FedCGR: Federated Cross-Domain Generative Recommendation

FedCGR: Federated Cross-Domain Generative Recommendation Zhuodong Liu

Hugen Lv

Xiangyu Li

Beijing Jiaotong University China 22711104@bjtu.edu.cn

Beijing Jiaotong University China 23722056@bjtu.edu.cn

Shanghai Jiao Tong University China xiangyuli@sjtu.edu.cn

Bohan Guo

Peiyu Hu∗

University of Malaya Malaysia 23114009@siswa.um.edu.my

Xi’an Jiaotong-Liverpool University China peiyuhu30@gmail.com

arXiv:2608.10929v1 [cs.AI] 11 Aug 2026

Abstract Cross-domain recommendation (CDR) transfers preference knowledge across related domains, but federated deployment makes crossdomain alignment difficult because the behavioral anchors that align item spaces, such as overlapping users and shared interaction signals, are often sparse, unavailable, or privacy-sensitive across clients. To address this tension, we revisit federated CDR as generation over a stable semantic item language. By representing items as discrete semantic ID (SID) sequences derived from public item-side metadata, cross-domain item alignment is induced by a shared vocabulary rather than by exchanging private interactions or aligning domain-specific embeddings. Directly federating SID-based generators, however, introduces two design constraints: the SID tokenizer must remain fixed to preserve cross-client token consistency, which creates a semantic-only bottleneck because local collaborative filtering (CF) signals cannot be globally shared or aligned; meanwhile, standard federated averaging can cause negative transfer under domain heterogeneity. To overcome these constraints, we propose FedCGR, a federated generative CDR framework that keeps the item language stable and makes adaptation explicit. FedCGR injects local CF evidence through a reliability-aware semantic interface and trains a prototype-personalized generator that selectively aggregates shared parameters according to domain relatedness while keeping domain-specific quantities local. Experiments on six Amazon cross-domain scenarios show that FedCGR consistently outperforms federated generative baselines and achieves competitive performance against strong sequential and federated CDR methods under both full-ranking and sampled evaluation protocols. Our code is publicly available at https://github.com/ZhuodLiu/FedCGR.

CCS Concepts • Information systems → Recommender systems; Information retrieval; • Computing methodologies → Machine learning.

Keywords Cross-Domain Recommendation, Federated Learning, Generative Recommendation, Semantic ID, Personalized Federation

∗ Corresponding author.

Figure 1: Comparison between (a) traditional federated CDR, where domain-specific representations are misaligned across clients, and (b) our approach, which aligns items through a shared SID vocabulary with personalized aggregation, while each client enriches representations with local CF signals that remain private.

1

Introduction

Cross-domain recommendation (CDR) transfers knowledge across related domains to alleviate sparsity and cold-start issues [52]. Existing CDR methods have made substantial progress through embedding mapping, cross-domain feature interaction, and transferable user representation learning, including CoNet [15], DDTCDR [26], CATN [56], RecGURU [24], and DisenCDR [32]. However, many of these methods assume centralized access to cross-domain interactions, which is difficult to satisfy when user behavior is held by different platforms or business units. Federated Learning (FL) [35] offers a natural alternative by allowing clients to collaboratively train models while keeping raw interaction sequences local. Yet federated CDR faces a fundamental tension: the behavioral anchors that make cross-domain item spaces alignable—overlapping users, co-occurrence graphs, or shared interaction signals—are precisely the signals that become sparse, unavailable, or privacy-sensitive under federated isolation.

CIKM ’26, 2026, Anonymous Review

Federated CDR methods have explored peer-to-peer collaboration [2], disentangled sequence modeling [53], graph or hypergraph modeling [41, 54], and prototype-based contrastive learning [42]. Despite these advances, most existing federated CDR models are still discriminative: they score candidate items in continuous representation spaces that are often domain-specific. As shown in Figure 1(a), cross-domain transfer therefore depends on making these private spaces comparable, creating a coupling between federated privacy and alignment effectiveness. As illustrated in Figure 1(b), SID-based generative recommendation changes what is shared across domains. Instead of forcing clients to align private continuous item spaces, it represents each item as a discrete token sequence produced by a shared tokenizer trained on public item-side metadata such as title, category, and description [37]. This effectively defines a shared item language: cross-domain item alignment becomes a property of the shared discrete vocabulary rather than an additional objective built on private behavioral data. Centralized generative CDR has already demonstrated the promise of unified semantic tokenization [17], but centralized interaction access is incompatible with the federated setting. However, a shared SID vocabulary is not a plug-and-play solution for federated CDR. Directly federating SID-based generators introduces two design constraints. First, the SID tokenizer must remain fixed to preserve cross-client token consistency; otherwise, the same token may refer to inconsistent semantic regions across domains. Fixing the tokenizer, however, creates a semantic-only bottleneck: domain-specific collaborative filtering (CF) signals cannot be incorporated through the tokenizer itself, and local CF embeddings trained independently on different clients are noisy, nonuniform in quality, and not coordinate-aligned across domains. Second, standard federated averaging is inadequate for autoregressive SID generation under domain heterogeneity. Since SID prediction decomposes an item into multiple sequential token decisions, perturbations in early token distributions can cascade through the prefix trie and cause negative transfer. Therefore, federated generative recommendation requires not only a shared token space, but also reliability-aware behavior injection and similarity-aware parameter sharing. We propose FedCGR, a federated cross-domain generative recommendation framework organized around one principle: keep the item language stable and make adaptation explicit. The stable part is a fixed SID tokenizer trained from item-side metadata, providing a shared discrete vocabulary across all clients. The adaptive part is implemented at two levels. First, FedCGR builds a reliabilityaware semantic interface: each interacted item is represented by its SID token embedding plus a residual CF signal whose strength is controlled by item-level reliability and a client-local gate. Second, FedCGR trains a prototype-personalized generator: shared generator parameters are aggregated more strongly from behaviorally related domains, while domain embeddings, gates, private experts, and local CF statistics stay on each client. This design merges what could appear as many components into two mechanisms that each address one of the two constraints above: reliable local evidence on a stable semantic interface, and personalized federated generation over that interface. The main contributions of this paper are summarized as follows:

Zhuodong Liu, Hugen Lv, Xiangyu Li, Bohan Guo, and Peiyu Hu

• We formulate federated CDR as generation over a stable semantic item language, where public item-side metadata defines a shared SID vocabulary while raw interactions remain local. • We propose FedCGR, which addresses two design constraints of federated SID generation—the semantic-only bottleneck from frozen tokenization, and negative transfer from uniform aggregation under domain heterogeneity. • FedCGR introduces a reliability-aware semantic interface that injects local CF evidence through item-level confidence and client-local gating, and a prototype-personalized generator with shared-private experts that selectively transfers parameters according to domain relatedness. • Experiments on six Amazon cross-domain scenarios show that FedCGR consistently improves over federated generative baselines, is competitive with strong sequential recommenders under full-ranking, and achieves the best results among federated CDR methods under sampled evaluation. Ablations reveal a cross-over pattern where CF enrichment and personalized aggregation become the binding constraint under different relatedness regimes.

2 Preliminaries 2.1 Federated CDR We consider 𝐾 domains {D1, . . . , D𝐾 } in a cross-silo federated setting, where each client corresponds to a domain, platform, or business unit rather than an individual user device. This setting matches practical CDR scenarios in which data silos are organizationally separated and cross-domain interactions cannot be centralized. Each domain D𝑖 maintains a user set U𝑖 , an item set V𝑖 , and private interaction sequences R𝑖 . Raw user-item interactions and user sequences are not shared across domains. The goal is to train a domain-specific next-item recommender 𝑝𝜃𝑖 (𝑣𝑇 +1 | 𝑆𝑢 ) for each domain while leveraging transferable knowledge from other domains through federated parameter sharing. In standard FedAvg [35], shared parameters are aggregated by data volume: 𝑊sh(𝑡 +1) =

𝐾 ∑︁ 𝑛𝑖 (𝑡 ) Í 𝑊𝑖,sh , 𝑛 𝑗 𝑗 𝑖=1

(1)

where 𝑛𝑖 is the number of local training instances. FedCGR replaces this single global aggregate with a personalized aggregate for each domain.

2.2

SID-Based Generative Recommendation

Given a user sequence 𝑆𝑢 = (𝑣 1, . . . , 𝑣𝑇 ), SID-based generative recommendation [11, 37] represents each item 𝑣 with a discrete code sequence c𝑣 = (𝑐 𝑣1 , . . . , 𝑐 𝐿𝑣 ). The codes are produced by Residual Quantization (RQ) [22] over item-side features. At level 𝑙, the tokenizer selects the closest codeword: 𝑐 𝑙𝑣 = arg min ∥r𝑙𝑣 − z𝑙𝑗 ∥ 2, 𝑗

𝑙 𝑙 r𝑙+1 𝑣 = r𝑣 − z𝑐 𝑙 .

