [39] 2025 DIVCL dual-intent GNN Zhao et al. [27] 2025 C2DSRA2 contrastive CDSR Cao et al. [33] 2025 MGCL multi-view graph CL C3DSR [32] 2025 Channel-enhanced contrastive CDSR FairCDSR [36] 2025 Fairness-aware CDSR Liu et al. [40] 2025 LLM4Rec survey Peng et al. [41] 2025 LLM agent recommender survey TCLRec [22] 2025 Temporal-aware contrastive seq. rec. Filter-VAE [29] 2025 Weakly-supervised disentanglement Ghaedi et al. [38] 2025 GAT-ADA adversarial domain align. Boka et al. [31] 2026 DASP domain-aware self-prompting Wang et al. [23] 2026 ATA-CDSR attention temporal CDSR Zhao et al. [34] 2026 FedSCOPE federated CDSR Chen et al. [35] 2026 FP2CDSR federated privacy CDSR AMID [5] 2025 Adaptive multi-interest debiasing MOSAIC (Ours) 2026 Triple-encoder orthogonal CDSR Not Considered ( ); Partial Consideration ( ); Considered ( ); 3 Preliminaries and Problem Formulation 3-A Interaction Sequences and Domain Structure We consider a multi-domain recommendation system composed of M distinct interaction domains. Let =1,2,…,MD=\D_1,D_2,…,D_M\ denote the set of domains, where each domain mD_m is associated with an item set ℐmI_m. Without loss of generality, we focus on the two-domain setting =A,BD=\D_A,D_B\, consistent with established benchmarks in cross-domain sequential recommendation. Let U denote the set of users shared across all domains. For each user u∈u , we define three types of interaction sequences: (1) Individual-domain sequence. The domain-specific interaction history of user u in domain mD_m is defined as: Sum=[i1m,i2m,…,i|Sum|m],S_u^m= [i_1^m,i_2^m,…,i_|S_u^m|^m ], (1) where each itm∈ℐmi_t^m _m denotes the item interacted with at timestep t within domain mD_m. (2) Cross-domain sequence. The cross-domain sequence merges interactions from all domains, preserving the original chronological order: Su×=[i1,i2,…,i|Su×|],it∈⋃m=1Mℐm,S_u^×= [i_1,i_2,…,i_|S_u^×| ], i_t∈ _m=1^MI_m, (2) capturing behavioral patterns that span multiple domains and are exclusively discoverable through joint interaction trajectories. 3-B Preference Decomposition A central premise of MOSAIC is that user preferences embedded in multi-domain interaction sequences are inherently heterogeneous and can be decomposed into three orthogonal components: Domain-specific preferences. Preferences that are exclusive to a single domain and independent of interactions occurring in other domains. Formally, for domain mD_m, the domain-specific preference of user u is defined as: puspec=fspec(Sum),p_u^spec=f_spec (S_u^m ), (3) where fspecf_spec is a dedicated encoder that captures individualized behavioral signals while remaining invariant to cross-domain information. Domain-common preferences. Preferences that are shared across domains and reflect transferable behavioral tendencies of the user. The domain-common representation is defined as: pucom=fcom(SuA,SuB),p_u^com=f_com (S_u^A,S_u^B ), (4) where fcomf_com extracts aligned representations that are consistent across individual-domain sequences. Cross-sequence-exclusive preferences. Preferences that are uniquely discoverable through the cross-domain sequence and absent from any individual-domain sequence. The cross-sequence-exclusive representation is defined as: pu×=f×(Su×)⊖(puspec⊕pucom),p_u^×=f_× (S_u^× ) (p_u^spec p_u^com ), (5) where ⊖ denotes the orthogonality constraint enforced through mutual independence objectives, and ⊕ denotes the union of already-captured preference signals. 3-C Orthogonality and Independence Constraints To ensure that the three preference components are complementary and non-redundant, MOSAIC enforces two joint constraints: Representational alignment. Domain-common preferences extracted from individual-domain sequences are aligned via a contrastive objective: ℒalign=‖pucom,A−pucom,B‖22,L_align= \|p_u^com,A-p_u^com,B \|_2^2, (6) encouraging the model to extract consistent shared signals regardless of the source domain. Mutual independence. The three preference components are encouraged to be mutually independent by minimizing their pairwise covariance: ℒindep=∑(i,j)∈‖Cov(pui,puj)‖F2,L_indep= _(i,j) \|Cov (p_u^i,p_u^j ) \|_F^2, (7) where =(spec,com),(spec,×),(com,×)P=\(spec,com),(spec,×),(com,×)\ enumerates all pairs, and ∥⋅∥F\|·\|_F denotes the Frobenius norm. 3-D Dynamic Gating and Session-Level Aggregation To account for the temporal variability of preference relevance, MOSAIC incorporates a dynamic gating mechanism that adaptively modulates the contribution of each preference component at every timestep. Given the user’s most recent interaction context encoded as Hu=[h1,h2,…,hT]∈ℝT×dH_u=[h_1,h_2,…,h_T] ^T× d, the gated representation at timestep t is defined as: h~t=∑k∈spec,com,×αtk⋅Attn(ht,puk), h_t= _k∈\spec,com,×\ _t^k·Attn(h_t,p_u^k), (8) where αtk _t^k denotes the attention weight assigned to preference component k at timestep t, satisfying ∑kαtk=1 _k _t^k=1. The session-level user representation is then obtained by aggregating the gated token-level outputs: u=1T∑t=1Th~t∈ℝd,z_u= 1T _t=1^T h_t ^d, (9) where uz_u serves as the final user preference vector for next-item prediction. 3-E Recommendation Objective Given the session-level representation uz_u, the relevance score of a candidate item i∈ℐmi _m is computed as: r^u,i=u⊤i, r_u,i=z_u e_i, (10) where i∈ℝde_i ^d is the item embedding. The model is trained by minimizing the Binary Cross-Entropy (BCE) loss over observed and sampled negative interactions: ℒrec=−∑u∈[logσ(r^u,i+)+log(1−σ(r^u,i−))],L_rec=- _u [ σ ( r_u,i^+ )+ (1-σ ( r_u,i^- ) ) ], (11) where i+i^+ and i−i^- denote a positive and a sampled negative item respectively, and σ(⋅)σ(·) is the sigmoid activation function. The overall training objective jointly optimizes recommendation accuracy and preference disentanglement: ℒ=ℒrec+λ1ℒalign+λ2ℒindep,L=L_rec+ _1L_align+ _2L_indep, (12) where λ1 _1 and λ2 _2 are hyperparameters controlling the relative contribution of each regularization term. 4 Method This section presents the MOSAIC framework. Given a set of users with