PaperlayerView exact revision on arXiv

Source-linked AI summary

Joint Item Recommendation and Attribute Inference: An Adaptive Graph Convolutional Network Approach

Le Wu, Yonghui Yang, Kun Zhang, Richang Hong, Yanjie Fu, Meng Wang

arXiv:2005.12021v1cs.IRcs.LG

TL;DR

Missing user and item attributes complicate recommendation and attribute inference, while prior work may assume complete attributes or use simple joint models. AGCN alternates graph embedding learning with updates to estimated attributes, and experiments report effectiveness on both tasks across three real-world datasets.

  • Problem

    User and item attributes are often incomplete, while prior approaches may assume completeness or model recommendation and attribute inference with simple methods.

    Source: §1 · p. 2, §3 · p. 3

  • Method

    AGCN jointly models recommendation and attribute inference by iterating graph embedding learning with approximated attributes and updates to missing attribute values.

    Source: §6 · pp. 9–10

  • Results

    AGCN improves over all baselines across three datasets and different evaluation metrics; on Amazon-Video Games, it improves about 8% on HR@10 and 7% on NDCG@10 over PinNGCF.

    Source: §5.2.2 · p. 8

  • Takeaways & Limitations

    Experiments on three real-world datasets show the effectiveness of AGCN for joint item recommendation and attribute inference.

    Source: §6 · pp. 9–10

Abstract

from arXiv · show

In many recommender systems, users and items are associated with attributes, and users show preferences to items. The attribute information describes users'(items') characteristics and has a wide range of applications, such as user profiling, item annotation, and feature-enhanced recommendation. As annotating user (item) attributes is a labor intensive task, the attribute values are often incomplete with many missing attribute values. Therefore, item recommendation and attribute inference have become two main tasks in these platforms. Researchers have long converged that user (item) attributes and the preference behavior are highly correlated. Some researchers proposed to leverage one kind of data for the remaining task, and showed to improve performance. Nevertheless, these models either neglected the incompleteness of user (item) attributes or regarded the correlation of the two tasks with simple models, leading to suboptimal performance of these two tasks. To this end, in this paper, we define these two tasks in an attributed user-item bipartite graph, and propose an Adaptive Graph Convolutional Network (AGCN) approach for joint item recommendation and attribute inference. The key idea of AGCN is to iteratively perform two parts: 1) Learning graph embedding parameters with previously learned approximated attribute values to facilitate two tasks; 2) Sending the approximated updated attribute values back to the attributed graph for better graph embedding learning. Therefore, AGCN could adaptively adjust the graph embedding learning parameters by incorporating both the given attributes and the estimated attribute values, in order to provide weakly supervised information to refine the two tasks. Extensive experimental results on three real-world datasets clearly show the effectiveness of the proposed model.

1 INTRODUCTION

Recommendation models often assume complete attributes, although real user and item attributes are frequently missing. AGCN jointly handles item recommendation and attribute inference by iteratively learning graph embeddings and updating estimated attributes; the reported results show gains on both tasks.

1 INTRODUCTION

  • Incomplete attributes challenge methods that assume complete features, while missing values also make attribute inference a central task in recommendation platforms.Attribute inference can support user or item search, profiling, and semantic understanding.Source: §1 · p. 2
  • Joint modeling is motivated by correlations between attributes and user-item behavior, but earlier joint approaches relied on shallow graph models and remained unsatisfactory.Attributes can be used in recommendation, and behavior can reflect attributes; modeling both tasks together may allow them to reinforce each other.Source: §1 · p. 2
  • About 7% improvement for item recommendation and more than 10% for attribute inference versus the best baselines are reported across three real-world datasets.The introduction presents these results as evidence of AGCN's effectiveness for both tasks.Source: §1 · p. 2
  • AGCN alternates graph-parameter learning using estimated attributes with updates that send newly approximated attributes back into the graph.This adaptive process incorporates given and estimated attributes into embedding learning for both tasks.Source: §1 · p. 2

2 RELATED WORK

Prior work addresses recommendation with attributes, graph convolution, or attribute inference, while joint methods have also combined recommendation and inference. This paper positions its approach as jointly modeling collaborative signals and incomplete node attributes.

