Paper deep dive

Generative Data Transformation: From Mixed to Unified Data

Jiaqing Zhang, Mingjia Yin, Hao Wang, Yuxin Tian, Yuyang Ye, Yawen Li, Wei Guo, Yong Liu, Enhong Chen

Year: 2026Venue: arXiv preprintArea: cs.AIType: PreprintEmbeddings: 61

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%

Last extracted: 7/20/2026, 9:59:19 AM

Summary

The paper introduces Taesar, a data-centric framework for cross-domain sequential recommendation that addresses data sparsity and negative transfer by regenerating mixed-domain sequences into unified target-domain sequences. Unlike model-centric approaches that use complex fusion architectures, Taesar employs a contrastive decoding mechanism with a base model, source-domain experts, and a target-domain expert to adaptively encode cross-domain context into target sequences, thereby improving recommendation performance without architectural modifications to downstream models.

Entities (9)

Taesar · framework · 98%Contrastive Decoding · technique · 94%Cross-Domain Sequential Recommendation · task · 92%Data-Centric Paradigm · approach · 91%Model-Centric Paradigm · approach · 91%Negative Transfer · problem · 90%Base Model · model-component · 88%Source Domain Expert · model-component · 88%Target Domain Expert · model-component · 88%

Relation Signals (8)

Taesar → targets → Cross-Domain Sequential Recommendation

confidence 97% · we present the first data-centric paradigm for CDSR... Taesar aims to shift the paradigm to a data-centric view.

Taesar → uses → Contrastive Decoding

confidence 96% · Taesar... employs a contrastive decoding mechanism to adaptively encode cross-domain context into target-domain sequences.

Taesar → addresses → Negative Transfer

confidence 95% · Taesar effectively combines the strengths of data- and model-centric paradigms... proactively mitigates negative transfer before model training.

Taesar → implements → Data-Centric Paradigm

confidence 93% · we propose a novel data-centric approach for cross-domain sequential recommendation... Taesar – target-aligned sequential regeneration.

Base Model → iscomponentof → Taesar

confidence 90% · Taesar first trains a base model on the mixed-domain sequences... Subsequently, domain-specific expert models are derived

Source Domain Expert → iscomponentof → Taesar

confidence 90% · domain-specific expert models are derived by adapting the base model to each domain’s individual sequences

Target Domain Expert → iscomponentof → Taesar

confidence 90% · the target-domain expert serves as a strong supervisor

Taesar → →

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Recommendation model performance is intrinsically tied to the quality, volume, and relevance of their training data. To address common challenges like data sparsity and cold start, recent researchs have leveraged data from multiple auxiliary domains to enrich information within the target domain. However, inherent domain gaps can degrade the quality of mixed-domain data, leading to negative transfer and diminished model performance. Existing prevailing \emph{model-centric} paradigm -- which relies on complex, customized architectures -- struggles to capture the subtle, non-structural sequence dependencies across domains, leading to poor generalization and high demands on computational resources. To address these shortcomings, we propose \textsc{Taesar}, a \emph{data-centric} framework for \textbf{t}arget-\textbf{a}lign\textbf{e}d \textbf{s}equenti\textbf{a}l \textbf{r}egeneration, which employs a contrastive decoding mechanism to adaptively encode cross-domain context into target-domain sequences. It employs contrastive decoding to encode cross-domain context into target sequences, enabling standard models to learn intricate dependencies without complex fusion architectures. Experiments show \textsc{Taesar} outperforms model-centric solutions and generalizes to various sequential models. By generating enriched datasets, \textsc{Taesar} effectively combines the strengths of data- and model-centric paradigms. The code accompanying this paper is available at~ \textcolor{blue}{this https URL}.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

59,603 characters extracted from source content.