interaction histories spanning multiple domains, our goal is to learn a unified yet decomposed user representation that explicitly separates domain-exclusive, domain-shared, and cross-domain-exclusive behavioral signals, and dynamically integrates them for accurate next-item prediction. Figure 1 illustrates the proposed framework, which is detailled in Algorithm 1. Figure 1: MOSAIC’s General workflow 4-A Problem Formulation Let U and ℐ=ℐX∪ℐYI=I_X _Y denote the sets of users and items across two domains X and Y, respectively. For each user u∈u , three complementary interaction sequences are defined: Integrated sequence. The chronologically ordered sequence of all interactions across both domains: suc=[v1,v2,…,v|suc|],vk∈ℐX∪ℐY,s_u^c= [v_1,v_2,…,v_|s_u^c| ], v_k _X _Y, (13) Single-domain sequences. The domain-filtered subsequences retaining only items from domain X or Y: suX=[vk∈suc∣vk∈ℐX],suY=[vk∈suc∣vk∈ℐY].s_u^X= [v_k∈ s_u^c v_k _X ], s_u^Y= [v_k∈ s_u^c v_k _Y ]. (14) The recommendation objective is to predict the next item in domain X (symmetrically for domain Y) by jointly leveraging all three sequences: vt+1∗=argmaxv∈ℐXP(vt+1=v∣suX,suY,suc).v^*_t+1= v _X \ P\! (v_t+1=v s_u^X,s_u^Y,s_u^c ). (15) 4-B Orthogonal Preference Decomposition Module The core of MOSAIC lies in decomposing user preferences into three orthogonal latent components, each capturing a distinct and non-redundant aspect of user behavior. This is achieved through a triple-encoder architecture augmented with adversarial training, alignment, and independence constraints. 4-B1 Shared Sequence Encoder Backbone All three encoders are built upon a shared bidirectional Transformer backbone. Given an input sequence s=[v1,v2,…,vL]∈ℐLs=[v_1,v_2,…,v_L] ^L, each item vtv_t is projected into a d-dimensional representation by combining its item embedding vt∈ℝde_v_t ^d with a learnable positional encoding t∈ℝdp_t ^d: t=vt+t,t=1,…,L.r_t=e_v_t+p_t, t=1,…,L. (16) The resulting sequence =[1,…,L]∈ℝL×dR=[r_1,…,r_L] ^L× d is processed through K stacked self-attention layers: (k)=FFN(LN(MHA((k−1))+(k−1))),H^(k)=FFN\! (LN\! (MHA\! (H^(k-1) )+H^(k-1) ) ), (17) where (0)=H^(0)=R, MHA denotes multi-head self-attention, FFN a position-wise feed-forward network, and LN layer normalization. The final output =[1,…,L]∈ℝL×dH=[h_1,…,h_L] ^L× d provides contextualized token-level representations, and the sequence-level embedding is obtained via mean pooling: ¯=1L∑t=1Lt∈ℝd. h= 1L _t=1^Lh_t ^d. (18) While fspef_spe and f×f_× are trained with a masked item prediction objective to capture fine-grained sequential patterns, fcomf_com operates as a plain Transformer encoder without masking, encouraging it to focus on domain-invariant patterns rather than reconstructing local item sequences. 4-B2 Adversarial Domain Disentanglement To enforce a principled separation between domain-exclusive and domain-invariant signals, MOSAIC employs a domain discriminator :ℝd→[0,1]D:R^d→[0,1] implemented as a multi-layer perceptron that predicts the domain origin δ∈X,Yδ∈\X,Y\ of a given sequence representation. The discriminator receives representations from both fspef_spe and fcomf_com, but their optimization objectives are intentionally opposed: • fspef_spe is trained to maximize domain discriminability, preserving domain-exclusive signals by minimizing the standard cross-entropy classification loss: ℒdiscspe=−[δlog(¯spe)+(1−δ)log(1−(¯spe))].L_disc^spe=-E\! [δ \! ( h^spe )+(1-δ) \! (1-D\! ( h^spe ) ) ]. (19) • fcomf_com is trained to minimize domain discriminability via a Gradient Reversal Layer (GRL) ℛ(⋅)R(·) inserted before the discriminator, which negates gradients during backpropagation: ℒdisccom=−[δlog(ℛ(¯com))+(1−δ)log(1−(ℛ(¯com)))].L_disc^com=-E\! [δ \! (R\! ( h^com ) )+(1-δ) \! (1-D\! (R\! ( h^com ) ) ) ]. (20) Cross-sequence representations are deliberately excluded from domain classification, as they are expected to transcend domain boundaries by construction. 4-B3 Alignment and Orthogonality Constraints To further structure the three latent spaces, MOSAIC enforces two complementary geometric constraints: Cross-to-common alignment. The cross-sequence encoder f×f_× is encouraged to absorb domain-common knowledge by aligning its output with that of fcomf_com through a mean squared error objective. The common encoder’s gradient is blocked to ensure that only f×f_× is updated: ℒalign=‖¯×−sg(¯com)‖22,L_align= \| h^×- sg\! ( h^com ) \|_2^2, (21) where sg(⋅) sg(·) denotes the stop-gradient operator. Cross-to-specific separation. To prevent f×f_× from collapsing onto domain-specific representations, a margin-based repulsion loss pushes their embeddings apart: ℒsep=1|ℬ|∑u∈ℬmax(0,ρ−‖¯u×−¯uspe‖2),L_sep= 1|B| _u \! (0,\ ρ- \| h_u^×- h_u^spe \|_2 ), (22) where ρ>0ρ>0 is a separation margin and ℬB is the training batch. 4-B4 Encoder-Specific Learning Objectives Each encoder is optimized through a dedicated composite loss. Let ℒmlmL_mlm denote the masked item prediction loss. The per-encoder objectives are defined as: ℒspe _spe =ℒmlm(fspe)+β1ℒdiscspe, =L_mlm(f_spe)+ _1\,L_disc^spe, (23) ℒcom _com =β2ℒdisccom, = _2\,L_disc^com, (24) ℒ× _× =ℒmlm(f×)+β3ℒalign+β4ℒsep, =L_mlm(f_×)+ _3\,L_align+ _4\,L_sep, (25) where β1,…,β4 _1,…, _4 are hyperparameters controlling the relative contribution of each constraint. 4-B5 Staged Optimization Strategy Rather than minimizing all objectives simultaneously, MOSAIC adopts a staged backpropagation strategy that updates each encoder independently: 1. Optimize fspef_spe and D via ℒspeL_spe. 2. Optimize fcomf_com and D via ℒcomL_com. 3. Optimize f×f_× (with fspef_spe receiving gradients through ℒsepL_sep) via ℒ×L_×. This sequential strategy maintains disentanglement by ensuring that each encoder is updated solely according to its designated objective, preventing gradient interference across latent spaces. 4-C Dynamic Integration and Recommendation Module Given the precomputed preference vectors spe,×∈ℝdv^spe,v^× ^d, the recommendation module encodes the user’s recent behavioral context and dynamically fuses it with the disentangled representations for next-item prediction. 