(2)

𝑣

The recommender predicts the next item by autoregressively generating its SID: 𝑝 (𝑣𝑇 +1 | 𝑆𝑢 ) =

𝐿 Ö 𝑙=1

𝑝𝜃 (𝑐𝑇𝑙 +1 | 𝑆𝑢 , 𝑐𝑇1:𝑙+1−1 ).

(3)

FedCGR: Federated Cross-Domain Generative Recommendation

CIKM ’26, 2026, Anonymous Review

During inference, generated SIDs are mapped back to valid items in the target domain through a prefix trie.

3 Methodology 3.1 Overview Figure 2 illustrates FedCGR. The central idea is to use a fixed SID space as the communication interface among domains, and to make everything built on top of this interface adaptive to reliability and domain relatedness. Item-side metadata is mapped to SIDs, which provide a shared discrete vocabulary across clients without using private interactions. Because this vocabulary must remain stable during federated training, FedCGR does not update the SID tokenizer after item-side pre-training. Instead, each client enriches SID token representations with local CF evidence extracted from its own interaction sequences. Since CF embeddings are trained independently on different clients, their coordinate systems are not comparable; therefore FedCGR keeps each client’s CF adapter and dense auxiliary head local rather than aggregating them. The residual CF injection is reliability-aware: item frequency estimates whether the local CF embedding is trustworthy, and a client-local gate calibrates the domain-level CF contribution. The resulting sequence representation is consumed by an autoregressive SID generator whose shared parameters are aggregated more strongly from behaviorally related domains via domain prototypes, while domain-specific quantities remain on each client. Thus, the method has two coupled parts rather than many independent modules: a reliability-aware semantic interface for representing user histories, and a prototype-personalized generator for federated transfer. We note that the frozen RQ-VAE tokenizer only determines item-to-SID assignments; the SID token embedding tables, Transformer parameters, prediction heads, and expert parameters are still trainable and can be optimized and aggregated without changing the SID assigned to any item. Privacy scope. FedCGR follows the standard cross-silo FL setting in which raw user sequences and local item-level CF statistics are not shared. Only shared generator parameters and lowdimensional domain prototypes are uploaded; domain embeddings, CF gates, CF adapters, dense heads, private experts, CF embeddings, and confidence scores remain on each client (Table 1). We do not claim formal differential privacy; secure aggregation and differential privacy are orthogonal mechanisms that can be integrated with FedCGR.

3.2

Reliability-Aware Semantic Interface

Stable SID vocabulary. For each item 𝑣, a frozen pre-trained text encoder, such as BGE (BAAI General Embedding) [3], maps public item-side metadata into e𝑣 ∈ R𝑑𝑒 . An RQ-VAE tokenizer quantizes e𝑣 into c𝑣 = (𝑐 𝑣1 , . . . , 𝑐 𝐿𝑣 ) and is trained with reconstruction, commitment, code utilization, and collision regularization: Lrqvae = Lrecon + 𝛽Lcommit + 𝜆𝑒 Lentropy + 𝜆𝑐 Lcollision .

(4)

After training, the tokenizer is fixed. This preserves cross-client token consistency: the same SID token always refers to the same semantic region throughout federated training. Fixing the tokenizer does not freeze the downstream SID token embedding tables used by the recommender.

In implementation, the target-domain prefix trie stores valid item SIDs. If multiple items in the same domain share a SID, the corresponding trie leaf keeps all associated items. For a generated SID Í𝐿 c, its sequence-level score is computed as 𝑠 sid (c) = 𝑙=1 log 𝑝 (𝑐 𝑙 | 1:𝑙 −1 𝑆𝑢 , 𝑐 ). All items sharing that leaf receive the same SID score; ties are broken by validation-set item popularity within the target domain. The collision regularization term Lcollision in the RQ-VAE objective encourages unique SID assignments, so collisions are rare in practice. Local CF evidence. The fixed SID tokenizer captures item-side semantics but does not encode domain-specific collaborative cooccurrence. Therefore, each client D𝑖 trains a lightweight sequential model (SASRec [21]; alternatives such as BERT4Rec [39] are also applicable) locally using only its private interaction sequences R𝑖 , and extracts a local CF item embedding table: 𝑑 cf {ecf | 𝑣 ∈ V𝑖 }. 𝑣,𝑖 ∈ R

(5)

The CF embeddings are kept local and fixed during federated generator training. Since CF embeddings trained independently on different clients are not guaranteed to be coordinate-aligned, FedCGR does not aggregate the CF adapter across clients. Instead, each client learns a local behavior adapter 𝐴𝑖 (·), analogous to parameterefficient adaptation modules [14], that maps its own CF embedding space into the hidden space of the SID generator:    cf cf cf 𝐴𝑖 (ecf . (6) 𝑣,𝑖 ) = Dropout LayerNorm 𝑊𝑖 e𝑣,𝑖 + b𝑖 The adapter parameters are client-local and are never uploaded. Since CF quality varies across items, we define an item-level reliability score from local interaction frequency: log(1 + 𝑛 𝑣,𝑖 ) 𝜌 𝑣,𝑖 = , (7) 𝜖 + max𝑣 ′ ∈ V𝑖 log(1 + 𝑛 𝑣 ′ ,𝑖 ) where 𝑛 𝑣,𝑖 is the interaction count of item 𝑣 in domain D𝑖 , and 𝜖 is a small constant for numerical stability. Head items with sufficient interactions obtain larger 𝜌 𝑣,𝑖 , while long-tail items receive smaller confidence. Reliability-aware residual fusion. For a history item 𝑣𝑡 in domain D𝑖 , FedCGR constructs the input representation as h𝑡 =

𝐿 ∑︁

E𝑙sid (𝑐 𝑙𝑣𝑡 ) + Epos (𝑡) + 𝜌 𝑣𝑡 ,𝑖 · 𝑔𝑖 · 𝐴𝑖 (ecf 𝑣𝑡 ,𝑖 ),

(8)

𝑙=1

where E𝑙sid is the trainable SID token embedding table at level 𝑙, Epos is the positional embedding, and 𝑔𝑖 = 𝜎 (𝛼𝑖 ) is a client-local gate. The residual form allows the model to enrich the stable SID representation with reliable local CF evidence, while suppressing noisy long-tail CF signals through 𝜌 𝑣𝑡 ,𝑖 and calibrating the domainlevel CF contribution through 𝑔𝑖 . A domain embedding e𝑖dom is prepended to the sequence: X𝑢 = [e𝑖dom ; h1 ; . . . ; h𝑇 ],

(9)

and remains client-local. Training-only dense regularization. To encourage the shared encoder memory to preserve reliable local collaborative evidence, each client uses a local dense prediction head 𝑓head,𝑖 (·) during training. Given encoder memories M1:𝑇 over item positions, the head predicts the target item’s local CF embedding:

ẑ𝑛 = norm 𝑓head,𝑖 MeanPool(M𝑛,1:𝑇 ) , (10)

CIKM ’26, 2026, Anonymous Review

Zhuodong Liu, Hugen Lv, Xiangyu Li, Bohan Guo, and Peiyu Hu

(a) Reliability-Aware Semantic Interface (Client i) (1) Unified SID Construction

2

4

(b) Prototype-Personalized Federated Generator (Client)

6

246+

Domain 1

…

𝑧%

𝑟!" - 2 = 𝑟# 𝑧%^

123456

Emb.

Domain 1 Encoder (Shared)

Domain 2 Original Textual Descriptions

+

Semantic CF Signal Positional ID 𝑝!,# , 𝑔# Embedding

Decoder

Encoder

Domain 2

(c) Prototype-Personalized Aggregation (Server)

Domain 2

…

Client 1

cos(𝐶' , 𝐶% )

cos(𝐶' , 𝐶& )

cos(𝐶% , 𝐶& )

Add & Norm

Cross Domain RQ-VAE

Personalized Aggregation

Shared-Private MoE

(2) Local CF Signal Extraction Domain 1

Client 3

Compute Similarity & Weights

Domain 2

Self-Attention

Client 2

𝑝!,'

Shared … Shared expert 1 expert n

𝑝!,%

Router

Domain 2

Private expert

Personalized Model Send to Local

…

𝑝!,(

Domain 2 Domain User Interactions

CF Model CF Emb. Item Reliability CF Adapter

Domain Prototype

Cloud

Local

Figure 2: The overall framework of FedCGR. A fixed RQ-VAE tokenizer maps item-side metadata into shared SIDs, forming a stable vocabulary across domains. Each client locally extracts CF evidence and reliability scores from private interactions, and injects reliable CF information into SID representations through a client-local behavior adapter. The autoregressive generator is trained federatively with prototype-personalized aggregation: shared parameters are selectively transferred according to domain relatedness, while domain-specific quantities remain local. where norm(·) denotes ℓ2 normalization. The target CF embedding is also normalized: z𝑛+ = norm(ecf ). 𝑣 + ,𝑖

(11)

𝑛

The auxiliary loss is a confidence-weighted local InfoNCE (Information Noise-Contrastive Estimation) objective: exp(ẑ𝑛⊤ sg(z𝑛+ )/𝜏) 1 ∑︁ Ldense = − 𝜌 + log Í , (12) ⊤ + |B| 𝑛∈ B 𝑣𝑛 ,𝑖 𝑚∈ B exp(ẑ𝑛 sg(z𝑚 )/𝜏) where sg(·) denotes stop-gradient and 𝜏 is the InfoNCE temperature. Positives and in-batch negatives are drawn only from the local client batch. The dense head is client-local and is discarded during inference.