2 RELATED WORK

  • Attribute-enhanced collaborative filtering models use user and item features, but assume attributes are complete and often require preprocessing to fill missing values.This paper instead learns attribute inference and item recommendation together.Source: §2 · pp. 2–3
  • Graph-based recommendation models use user-item structure for embedding learning, but prior GCN recommendation work generally focuses on collaborative filtering or content-based recommendation.The paper distinguishes its work by jointly propagating collaborative signals and node attributes with missing values.Source: §2 · p. 3
  • Attribute inference research includes label propagation, graph-manifold regularization, and GCNs, with prior approaches addressing graph or attribute correlations in different ways.The paper adopts GCNs for semi-supervised inference in the user-item behavior graph.Source: §2 · p. 3
  • Joint recommendation and attribute-inference approaches report gains over single-task modeling, but rely on shallow graph or random-walk methods.The related work says these methods do not well encode complex global structure and can have high time complexity.Source: §2 · p. 3

3 PROBLEM DEFINITION

The problem is defined over users and items, their preference matrix, and partially observed user and item attributes. The two goals are predicting preferences for unrelated items and inferring missing attributes.

3 PROBLEM DEFINITION

  • The setting includes user and item sets, an implicit-feedback preference matrix, and user and item attribute matrices.User-item connections are represented by entries in the rating matrix.Source: §3 · p. 3
  • Attribute indication matrices mark whether user and item attribute values are available, distinguishing observed values from missing ones.The user-side indicator uses 1 for an available attribute and 0 for a missing attribute.Source: §3 · p. 3
  • Item recommendation predicts preferences for unrelated items, while attribute inference predicts missing user or item attribute values from the attributed graph.The formulation denotes the predicted rating matrix and predicted user and item attribute matrices separately.Source: §3 · p. 3

4 THE PROPOSED MODEL

AGCN alternates graph learning and attribute updates to jointly support recommendation and attribute inference with incomplete attributes. Its graph module combines attribute and collaborative representations, propagates them over the bipartite graph, and feeds inferred missing values back into training.

4.1 Overall Architecture

  • AGCN iterates graph learning with current approximated attributes and attribute updates that return newly inferred values to the graph until convergence.The proposed model comprises graph-learning and attribute-update modules.Source: §4.1 · pp. 3–4, §4 · p. 3

4.2 Graph Learning Module

  • The graph-learning module fuses free user and item embeddings with approximated attribute embeddings, then propagates fused representations to capture higher-order graph structure.The fusion combines collaborative and content signals; propagation uses linear operations without nonlinear activations.Source: §4.2 · pp. 4–5
  • Each propagation layer combines a node's fusion embedding with connected users' or items' embeddings from the preceding layer.The matrix form updates all user and item fusion matrices simultaneously; the degree matrix smooths clustered neighbors' embeddings.Source: §4.2 · pp. 4–5

4.3 Attribute Update Module

  • The attribute-update module predicts preferences and attributes from final propagated embeddings, then replaces missing attribute values while preserving available values.Updated user and item attributes return to the graph-learning module for the next iteration.Source: §4.3 · p. 5

4.4 Model Optimization

  • The objective combines pairwise-ranking recommendation loss and cross-entropy attribute-inference loss, balanced by parameter γ.Training repeats the attribute-update process with Adam until convergence rather than using an end-to-end process.Source: §4.4 · p. 5

5.1 Experimental Settings

Experiments use Amazon-Video Games, Movielens-1M, and Movielens-20M, with attributes encoded according to label type and 90% of attribute values randomly deleted for evaluation.