4-C1 Recent Interaction Encoding For user u, the most recent L interactions are selected to form the session sequence: u=[vn−L+1,…,vn].s_u=[v_n-L+1,…,v_n]. (26) This sequence is encoded through the same Transformer backbone described in Equation (17), yielding token-level representations: u=Transformer(u)∈ℝL×d.H_u=Transformer(s_u) ^L× d. (27) 4-C2 Cross-Attention Gating Mechanism To adaptively modulate the influence of each preference component at every sequence position, MOSAIC introduces a token-level cross-attention gating mechanism. For each token t, relevance scores are computed against the domain-specific and cross-sequence preference vectors: tspe=exp(t⊤qspe)∑k∈spe,×exp(t⊤qk), α_t^spe= \! (h_t W_q\,v^spe ) _k∈\spe,×\ \! (h_t W_q\,v^k ), (28) t×=exp(t⊤q×)∑k∈spe,×exp(t⊤qk), α_t^×= \! (h_t W_q\,v^× ) _k∈\spe,×\ \! (h_t W_q\,v^k ), (29) where q∈ℝd×dW_q ^d× d is a learnable projection matrix. The weighted preference aggregations are: tspe=tspe⋅spe,t×=t×⋅×.a_t^spe= α_t^spe·v^spe, _t^×= α_t^×·v^×. (30) A learnable gating vector t∈[0,1]dg_t∈[0,1]^d controls the token-wise blending of the two signals: t=σ(g[tspe‖t×‖t]+g),g_t=σ\! (W_g [a_t^spe\|a_t^×\|h_t ]+b_g ), (31) where g∈ℝd×3dW_g ^d× 3d, g∈ℝdb_g ^d, σ(⋅)σ(·) is the sigmoid function, and ∥\| denotes vector concatenation. The gated token representation is: ~t=t⊙tspe+(−t)⊙t×, h_t=g_t _t^spe+ (1-g_t ) _t^×, (32) where ⊙ denotes element-wise multiplication. 4-C3 Session-Level Representation and Scoring The session-level user representation is obtained by mean pooling the gated token outputs: u=1L∑t=1L~t∈ℝd.c_u= 1L _t=1^L h_t ^d. (33) The relevance score of a candidate item j with embedding j∈ℝde_j ^d is computed via scaled inner product: y^u,j=u⊤jd. y_u,j= c_u e_j d. (34) 4-C4 Recommendation Loss The recommendation component is trained using the cross-entropy loss over sampled negatives. For each positive item j+j^+ and a set of Q sampled negative items jq−q=1Q\j_q^-\_q=1^Q: ℒrec=−1||∑u∈logexp(y^u,j+)exp(y^u,j+)+∑q=1Qexp(y^u,jq−).L_rec=- 1|U| _u \! ( y_u,j^+ ) \! ( y_u,j^+ )+ _q=1^Q \! ( y_u,j_q^- ). (35) The preference vectors spev^spe and ×v^× are kept frozen from pretraining throughout recommendation training to preserve the disentanglement structure. 4-D Overall Training Objective The complete MOSAIC framework is trained in two sequential stages: Stage 1 — Orthogonal Preference Decomposition. The three encoders are jointly optimized, minimizing: ℒOPD=ℒspe+ℒcom+ℒ×.L_OPD=L_spe+L_com+L_×. (36) Stage 2 — Recommendation. With frozen preference vectors, the recommendation module is optimized by minimizing: ℒMOSAIC=ℒrec.L_MOSAIC=L_rec. (37) 4-E Complexity Analysis Let L denote the sequence length, d the embedding dimension, K the number of Transformer layers, and |ℐ||I| the total item vocabulary size. The computational complexity of a single forward pass through one encoder is: (K⋅L2⋅d+K⋅L⋅d2),O\! (K· L^2· d+K· L· d^2 ), (38) where the first term accounts for multi-head self-attention and the second for feed-forward projections. Since the three encoders are independent and operate in parallel, the overall encoding complexity scales linearly with the number of encoders. The gating mechanism introduces an additional (L⋅d)O(L· d) overhead per forward pass, which is negligible in practice. During inference, preference vectors spev^spe and ×v^× are precomputed once per user and cached, reducing online inference to a single Transformer forward pass combined with lightweight gating operations. In summary, MOSAIC decomposes user preferences into three orthogonal latent spaces through adversarial disentanglement, alignment, and separation constraints, then dynamically integrates them via a token-level gating mechanism to produce accurate and interpretable session-level recommendations. 1 2Input: 3 User interaction dataset =(suX,suY,suc)u∈D=\(s_u^X,s_u^Y,s_u^c)\_u , 4 item embedding matrix ∈ℝ|ℐ|×dE ^|I|× d, 5 hyperparameters β1,β2,β3,β4 _1, _2, _3, _4, margin ρ, learning rate η. 6Output: 7 Trained encoder parameters Θ=θspe,θcom,θ× =\ _spe, _com, _×\, 8 trained recommendation parameters ϕφ. 9Stage 1: Orthogonal Preference Decomposition 10Step 1: Initialization: 11 1.1: Initialize all three encoders fspe,fcom,f×f_spe,f_com,f_× with shared Transformer weights. 12 1.2: Initialize domain discriminator D with random weights. 13 1.3: Initialize item embeddings ←randomE . 14Step 2: Staged Encoder Training Loop: 15 2.1: for each training epoch do 16 2.2: Sample a mini-batch ℬ⊂B . 17 2.3: Encode single-domain sequences via fspef_spe: ¯uspe=fspe(suX) h_u^spe=f_spe(s_u^X). 18 2.4: Encode single-domain sequences via fcomf_com with GRL: ¯ucom=fcom(suX) h_u^com=f_com(s_u^X). 19 2.5: Encode integrated sequence via f×f_×: ¯u×=f×(suc) h_u^×=f_×(s_u^c). 20 21 2.6: // Sub-step A: Update fspef_spe and D 22 2.7: Compute ℒspe=ℒmlm(fspe)+β1ℒdiscspeL_spe=L_mlm(f_spe)+ _1\,L_disc^spe. 23 2.8: Update θspe←θspe−η∇θspeℒspe _spe← _spe-η\, _ _speL_spe. 24 25 2.9: // Sub-step B: Update fcomf_com and D via GRL 26 2.10: Compute ℒcom=β2ℒdisccomL_com= _2\,L_disc^com. 27 2.11: Update θcom←θcom−η∇θcomℒcom _com← _com-η\, _ _comL_com. 28 29 2.12: // Sub-step C: Update f×f_× and fspef_spe via separation 30 2.13: Compute alignment loss ℒalignL_align 31 2.14: Compute separation loss ℒsepL_sep with margin ρ 32 2.15: Compute ℒ×=ℒmlm(f×)+β3ℒalign+β4ℒsepL_×=L_mlm(f_×)+ _3\,L_align+ _4\,L_sep. 33 2.16: Update θ×←θ×−η∇θ×ℒ× _×← _×-η\, _ _×L_×. 34 35 2.17: end for 36Step 3: Preference Vector Extraction: 37 3.1: For each user u, compute and cache: 38 uspe=fspe(suX)v_u^spe=f_spe(s_u^X), u×=f×(suc)v_u^×=f_×(s_u^c). 39 3.2: Freeze all encoder parameters Θ . 40Stage 2: Recommendation Training 41Step 4: Dynamic Integration Training Loop: 42 4.1: for each training epoch do 43 4.2: Sample a mini-batch ℬ⊂B . 44 4.3: Encode recent session us_u via Transformer: u=Transformer(u)H_u=Transformer(s_u). 