Personalized and Multi-View Representation for Federated Cold-Start Recommendation

Jaehyung Lim; Wonbin Kweon; Woojoo Kim; Junyoung Kim; Dongha Kim; Hwanjo Yu

Source-linked AI summary

Personalized and Multi-View Representation for Federated Cold-Start Recommendation

Jaehyung Lim, Wonbin Kweon, Woojoo Kim, Junyoung Kim, Dongha Kim, Hwanjo Yu

arXiv:2608.27826v1Aug 28, 2026cs.IRcs.LG

TL;DR

Federated cold-start recommendation must handle continuously arriving items while protecting client interactions and server-held attributes, but prior methods lack personalization, compositionality, and efficiency. PMFRec combines personalized generation, gated multi-view encoding, and single-representation exchange, achieving consistently strong cold-item results across datasets.

Problem
Prior federated cold-start methods face limited personalization, compositionality failure from single-space semantic encoding, and inefficient dual-representation alignment under dual-sided constraints.
Method
PMFRec uses a personalized representation generator, item-adaptive multi-view attribute encoding with orthogonality, and fused collaborative-attribute representations without explicit client-side regularization.
Results
PMFRec achieves the best performance on nearly all dataset-metric pairs, with relative gains over the strongest baseline reaching up to 7.56%.
Takeaways & Limitations
PMFRec provides user-specific cold-item representations while improving communication efficiency during training under dual-sided constraints.
Takeaways & Limitations
The server must retain item attribute features privately, while clients cannot access them under the dual-sided constraint.

Abstract

from arXiv · show

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).

1 Introduction

Federated cold-start recommendation must support continuously arriving items while preserving privacy on both sides. PMFRec addresses prior methods’ personalization, semantic compositionality, and efficiency limitations with personalized multi-view representations and single-representation exchange.

Most FedRec methods assume fixed item pools, while real services continuously introduce new items requiring cold-item recommendation for existing users.
Under dual-sided constraints, shared mappings produce essentially identical cold-item representations across users and dilute individualized preference signals.
A single encoder entangles heterogeneous semantics by forcing conflicting neighborhood structures into one latent space and uniform similarity notion.
Naïvely averaging multiple views still imposes a uniform mixture, so items require adaptive weighting of their most relevant semantic facets.
PMFRec generates user-specific cold-item representations with a personalized encoder and models complementary attribute semantics through view-specific projectors, item-wise gating, and orthogonality.
PMFRec fuses collaborative and attribute knowledge into one exchanged item representation, removing explicit client-side alignment regularization and reducing communication overhead.

2 Related Works

Related work spans federated recommendation, centralized cold-start recommendation, and federated cold-start recommendation. Existing federated approaches often assume fixed item pools, motivating methods for zero-interaction new items under dual-sided constraints.

FedRec keeps clients’ raw interaction data local while exchanging model parameters or updates, with prior work exploring personalization, graphs, and dual knowledge representations.
Centralized cold-start methods use collaborative filtering, item attributes, or hybrid signals, but typically assume user interactions are collected at a server.
Federated cold-start work addresses new items with insufficient interactions, but prior methods largely overlook zero-interaction items or face dual-sided information constraints.

3 Preliminaries

The problem partitions items into warm and cold sets, with private client-side user and warm-item parameters and server-side attribute features used to generate item representations.

Items are partitioned into warm items I_W and cold items I_C, each associated with a d_f-dimensional attribute feature vector.
Each client privately maintains user-specific parameters, including a user embedding, personalized scoring function, and warm-item embedding matrix.
The server does not store global item embeddings; instead, it generates warm-item representations from X_W and cold-item representations from X_C.
Dual-sided privacy prevents the server from observing client interactions or private parameters and prevents clients from accessing or inferring item attribute features.

4 Methodology

PMFRec’s training framework fuses collaborative knowledge into server-provided attribute-feature representations without client-side alignment regularization, while supporting server aggregation without storing all item embeddings.

PMFRec begins with regularizer-free warm training that fuses collaborative knowledge into server-provided attribute-feature-based representations.
The framework describes server aggregation of client-updated item embeddings without requiring the server to store all item embeddings.

