PaperlayerView exact revision on arXiv

Source-linked AI summary

FedHUR: Learning Hierarchical Utility-Guided Client Relations for Personalized Federated Recommendation

Mingzhe Han, Jiahao Liu, Dongsheng Li, Jiankui Zhou, Hansu Gu, Peng Zhang, Ning Gu, Tun Lu

arXiv:2609.11632v1cs.IR

TL;DR

Personalized federated recommendation must select useful information across clients, but existing relations may rely on predefined assumptions and a single global relation. FedHUR learns hierarchical utility-guided relations using item-item filters, and experiments on five real-world datasets report consistent gains over representative federated recommendation baselines.

  • Problem

    Personalized aggregation must account for different client contributions, while predefined relations may not identify useful information and a single global relation misses hierarchical relations.

    Source: §1 · pp. 1–2

  • Method

    FedHUR builds hierarchical item-item filters and uses target-client utility signals to retrieve candidate clients for personalized filter aggregation.

    Source: §1 · p. 2, §4.5 · p. 5

  • Results

    FedHUR consistently improves recommendation performance over representative federated recommendation baselines on five real-world datasets.

    Source: §6 · p. 10

  • Takeaways & Limitations

    The experiments demonstrate the effectiveness of hierarchical utility-guided client relation learning for personalized federated recommendation.

    Source: §6 · p. 10

    The reported results use full-ranking evaluation and are not directly comparable with results from sampled evaluation settings.

    Source: §5.1.4 · p. 8

Abstract

from arXiv · show

Federated recommendation enables collaborative model training while keeping user interaction data on local clients. A central problem in federated recommendation is how to aggregate useful information across clients for personalized recommendation. Existing personalized aggregation methods usually construct client relations from predefined parameter-based assumptions, such as parameter similarity or complementarity, and use these relations to determine aggregation weights. However, such methods construct a single global relation, which is insufficient to capture the hierarchical and multi-granularity nature of user relations in recommendation. Moreover, these predefined relations cannot directly reflect whether the related clients can improve prediction performance after aggregation. To address these limitations, we propose FedHUR, a federated recommendation framework for learning hierarchical utility-guided client relations. FedHUR takes item-item filters as the object for relation construction and aggregation. Specifically, it first aggregates and clusters each client's local information to obtain global hierarchical information. Each client computes hierarchical utility signals based on its local information and the global hierarchical information, indicating which collaborative information is useful for improving its prediction. The server uses these utility signals to retrieve clients that are useful to that client for further personalized aggregation. Extensive experiments on five real-world datasets show that FedHUR consistently outperforms existing federated recommendation baselines, demonstrating the effectiveness of hierarchical utility-guided client relation learning. Code is available at https://github.com/Mingzhe-Han/FedHUR.

1 Introduction

Federated recommendation keeps interaction data on local clients, but heterogeneous client information makes useful personalized aggregation difficult. FedHUR learns hierarchical client relations guided by estimated aggregation utility.

Motivation

  • Personalized aggregation is needed because clients have different data distributions and can provide different levels of useful information to a target client.Traditional weighted averaging mainly uses local data sizes, which may fail to aggregate the most useful information for each client.Source: §1 · pp. 1–2

Limitations of existing relations

  • Existing relations are limited because a single global relation misses hierarchical user relations, while parameter similarity or complementarity does not establish post-aggregation usefulness.In recommendation, clients may relate differently across item groups; useful information is defined by whether it improves performance after aggregation.Source: §1 · p. 2

Challenges

  • Defining hierarchical relations is difficult with sparse, noisy item information, while validating every candidate through aggregation and prediction would be costly.Whole-space relations can overlook item-group-specific links; isolated group relations can be unreliable, redundant, or conflicting.Source: §1 · p. 2

Graph-filtering perspective

  • Graph filtering offers coarse-to-fine item relations, with fine relations inheriting information from coarse groups to avoid noisy and redundant estimation.It can also estimate which items need external collaboration from filtering results, without repeated aggregation and validation.Source: §1 · p. 2

FedHUR

  • FedHUR uses item-item filters to build hierarchical relations and utility signals, then retrieves useful clients and aggregates their filters into personalized filters.Its two modules mine hierarchical relations and retrieve clients using utility scores; the introduction reports consistent gains over federated recommendation baselines.Source: §1 · p. 2

2 Related Work

Federated recommendation keeps interactions local but often relies on simple aggregation despite heterogeneous client behavior. Personalized aggregation uses client relations, and FedHUR differs by learning hierarchical, utility-guided relations.