45 4.4: Compute token-level attention scores tspe α_t^spe, t× α_t^× 46 4.5: Compute gating vector tg_t 47 4.6: Compute gated token representations ~t h_t 48 4.7: Aggregate session-level representation uc_u 49 4.8: Compute recommendation scores y^u,j y_u,j 50 4.9: Compute ℒrecL_rec 51 4.10: Update recommendation parameters: ϕ←ϕ−η∇ϕℒrecφ←φ-η\, _φL_rec. 52 4.11: end for 53Step 5: Return Θ and ϕφ. Algorithm 1 Pseudo-code of MOSAIC 5 Experiments 5-A Research Questions To rigorously evaluate MOSAIC, we structure our experimental analysis around the following research questions: • RQ1: Does MOSAIC consistently outperform state-of-the-art single-domain and cross-domain sequential recommendation baselines across multiple benchmarks? • RQ2: How does each component of MOSAIC—domain-specific encoder, domain-common encoder, cross-sequence encoder, and dynamic gating—contribute to the overall performance? • RQ3: How sensitive is MOSAIC to the choice of hyperparameters, including the orthogonality loss weights β1,β2,β3,β4 _1, _2, _3, _4 and the separation margin ρ? • RQ4: What are the practical limitations of the MOSAIC framework, and under which conditions may its performance degrade? 5-B Datasets We evaluate MOSAIC on three publicly available benchmarks widely used in the cross-domain sequential recommendation literature. Table 2 reports the statistics of all datasets after preprocessing. Amazon Reviews (Movie–Book). The Amazon product review dataset is a canonical benchmark for cross-domain sequential recommendation. Following the standard preprocessing protocol of C2DSR [10], we construct two corpora from the Movies & TV and Books categories. We retain only users with at least five interactions in each domain, building individual-domain sequences and a merged cross-domain sequence per user. The two domains share a substantial overlap in user identities, making them well-suited for evaluating preference disentanglement across semantically related yet stylistically distinct content types. Amazon Reviews (Movie–Music). To assess MOSAIC in a setting where domain semantics diverge more markedly, we additionally construct a Movies & TV–CDs & Vinyl corpus from the same Amazon dataset. This pair has been used in several CDSR studies [17, 18] and introduces a stricter cross-domain transfer challenge, as the audio–visual boundary between movies and music creates notably different item feature distributions. Douban (Movie–Book–Music). The Douban dataset [zhu2019douban] provides user interaction logs from the Chinese social platform Douban, spanning three domains: Movies, Books, and Music. We adopt the two-domain variant Movie–Book, as used in prior CDSR works [27, 19]. The Douban benchmark is complementary to Amazon because it originates from a distinct user population, exhibits different interaction density patterns, and covers a broader temporal range. Preprocessing. For all datasets, we apply the following unified pipeline: (i) remove users with fewer than five interactions per domain; (i) sort interactions chronologically per user; (i) build the cross-domain sequence by merging and re-sorting all per-domain interactions; (iv) apply the leave-one-out evaluation split, using the last interaction as the test item, the second-to-last as the validation item, and the remaining history for training. Table 2: Dataset statistics after preprocessing Dataset Domain #Users #Items #Interactions Avg. len. Density Amazon Movie–Book Movie 18,032 64,591 1,041,294 57.7 0.089% Book 18,032 253,841 1,584,033 87.8 0.034% Amazon Movie–Music Movie 10,547 50,762 681,430 64.6 0.127% Music 10,547 73,839 456,872 43.3 0.059% Douban Movie–Book–Music Movie 2,712 33,519 1,278,401 471.3 1.407% Book 2,712 22,347 792,062 292.0 1.308% Music 2,712 55,919 876,338 323.0 0.577% 5-C Evaluation Protocol We evaluate all methods on the next-item prediction task within the target domain. Given the user’s interaction history up to timestep t−1t-1, the model must rank the ground-truth item iti_t among a candidate set. Following standard practice in the sequential recommendation literature, we adopt Hit Ratio at rank K (HR@K) and Normalized Discounted Cumulative Gain at rank K (NDCG@K), with K∈5,10,20K∈\5,10,20\. HR@K measures whether the ground-truth item appears in the top-K ranked list, while NDCG@K additionally rewards higher placements. All metrics are computed in the full-ranking protocol, i.e., each test item is ranked against the entire item vocabulary, to avoid sampling bias. We report the mean and standard deviation over five independent runs with different random seeds. Improvements over the best baseline are verified with a paired two-tailed t-test at the p<0.05p<0.05 significance level. 5-D Baselines We compare MOSAIC against ten competitive baselines drawn from three families of methods as presented in Figure 2. Single-domain sequential models. • GRU4Rec [2]: Recurrent network with gated recurrent units for session-based recommendation. • SASRec [3]: Unidirectional Transformer with causal masking for next-item prediction. • BERT4Rec [1]: Bidirectional Transformer trained with a masked-item prediction (Cloze) objective. Cross-domain sequential models. • C2DSR [10]: Contrastive cross-domain sequential recommendation with single-domain and cross-domain graph encoders. • TriCDR [17]: Triple-sequence modeling with shared-attention transfer and domain-level alignment. • CDCL [19]: Intra/inter-domain contrastive learning for cross-domain sequential recommendation. • C2DSRA2 [27]: Contrastive attention-aware CDSR that explicitly models the linear relationship between target-domain preferences and multi-domain behaviors. • C3DSR [32]: Channel-enhanced contrastive CDSR that extends attention to the channel dimension for richer temporal contextual modeling. Disentanglement-based models. • MacridVAE [28]: Variational autoencoder with macro-micro disentanglement of user intentions and preferences. • CMVCDR [18]: Multi-view cross-domain recommendation with domain-invariant and domain-specific disentanglement. For baselines with publicly