4.1 Warm Training Phase: Client-side

During warm training, clients optimize personalized recommendation models locally using server-provided warm-item representations, while exchanging only one item-embedding matrix per round.

Clients learn private parameters and warm-item embeddings from local interaction data.
Each client initializes its warm-item embedding matrix with a potentially client-specific server-provided representation matrix Z_u.
Local optimization applies a Binary Cross Entropy implicit-feedback objective over warm items using personalized scoring.
Clients do not observe warm-item attributes or the server mapping; they receive Z_u only as initialization.
Exchanging one item-embedding type lets collaborative knowledge be absorbed without a client-side alignment regularizer, reducing communication overhead.

4.2 Warm Training Phase: Server-side

The server aggregates client-updated warm-item embeddings, learns personalized and global attribute-based representations, and combines them into scalable multi-view initializations without explicit client-side alignment.

Server aggregation and personalization: The server incrementally aggregates uploaded client embeddings into global collaborative knowledge while constructing a personalized representation generator for each client.
Server aggregation and personalization: Each personalized generator maps warm-item attributes to user-specific collaborative embeddings and can then interpret zero-interaction cold items from those users’ perspectives.
Server aggregation and personalization: PMFRec implements each personalized generator as a lightweight linear mapping followed by a nonlinear activation, with parameters W_u.
Efficient personalized fitting: A ridge-regression reformulation yields a closed-form personalized mapping whose server-side factor can be precomputed for efficient updates.
Fused initialization and deployment: Server-retained and offloaded deployment modes trade server-side storage against communication cost.
Global multi-view encoding: The global encoder decomposes attribute knowledge into K views, using view-specific projectors and gating weights to combine semantic perspectives.
Global multi-view encoding: Orthogonality reduces redundancy among view knowledge, while load balancing discourages collapse onto only a subset of views.
Fused initialization and deployment: The server sends fused personalized and multi-view warm representations, allowing clients to learn collaborative knowledge without an explicit regularization term.

4.3 Cold Inference Phase

In cold inference, PMFRec uses warm-trained mappings to construct user-specific multi-view representations for cold items and compute personalized preferences.

Cold-item representations are generated from warm-phase mappings, allowing different users to interpret cold items through personalized perspectives.
For a requesting user, the server constructs a user-specific and multi-view cold-item representation before prediction.
The predicted preference for each cold item is computed using the user’s personalized scoring function.

4.4 Additional Privacy Preservation

PMFRec can apply local differential privacy to client item-embedding gradients, addressing leakage from item indices and persistent update patterns.

Across communication rounds, consistent per-item update patterns may reveal likely positive items despite negative sampling.
Noise calibration provides (ε,δ)-LDP for the communicated per-round embedding change.
The privacy mechanism clips the item-embedding gradient and adds Gaussian noise before communicating updates.

5 Experiments

Experiments evaluate PMFRec on real-world cold-start datasets using top-K ranking metrics and ablations of personalization, multi-view encoding, gating, and regularization. PMFRec achieves broad performance gains, reduces user-level disparity and zero-score users, and shows that personalization and complementary multi-view design contribute distinct benefits.

Datasets and evaluation: Experiments use top-K Recall, Precision, and NDCG on CiteULike and XING cold-item recommendation datasets.CiteULike and XING provide item-side attribute information and use warm/cold item splits.
Main results: PMFRec achieves the best performance on nearly all dataset-metric pairs, with relative gains over the strongest baseline reaching 7.56%.Most improvements are statistically significant at p≤0.05 and remain stable across datasets with different scales and sparsity levels.
Main results: Improvements are consistent across Recall, Precision, and NDCG, including against baselines that directly exploit item attribute features.The results support jointly preserving user-specific preference signals and modeling multi-view representations under dual-sided constraints.
Ablation study: Removing personalization or using only personalized representations causes the largest ablation degradations, showing that effective performance requires both user-specific and shared collaborative knowledge.The personalized-only variant cannot leverage collaborative knowledge shared through the global item representation.
Ablation study: Single-view encoding and uniform view averaging reduce accuracy, supporting multi-view encoding and item-adaptive view selection for heterogeneous attribute semantics.The gating mechanism lets each item emphasize relevant semantic subspaces.
Ablation study: Removing either orthogonality or load-balancing regularization produces suboptimal results, while their joint removal can sometimes outperform removing only one term.This pattern suggests interdependence between the objectives in encouraging diverse and balanced view utilization.
Personalization and fairness: PMFRec reduces zero-score users and generally lowers Gini coefficients, while personalization consistently outperforms corresponding non-personalized settings.Combining personalization with multi-view representation usually provides the best overall performance, although it is not uniformly superior on every fairness metric.