2.1 Federated Recommendation

  • Federated recommendation methods adapt recommendation models to local training and exchange model information, but often use data-size-weighted averaging despite heterogeneous user behavior.Examples include FedRec, FedMF, and FedNCF; aggregating all updates into one global model may be suboptimal for individual clients.Source: §2.1 · p. 3

2.2 Personalized Aggregation in Federated Recommendation

  • Personalized aggregation assigns client-specific weights using relations, but most methods rely on predefined priors and a single global relation.FedHUR instead learns hierarchical utility-guided relations to retrieve candidate information for each target client at different levels.Source: §2.2 · p. 3

3 Preliminaries

The preliminaries define recommendation as preference-score prediction from interactions, then describe federated training and relation-based personalized aggregation. The framework keeps user parameters local while sharing item or other public information.

3.1 Recommendation Algorithms

  • Recommendation estimates a user's preference score for an item from observed interaction data, and embedding-based models map users and items to latent representations for prediction.The interaction dataset is represented as D = {(u, i, y_ui)}, where y_ui indicates whether an interaction is observed.Source: §3.1 · p. 3

3.2 Federated Recommendation Algorithms

  • In federated recommendation, clients train on private local datasets and upload item or public parameters while user parameters remain local.Uploaded parameters may include item embeddings, item parameters, or public prediction parameters; traditional aggregation can use weighted averaging.Source: §3.2 · p. 3

3.3 Relation-Based Personalized Aggregation

  • Relation-based personalized aggregation computes a relation score between target and candidate clients, normalizes scores into weights, and aggregates uploaded information for the target.The personalized aggregate is used alongside the target client's local user parameters.Source: §3.3 · p. 3

4 Method

FedHUR's round combines local embedding training, server-side global and hierarchical filter construction, local utility signals, and personalized filter aggregation. Its relation filters are built from coarse item groups toward finer groups that inherit parent-level information.

4.1 Overview

  • A communication round distributes hierarchical item-item filters, gathers locally computed utility signals, retrieves candidate clients, and returns personalized filters for local fine-tuning.The overview depicts one client for simplicity.Source: §4.1 · pp. 3–4, Figure 1 · p. 4

4.2 Local Training and Upload

  • Each client trains on private data and uploads item embeddings, while user parameters remain local and are never uploaded.The communicated item parameters are instantiated as local item embeddings.Source: §4.2 · p. 4

4.3 Global Information Aggregation

  • The server aggregates uploaded item embeddings into global embeddings and a global item-item filter, used respectively for item grouping and as the base for personalized aggregation.The global embeddings construct hierarchical item groups; the global filter serves as the global base filter.Source: §4.3 · p. 4

4.4 Hierarchical Relation Mining Module

  • FedHUR constructs filters at multiple granularities so relations can represent coarse common information and finer differences between clients.Finer filters inherit their corresponding coarse filter values, preserving coarse information while adding details.Source: §4.4 · p. 4

4.4.1 Coarse-to-Fine Item Grouping.

  • The server partitions global item embeddings into coarse groups, then recursively applies k-means within each group for finer levels.The number of levels is a hyperparameter, and other clustering methods may also be used.Source: §4.4.1 · pp. 4–5

4.4.2 Coarse-Level Relation Filters. For each coarse group G(1)

  • At the coarse level, each client averages item embeddings within a group to form group embeddings, computes filters between them, and the server averages local filters.A smaller level index denotes a coarser level.Source: §4.4.1 · p. 4, §4.4.2 · p. 5

4.4.3 Finer-Level Relation Filters.

  • At finer levels, clients compute filters between smaller groups and inherit the corresponding parent-level relation value before the server averages filters across clients.Repeating this construction preserves coarse relations while adding finer-level information for later utility-guided relation computation.Source: §4.4.3 · p. 5

4.5 Utility-Guided Client Retrieval Module

FedHUR replaces costly repeated aggregation-and-validation with retrieval guided by each target client’s estimated utility needs. It scores candidates against those needs and refines the scores before aggregation.

