PaperlayerView exact revision on arXiv

Source-linked AI summary

DGRec: Graph Neural Network for Recommendation with Diversified Embedding Generation

Liangwei Yang, Shengjie Wang, Yunzhe Tao, Jiankai Sun, Xiaolong Liu, Philip S. Yu, Taiqing Wang

arXiv:2211.10486v2cs.IR

TL;DR

GNN recommenders can produce accurate recommendations but may aggregate redundant neighbors and underrepresent diverse or long-tail items. DGRec modifies embedding generation with submodular neighbor selection, layer attention, and loss reweighting; experiments report improved diversity with comparable accuracy.

  • Problem

    GNN neighbor aggregation follows graph structure and can overlook neighbor redundancy, resulting in poor diversity in recommended lists.

  • Method

    DGRec combines submodular neighbor selection, layer attention, and loss reweighting to generate more diversified GNN embeddings.

  • Results

    DGRec achieves the best diversity while maintaining accuracy comparable to advanced accuracy-based recommender systems.

  • Takeaways & Limitations

    DGRec provides an accuracy-diversity trade-off through several hyper-parameters and can be applied to GNN-based recommenders with minimal accuracy cost.

Abstract

from arXiv · show

Graph Neural Network (GNN) based recommender systems have been attracting more and more attention in recent years due to their excellent performance in accuracy. Representing user-item interactions as a bipartite graph, a GNN model generates user and item representations by aggregating embeddings of their neighbors. However, such an aggregation procedure often accumulates information purely based on the graph structure, overlooking the redundancy of the aggregated neighbors and resulting in poor diversity of the recommended list. In this paper, we propose diversifying GNN-based recommender systems by directly improving the embedding generation procedure. Particularly, we utilize the following three modules: submodular neighbor selection to find a subset of diverse neighbors to aggregate for each GNN node, layer attention to assign attention weights for each layer, and loss reweighting to focus on the learning of items belonging to long-tail categories. Blending the three modules into GNN, we present DGRec(Diversified GNN-based Recommender System) for diversified recommendation. Experiments on real-world datasets demonstrate that the proposed method can achieve the best diversity while keeping the accuracy comparable to state-of-the-art GNN-based recommender systems.

1 INTRODUCTION

The paper frames diversified recommendation as a challenge for GNN-based systems, which can favor popular items and suffer from over-smoothing. DGRec addresses these issues through diversified neighbor selection, layer attention, and loss reweighting, aiming to improve diversity while limiting accuracy costs.

  • Motivation: Accuracy-only recommendation can create filter bubbles, while diversification seeks dissimilar recommendations that reflect users’ varied interests.An online A/B test reported increased user engagement and average time spent after diversification.
  • GNN-based recommendation: GNN recommenders represent interactions as user-item graphs and generate node embeddings by aggregating neighborhood information.High-order connections can provide personalized distant interests for diversification.
  • Challenges: Directly aggregating neighbors can submerge long-tail items, while stacked GNN layers can cause over-smoothing and reduce accuracy.Item occurrence and category sizes follow power-law distributions, making popular items and categories dominant during training.
  • Proposed approach: DGRec combines submodular neighbor selection, layer attention, and loss reweighting to diversify embeddings and address long-tail learning.The modules respectively select diverse neighbors, stabilize deep-layer training, and emphasize long-tail items and categories.
  • Contributions: The proposed system aims for the best trade-off between accuracy and diversity, with extensive real-world experiments evaluating effectiveness and module influence.The paper states that the modules can be applied to GNN-based methods with a small accuracy cost.

2 PRELIMINARIES

The preliminaries formulate diversified recommendation on a user-item bipartite graph and introduce GNN aggregation and submodular functions. Diversity is defined through category coverage, while submodularity provides a diminishing-returns basis for diversity-related selection.

  • Problem formulation: Diversified recommendation models users, items, categories, and observed interactions through a user-item bipartite graph.An edge connects a user and item when the interaction matrix records an observed interaction.
  • Problem formulation: The task recommends top-k items that are dissimilar to one another, commonly measuring diversity by the number of covered item categories.Category coverage is expressed as |∪ C(i)| over the recommended items.
  • Graph neural networks: A GNN learns node embeddings by aggregating information from neighboring nodes, making connected nodes tend to have similar representations.The general computation uses a neighborhood aggregation function and combines the result with the node’s embedding.
  • Submodular functions: A submodular function is a set function with a diminishing-returns property that can describe diversity among selected elements.The paper notes applications in summarization, sensor placement, training-data selection, and recommendation reranking.