Expand or collapse full text

by Generative Data Transformation: From Mixed to Unified Data Jiaqing Zhang 0009-0001-1039-9735 jiaqing.zhang@mail.ustc.edu.cn University of Science and Technology of China, Hefei, China , Mingjia Yin 0009-0005-0853-1089 mingjia-yin@mail.ustc.edu.cn University of Science and Technology of China, Hefei, China , Hao Wang 0000-0001-9921-2078 wanghao3@ustc.edu.cn University of Science and Technology of China, Hefei, China , Yuxin Tian 0009-0008-6141-6284 tyx682@mail.ustc.edu.cn University of Science and Technology of China, Hefei, China , Yuyang Ye 0000-0002-1513-7814 yeyuyang@mail.ustc.edu.cn University of Science and Technology of China, Hefei, China , Yawen Li 0000-0003-2662-3444 warmly0716@126.com Beijing University of Posts and Telecommunications, Beijing, China , Wei Guo guowei67@huawei.com 0000-0001-8616-0221 Huawei Noah’s Ark LabShenzhenChina , Yong Liu liu.yong6@huawei.com 0000-0001-9031-9696 Huawei Noah’s Ark LabShenzhenChina and Enhong Chen 0000-0002-4835-4102 cheneh@ustc.edu.cn University of Science and Technology of China, Hefei, China (2026) Abstract. Recommendation model performance is intrinsically tied to the quality, volume, and relevance of their training data. To address common challenges like data sparsity and cold start, recent researchs have leveraged data from multiple auxiliary domains to enrich information within the target domain. However, inherent domain gaps can degrade the quality of mixed-domain data, leading to negative transfer and diminished model performance. Existing prevailing model-centric paradigm – which relies on complex, customized architectures – struggles to capture the subtle, non-structural sequence dependencies across domains, leading to poor generalization and high demands on computational resources. To address these shortcomings, we propose Taesar, a data-centric framework for target-aligned sequential regeneration, which employs a contrastive decoding mechanism to adaptively encode cross-domain context into target-domain sequences. It employs contrastive decoding to encode cross-domain context into target sequences, enabling standard models to learn intricate dependencies without complex fusion architectures. Experiments show Taesar outperforms model-centric solutions and generalizes to various sequential models. By generating enriched datasets, Taesar effectively combines the strengths of data- and model-centric paradigms. The code accompanying this paper is available at https://github.com/USTC-StarTeam/Taesar. Sequential Recommendation, Data-Centric, Data Regeneration †journalyear: 2026†copyright: c†conference: Proceedings of the ACM Web Conference 2026; April 13–17, 2026; Dubai, United Arab Emirates†booktitle: Proceedings of the ACM Web Conference 2026 (W ’26), April 13–17, 2026, Dubai, United Arab Emirates†doi: 10.1145/3774904.3792124†isbn: 979-8-4007-2307-0/2026/04†ccs: Information systems Recommender systems Figure 1. Cosine similarity heatmap of gradient directions across four domains in multi-domain sequential recommendation. Naïvely combining data from multiple domains for training induces gradient conflicts and inconsistencies. 1. Introduction Figure 2. Motivation of Taesar. We aim to eliminate the domain semantic gap at the source level, prior to model training, thereby transforming cross-domain mixed data into unified target-domain data. Left: Source items with high transferability are mapped to semantically closest target items, while others are discarded. Right: Regenerated data aligns cross-domain behavioral patterns with target-domain relevance, improving model gradients and target-domain performance. Sequential recommendation seeks to model the temporal dynamics of user preferences based on chronological interaction sequences, thereby facilitating the prediction of users’ subsequent interactions, and enhancing personalized online experiences (Wu et al., 2024; Chen et al., 2024b; Wang et al., 2022; Xie et al., 2024; Petrov and Macdonald, 2024; Chong et al., 2023; Shen et al., ). Although considerable progress has been achieved in this field, real-world user behaviors are frequently dispersed across diverse platforms and domains. This fragmentation introduces two principal challenges: (i) exacerbated data sparsity within individual domains (Bao et al., 2025; Kim et al., 2023; Liu et al., 2023a; Yin et al., 2025; Xie et al., 2025; Xu et al., 2025a; 03-zhang2025killing; 04-ye2025fuxi; Zhou et al., 2025b; Wang et al., 2025b, a), which fundamentally limits the modeling capacity of single-domain recommenders, and (i) significant distributional heterogeneity across domains (Zang et al., 2022; Zhang et al., 2025a; Chen et al., 2024a; Cao et al., 2022; Xu et al., 2025b; Park et al., 2024), wherein domain-specific user interests consequently impede effective knowledge transfer, as shown in Figure 1. These challenges collectively prompt a critical research question: How can dispersed cross-domain user activities be effectively harnessed to reliably enhance target-domain recommendation while mitigating the risk of negative transfer ? A growing body of research has sought to address this challenge through cross-domain sequential recommendation (CDSR). The dominant paradigm involves jointly learning heterogeneous interaction data from multiple domains within a unified representation space. To mitigate negative transfer, several approaches (Hou et al., 2022b; Zhou et al., 2025a; Park et al., 2023b; Wu et al., 2025) incorporate auxiliary multimodal information—such as textual or visual content—as bridging signals across domains. For purely ID-based sequences, existing studies typically adopt a model-centric paradigm, developing sophisticated cross-domain knowledge transfer and fusion modules, often in conjunction with complex multi-task training strategies (Cao et al., 2022; Bian et al., 2025; Park et al., 2024). Despite their effectiveness, these approaches entail several inherent limitations. On one hand, unified multi-domain modeling incurs substantial optimization complexity: balancing multiple domains within a shared representation space increases model intricacy and training costs, often compromising domain-specific accuracy relative to single-domain counterparts. On the other hand, such methods exhibit limited generality and scalability: their cross-domain fusion modules are typically model-dependent, and multi-task learning setups require extensive hyperparameter tuning, hindering practical deployment. To address the limitations of the model-centric paradigm, we propose a novel data-centric approach for cross-domain sequential recommendation. Instead of designing complex transfer architectures, we focus on refining the data itself by eliminating detrimental inter-domain information and regenerating sequences using only target-domain items (Figure 2). Consequently, the challenge of multi-domain modeling is transformed into a precise single-domain problem. Building on this concept, we propose Taesar – target-aligned sequential regeneration. The core insight is that a target-domain model can effectively refine mixed-domain sequences by contrasting its predictions with those from specialized source-domain models, thereby enforcing a closer alignment between the regenerated sequences and target-domain semantics. To operationalize this idea, Taesar first trains a base model on the mixed-domain sequences to capture transferable, global cross-domain patterns. Subsequently, domain-specific expert models are derived by adapting the base model to each domain’s individual sequences by Domain-Specific Prediction strategy, allowing them to specialize in domain-specific behaviors. Guided by these models, Taesar employs Adaptative Contrastive Decoding: for every non-target-domain item within a mixed sequence, it dynamically