3.3

Prototype-Personalized Federated Generator

The generator is trained federatively, but shared and local quantities are explicitly separated. The shared parameter set 𝑊sh includes SID token embeddings, encoder/decoder attention layers, prediction heads, shared experts, and shared routing parameters. Domain embeddings, CF gates, CF adapters, dense heads, local CF statistics, and private experts remain local. Shared-private expert block. Each encoder layer uses a sharedprivate feed-forward block based on a Mixture-of-Experts (MoE) architecture:

router computes top-𝑘 routing weights over the union of shared and private experts: ∑︁ MoE𝑖 (x) = 𝑎𝑒 (x)𝐸𝑒 (x), (15) 𝑒 ∈ T𝑘 (x)

where T𝑘 (x) denotes the selected expert set and 𝑎𝑒 (x) is the normalized routing probability. Shared experts and shared routing parameters participate in personalized aggregation, while the private expert and private routing parameters remain local. A loadbalancing loss encourages stable expert utilization: Laux = 𝑁𝑒

𝑁𝑒 ∑︁

𝑓𝑗 𝑝 𝑗 ,

(16)

𝑗=1

where 𝑁𝑒 = 𝑁𝑠 + 1, 𝑓 𝑗 is the fraction of routed tokens, and 𝑝 𝑗 is the average routing probability for expert 𝑗 on the local client. Domain prototype. At the end of communication round 𝑡, client D𝑖 computes a domain prototype from the encoder memories of a local mini-batch subset S𝑖(𝑡 ) . To prevent the prototype from being dominated by the prepended domain token, we average only over item positions: ∑︁
1 p̄𝑖(𝑡 ) = (𝑡 ) MeanPool1:𝑇 sg(M𝑢,1:𝑇 ) . (17) |S𝑖 | (𝑡 ) (𝑢,𝑆𝑢 ) ∈ S𝑖

H′ = LayerNorm(H + MultiHeadAttn(H)),

(13)

H′′ = LayerNorm(H′ + MoE-FFN𝑖 (H′ )).

(14)

The prototype is updated with an exponential moving average (EMA) and normalized:   p𝑖(𝑡 ) = norm 𝛽p𝑖(𝑡 −1) + (1 − 𝛽) p̄𝑖(𝑡 ) . (18)

sh } and one clientThe block contains 𝑁𝑠 shared experts {𝐸 1sh, . . . , 𝐸 𝑁 𝑠 pri local private expert 𝐸𝑖 . For an input token representation x, the

The prototype is a compact domain-level summary used only for aggregation weighting.

FedCGR: Federated Cross-Domain Generative Recommendation

CIKM ’26, 2026, Anonymous Review

Prototype-personalized aggregation. After receiving shared (𝑡 +1) 𝐾 +1) 𝐾 } 𝑗=1 from clients, parameters {𝑊 𝑗,sh } 𝑗=1 and prototypes {p (𝑡 𝑗 the server constructs a personalized aggregate for each target domain: 𝐾 ∑︁ (𝑡 +1) (𝑡 +1) 𝑊¯ 𝑖,sh = 𝜔𝑖(𝑡𝑗 +1)𝑊 𝑗,sh , (19) 𝑗=1

Table 1: Parameter and statistic partition in FedCGR. The SID tokenizer is fixed after item-side pre-training; only shared generator parameters participate in federated aggregation. Category

(20)

Here 𝑛 𝑗 is the number of local training instances on client D 𝑗 , and 𝜏𝑎 controls the sharpness of similarity-aware aggregation. When 𝜏𝑎 → ∞, the similarity term becomes uniform and the aggregation reduces to data-volume-weighted FedAvg.

3.4

Autoregressive SID Prediction and Training

At SID level 𝑙, the decoder input is ( esos, 𝑙 = 1, 𝑙 q = 𝑙 −1 (𝑐 𝑙 −1 )]), 𝑙 > 1, 𝑓ctx ([E1sid (𝑐 1 ); . . . ; E𝑙sid

(21)

where esos is a learnable start token. After adding a level embedding e𝑙lvl , the Transformer decoder attends to the encoder memory M and predicts the next token: 𝑙 𝑝 (𝑐 𝑙 | 𝑆𝑢 , 𝑐 1:𝑙 −1 ) = softmax(𝑊head q̂𝑙 + b𝑙head ).

(22)

The SID generation loss is 𝐿

Lsid = −

1 ∑︁ log 𝑝𝜃 (𝑐𝑇𝑙 +1 | 𝑆𝑢 , 𝑐𝑇1:𝑙+1−1 ). 𝐿

(23)

𝑙=1

For client D𝑖 , the local training objective is 2 𝜇 (𝑡 ) L𝑖 = Lsid + 𝜆𝑑 Ldense + 𝜆𝑎 Laux + 𝑊𝑖,sh − 𝑊¯ 𝑖,sh . (24) 2 2 The proximal term is applied only to shared generator parameters. Client-local quantities, including domain embeddings, CF gates, CF adapters, dense heads, private experts, private routing parameters, CF embeddings, and confidence scores, are optimized or maintained locally and are not uploaded. Only shared generator parameters and domain prototypes are uploaded. Raw user sequences, local CF embeddings, confidence scores, gates, domain embeddings, CF adapters, dense heads, private experts, and private routing parameters remain local. At inference time, FedCGR builds a prefix trie over valid item SIDs in the target domain and applies domain-constrained beam search. The dense auxiliary head is not used during inference.

4

Experiments

In this section, we conduct extensive experiments to evaluate FedCGR. The experiments are designed to answer the following Research Questions (RQs): • (Effectiveness) RQ1: How does FedCGR perform against single-domain, federated discriminative CDR, and federated generative baselines? • (Ablation) RQ2: What is the contribution of each key component, including reliability-aware fusion and prototypepersonalized aggregation?

Comm.

Frozen tok. RQ-VAE encoder/codebooks, item-to-SID map Not updated

where +1) 𝑛 𝑗 exp(cos(p𝑖(𝑡 +1) , p (𝑡 )/𝜏𝑎 ) 𝑗 𝜔𝑖(𝑡𝑗 +1) = Í . (𝑡 +1) (𝑡 +1) 𝐾 , p𝑘 )/𝜏𝑎 ) 𝑘=1 𝑛𝑘 exp(cos(p𝑖

Parameters / Quantities

Federated shared

Encoder/decoder attention SID token embeddings Prediction heads Shared experts Shared router params

Pers. agg. Pers. agg. Pers. agg. Pers. agg. Pers. agg.

Domain emb. e𝑖dom CF gate 𝑔𝑖 / logit 𝛼𝑖 CF adapter 𝐴𝑖 Client-local Dense head 𝑓head,𝑖 Private expert Private router params Local stats

Not uploaded Not uploaded Not uploaded Train-only Not uploaded Not uploaded

CF embeddings ecf 𝑣,𝑖 Confidence scores 𝜌 𝑣,𝑖

Kept local Kept local

Algorithm 1 FedCGR Federated Training Require: Domains {D1, . . . , D𝐾 }, item metadata, federated rounds 𝑇fed , local epochs 𝐸 Ensure: Personalized generative recommender for each domain 1: Train and fix the RQ-VAE tokenizer from item-side metadata; assign SIDs to all items 2: for each domain D𝑖 in parallel do 3: Train a local SASRec model and extract {ecf 𝑣,𝑖 } 4: Compute local confidence scores {𝜌 𝑣,𝑖 } 5: end for (0) 6: Initialize shared generator parameters 𝑊sh 7: Initialize client-local parameters ¯ (0) ← 𝑊 (0) for all clients 8: Set 𝑊 𝑖,sh sh 9: for 𝑡 = 0, 1, . . . ,𝑇fed − 1 do (𝑡 ) 10: Server sends 𝑊¯ 𝑖,sh to each client D𝑖 11: for each domain D𝑖 in parallel do 12: Initialize local shared parameters from 𝑊¯ (𝑡 ) 𝑖,sh

13: 14: 15: 16: 17: 18: 19: 20: 21: 22: 23: 24:

for local epoch = 1, . . . , 𝐸 do for each local batch do Construct reliability-aware inputs by Eq. 8 Encode user history and autoregressively decode the target SID Update shared and local parameters using Eq. 24 end for end for Compute prototype p𝑖(𝑡 +1) (𝑡 +1) Upload shared parameters 𝑊𝑖,sh and prototype p𝑖(𝑡 +1) end for Server computes 𝜔𝑖(𝑡𝑗 +1) by Eq. 20 Server forms 𝑊¯ (𝑡 +1) by Eq. 19 𝑖,sh

25: end for

• (Cold-Start) RQ3: Does FedCGR improve recommendation for users with sparse interaction histories?