3 METHOD

DGRec modifies GNN embedding generation with diversified neighbor selection, layer attention, and category-based loss reweighting. These modules target redundant neighborhoods, over-smoothing, and the undertraining of long-tail categories.

  • Overall Framework: DGRec combines submodular neighbor selection, layer attention, and loss reweighting within a GNN-based recommender.The framework applies these modules during embedding generation and training.
  • Submodular Neighbor Selection: Submodular neighbor selection chooses a diverse, representative neighbor subset instead of aggregating all neighbors.It uses facility location over embedding similarities and selects up to k neighbors greedily.
  • Layer Attention: Layer attention combines embeddings from different GNN layers to exploit high-order neighbors while alleviating over-smoothing.The attention mechanism learns different weights for layer outputs.
  • Loss Reweighting: Loss reweighting increases the relative training emphasis on long-tail categories and decreases it for popular categories.The weights are based on category size, with larger β further reducing weights for popular categories.

4 EXPERIMENT

The experiments evaluate DGRec on real-world datasets through four research questions covering comparative performance, hyper-parameters, module necessity, and submodular-function choices.

  • Research Questions: The experiments ask whether DGRec outperforms existing diversified recommendation methods.
  • Research Questions: The study examines how hyper-parameters affect DGRec and the accuracy-diversity trade-off.
  • Research Questions: The study tests whether all three DGRec components are necessary for diversification.
  • Research Questions: The study compares the influence of different submodular functions.

4.1 Experimental Setup

DGRec is evaluated on TaoBao and Beauty using held-out test data, representative recommender baselines, accuracy and coverage metrics, and reported Top-100 and Top-300 retrieval results.

  • Datasets: DGRec is evaluated on two real-world datasets with category information: TaoBao and Beauty.TaoBao uses a 10-core setting, while Beauty uses a 5-core version.
  • Data Splits: Both datasets use random splits of 60% training, 20% validation, and 20% testing.Validation data supports hyper-parameter tuning and early stopping; final results use the test set.
  • Comparison Design: The study compares DGRec with baselines rather than re-ranking methods because DGRec is compatible with DPP, MMR, DUM, and Diversified PMF.
  • Baselines: The baselines include Popularity, MF-BPR, GCN, LightGCN, and DGCN.DGCN is identified as the current state-of-the-art diversified GNN recommender in the comparison.
  • Metrics: Recall and Hit Ratio measure accuracy, while Coverage measures diversity by counting covered recommendation categories.Results are reported for Top-100 and Top-300 retrieval.

4.2 Performance Evaluation (RQ1)

DGRec generally achieves the strongest diversity while maintaining accuracy close to the second-best methods. It also provides the best reported accuracy-diversity trade-off and surpasses DGCN across metrics.

  • Diversity: DGRec generally achieves the best Coverage@100 and Coverage@300, except for Coverage@300 on Beauty, where Popularity ranks higher.On Beauty, DGRec is only slightly below Popularity for Coverage@300.
  • Accuracy-Diversity Trade-off: LightGCN achieves the best Recall and Hit Ratio but the lowest Coverage, indicating a poor accuracy-diversity balance.
  • Accuracy-Diversity Trade-off: DGRec has similar Recall and Hit Ratio to the second-best method while achieving the best Coverage.The reported result indicates a small accuracy cost for increased diversity.
  • Baseline Comparison: DGRec surpasses DGCN on all reported metrics.The paper reports this as superiority over the state-of-the-art diversified GNN baseline in both accuracy and diversity.
  • Trade-off Visualization: DGRec occupies the upper-right position in the TaoBao accuracy-diversity comparison using Recall@300 and Coverage@300.

4.3 Parameter Sensitivity (RQ2)

DGRec’s parameter studies examine layer combination, loss reweighting, and submodular-selection settings as controls for the accuracy–diversity trade-off. Layer attention and long-tail-focused loss reweighting preserve diversity while limiting accuracy degradation.

