Federated recommendation (FedRec) enables personalized modeling without centralizing users' interaction histories, but most existing methods assume a fixed item pool and thus overlook the practical cold-item setting where new items continuously arrive. Under the dual-sided constraint, where the server cannot access clients' interactions while clients cannot access the server's proprietary item attribute features, prior federated cold-start recommendation approaches suffer from three structural limitations: a lack of personalization, compositionality failure caused by forcing heterogeneous semantics into a single embedding space, and training- and communication-inefficiency arising from explicit alignment between separate collaborative and attribute representations. To address these challenges, we propose Personalized and Multi-view Representation for Federated Cold-Start Recommendation (PMFRec). PMFRec learns a personalized representation generator to produce user-specific item representations from attribute features, and introduces a global multi-view encoder with item-adaptive gating and an orthogonality objective to capture complementary semantic views while reducing cross-view redundancy. In addition, PMFRec fuses collaborative and attribute knowledge into a single exchanged item representation, eliminating the need for an explicit client-side regularizer and reducing communication overhead. Extensive experiments on real-world datasets show that PMFRec consistently outperforms strong baselines in cold-item recommendation and further improves user-level fairness, warm-scenario adaptability, and robustness under Local Differential Privacy (LDP).
Figures & tables
Category
Drug
ML
Graph
Chemistry
Recommendation
Item A
✓
✓
✓
Item B
✓
✓
Item C
✓
✓
✓
Table 1. Toy example illustrating paper recommendation: each item corresponds to a paper, and the category indicates its research area.
Figure 1. Comparison between existing methods and PMFRec. Prior federated cold-start approaches rely on a global single encoder, resulting in identical representations. In contrast, PMFRec introduces a personalized and multi-view encoding to generate user-specific representations for same cold items.
Figure 2. The overall framework of PMFRec.
Dataset
Split
#Users
#Items
#Interactions
Sparsity
CiteULike
Train
5,551
13,584
164,210
99.78%
Val
5,551
1,018
13,037
–
Test
5,551
2,378
27,739
–
XING-5000
Train
5,000
11,290
120,944
99.79%
Val
5,000
1,869
21,713
–
Test
5,000
5,648
58,609
–
Table 2. Statistics of the four cold-start recommendation datasets. Sparsity is reported for the training split.
Methods
Metrics
CiteULike
XING-5000
XING-10000
XING-20000
Recall
Precision
NDCG
Recall
Precision
NDCG
Recall
Precision
NDCG
Recall
Precision
NDCG
FedMVMF
@5
0.1378
0.1078
0.1721
0.0083
0.0193
0.0181
0.0031
0.0076
0.0066
0.0067
0.0157
0.0144
@10
0.2228
0.0927
0.2154
0.0144
0.0168
0.0169
0.0054
0.0066
0.0064
0.0114
0.0136
0.0139
@20
0.3383
0.0751
0.2670
0.0337
0.0191
0.0307
0.0221
0.0108
0.0286
0.0239
0.0140
0.0213
FedWDR
@5
0.1585
0.1314
0.2049
0.0092
0.0229
0.0227
0.0123
0.0281
0.0276
0.0140
0.0350
0.0318
@10
0.2507
0.1080
0.2478
0.0173
0.0215
0.0230
0.0213
0.0248
0.0263
0.0236
0.0290
0.2960
Table 3. Performance comparison across four datasets: the best results are highlighted in bold, specifically used when our model’s result is larger than the best-performing baseline where Improv(%) represents the relative performance improvement over the best-performing baseline, and ∗ denotes statistical significance ( p≤0.05 ) for the paired t-test against the best baseline on each metric.
Methods
CiteULike
XING-5000
R@10
R@20
P@10
P@20
N@10
N@20
R@10
R@20
P@10
P@20
N@10
N@20
PMFRec
0.2929
0.4187
0.1309
0.0979
0.3001
0.3441
0.1530
0.2545
0.1718
0.1411
0.1850
0.2112
w/o Pu
0.2675
0.3900
0.1200
0.0916
0.2763
0.3234
0.1367
0.2304
0.1553
0.1294
0.1662
0.1952
w/ only Pu
0.2656
0.3988
0.1179
0.0922
0.2739
0.3244
0.1181
0.1927
0.1334
0.1093
0.1395
0.1625
w/ single-view
0.2886
0.4120
0.1286
0.0964
0.2936
0.3381
0.1498
0.2482
0.1694
0.1386
0.1816
0.2086
w/o gate
0.2895
0.4167
0.1284
0.0972
0.2947
0.3418
0.1523
0.2455
0.1699
0.1373
0.1816
0.2066
Table 4. Ablation study of PMFRec on CiteULike and XING-5000.
Dataset
Metric
IFedNCF
IPFedRec
w/o Pu
w/ single-view
PMFRec
CiteULike
Z@10
36.93%
37.15%
37.65%
34.89%
34.71 %
Z@20
24.23%
25.73%
25.24%
23.66%
22.92 %
XING-5000
Z@10
30.30%
32.00%
30.34%
28.94%
26.80 %
Z@20
12.12%
11.24%
11.86%
11.46%
11.04 %
Table 5. Comparison of zero-score ratio (Z@K).
Figure 3. Analysis for user-level fairness and performance distribution.
Figure 4. Effect of load balancing loss on CiteULike.
Figure 5. Analysis of cross-view redundancy on CiteULike.