4.5 Utility-Guided Client Retrieval Module

  • FedHUR estimates which information a target needs and retrieves helpful clients, avoiding repeated aggregation and validation that is computationally expensive and difficult to scale.The same procedure is applied to relation filters at each level.Source: §4.5 · p. 5
  • A target client builds a utility query from its local and global filters to describe the filter direction needed to improve local prediction.The query is projected and normalized to unit ℓ2 norm, retaining its main matching direction while reducing upload cost.Source: §4.5 · pp. 5–6

4.5.2 Utility-Guided Retrieval.

  • Retrieval scores measure how strongly each candidate representation aligns with the target’s utility directions, with larger scores indicating more-needed information.The score is based on projecting the candidate representation onto the target’s utility directions.Source: §4.5.2 · p. 6

4.5.3 Learned Retrieval Scorer.

  • A lightweight shared MLP refines direct retrieval scores using direction-level matching and candidate embedding-scale statistics.The final score combines direct and refined scores, and the scorer is shared across levels and item groups.Source: §4.5.3 · p. 6

4.6 Personalized Filter Aggregation

FedHUR normalizes retrieval scores into aggregation weights, then combines candidate residual filters across hierarchy levels to form each target client’s personalized filter.

4.6 Personalized Filter Aggregation

  • At each level, the server applies softmax to candidate retrieval scores to obtain aggregation weights, with temperature τ controlling the normalization.The level-specific personalized residual filter is then computed from the weighted candidate filters.Source: §4.6 · p. 6
  • FedHUR maps each level’s personalized residual filter back to the full item-item filter space and combines levels using βℓ.The mapping assigns item relations the value of their corresponding group relation.Source: §4.6 · p. 7
  • The resulting personalized filter preserves global collaborative information while incorporating utility-guided residual information from different levels.On large-scale datasets, parameter-based aggregation may optionally be used at the coarse level to reduce computation, while finer-level residual aggregation remains utility-guided.Source: §4.6 · p. 7

4.7 Local Fine-Tuning with Personalized Filter

The server sends each client its personalized filter for local item-embedding fine-tuning; user-side parameters remain local. FedHUR follows existing federated recommendation privacy practices.

4.7 Local Fine-Tuning with Personalized Filter

  • Each client locally fine-tunes its item embeddings by matching their item-item filter with the personalized filter received from the server.Only item-side embeddings are updated in this step; user-side parameters remain local and are not uploaded.Source: §4.7 · p. 7

4.8 Algorithm

  • Algorithm 1 summarizes one communication round of FedHUR.The supplied passage identifies the algorithm’s scope as one round.Source: §4.8 · p. 7

4.9 Privacy Protection

  • FedHUR uses the same privacy-protection strategy as existing federated recommendation methods, and secure aggregation or differential privacy can be applied to uploaded item information when needed.Its utility signals are computed from local filters and do not introduce additional privacy assumptions beyond existing methods.Source: §4.9 · p. 7

4.10 Communication Cost Discussion

FedHUR adds communication for hierarchical relation learning to FedCIA’s item-information exchange, using group-level filters and projected utility queries to keep the added messages lightweight.

4.10 Communication Cost Discussion

  • FedHUR adds hierarchical filters and projected utility queries to FedCIA’s item-information communication, with extra messages smaller than one additional full item-item filter.The filters contain group-level item-item information, and random projection compresses utility queries before upload.Source: §4.10 · p. 7
  • FedHUR follows FedCIA’s communication paradigm for collaborative aggregation while adding messages for hierarchical relation learning.The additional messages comprise hierarchical filters and projected utility queries.Source: §4.10 · p. 7
  • In the uncompressed setting, FedHUR’s total communication cost is less than twice FedCIA’s.The paper also states that FedHUR can adopt FedCIA’s SVD-based low-rank compression strategy.Source: §4.10 · p. 7

5 Experiments

The experiments evaluate whether FedHUR improves federated recommendation, learns more useful hierarchical relations, estimates useful clients, and does so at acceptable cost. FedHUR achieves the best or comparable performance across five datasets, while relation analyses, ablations, and efficiency tests examine how it works and what it costs.