contrasts the predictions of the corresponding source-domain expert, the target-domain expert, and the base model to determine whether and how to replace the item with a plausible target-domain item. Replacement probabilities are computed using both global and local contrastive scores alongside adaptive weights derived from information theory, yielding refined sequences that preserve the temporal order while containing only target-domain items, consequently enhancing recommendation performance. We highlight four key contributions as follows: • To the best of our knowledge, we present the first data-centric paradigm for CDSR, shifting the focus from model-level adaptation to semantic alignment at the data level. By regenerating sequences through contrastive decoding, our approach proactively mitigates negative transfer before model training. • We develop a unified generative framework comprising a pretraining and domain-adaptation module and a adaptive contrastive decoding module, which collectively address two fundamental challenges in CDSR: identifying translatable cross-domain items and preserving inherent behavioral patterns. • Taesar generates domain-purified sequences while retaining the original item ID structure, enabling seamless integration with existing sequential recommenders without any architectural or training modifications—thereby ensuring high practical compatibility and ease of deployment in CDSR systems. 2. Related works Cross-Domain Sequential Recommendation (CDSR) CDSR enhances recommendation systems by modeling user behaviors across domains for richer user understanding (Ma et al., 2024). The mainstream research trajectory has been dominated by model-centric paradigm. Early efforts, such as Pi-Net (Ma et al., 2019) and PSJNet (Sun et al., 2021), employed shared filters and transfer units to exchange information between domains, yet their performance was often constrained by domain similarity and susceptibility to behavioral noise. Subsequent works introduced more advanced transfer mechanisms, including graph-based propagation frameworks (e.g., MIFN (Ma et al., 2022), DA-GCN (Guo et al., 2021)) and attention-enhanced representation models (e.g., C2DSR (Cao et al., 2022), DREAM (Ye et al., 2023), MAN (Lin et al., 2024)) that jointly capture intra- and inter-domain dependencies. To improve generalization and scalability, several studies have explored domain-invariant and text-driven representations. UniCDR (Hou et al., 2022a) and UniSRec (Hou et al., 2022b) leveraged masking and contrastive pretraining to learn transferable user semantics, while RecGURU (Li et al., 2022a) incorporated adversarial learning to derive domain-agnostic embeddings. More recent approaches have addressed specific challenges, such as mitigating negative transfer via gradient regularization (Park et al., 2024), aligning semantic spaces for disjoint-user scenarios (Liu et al., 2023b), and enhancing task consistency through invariant LoRA modules (Bian et al., 2025). Despite progress, model-centric methods remain plagued by architectural complexity and optimization instability, limiting their adaptability. Therefore, to overcome these hurdles, Taesar aims to shift the paradigm to a data-centric view. Data Regeneration (DR) Machine learning is fundamentally framed as learning a mapping f:→f:X , where X denotes inputs (e.g., interaction history) and Y denotes targets (e.g., items) (Yin et al., 2024). However, the significant semantic gap between the noisy X and the concise Y makes direct optimization difficult. DR addresses this by introducing a latent transition: →′→X . In this framework, ′X acts as a distilled intermediate representation, simplifying the learning process and enhancing model convergence. This strategy differs fundamentally from existing paradigms such as data distillation (Sachdeva et al., 2023; Zhang et al., 2025b; Sachdeva et al., 2022; Wang et al., 2023; Wu et al., 2023), which compresses real data into compact synthetic representations; data generation (Liu et al., 2021; Kim et al., 2023; Yu et al., 2025), which synthesizes new samples from learned distributions; and other data augmentation approaches (Dang et al., 2024a, b; Liu et al., 2023a, 2021), which expand existing datasets through stochastic or semantic transformations. Building on the success of dataset regeneration methods like DR4SR (Yin et al., 2024) in capturing transition patterns, we propose to regenerate mixed-domain sequences into target-exclusive formats, a process designed to distill information from multi-domain and boost transferability. Contrastive Decoding (CD) It enhances language model generation by contrasting expert and amateur outputs to optimize fluency, diversity, and coherence (Li et al., 2022b). Recent advances demonstrate contrastive decoding’s versatility across diverse applications: it enhances reasoning in LLMs (O’Brien and Lewis, 2023) (despite persistent factual recall challenges), enables universal detoxification without model-specific tuning (Huimin et al., 2025), improves factual generation through APD’s asymptotic probability extrapolation (Chang et al., 2024), and suppresses hallucinations in multimodal models via ECD (Fieback et al., 2025). While effective for text generation, CD’s potential for data regeneration—especially in sequential recommendation—remains unexplored. Our tri-model CD framework addresses this gap for cross-domain sequential data regeneration. 3. Preliminaries Cross-Domain Sequential Recommendation A domain D is formally defined as a structured tuple =(,,ℰ)D=(U,V,E), where U denotes the set of users, V represents the set of items, and ℰE captures the set of observed interactions. The interaction set ℰE comprises a collection of sequences ii=1|ℰ|\x_i\_i=1^|E|, with each sequence i=[xij∈]j=1|i|x_i=[x_ij ]_j=1^|x_i| corresponding to an ordered list of items from V, thereby reflecting a user’s historical engagement. In the multi-domain paradigm, we consider M source domains 1,2,…,M\S_1,S_2,...,S_M\ and a singular target domain T. To facilitate domain differentiation, we denote the interaction sequences for a user u as um=(x1m,x2m,…,xtm)x_u^S_m=(x_1^S_m,x_2^S_m,...,x_t^S_m) for the m-th source domain, and u=(x1,x2,…,xt)x_u^T=(x_1^T,x_2^T,...,x_t^T) for the target domain. To enable cross-domain knowledge transfer, we construct a unified, or merged, interaction sequence uℳx_u^M by chronologically interleaving interactions from all source and target domains: (1) uℳ=Interleave⁡(u1,u2,…,uM,u), _u^M=Interleave(x_u^S_1,x_u^S_2,...,x_u^S_M,x_u^T), where the Interleave⁡(⋅)Interleave(·) operator maintains the inherent temporal ordering while consolidating interactions across domains. Given the observed sequences (1,…,M,,ℳ)u\(x^S_1,...,x^S_M,x^T,x^M)_u\, the cross-domain sequential recommendation task leverages the merged sequence ℳx^M to improve predictive performance, calculating probabilities based on the integrated interaction history: (2) xi+1=arg⁡maxxi+1∈P(xi+1|1:iℳ), x_i+1= x_i+1 ^Targ\,max\;\ P^T(x_i+1|x_1:i^M), thereby facilitating the vital knowledge transfer from the multiple source domains mm=1M\S_m\_m=1^M to the target domain T. Cross-Domain Data Regeneration Given an input dataset x where each sequence uℳx_u^M comprises items originating from multiple source domains and the target domain, data regeneration constructs a novel dataset x^N consisting exclusively of T-domain items. For a given sequence uℳx_u^M, the transformation is defined as: (3) uℳ _u^M =(x1d1,x2d2,…,x|uℳ|d|uℳ|),dj∈1,…,M,, =(x_1^d_1,x_2^d_2,...,x_|x_u^M|^d_|x_u^M|), d_j∈\S_1,...,S_M,T\, (4) u _u^N =(y1,y2,…,y|uℳ|)=Regenerate⁡(uℳ),s.t.