View
Title of literature
View 1
Evolving protein interaction networks through gene duplication
Light Microscopy Techniques for Live Cell Imaging
Widespread occurrence of antisense transcription in the human genome
View 2
Thirteen Ways to Look at the Correlation Coefficient
Information Literacy
A face(book) in the crowd: social Searching vs. social browsing
Table 6. Qualitative Case Study on CiteULike. The items assigned highest weights for different views reveal distinct semantic clusters (e.g., Biology vs. Social Science), supporting the interpretability of our multi-view decomposition.
Figure 6. Adaptability of personalized and multi-view representations to existing baselines.
Figure 7. Performance in the warm scenario.
Figure 8. Convergence curves on CiteULike.
Figure 9. Parameter sensitivity analysis on CiteULike and XING-5000.
Dataset
Method
IFedNCF
IFedNCF w/ PM
PMFRec
Metrics
R@10
N@10
R@10
N@10
R@10
N@10
CiteULike
w/o LDP
0.2774
0.2763
0.2954
0.3037
0.2929
0.3001
ε=4
0.0928
0.1029
0.1088
0.1147
0.2274
0.2424
ε=8
0.1556
0.1640
0.1734
0.1776
0.2381
0.2494
ε=10
0.1785
0.1892
0.1743
0.1781
0.2432
0.2531
XING-5000
w/o LDP
0.1429
0.1726
0.1695
0.2011
0.1530
0.1850
Table 7. Performance under (ε, δ)-LDP with varying privacy budgets ε and fixed δ .
Figure 10. Qualitative case study under LDP ( ε=10 ). Per-round diagnostics on CiteULike (top) and XING-5000 (bottom) for IFedNCF w/PM.
Federated recommendation facilitates collaborative model training across distributed clients while keeping sensitive user interaction data local. Conventional approaches typically rely on synchronizing high-dimensional item representations between the server and clients. This paradigm implicitly assumes that precise geometric alignment of embedding coordinates is necessary for collaboration across clients. We posit that establishing relative semantic relationships among items is more effective than enforcing shared representations. Specifically, global semantic relations serve as structural constraints for items. Within these constraints, the framework allows item representations to vary locally on each client, which flexibility enables the model to capture fine-grained user personalization while maintaining global consistency. To this end, we propose Cluster-Guided FedRec framework (CGFedRec), a framework that transforms uploaded embeddings into compact cluster labels. In this framework, the server functions as a global structure discoverer to learn item clusters and distributes only the resulting labels. This mechanism explicitly cuts off the downstream transmission of item embeddings, relieving clients from maintaining global shared item embeddings. Consequently, CGFedRec achieves the effective injection of global collaborative signals into local item representations without transmitting full embeddings. Extensive experiments demonstrate that our approach significantly improves communication efficiency while maintaining superior recommendation accuracy across multiple datasets.
Yuchun Tu, Zhiwei Li, Bingli Sun +2
Beihang University, Beijing, China · University of Technology Sydney, Sydney, Australia · University of Edinburgh, Edinburgh, UK
Cold-start item recommendation remains a persistent challenge in real-world systems due to the absence of interaction histories. While prior models attempt to bridge this gap using item content features, they universally suffer from the seesaw dilemma: enhancing performance for cold items inevitably degrades performance for warm items, and vice versa. We identify that this dilemma stems from a fundamental distributional disparity: warm item embeddings occupy a complex behavioral manifold" shaped by rich interaction signals, whereas cold item embeddings are constrained to a semantic manifold" derived solely from auxiliary content. Existing methods often force a rigid mapping between these inconsistent spaces, causing the model to sacrifice the precision of warm representations to accommodate cold ones. To address this, we propose DiffCold, a diffusion-based generative model that unifies warm and cold representations. Unlike GANs or VAEs, DiffCold leverages conditional diffusion to reconstruct warm item embeddings from content, preserving the underlying manifold structure without degradation. We further tailor this paradigm with two specific designs: a Retrieval-enhanced Aggregator that initializes generation using semantically similar warm items to bypass inefficient noise, and a Simulation-based Representation Alignment module that enforces distribution consistency between generated and real embeddings via contrastive learning. Experiments on three benchmarks confirm that DiffCold resolves the seesaw dilemma, consistently outperforming state-of-the-art methods across all metrics.
Kangning Zhang, Yingjie Qin, Weinan Zhang +2
Shanghai Jiao Tong University, Shanghai, China · Xiaohongshu Inc., Shanghai, China
Federated recommender systems enable collaborative model training while keeping user interaction data local and sharing only essential model parameters, thereby mitigating privacy risks. However, existing methods overlook a critical issue, i.e., the stable learning of a generalized item embedding throughout the federated recommender system training process. Item embedding plays a central role in facilitating knowledge sharing across clients. Yet, under the cross-device setting, local data distributions exhibit significant heterogeneity and sparsity, exacerbating the difficulty of learning generalized embeddings. These factors make the stable learning of generalized item embeddings both indispensable for effective federated recommendation and inherently difficult to achieve. To fill this gap, we propose a new federated recommendation framework, named Federated Recommendation with Generalized Embedding Learning (FedRecGEL). We reformulate the federated recommendation problem from an item-centered perspective and cast it as a multi-task learning problem, aiming to learn generalized embeddings throughout the training procedure. Based on theoretical analysis, we employ sharpness-aware minimization to address the generalization problem, thereby stabilizing the training process and enhancing recommendation performance. Extensive experiments on four datasets demonstrate the effectiveness of FedRecGEL in significantly improving federated recommendation performance. Our code is available at https://github.com/anonymifish/FedRecGEL.
Fengyuan Yu, Xiaohua Feng, Yuyuan Li +3
Zhejiang University, Hangzhou, China · Hangzhou Dianzi University, Hangzhou, China · OPPO Research Institute, Shenzhen, China