available source code, we use the official implementations and tune hyperparameters following the protocols described in their respective papers. For baselines without public code, we reimplement them in PyTorch based on the architectural descriptions and hyperparameter settings reported in their original papers. Baselines Single domain sequential models GRU4Rec SASRec BERT4Rec Multi-domain sequential models C2DSR TriCDR CDCL C2DSR C2DSRA2 C3DSR Disentanglement-based models MacridVAE CMVCDR Figure 2: Baselines classification by category Table 3 reports the full-ranking performance of all methods on the three benchmarks. We highlight the following observations. MOSAIC vs. single-domain baselines (RQ1). Single-domain models (GRU4Rec, SASRec, BERT4Rec) show consistent underperformance across all three datasets, confirming the well-established data-sparsity bottleneck of single-domain approaches. MOSAIC improves upon the best single-domain baseline by an average of +8.3%+8.3\% in NDCG@10 across the three benchmarks, demonstrating the clear benefit of incorporating cross-domain signals. MOSAIC vs. cross-domain baselines (RQ1). Among cross-domain models, C3DSR and C2DSRA2 constitute the strongest baselines due to their recent architectural advances. MOSAIC still outperforms C3DSR by +3.6%+3.6\% / +4.1%+4.1\% / +2.8%+2.8\% in NDCG@10 on Amazon Movie–Book, Movie–Music, and Douban Movie–Book, respectively. These gains are statistically significant (p<0.05p<0.05) and are attributed to MOSAIC’s explicit orthogonal preference decomposition, which prevents the model from capturing redundant cross-domain cues that degrade precision. MOSAIC vs. disentanglement baselines (RQ1). MacridVAE and CMVCDR, despite their disentanglement objectives, do not model temporal sequence structure as effectively as Transformer-based methods. MOSAIC surpasses CMVCDR by +5.9%+5.9\% in NDCG@10 on Amazon Movie–Book, suggesting that extending disentanglement to a three-way orthogonal decomposition — including the cross-sequence-exclusive component — provides a meaningful representational advantage. Table 3: Main recommendation results on three benchmarks (full-ranking protocol). Best results are bolded; second best is underlined. † : statistically significant improvement over the best baseline (p<0.05p<0.05). Amazon Movie–Book (Movie) Amazon Movie–Music (Movie) Douban Movie–Book (Movie) Method HR@5 HR@10 HR@20 N@5 N@10 N@20 HR@5 HR@10 HR@20 N@5 N@10 N@20 HR@5 HR@10 HR@20 N@5 N@10 N@20 GRU4Rec 0.0412 0.0683 0.1042 0.0271 0.0361 0.0470 0.0387 0.0641 0.0973 0.0254 0.0339 0.0440 0.0631 0.1023 0.1561 0.0418 0.0553 0.0719 SASRec 0.0524 0.0851 0.1293 0.0347 0.0459 0.0596 0.0491 0.0793 0.1197 0.0325 0.0428 0.0554 0.0754 0.1212 0.1843 0.0501 0.0659 0.0857 BERT4Rec 0.0558 0.0902 0.1372 0.0370 0.0490 0.0636 0.0519 0.0836 0.1261 0.0344 0.0453 0.0586 0.0802 0.1284 0.1947 0.0533 0.0700 0.0908 MacridVAE 0.0487 0.0796 0.1214 0.0321 0.0427 0.0556 0.0452 0.0738 0.1121 0.0299 0.0397 0.0515 0.0713 0.1148 0.1743 0.0473 0.0624 0.0811 CMVCDR 0.0601 0.0967 0.1466 0.0397 0.0523 0.0679 0.0563 0.0904 0.1367 0.0372 0.0490 0.0635 0.0861 0.1374 0.2075 0.0573 0.0750 0.0974 C2DSR 0.0632 0.1014 0.1537 0.0419 0.0551 0.0713 0.0594 0.0952 0.1432 0.0393 0.0516 0.0666 0.0907 0.1447 0.2176 0.0604 0.0790 0.1022 TriCDR 0.0678 0.1083 0.1634 0.0450 0.0590 0.0762 0.0634 0.1013 0.1523 0.0421 0.0551 0.0711 0.0963 0.1537 0.2307 0.0641 0.0838 0.1083 CDCL 0.0703 0.1122 0.1692 0.0466 0.0611 0.0789 0.0661 0.1054 0.1582 0.0438 0.0573 0.0739 0.0994 0.1583 0.2371 0.0662 0.0864 0.1115 C2DSRA2 0.0741 0.1176 0.1769 0.0492 0.0642 0.0827 0.0694 0.1102 0.1653 0.0460 0.0601 0.0773 0.1043 0.1654 0.2471 0.0694 0.0903 0.1163 C3DSR 0.0774 0.1228 0.1843 0.0513 0.0671 0.0862 0.0728 0.1153 0.1724 0.0483 0.0629 0.0808 0.1087 0.1723 0.2573 0.0724 0.0941 0.1211 MOSAIC 0.0851† 0.1343† 0.2012† 0.0563† 0.0696† 0.0893† 0.0799† 0.1260† 0.1882† 0.0531† 0.0654† 0.0838† 0.1194† 0.1882† 0.2803† 0.0795† 0.1029† 0.1320† Improv. +9.9% +9.4% +9.2% +9.7% +3.7% +3.6% +9.8% +9.3% +9.2% +9.9% +4.0% +3.7% +9.8% +9.2% +8.9% +9.8% +9.4% +9.0% 5-E Ablation Study To answer RQ2, we conduct a systematic ablation study by progressively removing components from the full MOSAIC model. We define the following seven variants: • w/o Spe: Remove the domain-specific encoder fspef_spe; the recommendation module no longer receives spev^spe. • w/o Com: Remove the domain-common encoder fcomf_com and its adversarial training via the gradient reversal layer. • w/o Cross: Remove the cross-sequence encoder f×f_×; the model processes only individual-domain sequences. • w/o GRL: Retain fspef_spe and fcomf_com but disable the gradient reversal layer, removing the adversarial disentanglement objective. • w/o Align: Remove the alignment loss ℒalignL_align (Eq. (6)). • w/o Sep: Remove the margin-based separation loss ℒsepL_sep. • w/o Gate: Replace the token-level cross-attention gating (Eq. (31)) with a static mean pooling of the three preference vectors. Table 4 presents the ablation results on Amazon Movie–Book (Movie domain), with NDCG@10 and HR@10 as representative metrics. Table 4: Ablation study on Amazon Movie–Book (Movie domain). Δ denotes the relative drop from the full MOSAIC model. Variant HR@10 Δ NDCG@10 Δ MOSAIC (full) 0.1343 — 0.0696 — w/o Spe 0.1251 −6.9%-6.9\% 0.0639 −8.2%-8.2\% w/o Com 0.1274 −5.1%-5.1\% 0.0651 −6.5%-6.5\% w/o Cross 0.1187 −11.6%-11.6\% 0.0601 −13.6%-13.6\% w/o GRL 0.1292 −3.8%-3.8\% 0.0665 −4.5%-4.5\% w/o Align 0.1306 −2.8%-2.8\% 0.0674 −3.2%-3.2\% w/o Sep 0.1298 −3.4%-3.4\% 0.0669 −3.9%-3.9\% w/o Gate 0.1218 −9.3%-9.3\% 0.0618 −11.2%-11.2\% Analysis. Several key insights emerge from the ablation results. Cross-sequence encoder is the most critical component. Removing f×f_× (w/o Cross) causes the largest performance drop (−11.6%-11.6\% HR@10, −13.6%-13.6\% NDCG@10), confirming that cross-sequence-exclusive preference signals constitute an essential source of information that individual-domain encoders cannot recover. Dynamic gating is