GRCN: Graph-Refined Convolutional Network for Multimedia Recommendation with Implicit Feedback

Wei Yinwei; Wang Xiang; Nie Liqiang; He Xiangnan; Chua Tat-Seng

Source-linked AI summary

GRCN: Graph-Refined Convolutional Network for Multimedia Recommendation with Implicit Feedback

Wei Yinwei, Wang Xiang, Nie Liqiang, He Xiangnan, Chua Tat-Seng

arXiv:2111.02036v1Nov 3, 2021cs.IRcs.MM

TL;DR

Implicit-feedback graphs can contain false-positive interactions that contaminate GCN message propagation and user-preference representations. GRCN adaptively identifies and softly prunes likely noisy edges before graph convolution; experiments on three datasets show effectiveness and outperforming state-of-the-art baselines. The paper also notes that user intent remains insufficiently explored.

Problem
Implicit-feedback interaction graphs may include false-positive behaviors that negatively influence neighborhoods and contaminate user-preference signals during GCN propagation.
Method
GRCN uses item content and historical user behavior to adaptively refine the interaction graph, softly pruning edges with high confidence of being false-positive interactions before graph convolution.
Results
Experiments on three real-world datasets validate GRCN's effectiveness and show that it outperforms state-of-the-art baselines including MMGCN, DisenGCN, and GAT.
Takeaways & Limitations
Refining the user-item graph provides a way to reduce the effect of noisy implicit feedback in GCN-based recommendation.
Takeaways & Limitations
The implicit-feedback problem remains incompletely solved because the motivation behind user behavior, or user intent, is insufficiently explored.

Abstract

from arXiv · show

Reorganizing implicit feedback of users as a user-item interaction graph facilitates the applications of graph convolutional networks (GCNs) in recommendation tasks. In the interaction graph, edges between user and item nodes function as the main element of GCNs to perform information propagation and generate informative representations. Nevertheless, an underlying challenge lies in the quality of interaction graph, since observed interactions with less-interested items occur in implicit feedback (say, a user views micro-videos accidentally). This means that the neighborhoods involved with such false-positive edges will be influenced negatively and the signal on user preference can be severely contaminated. However, existing GCN-based recommender models leave such challenge under-explored, resulting in suboptimal representations and performance. In this work, we focus on adaptively refining the structure of interaction graph to discover and prune potential false-positive edges. Towards this end, we devise a new GCN-based recommender model, \emph{Graph-Refined Convolutional Network} (GRCN), which adjusts the structure of interaction graph adaptively based on status of model training, instead of remaining the fixed structure. In particular, a graph refining layer is designed to identify the noisy edges with the high confidence of being false-positive interactions, and consequently prune them in a soft manner. We then apply a graph convolutional layer on the refined graph to distill informative signals on user preference. Through extensive experiments on three datasets for micro-video recommendation, we validate the rationality and effectiveness of our GRCN. Further in-depth analysis presents how the refined graph benefits the GCN-based recommender model.

1 INTRODUCTION

GRCN addresses noisy false-positive interactions in implicit-feedback graphs by adaptively refining the graph before GCN propagation. Experiments on three real-world datasets support its effectiveness and show why graph refinement benefits recommendation.

Motivation: Implicit-feedback graphs may contain accidental or low-interest interactions that contaminate neighborhood propagation and user-preference signals.False-positive edges are treated like true-positive interactions, allowing noise to affect neighboring representations.
Motivation: False-positive interactions can significantly degrade GCN-based recommendation, even when only a few noisy edges are added.The problem reflects GCN vulnerability to graph-structure perturbations.
Present Work: GRCN adaptively refines the user-item interaction graph to identify and softly prune edges likely to represent false-positive interactions.Its graph-refining layer uses item content and historical user behavior to model preference and refine the graph.
Present Work: The model combines graph refining, graph convolutional, and prediction layers to distill preference signals from the refined graph.The graph-refining component uses a neighbor-routing mechanism to highlight user preference toward item prototypes across modalities.
Empirical Validation: Experiments on three real-world datasets show that GRCN outperforms state-of-the-art baselines including MMGCN, DisenGCN, and GAT.Embedding visualizations provide explanations for how graph-refining operations improve GCN-based recommendation.