CIKM ’26, 2026, Anonymous Review

Zhuodong Liu, Hugen Lv, Xiangyu Li, Bohan Guo, and Peiyu Hu

Table 2: Domain statistics after 5-core filtering (upper) and cross-domain scenario composition (lower). Overlap denotes the ratio of users appearing in all domains of a scenario to the total unique users across those domains, computed after independent per-domain 5-core filtering. Domain

#Users

#Items

#Inter.

Avg Len

Density

Food Kitchen Grocery Beauty Sports

9,215 7,536 6,423 8,159 7,102

3,742 4,116 2,911 4,727 5,166

98,866 78,589 63,812 82,438 68,747

10.73 10.43 9.93 10.10 9.68

0.29% 0.25% 0.34% 0.21% 0.19%

Scenario

Domains

FK GB GS FKB GBS GKBS

Food, Kitchen Grocery, Beauty Grocery, Sports Food, Kitchen, Beauty Grocery, Beauty, Sports Grocery, Kitchen, Beauty, Sports

Overlap 4.2% 2.1% 0.8% 1.5% 0.6% 0.3%

• (Sensitivity) RQ4: How sensitive is FedCGR to key hyperparameters, and what do the learned reliability signals reveal? • (Efficiency) RQ5: What is the communication and training overhead of FedCGR?

4.1

Experimental Setup

4.1.1 Datasets. We evaluate FedCGR on cross-domain scenarios constructed from Amazon Review datasets [36]. We select five product domains—Grocery, Beauty, Sports, Food, and Kitchen— and compose six scenarios that span different domain counts and relatedness levels. Two-domain scenarios. FK (Food–Kitchen) pairs two household consumable domains with strong semantic overlap, representing high-affinity transfer; GB (Grocery–Beauty) combines two moderately related categories, representing medium-affinity transfer; GS (Grocery–Sports) combines two largely unrelated categories, representing low-affinity transfer. These three endpoints allow us to measure how domain relatedness affects federated generative recommendation. Multi-domain scenarios. FKB (Food–Kitchen–Beauty) mixes three domains where two are closely related (Food and Kitchen) while the third (Beauty) is more distant, testing whether prototypepersonalized aggregation can selectively transfer among unevenly related clients. GBS (Grocery–Beauty–Sports) mixes three domains of varying relatedness within a single federation, further testing selective transfer. GKBS (Grocery–Kitchen–Beauty–Sports) increases heterogeneity to four domains and tests scalability. For each domain, we apply 5-core filtering and split each user sequence by leave-one-out evaluation. Table 2 summarizes the statistics. 4.1.2 Baselines. Because full-ranking and sampled evaluation require different candidate-generation interfaces, and because several federated CDR repositories report results only under the 999negative protocol, we organize baselines by evaluation protocol.

Full-ranking baselines (Table 3). (1) Single-domain sequential recommenders (GRU4Rec [10], SASRec [21]) train each domain independently without cross-domain transfer, serving as a lower reference for the value of federation. (2) Single-domain generative recommender (TIGER [37]) uses SID-based autoregressive generation but trains per domain, isolating the benefit of semantic tokenization from federation. (3) Federated generative baselines (TIGER+FedAvg, TIGER+FedProx) directly federate TIGER with standard aggregation strategies, testing whether naïve federation already helps generative models. 999-neg baselines (Table 4). (4) Local sequential recommenders (LocalSASRec, LocalCL4SRec, LocalDuoRec) train contrastive or dual-objective sequential models independently per domain. (5) Federated discriminative CDR methods (FedSASRec, FedCL4SRec, FedDCSR [53]) federate discriminative sequential recommenders with varying degrees of disentanglement, following their original sampled evaluation protocol for direct comparison. 4.1.3 Evaluation Protocol. For each user sequence, the last item is used for testing, the second-to-last item for validation, and the remaining interactions for training. We report Hit Rate at rank 10 (H@10) and Normalized Discounted Cumulative Gain at rank 10 (NDCG@10, abbreviated N@10). Our primary evaluation is alluser full-ranking, where each test user is evaluated against the entire item catalogue of the target domain. For compatibility with prior federated CDR repositories [53, 54], we additionally conduct sampled evaluation with 999 negatives: each test user ranks the ground-truth item among 999 randomly sampled negatives. The negative set is sampled once per user and shared across all methods to ensure fair comparison; we use random seed 42 and report results averaged over three independent runs. 4.1.4 Implementation Details. The SID tokenizer uses BGE-baseen-v1.5 as the text encoder (𝑑𝑒 = 768), 𝐿 = 3 quantization levels, and codebook size 𝐶 = 256. The local CF extractor is SASRec with 𝑑 cf = 128, 2 layers, 2 heads, and 15 training epochs per domain. After local pre-training, the CF item embeddings are fixed and kept on the corresponding client. The federated generator uses 𝑑 model = 256, 4 attention heads, 𝑁 enc = 𝑁 dec = 3, dropout 0.15, and maximum sequence length 50. The shared-private expert block uses 𝑁𝑠 = 2 shared experts and one private expert per client, with expert hidden size 512 and top-𝑘 = 2 routing. Shared experts and shared routing parameters participate in personalized aggregation, while private experts and private routing parameters remain local. Each client uses a local CF adapter 𝐴𝑖 implemented as Linear(128 → 256)+LayerNorm+Dropout(0.1), and a local training-only dense head implemented as Linear(256 → 64)+GELU+Linear(64 → 128). The CF adapter and dense head are optimized locally during federated generator training but are not uploaded to the server. We set 𝜆𝑑 = 0.1, 𝜆𝑎 = 0.01, InfoNCE temperature 𝜏 = 0.07, proximal coefficient 𝜇 = 0.01, prototype momentum 𝛽 = 0.9, local epochs 𝐸 = 2, federated rounds 𝑇fed = 40, batch size 1024, learning rate 3 × 10−4 with cosine decay and 5-round warmup, label smoothing 0.1, gradient clip norm 1.0, and beam size 𝐵 = 50. The aggregation temperature 𝜏𝑎 is tuned on validation data. All experiments are repeated with three random seeds (42, 123, 456); standard deviations are below 0.003 on all main metrics and are omitted for readability. All experiments are conducted on NVIDIA RTX 4090 GPUs.

FedCGR: Federated Cross-Domain Generative Recommendation

4.2

Overall Performance (RQ1)

We evaluate FedCGR from two complementary perspectives. Table 3 reports per-domain results under full-ranking evaluation, where each test user is evaluated against the entire item catalogue of the target domain. Table 4 adopts the 999-neg sampled protocol used by prior federated CDR work [53, 54], enabling direct comparison with federated discriminative baselines under their own evaluation paradigm. 4.2.1 Full-ranking analysis (Table 3). FedCGR is the strongest federated generative method. Across all 22 domain-metric cells, FedCGR consistently outperforms both TIGER+FedAvg and TIGER+ FedProx. The improvement is most pronounced on GKBS, the most heterogeneous scenario: on GKBS Grocery N@10, FedCGR achieves 0.074, improving over TIGER+FedProx (0.062, +19.4%) and TIGER+FedAvg (0.059, +25.4%); on GKBS Kitchen H@10, FedCGR reaches 0.079 versus TIGER+FedProx (0.071) and TIGER+FedAvg (0.056). These gains confirm that prototype-personalized aggregation and reliability-aware CF residuals provide consistent benefits over naïve federation strategies. Federation progressively improves generative models. A clear progression from TIGER → TIGER+FedAvg → TIGER+FedProx → FedCGR is visible in most domains. On GKBS Kitchen H@10: 0.051 → 0.056 → 0.071 → 0.079; on GS Grocery H@10: 0.145 → 0.159 → 0.167 → 0.175; on FK Kitchen H@10: 0.033 → 0.033 → 0.034 → 0.061. This demonstrates that (1) cross-domain knowledge transfer benefits SID-based generation, (2) proximal regularization helps but is insufficient under domain heterogeneity, and (3) the reliability-aware interface and personalized aggregation of FedCGR provide further substantial gains. FedCGR is competitive with non-federated discriminative baselines. GRU4Rec and SASRec train independently with full access to local interactions. FedCGR achieves the best result in the majority of domain-metric cells, including all eight cells in GKBS, all four in FK, and both domains in GS. On FK Kitchen, FedCGR (H@10: 0.061) surpasses both SASRec (0.055) and GRU4Rec (0.053), despite TIGER and its federated variants falling well below discriminative methods on this domain. In domains with dense interactions and strong collaborative signals, such as GBS Beauty (SASRec H@10: 0.191 vs. FedCGR 0.162), discriminative methods retain an advantage, as embedding-based scoring benefits directly from rich interaction data. Crucially, FedCGR achieves this competitive performance while keeping all raw user interactions private—a constraint that non-federated methods do not face. 4.2.2 999-neg analysis (Table 4). Under the sampled evaluation protocol adopted by prior federated CDR methods, FedCGR achieves the best results across all six scenarios on both metrics. FedCGR consistently outperforms all federated baselines. Compared with FedDCSR, the strongest baseline, FedCGR improves H@10 from 0.180 to 0.186 (+3.3%) on GKBS and from 0.170 to 0.208 (+22.4%) on GS. The improvement is largest on GS, where the low behavioral affinity between Grocery and Sports makes uniform aggregation particularly harmful. Naïve federation causes negative transfer. FedSASRec and FedCL4SRec consistently underperform their non-federated counterparts: on GKBS, H@10 drops from 0.144 (LocalSASRec) to 0.129

CIKM ’26, 2026, Anonymous Review

(FedSASRec) after federation. This confirms that negative transfer is a general challenge in federated CDR, not specific to generative models. FedDCSR partially addresses this through disentangled representations, but FedCGR achieves further gains by combining a unified SID vocabulary with prototype-personalized aggregation, yielding consistent improvements across scenarios of varying size and domain affinity.