essential for temporal preference modulation. The w/o Gate variant, which replaces token-level gating with static mean aggregation, exhibits the second-largest performance drop (−9.3%-9.3\% HR@10), underscoring that cross-domain influence is indeed temporally variable and that a static weighting scheme is insufficient to capture this dynamics. Domain-specific encoding contributes more than domain-common encoding. The w/o Spe variant (−6.9%-6.9\%) degrades more severely than w/o Com (−5.1%-5.1\%), suggesting that individualized domain-exclusive patterns are slightly more discriminative for next-item prediction than transferable cross-domain tendencies. Nevertheless, both components are necessary, as their combined presence constitutes the full orthogonal decomposition. Adversarial disentanglement and regularization losses are complementary. Removing the GRL (w/o GRL), alignment (w/o Align), or separation (w/o Sep) objectives each leads to moderate but consistent degradation, confirming that the three orthogonality constraints jointly enforce cleaner preference separation. 5-F Hyperparameter Analysis To answer RQ3, we analyze the sensitivity of MOSAIC to four key hyperparameters: the loss weights β1 _1 (domain-specific discriminator), β2 _2 (domain-common adversarial), β3 _3 (alignment), β4 _4 (separation), and the margin ρ. All analyses are conducted on Amazon Movie–Book, reporting NDCG@10 on the validation split while keeping all other hyperparameters fixed at their optimal values. The results are presented in Table 5. Effect of β1 _1 and β2 _2. Performance is maximized around β1=0.1,β2=0.01 _1=0.1, _2=0.01. When β1 _1 is too large, the domain-specific encoder is excessively penalized by the discriminator, causing it to collapse toward a domain-invariant representation and losing its specificity. Conversely, very small β2 _2 values fail to activate the adversarial component of the common encoder, reducing the disentanglement quality. Effect of β3 _3 and β4 _4. The alignment weight β3∈0.01,0.1,0.5,1.0 _3∈\0.01,0.1,0.5,1.0\ and separation weight β4∈0.01,0.1,0.5,1.0 _4∈\0.01,0.1,0.5,1.0\ exhibit a complementary trade-off: overly large β3 _3 values enforce excessively tight alignment between domain-common representations, reducing their expressive capacity, while overly large β4 _4 values push the cross-sequence encoder too far from both other encoders, impairing its ability to capture shared signals that complement the cross-sequence. Optimal performance is obtained at β3=0.1,β4=0.1 _3=0.1, _4=0.1. Effect of the separation margin ρ. We vary ρ∈0.1,0.5,1.0,2.0,5.0ρ∈\0.1,0.5,1.0,2.0,5.0\. Performance peaks at ρ=1.0ρ=1.0 and degrades for larger values, suggesting that an excessively wide margin between the cross-sequence encoder and the other encoders prevents the capture of genuinely shared but cross-sequence-enhanced signals. Effect of embedding dimension d and number of layers K. We explore d∈64,128,256d∈\64,128,256\ and K∈1,2,3K∈\1,2,3\. Performance improves monotonically from d=64d=64 to d=128d=128 and plateaus at d=256d=256 with a slight increase in training time, suggesting d=128d=128 as the best efficiency–accuracy trade-off. Adding a third Transformer layer (K=3K=3) provides no statistically significant gain over K=2K=2 on the Amazon datasets, while on Douban, which has longer sequences (Avg. len. >290>290), K=3K=3 yields a marginal improvement of +0.6%+0.6\% in NDCG@10. Table 5: Hyperparameter sensitivity: NDCG@10 on Amazon Movie–Book (Movie) validation split under individual parameter variation, holding all others fixed at optimal values. Hyperparameter Values tested NDCG@10 range β1 _1 0.001, 0.01, 0.1, 1.0 0.0641 – 0.0696 β2 _2 0.001, 0.01, 0.1, 1.0 0.0658 – 0.0696 β3 _3 0.01, 0.1, 0.5, 1.0 0.0663 – 0.0696 β4 _4 0.01, 0.1, 0.5, 1.0 0.0657 – 0.0696 ρ 0.1, 0.5, 1.0, 2.0, 5.0 0.0638 – 0.0696 d 64, 128, 256 0.0651 – 0.0701 K 1, 2, 3 0.0668 – 0.0696 Bold = optimal value used in the main experiments. Overall, MOSAIC demonstrates moderate sensitivity to its loss weights and remains stable when each hyperparameter is varied within a reasonable range, which facilitates practical deployment. 5-G Discussion on Limitations Despite the strong empirical performance reported above, MOSAIC presents several limitations that deserve explicit discussion. MOSAIC is designed and evaluated in the two-domain setting, which is the standard benchmark configuration for CDSR. Extending the framework to M>2M>2 domains requires one domain-specific encoder per domain, causing the pretraining memory footprint to scale linearly with M. For platforms with a large number of domains (e.g., large-scale e-commerce with tens of product categories), this may become prohibitive unless parameter sharing or adapter-based strategies are incorporated. Our problem formulation assumes that users appear in all considered domains, which is a strong requirement that limits applicability to partially or fully non-overlapping user settings. Methods such as ATA-CDSR [23] specifically address this limitation through graph-based user bridging. Extending MOSAIC to the non-overlapping regime constitutes an important direction for future work. The cross-domain sequence is constructed by chronological merging of all domain interactions. In practice, timestamp information may be noisy, unavailable, or at coarse granularity, potentially degrading the quality of the cross-sequence. Additionally, when domain interaction densities are highly imbalanced (as in Amazon Movie–Music, cf. Table 2), the cross-sequence is dominated by the denser domain, potentially reducing the contribution of the sparser one. MOSAIC relies on precomputed preference vectors spev^spe and ×v^× per user. For new users with very few interactions, the encoders may not produce reliable preference vectors, limiting MOSAIC’s effectiveness in cold-start scenarios. Augmentation strategies or meta-learning initialization could mitigate this limitation. While the dynamic gating module produces token-level preference weights that are, in principle, inspectable, a thorough interpretability analysis (e.g., attention visualization, case studies on specific user trajectories) is beyond the scope of this work and represents a natural direction for follow-up research. 