2 METHODOLOGY

GRCN adaptively refines the user-item interaction graph before graph convolution, using multimodal content and user preferences to softly downweight noisy edges and enrich node representations.

2.1 Preliminary: GRCN constructs a user-item interaction graph from implicit feedback and represents users and items with trainable node embeddings.
2.2 Model Overview: The model contains graph refining, graph convolutional, and prediction layers that respectively adjust graph structure, propagate information, and infer interactions.
2.2.1 Graph Refining Layer: GRCN assumes false-positive items are distant from user preferences in content space, then uses a prototypical network with iterative neighbor routing to estimate those preferences.
2.2.1 Graph Refining Layer: Multimodal affinity scores and user/item base vectors produce edge weights that softly prune interactions likely to be noisy.
2.2.2 Graph Convolutional Layer: Weighted edges control message passing across stacked graph convolutional layers, aggregating collaborative signals from l-hop neighbors into enriched embeddings.
2.2.3 Prediction Layer: The final user and item representations concatenate multimodal features with enriched ID embeddings, and their inner product estimates preference for each item.

3 EXPERIMENTS

Experiments on three public datasets address GRCN’s comparative performance, component effects, and representation benefits against GCN-based recommendation models.

3 EXPERIMENTS: The experiments evaluate GRCN against state-of-the-art GCN-based recommendation models on three public datasets.They also study the effects of the prototypical network and pruning operations, and how the refined graph benefits learned representations.

3.1 Experiments Settings

The evaluation uses Movielens, Tiktok, and Kwai with implicit-feedback interactions, multimodal item features where available, ranking metrics, and comparable GCN-based baselines.

Datasets: Movielens, Tiktok, and Kwai provide the three evaluation datasets, with multimodal item features available differently across datasets.Kwai lacks supplied audio and textual features, while Movielens and Tiktok include visual, acoustic, and textual information.
Evaluation Protocol: Historical interactions are randomly split per user in an 8:1:1 ratio into training, validation, and testing sets.Negative sampling is used for training, while validation and testing support hyperparameter tuning and performance evaluation.
Evaluation Protocol: Precision@K, Recall@K, and NDCG@K evaluate ranked user–micro-video recommendations against unconsumed items treated as negatives.During validation and testing, candidate interactions are scored and ranked in descending order.
Baselines: Baselines are grouped into message-nonadaptive methods, GraphSAGE and MMGCN, and message-adaptive methods, NGCF, GAT, and DisenGCN.Two graph convolutional layers are used for GRCN and the baselines for fair comparison.
Evaluation Settings: Figure 3 reports Recall@10 and NDCG@10 across different prototypical-network routing iterations on Movielens, Tiktok, and Kwai.The figure examines how iterative preference modeling relates to recommendation performance.

3.2 Performance Comparison (RQ1)

GRCN consistently achieves the best performance across the three datasets, with its gains attributed to refining noisy interaction-graph edges before graph convolution.

Overall Comparison: GRCN consistently achieves the best performance on Movielens, Tiktok, and Kwai.The comparison is against the evaluated state-of-the-art GCN-based recommendation baselines.
Overall Comparison: 9.55%, 17.62%, and 11.56% Recall@10 improvements over the strongest baselines are reported for Movielens, Tiktok, and Kwai, respectively.These values correspond to the three datasets in that order.
Interpretation: The reported improvements could be attributed to graph refining, supporting the value of identifying and pruning noisy interaction-graph edges.The comparison uses similar graph convolutional operations across models.
Interpretation: Message-adaptive methods outperform GraphSAGE, indicating that implicit-feedback graph structure can convey meaningless or harmful signals during message passing and aggregation.The passage frames local graph structure as insufficient to directly reflect user-preference patterns.