4.3

Ablation Study (RQ2)

We ablate the two design principles of FedCGR: the reliability-aware semantic interface and the prototype-personalized generator. We select FK (high-affinity, 2-domain) and GKBS (high-heterogeneity, 4-domain) to examine component contributions under contrasting conditions. Table 5 reports N@10 (scenario-level macro-average). Reliability-aware semantic interface. Removing the CF residual entirely (SID only) causes the largest drop in this group (−12.8% on average), confirming that local collaborative signals provide complementary information beyond item-side semantics. Removing item confidence 𝜌 𝑣 or the domain gate 𝑔𝑖 weakens the model’s ability to distinguish reliable from noisy CF signals: without 𝜌 𝑣 , long-tail items inject unreliable embeddings; without 𝑔𝑖 , the model cannot calibrate the overall CF contribution at the domain level. The dense head and confidence-weighted loss each contribute moderately (−3.7% and −4.8%), validating the auxiliary alignment mechanism that encourages the shared encoder to preserve reliable CF evidence during training. Prototype-personalized generator. We distinguish two aggregation ablations. FedAvg aggregation (†) replaces both personalized weighting and the shared-private split with a single data-volumeweighted global model, averaging all experts across clients. In contrast, w/o prototype weights (‡) keeps the shared-private MoE structure with private experts remaining local, but sets similarity weights uniform (𝜏𝑎 → ∞), so only shared parameters are averaged by data volume. Replacing the entire personalized pipeline with standard FedAvg causes a substantial drop (−11.2%). The degradation is more pronounced on GKBS (0.079 vs. 0.092, −14.1%) than on FK (0.044 vs. 0.048, −8.3%), consistent with the expectation that personalized aggregation matters more when domains are behaviorally diverse. Keeping the shared-private MoE structure but removing only the cosine similarity weights (w/o prototype weights) recovers part of the loss (GKBS: 0.081 vs. 0.079), confirming that similarity-aware aggregation and domain-specific capacity are complementary rather than redundant. Division of labor between the two principles. The ablations reveal a cross-over pattern that clarifies the respective roles of the two design principles. On the high-affinity FK scenario, removing local CF evidence (−12.5%) is more harmful than replacing personalized aggregation with FedAvg (−8.3%), indicating that once domains already share a useful semantic language, the main bottleneck becomes recovering domain-specific collaborative behavior. On the heterogeneous GKBS scenario, the pattern reverses: FedAvg (−14.1%) is more harmful than CF removal (−13.0%), showing that negative transfer becomes the binding constraint as domain relatedness decreases. This cross-over supports the central thesis of

CIKM ’26, 2026, Anonymous Review

Zhuodong Liu, Hugen Lv, Xiangyu Li, Bohan Guo, and Peiyu Hu

Table 3: Per-domain results under full-ranking evaluation (H@10 and N@10). Best result per row is bold; second-best is underlined. Single-domain methods train independently per scenario; federated methods share parameters across domains. Due to scenario-specific data filtering (Section 4.1.1), single-domain results may vary across scenarios for the same domain. Single-Domain Discr. Metric

GRU4Rec

SASRec

TIGER

TIGER+FedAvg

TIGER+FedProx

FedCGR

Food

H@10 N@10

0.104 0.059

0.107 0.057

0.114 0.056

0.119 0.061

0.122 0.069

0.179 0.077

Kitchen

H@10 N@10

0.053 0.027

0.055 0.027

0.033 0.017

0.033 0.017

0.034 0.017

0.061 0.029

Grocery

H@10 N@10

0.144 0.081

0.164 0.092

0.145 0.089

0.159 0.095

0.167 0.095

0.175 0.104

Sports

H@10 N@10

0.198 0.110

0.105 0.081

0.201 0.079

0.204 0.081

0.208 0.089

0.216 0.119

Grocery

H@10 N@10

0.144 0.081

0.164 0.092

0.145 0.101

0.149 0.101

0.151 0.112

0.157 0.119

Beauty

H@10 N@10

0.184 0.113

0.191 0.111

0.106 0.055

0.137 0.076

0.141 0.062

0.162 0.098

Sports

H@10 N@10

0.208 0.117

0.177 0.104

0.163 0.125

0.201 0.129

0.204 0.136

0.209 0.146

Grocery

H@10 N@10

0.098 0.054

0.111 0.059

0.101 0.055

0.115 0.059

0.118 0.062

0.122 0.074

Kitchen

H@10 N@10

0.039 0.021

0.047 0.024

0.051 0.031

0.056 0.033

0.071 0.039

0.079 0.045

Beauty

H@10 N@10

0.118 0.073

0.136 0.080

0.122 0.094

0.135 0.099

0.136 0.107

0.142 0.122

Sports

H@10 N@10

0.120 0.110

0.124 0.099

0.129 0.117

0.134 0.117

0.139 0.121

0.148 0.125

GS

GKBS

FedCGR: a stable item language needs both local evidence interfaces and personalized sharing, and their relative importance is governed by the domain-relatedness regime. Combined ablation. Removing both the CF residual and personalized aggregation simultaneously yields the lowest performance (−14.9%), approaching TIGER+FedAvg in Table 3. This confirms that the two principles are complementary: the semantic interface recovers local behavioral evidence, while personalized federation reduces negative transfer, and their combination produces gains beyond what either achieves alone.

0.16 0.14

0.12 FedProx SID-only Full

0.12 0.10 0.082

0.08 0.06

0.10

0.128

NDCG@10

FK

GBS

Federated Generative

Domain

NDCG@10

Scenario

Single-Domain Gen.

0.091

0.08 0.060

0.06

0.04

0.048

FedProx SID-only Full

0.037

0.04

4.4

Cold-Start Analysis (RQ3)

Since cold-start alleviation is a central motivation for CDR, we examine how FedCGR performs across users with different activity levels. We partition test users into three groups by training sequence length: cold (≤ 7), warm (8–15), and active (≥ 16). On GB, the cold/warm/active groups account for 45%, 28%, and 27% of test users, respectively; on FKB, the proportions are 57%, 25%, and 18%. Figure 3 reports NDCG@10 on GB and FKB for FedCGR, its SIDonly ablation (w/o CF residual), and TIGER+FedProx. The results separate two sources of benefit. The first source is the stable SID vocabulary. On cold users, both FedCGR variants outperform TIGER+FedProx, confirming that the semantic item language provides a strong cold-start foundation because item

0.02

Cold

Warm

(a) GB

Active

Cold

Warm

Active

(b) FKB

Figure 3: Cold-start analysis: NDCG@10 across user activity groups on (a) GB and (b) FKB. The stable SID vocabulary provides the main cold-start gain; the reliability-aware CF residual adds an increasing benefit as user activity grows.

representations are grounded in metadata rather than interaction counts. The SID-only variant does not degrade on cold users relative to TIGER+FedProx, showing that the reliability modulation (𝜌 𝑣

FedCGR: Federated Cross-Domain Generative Recommendation

CIKM ’26, 2026, Anonymous Review

(a) λd

Table 4: Results under 999-neg sampled evaluation (scenariolevel macro-average). Best is bold; second-best is underlined.