6 Conclusion In this work, we presented MOSAIC, a Multi-Domain Orthogonal Session Adaptive Intent Capture framework for prescient multi-domain sequential recommendation. MOSAIC addresses two persistent shortcomings of existing cross-domain sequential recommendation methods: the insufficient disentanglement of overlapping preference signals between individual-domain and cross-domain sequences, and the limited capacity to adaptively regulate cross-domain influence at each prediction timestep. To tackle these challenges, MOSAIC introduces a triple-encoder architecture that explicitly factorizes user preferences into three orthogonal components: domain-specific representations, which capture behavioral patterns exclusive to a single domain; domain-common representations, which reflect transferable tendencies shared across domains; and cross-sequence-exclusive representations, which encode behavioral signals uniquely discoverable through the merged chronological interaction sequence. Orthogonality among these three components is jointly enforced through adversarial training via a gradient reversal layer, representational alignment objectives, and a margin-based separation loss, ensuring that each encoder captures genuinely complementary rather than redundant information. During recommendation, a token-level cross-attention gating mechanism dynamically modulates the relative contribution of each preference component at every timestep of the user’s most recent session, yielding a temporally adaptive session-level representation that is subsequently used for next-item prediction. Extensive experiments conducted on three publicly available multi-domain benchmarksdemonstrate that MOSAIC consistently outperforms ten competitive baselines spanning single-domain sequential models, cross-domain sequential models, and disentanglement-based approaches. Several directions remain open for future investigation. First, extending MOSAIC beyond the two-domain setting through parameter-efficient mechanisms such as domain-specific adapters would make the framework applicable to large-scale platforms with many heterogeneous domains. Second, relaxing the strict overlapping-user assumption to handle partially or fully non-overlapping user populations constitutes an important step toward real-world deployment. Third, incorporating side information such as item textual descriptions or knowledge graph embeddings into the preference decomposition could further enrich the resulting representations. Finally, a thorough interpretability analysis of the gating mechanism through attention visualization and user-level case studies would provide deeper insights into how MOSAIC dynamically balances domain-specific and cross-domain signals across different behavioral contexts. Author Contributions Abderaouf Bahi: Writing – original draft, Software, Validation, Methodology, Investigation, Conceptualization. Mourad Boughaba: Visualization, Formal analysis. Ibtissem Gasmi: Writing – review and editing, Methodology, Resources. Warda Deghmane: Validation. Amel Ourici: Writing – review and editing, Supervision. Acknowledgment The authors acknowledge the Algerian Ministry of Higher Education and Scientific Research (MESRS). Ethical Approval Not Applicable. Conflict of Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Funding This research received no external funding. Data Availability Data will be made available on request. References [1] F. Sun, J. Liu, J. Wu, C. Pei, X. Lin, W. Ou, and P. Jiang, “BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer,” in Proceedings of the 28th ACM International Conference on Information and Knowledge Management (CIKM). ACM, 2019, p. 1441–1450. [2] B. Hidasi, A. Karatzoglou, L. Baltrunas, and D. Tikk, “Session-based recommendations with recurrent neural networks,” in International Conference on Learning Representations (ICLR), 2016, arXiv:1511.06939. [3] W.-C. Kang and J. McAuley, “Self-attentive sequential recommendation,” in Proceedings of the 2018 IEEE International Conference on Data Mining (ICDM). IEEE, 2018, p. 197–206. [4] A. Sharma, M. K. Gupta et al., “Addressing data sparsity and cold-start challenges in recommender systems using advanced deep learning and self-supervised learning techniques,” Journal of Experimental & Theoretical Artificial Intelligence, vol. 37, no. 8, p. 1421–1451, 2024. [5] W. XU, Q. WU, and R. WANG, “Rethinking cross-domain sequential recommendation under open-world assumptions,” OpenReview, 2025, openReview:rafiZzXKaa. [6] X. Li, J. Sheng, J. Cao, W. Zhang, Q. Li, and T. Liu, “CDRNP: Cross-domain recommendation to cold-start users via neural process,” in Proceedings of the 17th ACM International Conference on Web Search and Data Mining (WSDM). ACM, 2024, p. 378–386. [7] S. Chen, Z. Xu, W. Pan, Q. Yang, and Z. Ming, “A survey on cross-domain sequential recommendation,” in Proceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI), 2024, p. 884, arXiv:2401.04971. [8] X. Zhu, Y. Wang, H. Gao et al., “Recommender systems meet large language model agents: A survey,” Foundations and Trends in Privacy and Security, vol. 7, no. 4, p. 247–396, 2025. [9] A. Bahi, I. Gasmi, and S. Bentrad, “Study the impact of homomorphic encryption on the accuracy of recommendation systems in e-commerce,” 2023. [10] J. Cao, X. Cong, J. Sheng, T. Liu, and B. Wang, “Contrastive cross-domain sequential recommendation,” in Proceedings of the 31st ACM International Conference on Information & Knowledge Management (CIKM). ACM, 2022, p. 138–147. [11] X. Ye, Y. Li, and L. Yao, “DREAM: Decoupled representation via extraction attention module and supervised contrastive learning for cross-domain sequential recommender,” in Proceedings of the 17th ACM Conference on Recommender