3.3 Ablation Study (RQ2)

The ablation studies examine prototypical-network routing and pruning choices, while t-SNE visualizations compare representations learned by GAT and GRCN.

Ablation Design: The ablation study evaluates the prototypical network and pruning operations as two components of GRCN’s graph refining layer.The experiments assess how these designs affect user-preference modeling and recommendation performance.
Effects of Prototypical Network: Increasing routing iterations improves Recall@10 and NDCG@10 on Movielens and Tiktok, but both metrics tend to decline on Kwai.The Kwai decline is suggested to reflect smoother user representations that capture fewer discriminative features.
Effects of Prototypical Network: GRCN outperforms GRCN-ID, while GRCN-ID still exceeds the earlier GCN-based baselines.The comparison supports contributions from preference-enriched user representations and item features, alongside the refined graph.
Effects of Pruning Operations: Multimodal pruning variants generally outperform single-modality variants, and GRCN_max outperforms GRCN_mean in the reported comparisons.The pruning study compares modality-specific variants, multimodal fusion operations, base-value incorporation, and hard pruning.
Representation Visualization: Figure 4 visualizes t-SNE-transformed user and item embeddings learned by GAT and GRCN for randomly selected Tiktok users and their interacted items.Stars mark users, circles mark items, and links represent interactions; GRCN shows more discernible item clustering around users.
Effects of Pruning Operations: GRCN yields the best pruning-variant results because soft pruning weakens noisy messages while enhancing messages from true-positive interactions.This differs from hard pruning, which completely interrupts messages from false-positive interactions.

3.4 Visualization

The visualization compares user and item ID embeddings learned by GAT and GRCN, showing more discernible item clustering around users with GRCN.

The t-SNE visualization compares two-dimensional ID embeddings learned from GAT and GRCN for randomly selected TikTok users and their consumed items.Colors denote edges from each user to interacted items.
GRCN produces discernible clustering of item nodes around user nodes, indicating more discriminative user representations despite shared consumed items.The paper attributes this to graph refinement assigning edge weights according to user-preference and item-content affinities.
Graph refining operations are credited with eliminating noise from false-positive interactions by assigning different weights to edges.

4 RELATED WORK

Related work covers implicit-feedback recommendation and GCN-based personalized recommendation, while identifying noisy implicit interactions as an unresolved issue for graph-based models.

Recommendation with Implicit Feedback: Implicit-feedback recommendation must distinguish positive from negative instances, using unobserved items as negatives or sampling them for pairwise ranking.BPR constructs user–positive-item–negative-item triples, while confidence can quantify implicit feedback.
Recommendation with Implicit Feedback: Multimedia recommenders such as VBPR and ACF integrate item multimedia content with ID embeddings, but primarily use direct user–item interactions.
GCN-based Personalized Recommendation: The proposed approach measures similarity between user preference and item content to discover false-positive feedback in historical records.
GCN-based Personalized Recommendation: GCN-based recommenders model user–item graphs and propagate collaborative or high-order connectivity signals, but existing models ignore implicit-feedback effects.Iterative graph convolution can worsen representation disruption caused by distorted graph structure, motivating graph refinement.

5 CONCLUSION AND FUTURE WORK

The paper proposes a model that refines user–item graphs by identifying and pruning false-positive feedback, with empirical support from three public benchmarks. It also identifies user intent as an important direction for future work.

The proposed Structure-Refined Graph Convolutional Networks model identifies false-positive feedback and prunes corresponding noisy edges in the interaction graph.
Empirical results on three public benchmarks demonstrate the efficiency of the proposed model.
The paper states that implicit feedback remains incompletely solved because user intent, beyond observed preference and behavior, is insufficiently explored.Future work will study how to learn and leverage user intents.