FK

GS

GB

FKB

GBS

GKBS

LocalSASRec LocalCL4SRec LocalDuoRec

.130 .134 .139

.153 .154 .155

.175 .173 .173

.151 .155 .160

.162 .160 .168

.144 .141 .148

FedSASRec FedCL4SRec FedDCSR

.127 .127 .179

.149 .148 .170

.167 .171 .194

.142 .148 .187

.148 .152 .177

.129 .132 .180

FedCGR

.187

.208

.207

.200

.191

.186

LocalSASRec LocalCL4SRec LocalDuoRec

.077 .080 .081

.096 .095 .094

.110 .109 .109

.092 .095 .097

.102 .100 .104

.087 .086 .089

FedSASRec FedCL4SRec FedDCSR

.075 .074 .098

.092 .091 .096

.107 .110 .111

.087 .089 .104

.095 .097 .103

.079 .080 .100

FedCGR

.103

.112

.124

.112

.111

.106

NDCG@10

Method

0.08

0.10 0.09 0.08 0.07

0.0

Variant

FK

GKBS

Δ avg

FedCGR full

0.048

0.092

–

0.042 0.044 0.045 0.046 0.046

0.080 0.085 0.086 0.089 0.087

−12.8% −8.0% −6.4% −3.7% −4.8%

0.044 0.044 0.045 0.046

0.079 0.081 0.084 0.088

−11.2% −10.1% −7.5% −4.3%

0.041

0.078

−14.9%

GB

GBS

0.05

0.1

FKB

0.2

0.5

0.03

GB

FKB

0.05

0.10

(c) nshared

GKBS

0.20

∞

(d) L

0.11 0.10

0.10 0.09 0.08 0.07 0.06

GBS

1

Table 5: Ablation study (N@10, scenario-level macro-average). Δ is the relative change versus the full model, averaged over both scenarios.

0.07

0.06

0.06

NDCG@10

4-Dom.

NDCG@10

3-Domain

0.11

NDCG@10

N@10

H@10

2-Domain

(b) τa

GKBS

2

3

FKB

4

0.08 0.06 0.04

FK

2

GBS

3

GKBS

4

5

Figure 4: Sensitivity analysis: (a) NDCG@10 vs. dense loss weight 𝜆𝑑 , (b) NDCG@10 vs. aggregation temperature 𝜏𝑎 , (c) NDCG@10 vs. number of shared experts 𝑛 shared , and (d) NDCG@10 vs. number of RQ levels 𝐿. The dense loss weight and aggregation temperature exhibit a clear optimum around 0.10, while performance is relatively stable across 𝑛 shared ∈ {2, 3, 4} and 𝐿 ∈ {3, 4}.

Reliability-aware semantic interface w/o CF residual (SID only) w/o item confidence 𝜌 𝑣 w/o domain gate 𝑔𝑖 w/o dense head 𝑓head,𝑖 w/o confidence-weighted dense loss Prototype-personalized generator FedAvg aggregation† w/o prototype weights‡ w/o shared-private split (shared FFN only) w/o private expert Combined w/o CF residual + FedAvg†

and 𝑔𝑖 ) successfully suppresses noisy CF signals rather than injecting harmful noise. The second source is the reliability-aware CF residual, whose benefit grows with user activity: the gap between FedCGR full and the SID-only variant widens from cold to active users because richer interaction histories make local CF embeddings more reliable. Thus, FedCGR does not rely on CF signals for cold users; instead, it falls back to the semantic item language and progressively activates CF evidence as it becomes trustworthy.

4.5

Sensitivity and Diagnostic Analysis (RQ4)

We further analyze how FedCGR behaves under different reliability and personalization settings. 4.5.1 Hyper-parameter sensitivity. Figure 4 reports validation NDCG@10 under varying dense loss weights 𝜆𝑑 , aggregation temperatures 𝜏𝑎 , numbers of shared experts 𝑛 shared , and RQ levels 𝐿. The dense loss weight 𝜆𝑑 shows a clear optimum around 0.10; too small a weight loses the alignment benefit, while too large a weight causes the auxiliary objective to dominate the SID generation loss. The aggregation temperature 𝜏𝑎 also exhibits a moderate optimum at 0.10: very low temperatures concentrate aggregation weights on the single most similar domain and reduce diversity, while 𝜏𝑎 = ∞ collapses to data-volume averaging and suffers from negative transfer. Notably, the performance gap between 𝜏𝑎 = 0.10 and 𝜏𝑎 = ∞ is larger on the four-domain scenario GKBS (14.8% relative drop) than on GB (11.0%), confirming that personalized aggregation becomes more important as domain heterogeneity increases. The number of shared experts 𝑛 shared and the number of RQ levels 𝐿 are relatively robust: performance is stable across 𝑛 shared ∈ {2, 3, 4} and 𝐿 ∈ {3, 4}, with slight degradation at extreme values. 4.5.2 Diagnostic statistics. Table 6 summarizes the learned gates and confidence statistics after convergence.

CIKM ’26, 2026, Anonymous Review

Zhuodong Liu, Hugen Lv, Xiangyu Li, Bohan Guo, and Peiyu Hu

Table 6: Diagnostic statistics for reliability-aware fusion. 𝑔𝑖 is the learned client-local gate; 𝜌¯ is the mean item confidence. Domain

𝑔𝑖

𝜌¯

Interpretation

Food Grocery Kitchen Beauty Sports

0.67 0.61 0.58 0.52 0.41

0.42 0.38 0.35 0.31 0.25

Strong CF quality Moderate-strong CF quality Moderate CF quality Moderate CF quality Sparse / noisy CF

0.12 0.08 target = 0.07

target = 0.09

0.08

NDCG@10

NDCG@10

0.10

0.06

0.06

0.04

0.04 FedAvg FedProx FedCGR

0.02 0.00

0

250

500

750

Cumulative Communication (MB) (a) GBS

0.02

FedAvg FedProx FedCGR

0.00 0

250

500

750

Cumulative Communication (MB) (b) GKBS

Figure 5: NDCG@10 vs. cumulative upload communication on (a) GBS and (b) GKBS. Dashed lines mark reference targets. Despite higher per-round cost, FedCGR reaches each target with less total communication than both baselines. Domains with higher average interactions per item—Food (26.4), Grocery (21.9)—learn higher gate values and item confidence, allowing more CF evidence into the SID representation. Sparser domains such as Sports (13.3 avg interactions per item) learn lower gates, relying more on the SID-only representation. This confirms that the gate 𝑔𝑖 and confidence 𝜌 𝑣 learn to calibrate the domain-level CF contribution without manual tuning, consistent with the design motivation of the reliability-aware interface.

4.6

Efficiency Analysis (RQ5)

Figure 5 plots NDCG@10 against cumulative upload communication on GBS and GKBS. Per-round cost. FedCGR uploads 5.1M shared parameters (20.6 MB) per client per round, approximately 18% more than TIGER+FedAvg and TIGER+FedProx (4.3M, 17.4 MB), due to shared experts and the dense auxiliary loss. The prototype upload is negligible (<1 KB). Download communication equals the upload size for all three methods: TIGER+FedAvg and TIGER+FedProx broadcast one global model (4.3M), while FedCGR sends each client a personalized aggregate of the same shared parameter size (5.1M). Per-round training time is 4.3 min for FedCGR vs. 3.5 min for TIGER+FedAvg. The reported time excludes one-time local CF pre-training (approximately 8 minutes per domain), which is performed independently before federated generator training. Accuracy-per-communication. The relevant efficiency metric is not per-round payload alone, but how much total communication is needed to reach a given accuracy target. Although FedCGR

costs more per round, it converges faster and reaches a higher plateau. As shown in Figure 5(a), on GBS FedCGR reaches the 0.09 N@10 target at round 14 with 288 MB cumulative upload, whereas TIGER+FedAvg requires 435 MB (round 25) and TIGER+FedProx requires 365 MB (round 21)—a 34% reduction relative to FedAvg. On the more heterogeneous GKBS (Figure 5(b)), FedCGR crosses the 0.07 target at 288 MB (round 14), while TIGER+FedAvg needs 522 MB (round 30) and TIGER+FedProx needs 365 MB (round 21)—a 45% reduction relative to FedAvg. The communication savings grow with domain heterogeneity, confirming that prototype-personalized aggregation not only improves accuracy but also accelerates convergence, turning a modest per-round overhead into a net reduction in total communication budget.

5

Related Work

CDR and Federated CDR. CDR transfers knowledge across domains to alleviate sparsity and cold-start problems, from early shared-factor or mapping-based methods [34, 38] to neural transfer models such as CoNet [15], DDTCDR [26], and RecGURU [24], as well as graph-based [30, 55] and contrastive [47] cross-domain methods. Recent studies further address distribution shift and negative transfer through source filtering, causal disentanglement, and robust optimization [6, 25, 44], while GenCDR [17] introduces unified SIDs for centralized generative CDR. Federated CDR instead keeps raw interactions local [35], with existing methods exploring decentralized transfer and privacy protection [31, 46], disentangled or graph-based representation learning [2, 42, 50, 53, 54], and semantic bridging without overlapping users [8, 9, 18, 33, 58]. General personalized FL has also studied regularization-based, sharedrepresentation, and clustered aggregation strategies [1, 4, 5, 7, 48]. Unlike these mostly discriminative approaches, FedCGR studies federated CDR as autoregressive generation over shared SIDs, where item alignment is induced by a stable semantic vocabulary before federated training. Generative Recommendation with Semantic IDs. Generative recommendation formulates retrieval as sequence generation over discrete item tokens [16, 19, 27], and has been extended with multimodal item representations [45]. VQ-Rec [12] learns vectorquantized transferable item codes, and TIGER [37] establishes RQVAE-based SID generation for next-item recommendation. Later work improves SID construction through end-to-end, contrastive, hierarchical, textual, or industrial-scale alignment objectives [20, 40, 43, 51, 57], and extends the paradigm with order-agnostic identifiers, parallel SID prediction, temporal modeling, speculative decoding, and dense-generative retrieval unification [13, 23, 28, 29, 49]. GenCDR [17] applies SID-based generation to centralized CDR. In contrast, FedCGR targets federated deployment, where the tokenizer must remain fixed, CF evidence is local, and aggregation must handle domain heterogeneity; we address these constraints with reliability-aware CF injection and prototype-personalized federated generation.