∀yk∈. =(y_1^T,y_2^T,...,y_|x_u^M|^T)=Regenerate(x_u^M), .t. ∀~y_k ^T. The regeneration operator Regenerate(⋅)Regenerate(·) is implemented as: (5) yk=xkif xk∈,fm(xk)if xk∈m,m=1,…,M,y_k= casesx_k&if x_k ^T,\\ f_m(x_k)&if x_k ^S_m,~m=1,...,M, cases where each function fm(⋅):m→∪∅f_m(·):S_m ∪\ \ either maps (transforms) or discards (replaces with a null item) the item from the m-th source domain into the target-domain item space. The composite dataset is subsequently generated as: (6) =∪, =x^N ^T, where x^T retains the original target-domain sequences, while x^N contains the regenerated target-domain sequences, together constituting an enriched collection of user interaction sequences. 4. Cross-Domain Sequential Data Regeneration Figure 3. Overview of Taesar. (1) In the first stage, Tri-model Pretraining, we construct three views of decoder models: a base model (θℳ _M), distinct source domain experts (θ _S), and a target domain expert (θ _T). (2) We select the base model, the target domain expert, and the source domain expert corresponding to the item to be transformed to regenerate mixed cross-domain sequences using global and local contrastive decoding among the three models. (3) Based on the two types of contrastive scores, we decide whether to replace a source-domain item with a target-domain item or to discard the source-domain item entirely. In this section, we formally present Taesar. As depicted in Figure 3, the framework comprises two principal stages: 1) Pretraining and Domain-Specific Adaptation Stage, designed to preserve cross-domain behavioral patterns and domain-specific semantics; and 2) Adaptive Contrastive Decoding Stage, which focuses on decoupling domain-shared knowledge and eliminating source-domain biases. This dual-mechanism design yields regenerated data that facilitates seamless integration with various downstream recommendation models and ensures strong generalization across subsequent tasks. 4.1. Pretraining and Domain-Specific Adaptation Mixed-domain sequences contain items from multiple domains, whose contextual dependencies and behavioral patterns are inherently complex and heterogeneous. Directly applying contrastive decoding to such sequences would make it difficult for the decoder to capture the latent semantic correspondences across domains. Our goal is twofold: (1) to learn the shared cross-domain structural patterns that are transferable across domains, and (2) to develop domain-specific expertise for accurate target-domain predictions. 4.1.1. Cross-Domain Encoder Pretraining We pretrain a base model θℳ _M on mixed-domain sequences to capture shared sequential dynamics and long-term user preferences across all domains. Formally, the pretraining objective is: (7) θℳ∗ _M^* ≜arg⁡minθ⁡ℒℳ(θ|θℳ0), _θL_M(θ~|~θ^0_M), (8) ℒℳ(θ) _M(θ) ≜∼ℳ[∑i=1||−1ℓ(Φθ(1:i),xi+1)], _x ^M [ _i=1^|x|-1 ( _θ(x_1:i),x_i+1 ) ], where Φθ _θ denotes a Transformer-like decoder, ℓ(⋅) (·) is the cross-entropy loss, and 1:ix_1:i is the subsequence of the first i tokens in x. This pretraining stage achieves cross-domain pattern discovery by capturing generalizable behavioral regularities and performs representation alignment by mapping items from all domains into a shared, structured semantic space, providing a structured foundation for the subsequent contrastive decoding stage. 4.1.2. Domain-Specific Decoder Adaptation After obtaining the pretrained encoder θℳ∗ _M^*, we fine-tune domain-specific decoders for each domain (1,2,…,M,S_1,S_2,…,S_M,T). To preserve cross-domain knowledge while specializing for a target domain T, we employ a Domain-Specific Prediction (DSP) strategy: the model receives the full mixed-domain sequence as input but predicts only items belonging to the current domain. Formally, the adaptation objective is: (9) θ∗ _T^* ≜arg⁡minθ⁡ℒDSP(θ|θℳ∗), _θL_T^DSP(θ~|~ _M^*), (10) ℒDSP(θ) _T^DSP(θ) ≜∼ℳ[∑(i,t)∈()ℓ(Φθ(1:i),xt)], _x ^M [ _(i,t) _T(x) ( _θ(x_1:i),x_t ) ], where ()P_T(x) contains all valid prediction pairs of target-domain items, that can be formalized as follows: (11) ()=(i,t)∣ _T(x)= \(i,t) xi,xt∈,t>i,∄j∈(i,t) s.t. xj∈. ~x_i,x_t ^T,~t>i,~ ∃ j∈(i,t) s.t. x_j ^T \. In words, for each target-domain item xix_i (except the last in the sequence), the decoder predicts the next occurring target-domain item xtx_t, using the full mixed-domain history 1:ix_1:i as context. This approach allows the model to fully utilize cross-domain patterns learned by the encoder while enabling domain-specific decoders to develop the specialized expertise necessary for accurate target-domain predictions. The same SDP strategy is also implemented for every source domain during their respective adaptation stages. 4.2. Adaptive Contrastive Decoding This process is designed to address two key challenges: (1) determining whether an item from a mixed-domain sequence should be converted into a target-domain item, and (2) if transformation is appropriate, deciding which target-domain item it should be transformed into. After obtaining the cross-domain decoder and the domain-specific decoder in the previous stage, we draw inspiration from the conventional ”Expert-Amateur” contrastive decoding paradigm. In this setup, the target-domain expert θ∗θ^*_T serves as a strong supervisor, providing high-quality domain knowledge. The source-domain experts function as Amateurs relative to the target domain, providing guidance on filtering potentially harmful cross-domain biases. Meanwhile, a third-party base model θℳ∗θ^*_M is employed to quantify knowledge that is transferable across domains. 4.2.1. Contrastive Score Calculation When xi+1x_i+1 originates from the source domain k within a mixed sequence, the historical sequence 1:ix_1:i is fed in parallel to the source-domain expert kP^S_k, the target-domain expert P^T, and the base model ℳP^M, yielding unnormalized prediction scores. To balance cross-domain transferability and target-domain specificity, we first compute the Global Contrastive Score over the entire item space V: (12) Global-scorei+1=αg⋅−βg⋅k, -score_i+1= _g·P^T- _g·P^S_k~, where αg _g and βg _g are the domain confidence and transfer reward coefficients, respectively. A source-domain item whose global score is maximized for a target-domain item indicates that the target-domain expert is highly confident about the transformation, and that the item also possesses transferable patterns across domains. Otherwise, the item is unlikely to require transformation. For items deemed convertible, we further compute the Local Contrastive Score restricted to the target-domain item set V^T: (13) Local-scorei+1=αl⋅−βl⋅k. -score_i+1= _l·P^T_V^T- _l·P^S_k_V^T~. By focusing exclusively on target-domain candidates, the local score precisely identifies the most suitable item for transformation, effectively combining domain-specific confidence and cross-domain transferability. This two-level contrastive design enables the model to distill transferable patterns while retaining domain-specialized discriminative power. 4.2.2. Adaptive Weight Decision To enhance the adaptivity of knowledge integration, we define the domain confidence coefficient α and the transfer reward coefficient β from an information-theoretic perspective, leveraging entropy and divergence principles. Specifically, α quantifies the informational certainty of the target-domain expert using the Shannon entropy of its output distribution. Lower entropy indicates a peaked, confident distribution, where the expert assigns high probability to specific items. Accordingly, positions with lower entropy should contribute more during decoding. To normalize α into [0,1][0,1], we define: (14) αg=1−H((v∣1:i)))1−maxv∈⁡H((v∣1:i)). _g= 1-H(P^T(v _1:i)))1- _v H(P^T(v _1:i)). From this viewpoint, α serves as an information-certainty weight, amplifying the target expert’s influence in confident regions while reducing its effect in uncertain areas. Meanwhile, β captures the transferability of cross-domain knowledge via the Jensen–Shannon Divergence (JSD) between the base model and the source-domain expert. As a symmetrized and bounded measure, lower JSD indicates stronger alignment and hence higher transferability, whereas higher divergence highlights domain-specific conflicts. For the prediction distributions ℳP^M and kP^S_k at the current position, βg _g is defined as following: (15) βg=12[KL(ℳ∥ℳ+k2)+KL(k∥ℳ+k2)]. _g= 12 [KL (P^M\,\|\, P^M+P^S_k2 )+KL (P^S_k\,\|\, P^M+P^S_k2 ) ]. Thus, β serves as a transfer-alignment weight, reinforcing knowledge in regions where the base model and source expert are strongly aligned while attenuating contributions from positions with high cross-domain conflict. The same principles are applied locally to αl _l and βl _l, computed are restricted over the target-domain vocabulary. 