5.1 Experimental Settings

  • Across five datasets and 100 clients per dataset, FedHUR is compared with federated baselines using Recall@10, MRR@10, and NDCG@10 under full-ranking evaluation.The authors caution that these results are not directly comparable with sampled-candidate evaluation results.Source: §5.1.1 · p. 8, §5.1.2 · p. 8, §5.1.3 · p. 8

5.2 Overall Performance (RQ1)

  • FedHUR achieves the best or comparable recommendation performance across five datasets, and the authors report improvements over baselines on all datasets.They attribute this to hierarchical utility-guided relations that retrieve more useful candidate information for each target client.Source: §5.2 · p. 8

5.3 Relation Analysis (RQ2)

  • Under a fixed aggregation framework, utility-guided relations outperform predefined relations, and hierarchical relations generally outperform single relations; the hierarchical utility relation achieves the best overall performance.The comparisons include no relation, similarity, complementarity, and combined similarity-plus-complementarity relations.Source: Table 3 · p. 9, §5.3 · p. 9

5.4.1 Effect of utility query construction.

  • The symmetric utility query performs best across metrics, while a local-filter-only query performs clearly worse and residual-query performance is unstable.The authors report that the query should account for information already captured by the current base filter.Source: Table 4 · p. 9, §5.4.1 · p. 9

5.4.2 Effect of scorer input features.

  • The full scorer achieves the best results across metrics by combining direction matching, embedding-norm statistics, and total matching score.The score-only variant performs much worse, while removing either direction or norm features weakens performance.Source: Table 5 · p. 9, §5.4.2 · pp. 9–10

5.4.3 Correlation analysis.

  • On ML-100K, utility scores are consistently positively correlated with validation and test performance, while similarity-score correlations are unstable and mostly negative.The authors interpret utility scores as better aligned with the target client's current aggregation need than general similarity.Source: Table 6 · p. 9, §5.4.3 · p. 10

5.4.4 Visualization of Hierarchical Relations.

  • Figure 2 shows relations refining from broad item groups to localized fine groups; Figure 3 reports best performance with three levels and diminishing gains from larger projection dimensions.The authors associate overly deep hierarchies with noisy or redundant relations and say moderate projection size is sufficient to preserve utility directions while reducing communication cost.Source: Figure 2 · p. 10, §5.4.4 · p. 10, Figure 3 · p. 10

5.5 Efficiency Analysis (RQ4)

  • FedHUR's total communication cost is about 27% higher than FedCIA, and its 700.05 s training time is about 14.8% higher.The authors characterize the additional communication and computation cost as limited while achieving better recommendation performance.Source: Table 7 · p. 10, §5.5 · p. 10

6 Conclusion

FedHUR learns hierarchical utility-guided client relations for personalized aggregation. Its two modules model relations from coarse to fine and retrieve clients matching each target’s utility direction; experiments on five real-world datasets show improved performance over representative federated recommendation baselines.

6 Conclusion

  • Experiments on five real-world datasets show FedHUR consistently improves recommendation performance over representative federated recommendation baselines.Source: §6 · p. 10
  • FedHUR’s hierarchical relation mining module constructs multi-level relations, enabling modeling from coarse to fine.Source: §6 · p. 10
  • Its utility-guided client retrieval module derives signals from each target’s local information and retrieves candidate clients matching the target’s utility direction.Source: §6 · p. 10
  • FedHUR is a federated recommendation framework that learns hierarchical utility-guided client relations for personalized aggregation.Source: §6 · p. 10

GenAI Usage Disclosure

The authors report using generative AI only for language polishing and grammar checking. They state that the technical work was produced and verified by the authors, who reviewed the AI-assisted text and take responsibility for the final manuscript.

GenAI Usage Disclosure

  • Generative AI tools were used only for language polishing and grammar checking during manuscript preparation.Source: GenAI Usage Disclosure · p. 11
  • The authors state that all technical ideas, method design, experimental results, analyses, and conclusions were produced and verified by them.Source: GenAI Usage Disclosure · p. 11
  • The authors reviewed and revised all AI-assisted text and take responsibility for the final manuscript.Source: GenAI Usage Disclosure · p. 11
Loading 2609.11632v1…