6

Conclusion

This paper proposed FedCGR, which formulates federated CDR as generation over a stable semantic item language. A reliabilityaware semantic interface injects local CF evidence into fixed SID

FedCGR: Federated Cross-Domain Generative Recommendation

representations through confidence-gated residuals; a prototypepersonalized generator selectively aggregates shared parameters according to domain relatedness. Experiments on six Amazon scenarios confirm consistent improvements over federated generative baselines and competitive performance against strong sequential and federated CDR methods. Ablations reveal a cross-over pattern: CF enrichment dominates in high-affinity scenarios while personalized aggregation becomes the binding constraint under high heterogeneity. Despite an 18% higher per-round communication, faster convergence reduces the total budget by 34–45%. Future work may integrate secure aggregation, learn confidence estimates end-to-end, and extend the framework to scenarios where new domains or items arrive incrementally during federation.

CIKM ’26, 2026, Anonymous Review

CIKM ’26, 2026, Anonymous Review

Generative AI Disclosure Generative AI tools were used only for language polishing and did not contribute to the research ideas, method design, experiments, analysis, or conclusions.

References [1] Durmus Alp Emre Acar, Yue Zhao, Ruizhao Zhu, Ramon Matas, Matthew Mattina, Paul Whatmough, and Venkatesh Saligrama. 2021. Debiasing Model Updates for Improving Personalized Federated Training. In Proceedings of the 38th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 139). PMLR, 21–31. [2] Gaode Chen, Xinghua Zhang, Yijun Su, Yantong Lai, Ji Xiang, Junbo Zhang, and Yu Zheng. 2023. Win-Win: A Privacy-Preserving Federated Framework for Dual-Target Cross-Domain Recommendation. In Proceedings of the 37th AAAI Conference on Artificial Intelligence. 4149–4156. [3] Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation. In Findings of the Association for Computational Linguistics: ACL 2024. [4] Liam Collins, Hamed Hassani, Aryan Mokhtari, and Sanjay Shakkottai. 2021. Exploiting Shared Representations for Personalized Federated Learning. In Proceedings of the 38th International Conference on Machine Learning (ICML). [5] Canh T. Dinh, Nguyen H. Tran, and Tuan Dung Nguyen. 2020. Personalized Federated Learning with Moreau Envelopes. In Advances in Neural Information Processing Systems (NeurIPS). [6] Jing Du, Zesheng Ye, Bin Guo, Zhiwen Yu, and Lina Yao. 2024. Identifiability of Cross-Domain Recommendation via Causal Subspace Disentanglement. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. [7] Avishek Ghosh, Jichan Chung, Dong Yin, and Kannan Ramchandran. 2020. An Efficient Framework for Clustered Federated Learning. In Advances in Neural Information Processing Systems (NeurIPS). [8] Lei Guo, Ziang Lu, Junliang Yu, Quoc Viet Hung Nguyen, and Hongzhi Yin. 2024. Prompt-enhanced Federated Content Representation Learning for Crossdomain Recommendation. In Proceedings of the ACM Web Conference (WWW). doi:10.1145/3589334.3645337 [9] Xinrui He, Ting-Wei Li, Tianxin Wei, Xuying Ning, Xinyu He, Wenxuan Bao, Hanghang Tong, and Jingrui He. 2026. FeDecider: An LLM-Based Framework for Federated Cross-Domain Recommendation. In Proceedings of the ACM Web Conference (WWW). doi:10.1145/3774904.3792294 [10] Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. 2016. Session-based Recommendations with Recurrent Neural Networks. In Proceedings of the International Conference on Learning Representations (ICLR). [11] Min Hou, Le Wu, Yuxin Liao, Yonghui Yang, Zhen Zhang, Yu Wang, Changlong Zheng, Han Wu, and Richang Hong. 2026. A Survey on Generative Recommendation: Data, Model, and Tasks. arXiv preprint arXiv:2510.27157 (2026). [12] Yupeng Hou, Zhankui He, Julian J. McAuley, and Wayne Xin Zhao. 2023. Learning Vector-Quantized Item Representation for Transferable Sequential Recommenders. In Proceedings of the ACM Web Conference (WWW). doi:10.1145/3543507. 3583434 [13] Yupeng Hou, Jiacheng Li, Ashley Shin, Jinsung Jeon, Abhishek Santhanam, Wei Shao, Kaveh Hassani, Ning Yao, and Julian McAuley. 2025. Generating Long Semantic IDs in Parallel for Recommendation. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Vol. 2. 956–966. [14] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. arXiv preprint arXiv:2106.09685 (2022). [15] Guangneng Hu, Yu Zhang, and Qiang Yang. 2018. CoNet: Collaborative Cross Networks for Cross-Domain Recommendation. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management. 667–676. doi:10.1145/3269206.3271684 [16] Peiyu Hu, Weihai Lu, Siying Gu, Elliott Wen, Changyu Zeng, Senzhang Wang, and Jia Wang. 2026. A Modular Survey for Semantic ID-Based Generative Recommendation. Preprints (2026). [17] Peiyu Hu, Wayne Lu, and Jia Wang. 2026. From IDs to Semantics: A Generative Framework for Cross-Domain Recommendation with Adaptive Semantic Tokenization. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 14874–14882. [18] Deling Huang and Qilong Feng. 2025. Federated Cross-Domain Recommendation Framework With Graph Neural Network. Expert Systems 42, 8 (2025), e70087. [19] Jianchao Ji, Zelong Li, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, Juntao Tan, and Yongfeng Zhang. 2024. GenRec: Large Language Model for Generative Recommendation. In Proceedings of the European Conference on Information Retrieval (ECIR). 494–502.