5. Experiment In this section, we empirically evaluate Taesar in terms of its effectiveness compared with mode-centric methods, generalizability across different architectures, and the rationality of its design. 5.1. Experimental Settings 5.1.1. Datasets. Table 1. Statistics of experimental datasets. # user # item # inter Avg. Dataset ( |||U| ) ( |||V| ) ( ∑N _xN_x ) length sparsity(%) Books 70,672 526,955 13.12 99.98 Electronics 44,278 501,759 12.49 99.97 Sports 38,030 355,807 8.86 99.98 Tools 40,165 31,880 331,230 8.25 99.97 Publicly available Amazon review datasets were employed in our experiments111http://snap.stanford.edu/data/amazon/productGraph/categoryFiles/. We randomly selected four domains: Books, Electronics, Sports and Outdoors, and Tools and Home Improvement. For clarity, these are referred to as Books, Electronics, Sports, and Tools, respectively and follow the preprocessing in (Park et al., 2024; Bian et al., 2025) to guarantee a minimum of 3 interactions associated with each user and item, then, retain only the 128 most recent interactions for longer sequences. Finally, user interactions are formatted chronologically. Table 1 summarizes the statistics of the resulting datasets. Table 2. Overall performance of Taesar compared with state-of-the-art multi-domain sequential recommendation methods, and its generalization across single-domain models with different architectures. Best results are in bold, second-best underlined. BEST: Books, Electronics, Sports, Tools Books Electronics Method HR@10 HR@20 NG@10 NG@20 MRR@10 MRR@20 HR@10 HR@20 NG@10 NG@20 MRR@10 MRR@20 SASRec 0.0563 0.0660 0.0371 0.0395 0.0311 0.0317 0.0350 0.0430 0.0257 0.0277 0.0228 0.0234 GRU4Rec 0.0384 0.0450 0.0283 0.0300 0.0253 0.0257 0.0247 0.0325 0.0184 0.0204 0.0165 0.0170 GCE-GNN 0.0598 0.0733 0.0447 0.0481 0.0400 0.0409 0.0353 0.0424 0.0286 0.0304 0.0266 0.0271 CL4SRec 0.0585 0.0708 0.0442 0.0473 0.0398 0.0406 0.0347 0.0424 0.0278 0.0298 0.0257 0.0262 SASRec4SASRec_4 0.0504 0.0575 0.0342 0.0359 0.0291 0.0295 0.0311 0.0389 0.0240 0.0260 0.0218 0.0224 ABXI 0.0489 0.0549 0.0341 0.0356 0.0369 0.0370 0.0352 0.0423 0.0290 0.0308 0.0246 0.0248 SyNCRec 0.0455 0.0564 0.0337 0.0367 0.0259 0.0262 0.0314 0.0402 0.0252 0.0273 0.0193 0.0197 CGRec 0.0413 0.0596 0.0316 0.0348 0.0218 0.0221 0.0338 0.0429 0.0203 0.0247 0.0190 0.0194 SASRec+Taesar 0.0613 0.0738 0.0406 0.0438 0.0342 0.0351 0.0441 0.0527 0.0291 0.0312 0.0243 0.0249 GRU4Rec + Taesar 0.0407 0.0487 0.0307 0.0328 0.0277 0.0282 0.0288 0.0346 0.0233 0.0248 0.0216 0.0220 GCE-GNN+Taesar 0.0774 0.0945 0.0566 0.0609 0.0502 0.0514 0.0404 0.0499 0.0319 0.0343 0.0293 0.0300 CL4SRec + Taesar 0.0772 0.0912 0.0567 0.0603 0.0504 0.0514 0.0408 0.0506 0.0317 0.0342 0.0289 0.0296 Sports Tools Method HR@10 HR@20 NG@10 NG@20 MRR@10 MRR@20 HR@10 HR@20 NG@10 NG@20 MRR@10 MRR@20 SASRec 0.0399 0.0472 0.0295 0.0314 0.0262 0.0267 0.0333 0.0408 0.0249 0.0268 0.0222 0.0227 GRU4Rec 0.0262 0.0319 0.0202 0.0216 0.0183 0.0187 0.0237 0.0286 0.0188 0.0201 0.0173 0.0177 GCE-GNN 0.0379 0.0460 0.0318 0.0339 0.0300 0.0305 0.0332 0.0389 0.0270 0.0285 0.0251 0.0255 CL4SRec 0.0413 0.0471 0.0335 0.0350 0.0311 0.0315 0.0325 0.0391 0.0268 0.0285 0.0251 0.0255 SASRec4SASRec_4 0.0373 0.0419 0.0281 0.0292 0.0251 0.0254 0.0319 0.0362 0.0236 0.0247 0.0210 0.0213 ABXI 0.0386 0.0438 0.0306 0.0320 0.0254 0.0255 0.0309 0.0357 0.0271 0.0283 0.0232 0.0234 SyNCRec 0.0368 0.0395 0.0306 0.0343 0.0244 0.0247 0.0371 0.0431 0.0307 0.0329 0.0206 0.0230 CGRec 0.0344 0.0384 0.0299 0.0316 0.0224 0.0228 0.0344 0.0399 0.0202 0.0237 0.0215 0.0228 SASRec+Taesar 0.0568 0.0667 0.0354 0.0379 0.0287 0.0294 0.0465 0.0573 0.0288 0.0315 0.0232 0.0240 GRU4Rec + Taesar 0.0322 0.0379 0.0252 0.0266 0.0231 0.0235 0.0267 0.0326 0.0217 0.0231 0.0201 0.0205 GCE-GNN+Taesar 0.0582 0.0676 0.0445 0.0469 0.0403 0.0409 0.0482 0.0568 0.0377 0.0398 0.0344 0.0350 CL4SRec + Taesar 0.0602 0.0698 0.0464 0.0488 0.0421 0.0428 0.0491 0.0568 0.0385 0.0405 0.0352 0.0358 5.1.2. Baselines and Target Models. To fully demonstrate the cross-architecture generalizability of Taesar, we select target models from various model architecture categories: • RNN-Based: GRU4Rec(Hidasi et al., 2015) models user interaction sequences with item embeddings processed by GRUs and optimizes ranking-based objectives to improve next-item prediction accuracy. • GNN-Based: GCE-GNN(Wang et al., 2020) learns item embeddings from both the current session and all global sessions then uses GNN on each graph and combines them with a soft attention mechanism. • Attention-Based: SASRec(Kang and McAuley, 2018) models user sequences with self-attention to capture long-term dependencies. • Contrastive-Learning-Based: CL4SRec(Xie et al., 2022) stands out as a powerful contrastive sequential recommender by introducing three effective unique sequence-level augmentation strategies. To verify the Taesar’s superiority compared with cross-domain methods, we select the following representative baselines: • SASRecmSASRec_m is a directly variant using SASRec (Kang and McAuley, 2018) trained on naive mixing data to revealing performance degradation. • ABXI(Bian et al., 2025) introduces a task-guided alignment mechanism to address the prediction mismatches and employs two types of LoRA to adaptively integrates domain-invariant interests. • CGRec(Park et al., 2023a) reweights domain loss via hierarchical contrastive learning to capture cross-domain correlations. • SyNCRec(Park et al., 2024) uses single-domain and multi-domain contrastive learning to control negative transfer and an auxiliary mutual information loss to enhance representation transfer across tasks. 5.1.3. Evaluation Protocols. We rigorously evaluate the next-item prediction performance using the widely adopted leave-one-out split strategy. Specifically, for each sequence, the most recent interaction is reserved for testing, the second-to-last for validation, and the remainder for training. Consistent with standard cross-domain practices, we construct domain-specific validation and test sets based on the domain of the last item in each sequence. Performance is assessed using three established ranking metrics: Hit Rate (HR@kHR@k), Normalized Discounted Cumulative Gain (NG@kNG@k), and Mean Reciprocal Rank (MRR@kMRR@k), with k∈10,20k∈\10,20\. Moreover, to address the observation by Krichene (Krichene and Rendle, 2020) that ”sampled metrics” (metrics computed using sampled negative items) can yield results inconsistent with full metrics, meaning the relative performance ranking between different algorithms may not remain stable, we adopt the entire item set as the candidate item set during evaluation. 