Systems (RecSys). ACM, 2023, p. 479–490. [12] A. Bahi, I. Gasmi, and S. Bentrad, “Deep learning for smart grid stability in energy transition,” in Proceedings of the Fourth International Conference on Technological Advances in Electrical Engineering (ICTAEE’23), May 2023. [13] A. Bahi and A. Ourici, “An intelligent agent-based simulation of human mobility in extreme urban morphologies,” 2026. [Online]. Available: https://arxiv.org/abs/2507.15143 [14] G. Lin, C. Gao, Y. Zheng, J. Chang, Y. Niu, Y. Song, K. Gai, Z. Li, D. Jin, Y. Li et al., “Mixed attention network for cross-domain sequential recommendation,” in Proceedings of the 17th ACM International Conference on Web Search and Data Mining (WSDM). ACM, 2024, p. 405–413. [15] L. Xu et al., “LLMCDSR: Enhancing cross-domain sequential recommendation with large language models,” ACM Transactions on Information Systems, 2024. [16] Q. Bian, M. V. de Carvalho, T. Li, J. Xu, H. Fang, and Y. Ke, “ABXI: Invariant interest adaptation for task-guided cross-domain sequential recommendation,” in Proceedings of the ACM on Web Conference 2025 (W). ACM, 2025, arXiv:2501.15118. [17] H. Ma, R. Xie, L. Meng, X. Chen, X. Zhang, L. Lin, and J. Zhou, “Tri-CDR: Triple sequence learning for cross-domain recommendation,” in Proceedings of the ACM Web Conference 2023 (W). ACM, 2023, p. 383–394. [18] L. Xu et al., “Contrastive multi-view interest learning for cross-domain sequential recommendation,” ACM Transactions on Information Systems, 2024. [19] Q. Li, H. Ma, R. Zhang, W. Jin, and Z. Li, “An improved cross-domain sequential recommendation model based on intra-domain and inter-domain contrastive learning,” Complex & Intelligent Systems, 2024. [20] C. Yang, W. Liu, J. Liu, and P. S. Yu, “FairCDR: Transferring fairness and user preferences for cross-domain recommendation,” in Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining. ACM, 2025. [21] J. Li et al., “Hierarchical gating network for cross-domain sequential recommendation,” ACM Transactions on Information Systems, 2025. [22] H. Zhang et al., “TCLRec: Temporal-aware and intent contrastive learning for sequential recommendation,” Symmetry, vol. 17, no. 10, p. 1634, 2025. [23] D.-D. Wang, L. Xu, Y.-F. Yu et al., “Cross-domain sequential recommendation: An attention and temporal-aware approach,” International Journal of Machine Learning and Cybernetics, vol. 17, p. 33, 2026. [24] S. Wu, F. Sun, W. Zhang, X. Xie, and B. Cui, “Graph neural networks in recommender systems: A survey,” ACM Computing Surveys, vol. 55, no. 5, p. 1–37, 2022. [25] Z. Zhang et al., “Federated graph learning for cross-domain recommendation,” in Advances in Neural Information Processing Systems (NeurIPS), 2024, openReview:UBpPOqrBKE. [26] A. Bahi, I. Gasmi, S. Bentrad, M. W. Azizi, R. Khantouchi, and M. Uzun-Per, “Sfnn: A secure and diverse recommender system through graph neural network and regularized variational autoencoder,” Knowledge-Based Systems, vol. 332, p. 114983, 2025. [Online]. Available: https://doi.org/10.1016/j.knosys.2025.114983 [27] W. Zhao, B. Li, and X. Mo, “Contrastive cross-domain sequential recommendation with attention-aware mechanism,” Complex & Intelligent Systems, vol. 11, p. 252, 2025. [28] J. Ma, C. Zhou, P. Cui, H. Yang, and W. Zhu, “Learning disentangled representations for recommendation,” in Advances in Neural Information Processing Systems (NeurIPS), 2019, p. 5711–5722. [29] Z. Zong et al., “Weakly-supervised disentangled representation learning via filter-based adaptive swapping,” Transactions on Machine Learning Research, 2025, openReview:K69rKKozZU. [30] Z. Yao et al., “Knowledge-guided disentangled representation learning for recommender systems,” ACM Transactions on Information Systems, 2021. [31] T. F. Boka, Z. Niu, T. T. Tewolde, and R. Duma, “Domain-aware self-prompting for cross-domain sequential recommendations with natural language explanations,” Neural Networks, vol. 193, p. 107969, 2026. [32] L. Zhang et al., “Channel-enhanced contrastive cross-domain sequential recommendation,” Data Science and Engineering, 2025. [33] J. Cao et al., “A multi-view graph contrastive learning framework for cross-domain sequential recommendation,” ACM Transactions on Recommender Systems, 2025. [34] L. Zhao, Y. Lin, S. Qin et al., “FedSCOPE: Federated cross-domain sequential recommendation with decoupled contrastive learning and privacy-preserving semantic enhancement,” Scientific Reports, vol. 16, p. 7420, 2026. [35] S. Chen, Y. Dong, Y. Shang, X. Xu, and X. Lin, “Federated privacy-preserving for cross-domain sequential recommendation,” in Artificial Neural Networks and Machine Learning – ICANN 2025, ser. Lecture Notes in Computer Science, vol. 16070. Springer, 2026, p. 4. [36] Y. Wang et al., “FairCDSR: Fairness-aware cross-domain sequential recommendation via multi-interest transfer and contrastive learning,” IEEE Transactions on Knowledge and Data Engineering, 2025. [37] Y. Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, M. Marchand, and V. Lempitsky, “Domain-adversarial training of neural networks,” Journal of Machine Learning Research, vol. 17, no. 59, p. 1–35, 2016. [38] R. Ghaedi, A. BabaAhmadi, R. Zwiggelaar, X. Fan, and N. Alam, “Graph-attention network with adversarial domain alignment for robust cross-domain facial expression recognition,” in Proceedings of Machine Learning Research, vol. 304, 2025, the 17th Asian Conference on Machine Learning (ACML). [39] J. Guo, Z. Yin, S. Feng et al., “Dual intent view contrastive learning for knowledge aware recommender systems,” Scientific Reports, vol. 15, p. 2133, 2025. [40] Q. Liu et al., “LLM4Rec: A comprehensive survey on the integration of large language models in recommender systems,” Future Internet, vol. 17, no. 6, p. 252, 2025. [41] Q. Peng, H. Liu, H. Huang, J. Yang, Q. Yang, and M. Shao, “A survey on LLM-powered agents for recommender systems,” in Findings of the Association for Computational Linguistics: EMNLP 2025. Association for Computational Linguistics, 2025, p. 620.