5.5 Multi-view Analysis (RQ4)

PMFRec’s multi-view encoder learns complementary, non-collapsed, and interpretable semantic views, while its representations transfer to existing baselines and warm recommendation. The method also shows stable convergence, parameter sensitivity, and improved robustness under LDP by avoiding conflicting client-side objectives.

Multi-view decomposition: Orthogonality regularization lowers cross-view similarity, producing less redundant and more complementary representations from unseen cold-item attributes.Without it, view representations remain highly correlated and collapse into overlapping feature subspaces.
Multi-view decomposition: Highest-weight item groups form distinct semantic clusters, such as Biology versus Social Science, supporting interpretable multi-view decomposition.The qualitative case study reports different semantics for representative views.
Multi-view decomposition: Load balancing keeps routing masses more even and stable, preventing dominant-view collapse and ensuring that all views participate.Without load balancing, routing progressively skews toward a subset of views.
Adaptability and warm scenario: 0.68% to 18.61% performance gains result when PMFRec representations replace item representations in existing baselines.The augmented baselines retain their original architectures, but separately transmitting the two representation types can reduce communication efficiency.
Adaptability and warm scenario: Up to a 25.8% performance improvement is achieved in the warm scenario, where multi-view encoding is more critical than personalization.Removing multi-view encoding causes a substantial warm-scenario performance drop, unlike the cold setting.
LDP robustness: PMFRec remains more robust under smaller privacy budgets, whereas regularizer-based baselines experience sharper performance declines as LDP noise increases.The analysis attributes this gap to conflicting recommendation and regularization gradients, which PMFRec avoids by eliminating the client-side regularizer.

6 Conclusion

PMFRec addresses federated cold-start recommendation under dual-sided constraints with personalized and multi-view item representations. The framework combines complementary semantic views, single-representation exchange, and broad empirical improvements across fairness, warm adaptation, and LDP robustness.

Framework: PMFRec generates user-specific cold-item representations by combining a personalized encoder with a global multi-view attribute encoder and item-adaptive gating.The multi-view design models heterogeneous attribute semantics through complementary views and selectively composes them for each item.
Framework: PMFRec integrates collaborative and attribute knowledge into one exchanged item representation, eliminating the explicit client-side regularizer and improving communication efficiency.This replaces the separate-representation alignment design used by prior approaches.
Empirical findings: Experiments and analyses report consistent baseline improvements, better user-level fairness, warm-scenario transfer, and stronger robustness under LDP.Ablation studies support the importance of personalization, multi-view encoding, adaptive gating, and view regularization.

A Why Clipping + Gaussian Noise Satisfies (𝜀,𝛿)-LDP

The appendix explains that clipping bounds local gradients, Gaussian noise is calibrated to the resulting sensitivity, and post-processing preserves the per-step LDP guarantee.

Sensitivity bound: Clipping maps each local gradient to a Frobenius norm no greater than C, which bounds the vectorized ℓ2-sensitivity by 2C.The clipped mapping is defined as f(G)=G·min(1,C/∥G∥F).
Gaussian calibration: The Gaussian mechanism adds noise with calibrated standard deviation sufficient to satisfy (ε,δ)-LDP for a single client’s local mechanism.The calibration uses the sensitivity bound and privacy parameters ε and δ.
Privacy preservation: Releasing the privatized gradient and applying the server update are deterministic post-processing, so the same per-step (ε,δ)-LDP guarantee is preserved.The server observes ΔQ_u=−ηG̃ after receiving the privatized gradient.