5.1.4. Implementation Details. We implemented Taesar using PyTorch and developed single-domain sequential recommendation models based on the Recbole library. For the multi-domain baselines, we utilized their publicly available source code. Crucially, we replaced the sampled metrics with full metrics to ensure alignment with our evaluation methodology, as discussed previously. Notably, although ABXI was initially demonstrated in a two-domain setting, the original work established its straightforward scalability to multiple domains. Consequently, we adapted and extended both methods to effectively operate in our multi-domain environment. To ensure a fair comparison, all single-domain models were uniformly configured across all datasets (including the Taesar-regenerated dataset) with the following fixed hyperparameters: 22 attention heads and layers, hidden and inner sizes of 128128, and dropout rates of 0.20.2. Conversely, the multi-domain models were subjected to hyperparameter search over the following space: attention heads ∈1,2,3∈\1,2,3\, layers ∈1,2,3∈\1,2,3\, hidden size ∈64,128,256∈\64,128,256\, and inner size ∈64,128,256∈\64,128,256\. Specifically for ABXI, the LoRA rank ∈8,16,32∈\8,16,32\, and for SyNCRec, the number of experts ∈4,6,8∈\4,6,8\. All models utilize the Adam optimizer with a learning rate ∈10−3,3×10−4,10−4∈\10^-3,3× 10^-4,10^-4\. The training process runs for a maximum of 300300 epochs, and early stopping is employed if the NG@10NG@10 on the validation data fails to improve for 3030 consecutive epochs. 5.2. Overall Performance 5.2.1. How superior is Taesar to model-centric methods? Experimental results reveal that cross-domain recommendation performance is highly sensitive to data processing strategies. As shown in Table 2: (1) Naïve multi-domain merging leads to negative transfer. The performance of SASRec4 falls short of the single-domain baseline SASRec, indicating that simply merging data from multiple domains—without any adaptation—induces severe negative transfer due to distributional heterogeneity, thereby compromising the target domain’s performance. (2) Model-centric improvements are limited under weak inter-domain semantics. Traditional model-centric approaches occasionally outperform both SASRec4 and single-domain baselines on specific metrics; however, their overall performance tends to degrade. This suggests that when semantic correlations across domains are weak, architectural enhancements alone cannot effectively bridge inter-domain discrepancies. (3) Taesar effectively mitigates inter-domain interference. In contrast, our data-centric framework Taesar achieves consistent and significant improvements over both SASRec4 and single-domain models. This demonstrates that Taesar alleviates inter-domain interference through data-level refinement, enabling precise and domain-specific modeling without being hindered by heterogeneous knowledge. Collectively, these findings underscore that as model-level optimizations approach diminishing returns, structured data regeneration—the core of Taesar—offers a more robust and interpretable solution for cross-domain recommendation. Table 3. Ablation results of Taesar on different designs of Pretraining, Domain-Specific Adaptation, and Adaptive Contrastive Decoding strategies. Electronics Variants HR@10 HR@20 NG@10 NG@20 MRR@10 MRR@20 w/o DSA 0.0352 0.0434 0.0237 0.0258 0.0202 0.0207 DSA w/o DSP 0.0350 0.0450 0.0241 0.0266 0.0208 0.0215 DSP w/o SDE 0.0398 0.0472 0.0282 0.0300 0.0239 0.0247 DSP w/o GCS 0.0439 0.0523 0.0289 0.0310 0.0242 0.0247 DSP w/o LCS 0.0431 0.0512 0.0278 0.0299 0.0231 0.0236 Taesar Full 0.0441 0.0527 0.0291 0.0312 0.0243 0.0249 5.2.2. How versatile is regenerated data from Taesar? Although Taesar employs a decoder-based architecture to perform contrastive data regeneration, the regenerated data itself remains model-agnostic. To assess its generality, we train multiple models with distinct architectural designs on the regenerated dataset. As shown in Table 2, (1) Regenerated data generalizes across architectures. Models with diverse architectures trained on Taesar’s regenerated data exhibit consistent and substantial performance improvements, confirming the model-agnostic generalization capability of the regenerated dataset. (2) More expressive architectures yield greater improvements. When trained on the regenerated data, models with stronger representational capacity exhibit more pronounced performance gains. This observation suggests that the high-quality, semantically consistent data produced by Taesar allows expressive architectures to more effectively leverage their modeling capacity, thereby amplifying their advantage in capturing subtle user–item interactions. (3) Data-centric regeneration complements model-centric design. The consistent cross-architecture gains underscore the inherent complementarity between paradigms: model-centric approaches enhance structural expressiveness, whereas Taesar’s data-centric paradigm fundamentally strengthens the learning signal, providing a more robust basis for future advancing recommender systems. 5.3. Further Analysis 5.3.1. How indispensable is the architecture of Taesar ? Table 3 presents a comprehensive ablation study evaluating the design choices of Taesar. For the first stage, Pretraining and Domain-Specific Adaptation, we investigate the necessity of domain-specific adaptation (DSA) and the effectiveness of domain-specific prediction (DSP) within DSA. Specifically, ‘w/o DSA’ refers to training the domain-specific experts directly using DSP without adapting from the base model, whereas ‘DSA w/o DSP’ denotes adapting from the base model using only single-domain sequence data without employing DSP. The results indicate that DSA coupled with DSP consistently yields the best performance, highlighting the critical role of this combination. For the second stage, Adaptive Contrastive Decoding, we examine the contributions of source-domain experts (Amateur) and the two-level contrastive design. Here, ‘w/o SDE’ removes the source-domain experts during decoding, relying solely on the target-domain expert’s prediction distribution for cross-domain item transformation. ‘w/o GCS’ and ‘w/o LCS’ correspond to using only the local or global contrastive scores, respectively. Across all these variants, performance consistently drops compared to the full Taesar model, demonstrating the importance of both the source-domain knowledge and the two-level contrastive mechanism. Overall, these ablation results clearly illustrate that each architectural component of Taesar—from DSA with DSP to adaptive contrastive decoding—is essential for achieving optimal cross-domain sequential recommendation performance. 5.3.2. Is domain-shared information successfully detected and effectively transferred? Figure 4. Schematic illustration of the adaptive contrastive decoding phase: (a) selection of transferable items from source domain S (Electronics), and (b) distribution-level mapping to target domain T (Books). To assess whether Taesar effectively identifies and transfers domain-shared information, we analyze the item-level mappings produced during adaptive contrastive decoding. Specifically, we extract the mapping dictionary from source items to their corresponding target items and visualize the learned item embeddings from both domains. As shown in Figure 4, blue points represent source-domain items, with solid blue dots indicating items selected for transformation into target-domain counterparts, corresponding to solid red dots. The visualization reveals that these selected source items have embeddings closely aligned with those of the target items, suggesting that they capture transferable, domain-shared semantics. This alignment demonstrates that Taesar successfully identifies and transforms transferable items, highlighting its effectiveness in utilizing cross-domain knowledge. Figure 5. Comparison of data statistics between the original and regenerated Books domain dataset: (a) item frequency, (b) long-tail coverage, and (c) sequence length distribution. 