Zhuodong Liu, Hugen Lv, Xiangyu Li, Bohan Guo, and Peiyu Hu

[20] Bowen Jin, Hansi Zeng, Guoyin Wang, Xiusi Chen, Tianxin Wei, Ruirui Li, Zhengyang Wang, Zheng Li, Yang Li, Hanqing Lu, Suhang Wang, Jiawei Han, and Xianfeng Tang. 2024. Language Models as Semantic Indexers. In Proceedings of the 41st International Conference on Machine Learning (ICML). [21] Wang-Cheng Kang and Julian J. McAuley. 2018. Self-Attentive Sequential Recommendation. In Proceedings of the IEEE International Conference on Data Mining (ICDM). 197–206. [22] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. 2022. Autoregressive Image Generation using Residual Quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 11523–11532. [23] Sunkyung Lee, Seongmin Park, Jonghyo Kim, Mincheol Yoon, and Jongwuk Lee. 2025. Enhancing Time Awareness in Generative Recommendation. In Findings of the Association for Computational Linguistics: EMNLP 2025. 23917–23933. [24] Chenglin Li, Mingjun Zhao, Huanming Zhang, Chenyun Yu, Lei Cheng, Guoqiang Shu, BeiBei Kong, and Di Niu. 2022. 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. 571–581. doi:10.1145/3488560.3498388 [25] Hanyu Li, Weizhi Ma, Peijie Sun, Jiayu Li, Cunxiang Yin, Yancheng He, Guoqiang Xu, Min Zhang, and Shaoping Ma. 2024. Aiming at the Target: Filter Collaborative Information for Cross-Domain Recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. [26] Pan Li and Alexander Tuzhilin. 2020. DDTCDR: Deep Dual Transfer Cross Domain Recommendation. In Proceedings of the 13th International Conference on Web Search and Data Mining (WSDM). 331–339. [27] Xiaopeng Li, Bo Chen, Junda She, Shiteng Cao, You Wang, Qinlin Jia, Haiying He, Zheli Zhou, Zhao Liu, Ji Liu, Zhiyang Zhang, Yu Zhou, Guoping Tang, Yiqing Yang, Chengcheng Guo, Si Dong, Kuo Cai, Pengyue Jia, Maolin Wang, Wanyu Wang, Shiyao Wang, Xinchen Luo, Qigen Hu, Qiang Luo, Xiao Lv, Chaoyi Ma, Ruiming Tang, Kun Gai, Guorui Zhou, and Xiangyu Zhao. 2025. A Survey of Generative Recommendation from a Tri-Decoupled Perspective: Tokenization, Architecture, and Optimization. Preprints (2025). doi:10.20944/preprints202512.0203.v1 [28] Xinyu Lin, Haokai Shi, Wenjie Wang, Fuli Feng, Qiang Wang, See-Kiong Ng, and Tat-Seng Chua. 2025. Order-agnostic Identifier for Large Language Model-based Generative Recommendation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1923–1933. doi:10.1145/3726302.3730053 [29] Xinyu Lin, Chaoqun Yang, Wenjie Wang, Yongqi Li, Cunxiao Du, Fuli Feng, See-Kiong Ng, and Tat-Seng Chua. 2025. Efficient Inference for Large Language Model-based Generative Recommendation. In Proceedings of the International Conference on Learning Representations (ICLR). 91672–91697. [30] Meng Liu, Jianjun Li, Guohui Li, and Peng Pan. 2020. Cross Domain Recommendation via Bi-directional Transfer Graph Collaborative Filtering Networks. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management. 885–894. doi:10.1145/3340531.3412012 [31] Shuchang Liu, Shuyuan Xu, Wenhui Yu, Zuohui Fu, Yongfeng Zhang, and Amélie Marian. 2021. FedCT: Federated Collaborative Transfer for Recommendation. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 716–725. [32] X. Liu, B. Li, Y. Chen, X. Li, S. Xu, and H. Yin. 2022. Disentangled Representations for Cross-Domain Recommendation via Heterogeneous Graph Contrastive Learning. arXiv preprint arXiv:2206.03907 (2022). [33] Ziang Lu, Lei Guo, Xu Yu, Zhiyong Cheng, Xiaohui Han, and Lei Zhu. 2025. Federated Semantic Learning for Privacy-preserving Cross-domain Recommendation. ACM Transactions on Information Systems (2025). doi:10.1145/3728359 [34] Tong Man, Huawei Shen, Xiaolong Jin, and Xueqi Cheng. 2017. Cross-Domain Recommendation: An Embedding and Mapping Approach. In Proceedings of the 26th International Joint Conference on Artificial Intelligence (IJCAI). 2464–2470. [35] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. 2017. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS). 1273–1282. [36] Jianmo Ni, Jiacheng Li, and Julian J. McAuley. 2019. Justifying Recommendations using Distantly-Labeled Reviews and Fine-Grained Aspects. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 188–197. [37] Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan H. Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Maheswaran Sathiamoorthy. 2023. Recommender Systems with Generative Retrieval. In Advances in Neural Information Processing Systems (NeurIPS). [38] Ajit P. Singh and Geoffrey J. Gordon. 2008. Relational Learning via Collective Matrix Factorization. In Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 650–658.

FedCGR: Federated Cross-Domain Generative Recommendation

[39] Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. 2019. BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformers. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (CIKM). 1441–1450. [40] Juntao Tan, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, Zelong Li, and Yongfeng Zhang. 2024. IDGenRec: LLM-RecSys Alignment with Textual ID Learning. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 355–364. [41] Changxin Tian, Yuexiang Xie, Xu Chen, Yaliang Li, and Xin Zhao. 2024. Privacypreserving Cross-domain Recommendation with Federated Graph Learning. ACM Transactions on Information Systems 42, 5 (2024). [42] Li Wang, Qiang Wu, and Min Xu. 2025. FedPCL-CDR: A Federated PrototypeBased Contrastive Learning Framework for Privacy-Preserving Cross-Domain Recommendation. Neural Networks (2025), 108380. [43] Wenjie Wang, Honghui Bao, Xinyu Lin, Jizhi Zhang, Yongqi Li, Fuli Feng, SeeKiong Ng, and Tat-Seng Chua. 2024. Learnable Item Tokenization for Generative Recommendation. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management (CIKM). doi:10.1145/3627673.3679569 [44] Yibo Wang, Yingchun Jian, Wenhao Yang, Shiyin Lu, Lei Shen, Bing Wang, Xiaoyi Zeng, and Lijun Zhang. 2025. Towards Unbiased Information Extraction and Adaptation in Cross-Domain Recommendation. In Proceedings of the 39th AAAI Conference on Artificial Intelligence. [45] Yuhao Wang, Junwei Pan, Xinhang Li, Maolin Wang, Yuan Wang, Yue Liu, Dapeng Liu, Jie Jiang, and Xiangyu Zhao. 2025. Empowering Large Language Model for Sequential Recommendation via Multimodal Embeddings and Semantic IDs. In Proceedings of the 34th ACM International Conference on Information and Knowledge Management (CIKM). 3209–3219. [46] Meihan Wu, Le Li, Chang Tao, Eric Rigall, Xiaodong Wang, and Cheng-Zhong Xu. 2022. FedCDR: Federated Cross-Domain Recommendation for Privacy-Preserving Rating Prediction. In Proceedings of the 31st ACM International Conference on Information and Knowledge Management (CIKM). 2179–2188. [47] Ruobing Xie, Qi Liu, Liangwei Wang, Shukai Liu, Bo Zhang, and Leyu Lin. 2022. Contrastive Cross-Domain Recommendation in Matching. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4226– 4236. [48] Jian Xu, Xinyi Tong, and Shao-Lun Huang. 2023. Personalized Federated Learning with Feature Alignment and Classifier Collaboration. arXiv preprint arXiv:2306.11867 (2023). [49] Liu Yang, Fabian Paischer, Kaveh Hassani, Jiacheng Li, Shuai Shao, Zhang Gabriel Li, Yun He, Xue Feng, Nima Noorshams, Sem Park, Bo Long, Robert D. Nowak, Xiaoli Gao, and Hamid Eghbalzadeh. 2025. Unifying Generative and Dense Retrieval for Sequential Recommendation. In Proceedings of the International Conference on Learning Representations (ICLR). [50] Ziqi Yang, Zhaopeng Peng, Zihui Wang, Jianzhong Qi, Chaochao Chen, Weike Pan, Chenglu Wen, Cheng Wang, and Xiaoliang Fan. 2024. Federated Graph Learning for Cross-Domain Recommendation. In Advances in Neural Information Processing Systems (NeurIPS). [51] Wencai Ye, Mingjie Sun, Shaoyun Shi, Peng Wang, Wenjin Wu, and Peng Jiang. 2025. DAS: Dual-Aligned Semantic IDs Empowered Industrial Recommender System. In Proceedings of the 34th ACM International Conference on Information and Knowledge Management (CIKM). 6217–6224. doi:10.1145/3746252.3761529 [52] Tianzi Zang, Yanmin Zhu, Haobing Liu, Ruohan Zhang, and Jiadi Yu. 2022. A Survey on Cross-Domain Recommendation: Taxonomies, Methods, and Future Directions. ACM Transactions on Information Systems 41, 2 (2022), 1–39. [53] Hongyu Zhang, Dongyi Zheng, Xu Yang, Jiyuan Feng, and Qing Liao. 2024. FedDCSR: Federated Cross-domain Sequential Recommendation via Disentangled Representation Learning. In Proceedings of the SIAM International Conference on Data Mining (SDM). 541–549. [54] Hongyu Zhang, Dongyi Zheng, Luyao Zhong, Xu Yang, Jiyuan Feng, Yunqing Feng, and Qing Liao. 2024. FedHCDR: Federated Cross-Domain Recommendation with Hypergraph Signal Decoupling. In Proceedings of the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD). 348–365. [55] Cheng Zhao, Chenliang Li, and Cong Fu. 2019. Cross-Domain Recommendation via Preference Propagation GraphNet. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (CIKM). 2165–2168. [56] Cheng Zhao, Chenliang Li, Rong Xiao, Hongbo Deng, and Aixin Sun. 2020. CATN: Cross-Domain Recommendation for Cold-Start Users via Aspect Transfer Network. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. 229–238. doi:10.1145/3397271.3401169 [57] Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting Large Language Models by Integrating Collaborative Semantics for Recommendation. In 2024 IEEE 40th International Conference on Data Engineering (ICDE). 1435–1448. doi:10.1109/ICDE60146.2024. 00118 [58] Dongyi Zheng, Hongyu Zhang, Jianyang Zhai, Lin Zhong, Lingzhi Wang, Jiyuan Feng, Xiangke Liao, Yonghong Tian, Nong Xiao, and Qing Liao. 2025. FedCSR: A

CIKM ’26, 2026, Anonymous Review

Federated Framework for Multi-Platform Cross-Domain Sequential Recommendation with Dual Contrastive Learning. In Proceedings of the 31st International Conference on Computational Linguistics (COLING).

Related documents

• FedGMI: Generative Model-Driven Federated Learning for Probabilistic Mixture Inference#175222
• One Pool, Two Caches: Adaptive HBM Partitioning for Accelerating Generative Recommender Serving#168287
• FedReLa: Imbalanced Federated Learning via Re-Labeling#307006
• FedSPM: Routing-Enabled Federated Learning under Dual Heterogeneity via Semiparametric Mixture#343439
• Win-Win: A Privacy-Preserving Federated Framework for Dual-Target Cross-Domain Recommendation#781759