  • Layer number: Layer attention avoids the sharp Recall@300 decline seen with mean aggregation as layers increase, while maintaining high Coverage@300.At comparable Recall@300 or Coverage@300, layer attention provides the stronger accuracy–diversity balance.
  • Loss reweighting: Increasing β lowers accuracy while increasing diversity by focusing training more on items in long-tail categories.At β=0.95, DGRec achieves Coverage@300 above 105 and Recall@300 above 0.086.
  • Submodular selection: DGRec remains stable across σ values from 0.01 to 100, although the best Coverage@300 at σ=10 coincides with the worst Recall@300.The σ parameter controls pairwise neighbor-similarity computation.
  • Submodular selection: Increasing k raises Coverage@300 and lowers Recall@300 because submodular selection aggregates a larger, more diverse neighbor set.Recall@300 does not drop substantially as diversity increases.
  • Submodular selection: The submodular-selection module is not highly sensitive and allows accuracy and diversity to be balanced through σ and k.The experiments report no dramatic performance change from this module.

4.4 Ablation Study (RQ3)

The ablation study shows that DGRec’s three modules jointly improve diversity and balance it against accuracy. Submodular neighbor selection adds diversity at minimal accuracy cost, while loss reweighting has the largest overall impact.

  • Overall ablation: The intact DGRec achieves the best C@300, confirming that combining the proposed modules increases diversity.The ablation is conducted on the TaoBao dataset.
  • Submodular neighbor selection: Removing submodular neighbor selection lowers C@300 from 89.1684 to 84.9129 with only tiny changes in Recall@300 and HR@300.This isolates a diversity gain with minimal accuracy cost.
  • Layer attention: Removing layer attention decreases C@300 while increasing R@300 and HR@300, indicating its role in balancing accuracy and diversity.The ablation therefore trades diversity for higher accuracy when layer attention is removed.
  • Loss reweighting: Removing loss reweighting causes large declines in R@300, HR@300, and C@300, making it the most influential module in the ablation.The paper attributes its impact to balancing long-tail-category training and guiding layer-attention learning.

4.5 Choice of Submodular Functions (RQ4)

The choice of submodular function substantially changes the accuracy–diversity profile. DGRec selects facility location because it nearly maximizes diversity without requiring category information during aggregation.

  • Function comparison: Bucket coverage achieves much higher Recall@300 but much lower Coverage@300 than category coverage and facility location.Thus, the selection function materially changes the model’s accuracy–diversity outcome.
  • Function comparison: Category coverage and facility location produce similar Recall@300 and Coverage@300 results.The paper interprets this similarity as evidence that facility location captures category information through the learned embedding.
  • Function choice: DGRec uses facility location because it nearly achieves the best diversity and does not require category information during aggregation.This broadens applicability when category information is unobserved.

5 RELATED WORK

Related work spans GNN-based recommendation and diversified recommendation. DGRec extends GNN recommenders at the retrieval stage by adding embedding-generation modules that address their diversity shortcoming.

  • GNN-based recommender systems: GNN-based recommenders represent user interactions as a bipartite graph and exploit high-order connectivity through graph-based aggregation.Prior systems include GCMC, SpectralCF, PinSAGE, NGCF, LightGCN, and UltraGCN.
  • GNN-based recommender systems: Earlier GNN-based recommenders primarily optimize accuracy, which can result in poor diversity.DGRec is built on GNNs and adds three modules to address this limitation.
  • Diversified recommendation: Diversified recommendation methods include retrieval-time selection, propagation-based methods, re-ranking, submodular selection, distance-based losses, and determinantal point processes.The cited methods pursue diversity through different stages or objectives.
  • DGRec’s position: DGRec diversifies GNN-based recommendation during retrieval, while methods such as DPP and DUM operate through compatible re-ranking procedures.DGCN is identified as the most similar prior work because both target diversity in GNN methods.

6 CONCLUSIONS

DGRec diversifies GNN-based recommender systems through diversified embedding generation while preserving comparable accuracy. Its hyperparameters allow users to adjust the accuracy–diversity trade-off.

  • DGRec targets diversified embedding generation for GNN-based recommender systems with minimal accuracy cost.
  • DGRec surpasses the state-of-the-art diversified recommender system on diversity while achieving comparable accuracy to the most advanced accuracy-based recommender system.
  • Several hyperparameters enable DGRec to trade off accuracy and diversity, and experiments on real-world datasets examine the influence of its modules.
Loading 2211.10486v2…