5.3.3. How regenerated dataset enhances diversity and context? Analysis of the regenerated dataset in the Books domain reveals that our data-centric regeneration substantially improves target-domain distributions and sequence context, explaining the observed performance gains. Compared with the original dataset, regenerated dataset exhibits a flatter item frequency curve (Figure 5 (a)), activating more low-frequency items and mitigating head dominance. Long-tail coverage is also enhanced (Figure 5 (b)), with reduced Top 10% and increased Mid 40% and Tail 50% proportions, indicating the injection of diverse, previously underrepresented interactions. Moreover, user sequences in the regenerated dataset are longer and denser (Figure 5 (c)), providing richer temporal context and enabling models to capture extended sequential patterns. Collectively, these improvements in item-level diversity and sequence-level context highlight how data-centric regeneration enables robust and interpretable performance gains. 6. Conclusion In this work, we present Taesar, a novel data-centric framework for cross-domain sequential recommendation. By prioritizing data transformation over complex transfer architecture designs, Taesar effectively mitigates negative transfer through target-aligned sequence regeneration. Specifically, it leverages adaptive contrastive decoding to identify transferable items from source domains and transform them into semantically aligned target-domain items, preserving temporal dynamics while eliminating potentially detrimental inter-domain information. Extensive experiments across multiple datasets and model architectures demonstrate that Taesar not only attains SOTA performance in cross-domain recommendation but also generalizes seamlessly to existing single-domain sequential recommenders without requiring modifications to their architectures. These findings highlight the synergy between data-centric and model-centric paradigms, providing a practical, theoretically grounded framework for cross-domain knowledge transfer. Despite these advances, important limitation remain. A deeper theoretical understanding of the limits of information preservation during the regeneration process is needed. Future work will focus on addressing above challenge, advancing both the theoretical foundations and practical applicability of data-centric recommendation. References K. Bao, J. Zhang, W. Wang, Y. Zhang, Z. Yang, Y. Luo, C. Chen, F. Feng, and Q. Tian (2025) A bi-step grounding paradigm for large language models in recommendation systems. ACM Transactions on Recommender Systems 3 (4), p. 1–27. Cited by: §1. Q. Bian, M. de Carvalho, T. Li, J. Xu, H. Fang, and Y. Ke (2025) ABXI: invariant interest adaptation for task-guided cross-domain sequential recommendation. In Proceedings of the ACM on Web Conference 2025, p. 3183–3192. Cited by: §1, §2, 2nd item, §5.1.1. J. Cao, X. Cong, J. Sheng, T. Liu, and B. Wang (2022) Contrastive cross-domain sequential recommendation. In Proceedings of the 31st ACM international conference on information & knowledge management, p. 138–147. Cited by: §1, §1, §2. H. Chang, N. Peng, M. Bansal, A. Ramakrishna, and T. Chung (2024) Explaining and improving contrastive decoding by extrapolating the probabilities of a huge and hypothetical lm. arXiv preprint arXiv:2411.01610. Cited by: §2. S. Chen, Z. Xu, W. Pan, Q. Yang, and Z. Ming (2024a) A survey on cross-domain sequential recommendation. arXiv preprint arXiv:2401.04971. Cited by: §1. S. Chen, Z. Xu, W. Pan, Q. Yang, and Z. Ming (2024b) A survey on cross-domain sequential recommendation. arXiv preprint arXiv:2401.04971. Cited by: §1. L. Chong, X. Liu, R. Zheng, L. Zhang, X. Liang, J. Li, L. Wu, M. Zhang, and L. Lin (2023) Ct4rec: simple yet effective consistency training for sequential recommendation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, p. 3901–3913. Cited by: §1. Y. Dang, Y. Liu, E. Yang, G. Guo, L. Jiang, X. Wang, and J. Zhao (2024a) Repeated padding for sequential recommendation. In Proceedings of the 18th ACM Conference on Recommender Systems, p. 497–506. Cited by: §2. Y. Dang, Y. Liu, E. Yang, G. Guo, L. Jiang, X. Wang, and J. Zhao (2024b) Repeated padding for sequential recommendation. In Proceedings of the 18th ACM Conference on Recommender Systems, p. 497–506. Cited by: §2. L. Fieback, N. Balar, J. Spiegelberg, and H. Gottschalk (2025) Efficient contrastive decoding with probabilistic hallucination detection-mitigating hallucinations in large vision language models. arXiv preprint arXiv:2504.12137. Cited by: §2. L. Guo, L. Tang, T. Chen, L. Zhu, Q. V. H. Nguyen, and H. Yin (2021) DA-gcn: a domain-aware attentive graph convolution network for shared-account cross-domain sequential recommendation. arXiv preprint arXiv:2105.03300. Cited by: §2. B. Hidasi, A. Karatzoglou, L. Baltrunas, and D. Tikk (2015) Session-based recommendations with recurrent neural networks. arXiv preprint arXiv:1511.06939. Cited by: 1st item. Y. Hou, S. Mu, W. X. Zhao, Y. Li, B. Ding, and J. Wen (2022a) Towards universal sequence representation learning for recommender systems. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, p. 585–593. Cited by: §2. Y. Hou, S. Mu, W. X. Zhao, Y. Li, B. Ding, and J. Wen (2022b) Towards universal sequence representation learning for recommender systems. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, p. 585–593. Cited by: §1, §2. L. Huimin, M. Isonuma, J. Mori, and I. Sakata (2025) Unidetox: universal detoxification of large language models via dataset distillation. In The Thirteenth International Conference on Learning Representations, Cited by: §2. W. Kang and J. McAuley (2018) Self-attentive sequential recommendation. In 2018 IEEE international conference on data mining (ICDM), p. 197–206. Cited by: 3rd item, 1st item. K. Kim, D. Hyun, S. Yun, and C. Park (2023) Melt: mutual enhancement of long-tailed user and item for sequential recommendation. In Proceedings of the 46th international ACM SIGIR conference on Research and development in information retrieval, p. 68–77. Cited by: §1, §2. W. Krichene and S. Rendle (2020) On sampled metrics for item recommendation. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, p. 1748–1757. Cited by: §5.1.3. C. Li, M. Zhao, H. Zhang, C. Yu, L. Cheng, G. Shu, B. Kong, and D. Niu (2022a) RecGURU: adversarial learning of generalized user representations for cross-domain recommendation. In Proceedings of the fifteenth ACM international conference on web search and data mining, p. 571–581. Cited by: §2. X. L. Li, A. Holtzman, D. Fried, P. Liang, J. Eisner, T. Hashimoto, L. Zettlemoyer, and M. Lewis (2022b) Contrastive decoding: open-ended text generation as optimization. arXiv preprint arXiv:2210.15097. Cited by: §2. G. Lin, C. Gao, Y. Zheng, J. Chang, Y. Niu, Y. Song, K. Gai, Z. Li, D. Jin, Y. Li, et al. (2024) Mixed attention network for cross-domain sequential recommendation. In Proceedings of the 17th ACM international conference on web search and data mining, p. 405–413. Cited by: §2. Q. Liu, F. Yan, X. Zhao, Z. Du, H. Guo, R. Tang, and F. Tian (2023a) Diffusion augmentation for sequential recommendation. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, p. 1576–1586. Cited by: §1, §2. W. Liu, X. Zheng, C. Chen, J. Su, X. Liao, M. Hu, and Y. Tan (2023b) Joint internal multi-interest exploration and external domain alignment for cross domain sequential recommendation. In Proceedings of the ACM web conference 2023, p. 383–394. Cited by: §2. Z. Liu, Z. Fan, Y. Wang, and P. S. Yu (2021) Augmenting sequential recommendation with pseudo-prior items via reversely pre-training transformer. In Proceedings of the 44th international ACM SIGIR conference on Research and development in information retrieval, p. 1608–1612. Cited by: §2. H. Ma, R. Xie, L. Meng, X. Chen, X. Zhang, L. Lin, and J. Zhou (2024) Triple sequence learning for cross-domain recommendation. ACM Transactions on Information Systems 42 (4), p. 1–29. Cited by: §2. M. Ma, P. Ren, Z. Chen, Z. Ren, L. Zhao, P. Liu, J. Ma, and M. de Rijke (2022) Mixed information flow for cross-domain sequential recommendations. ACM Transactions on Knowledge Discovery from Data (TKDD) 16 (4), p. 1–32. Cited by: §2. M. Ma, P. Ren, Y. Lin, Z. Chen, J. Ma, and M. d. Rijke (2019) π-Net: a parallel information-sharing network for shared-account cross-domain sequential recommendations. In Proceedings of the 42nd international ACM SIGIR conference on research and development in information retrieval, p. 685–694. Cited by: §2. S. O’Brien and M. Lewis (2023) Contrastive decoding improves reasoning in large language models. arXiv preprint arXiv:2309.09117. Cited by: §2. C. Park, T. Kim, T. Choi, J. Hong, Y. Yu, M. Cho, K. Lee, S. Ryu, H. Yoon, M. Choi, et al. (2023a) Cracking the code of negative transfer: a cooperative game theoretic approach for cross-domain sequential recommendation. In Proceedings of the 32nd ACM international conference on information and knowledge management, p. 2024–2033. Cited by: 3rd item. C. Park, T. Kim, T. Choi, J. Hong, Y. Yu, M. Cho, K. Lee, S. Ryu, H. Yoon, M. Choi, et al. (2023b) Cracking the code of negative transfer: a cooperative game theoretic approach for cross-domain sequential recommendation. In Proceedings of the 32nd ACM international conference on information and knowledge management, p. 2024–2033. Cited by: §1. C. Park, T. Kim, H. Yoon, J. Hong, Y. Yu, M. Cho, M. Choi, and J. Choo (2024) Pacer and runner: cooperative learning framework between single-and cross-domain sequential recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, p. 2071–2080. Cited by: §1, §1, §2, 4th item, §5.1.1. A. Petrov and C. Macdonald (2024) RSS: effective and efficient training for sequential recommendation using recency sampling. ACM Transactions on Recommender Systems 3 (1), p. 1–32. Cited by: §1. N. Sachdeva, M. Dhaliwal, C. Wu, and J. McAuley (2022) Infinite recommendation networks: a data-centric approach. Advances in Neural Information Processing Systems 35, p. 31292–31305. Cited by: §2. N. Sachdeva, Z. He, W. Kang, J. Ni, D. Z. Cheng, and J. McAuley (2023) Farzi data: autoregressive data distillation. arXiv preprint arXiv:2310.09983. Cited by: §2. [35] T. Shen, H. Wang, C. Wu, J. Y. Chin, W. Guo, Y. Liu, H. Guo, D. Lian, R. Tang, and E. Chen P-law: predicting quantitative scaling law with entropy guidance in large recommendation models. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, Cited by: §1. W. Sun, M. Ma, P. Ren, Y. Lin, Z. Chen, Z. Ren, J. Ma, and M. De Rijke (2021) Parallel split-join networks for shared account cross-domain sequential recommendations. IEEE Transactions on Knowledge and Data Engineering 35 (4), p. 4106–4123. Cited by: §2. C. Wang, J. Sun, Z. Dong, R. Li, and R. Zhang (2023) Gradient matching for categorical data distillation in ctr prediction. In Proceedings of the 17th ACM Conference on Recommender Systems, p. 161–170. Cited by: §2. H. Wang, W. Guo, L. Zhang, J. Y. Chin, Y. Ye, H. Guo, Y. Liu, D. Lian, R. Tang, and E. Chen (2025a) Generative large recommendation models: emerging trends in llms for recommendation. In Companion Proceedings of the ACM on Web Conference 2025, p. 49–52. Cited by: §1. K. Wang, H. Wang, K. Song, W. Guo, K. Cheng, Z. Li, Y. Liu, D. Lian, and E. Chen (2025b) A universal framework for compressing embeddings in ctr prediction. In International Conference on Database Systems for Advanced Applications, p. 84–100. Cited by: §1. S. Wang, Q. Zhang, L. Hu, X. Zhang, Y. Wang, and C. Aggarwal (2022) Sequential/session-based recommendations: challenges, approaches, applications and opportunities. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, p. 3425–3428. Cited by: §1. Z. Wang, W. Wei, G. Cong, X. Li, X. Mao, and M. Qiu (2020) Global context enhanced graph neural networks for session-based recommendation. In Proceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval, p. 169–178. Cited by: 2nd item. J. Wu, Q. Liu, H. Hu, W. Fan, S. Liu, Q. Li, X. Wu, and K. Tang (2023) TF-dcon: leveraging large language models (llms) to empower training-free dataset condensation for content-based recommendation. arXiv preprint arXiv:2310.09874. Cited by: §2. L. Wu, Z. Zheng, Z. Qiu, H. Wang, H. Gu, T. Shen, C. Qin, C. Zhu, H. Zhu, Q. Liu, et al. (2024) A survey on large language models for recommendation. World Wide Web 27 (5), p. 60. Cited by: §1. W. Wu, Z. Chen, X. Qiu, S. Song, X. Huang, F. Ma, and J. Xiao (2025) LLM-enhanced multimodal fusion for cross-domain sequential recommendation. arXiv preprint arXiv:2506.17966. Cited by: §1. W. Xie, H. Wang, M. Fang, R. Yu, W. Guo, Y. Liu, D. Lian, and E. Chen (2025) Breaking the bottleneck: user-specific optimization and real-time inference integration for sequential recommendation. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, p. 3333–3343. Cited by: §1. W. Xie, H. Wang, L. Zhang, R. Zhou, D. Lian, and E. Chen (2024) Breaking determinism: fuzzy modeling of sequential recommendation using discrete state space diffusion model. arXiv preprint arXiv:2410.23994. Cited by: §1. X. Xie, F. Sun, Z. Liu, S. Wu, J. Gao, J. Zhang, B. Ding, and B. Cui (2022) Contrastive learning for sequential recommendation. In 2022 IEEE 38th international conference on data engineering (ICDE), p. 1259–1273. Cited by: 4th item. X. Xu, H. Wang, W. Guo, L. Zhang, W. Yang, R. Yu, Y. Liu, D. Lian, and E. Chen (2025a) Multi-granularity interest retrieval and refinement network for long-term user behavior modeling in ctr prediction. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1, p. 2745–2755. Cited by: §1. Z. Xu, X. Chen, W. Pan, and Z. Ming (2025b) Heterogeneous graph transfer learning for category-aware cross-domain sequential recommendation. In Proceedings of the ACM on Web Conference 2025, p. 1951–1962. Cited by: §1. X. Ye, Y. Li, and L. Yao (2023) 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, p. 479–490. Cited by: §2. M. Yin, J. Pan, H. Wang, X. Wang, S. Zhang, J. Jiang, D. Lian, and E. Chen (2025) From feature interaction to feature generation: a generative paradigm of ctr prediction models. arXiv preprint arXiv:2512.14041. Cited by: §1. M. Yin, H. Wang, W. Guo, Y. Liu, S. Zhang, S. Zhao, D. Lian, and E. Chen (2024) Dataset regeneration for sequential recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, p. 3954–3965. Cited by: §2. H. Yu, Y. Wu, H. Wang, W. Guo, Y. Liu, Y. Li, Y. Ye, J. Du, and E. Chen (2025) Thought-augmented planning for llm-powered interactive recommender agent. arXiv preprint arXiv:2506.23485. Cited by: §2. T. Zang, Y. Zhu, H. Liu, R. Zhang, and J. Yu (2022) A survey on cross-domain recommendation: taxonomies, methods, and future directions. ACM Transactions on Information Systems 41 (2), p. 1–39. Cited by: §1. H. Zhang, M. Cheng, Q. Liu, J. Jiang, X. Wang, R. Zhang, C. Lei, and E. Chen (2025a) A comprehensive survey on cross-domain recommendation: taxonomy, progress, and prospects. arXiv preprint arXiv:2503.14110. Cited by: §1. J. Zhang, M. Yin, H. Wang, Y. Li, Y. Ye, X. Lou, J. Du, and E. Chen (2025b) Td3: tucker decomposition based dataset distillation method for sequential recommendation. In Proceedings of the ACM on Web Conference 2025, p. 3994–4003. Cited by: §2. D. Zhou, X. Cai, and W. Pan (2025a) Contrastive text-enhanced transformer for cross-domain sequential recommendation. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, p. 4110–4119. Cited by: §1. R. Zhou, H. Wang, W. Guo, Q. Jia, W. Xie, X. Xu, Y. Liu, D. Lian, and E. Chen (2025b) MIT: a multi-tower information transfer framework based on hierarchical task relationship modeling. In Companion Proceedings of the ACM on Web Conference 2025, p. 651–660. Cited by: §1.

Instant research discovery