5.1.1 Datasets Description.

  • Amazon-Video Games, Movielens-1M, and Movielens-20M provide the three evaluation datasets, covering product attributes, user attributes, and movie genres.The Amazon dataset uses price, platform, and theme; Movielens-1M uses gender, age, and occupation; Movielens-20M uses genres.Source: §5.1.1 · p. 6
  • Price is treated as a 10-interval single-label attribute, while platform, theme, and movie genres are multi-label attributes.Single-label attributes use one-hot encoding and multi-label attributes use binarization encoding.Source: §5.1.1 · p. 6
  • Historical interactions are split 8:1:1 into training, validation, and test sets, and 90% of attribute values are randomly deleted for testing.Users with fewer than five rating records are filtered out.Source: §5.1.1 · p. 6

5.1.2 Parameter Settings.

  • AGCN uses embedding size 32 and regularization parameter 0.01 after testing the listed alternatives.Embedding sizes tested were 16, 32, 64, and 128; regularization values tested were 0.1, 0.01, 0.001, and 0.0001.Source: §5.1.2 · p. 6

5.2 Performance on Item Recommendation

AGCN improves item-recommendation performance over the baselines across three datasets and evaluation metrics. Its advantage is especially pronounced among users with sparse interaction histories, where graph-based models show larger improvements.

5.2.1 Baselines and Evaluation Metrics.

  • The evaluation compares recommendation baselines including BPR, FM, NGCF, PinNGCF, and BLA, using HR@N and NDCG@N for top-N ranking.HR counts relevant items in the top-N list; NDCG rewards relevant items ranked higher.Source: §5.2.1 · pp. 6–7

5.2.2 Overall Performance.

  • Graph-based NGCF and PinNGCF outperform behavior-only BPR, PinNGCF outperforms FM, and GCN-based baselines outperform BLA.The authors attribute GCNs’ performance relative to BLA to their use of higher-order graph structure for user and item embeddings.Source: §5.2.2 · pp. 7–8
  • On Amazon-Video Games, AGCN improves about 8% over PinNGCF on HR@10 and 7% on NDCG@10, and improves over all baselines across three datasets and metrics.The overall comparisons are reported for HR@N and NDCG@N.Source: Table 2 · p. 7, Table 3 · p. 7, §5.2.2 · p. 8

5.2.3 Sparsity Analysis.

  • Across user groups, recommendation performance rises with more training records, while graph-based models show larger improvements for very sparse users.The reported group comparison uses NDCG@10; the first-group Amazon comparisons are reported against BPR for FM, NGCF, PinNGCF, and AGCF.Source: Figure 2 · p. 7, §5.2.3 · p. 8

5.3 Performance on Attribute Inference

This section evaluates attribute inference against four baselines and reports AGCN's best performance across the tested attributes and datasets. It also describes differences among the baselines and the metrics used.

5.3.1 Baselines and Evaluation Metrics.

  • ACC evaluates single-label attribute inference and MAP evaluates multi-label inference; larger values indicate better performance.Because of data limits, each dataset does not support both user- and item-attribute inference.Source: §5.3.1 · p. 8

5.3.2 Overall Performance.

  • AGCN exceeds the strongest baseline by 4.8%, 28.0%, and 11.6% on Amazon-Video Games price, platform, and theme attributes.The reported comparison is against the strongest baseline for each attribute.Source: §5.3.2 · p. 8
  • The model improves over the strongest baseline by 5.1%, 13.8%, and 16.4% on gender, age, and occupation attributes, respectively; on Movielens-20M, it improves by 6.6% on genres.Source: §5.3.2 · p. 8
  • AGCN improves over the strongest baseline by 6.6% on the Movielens-20M genres attribute.Source: §5.3.2 · p. 8
  • GR and Semi-GCN outperform LP in most situations, while LP is the strongest baseline for Amazon-Video Games price prediction.The authors suggest that price may be less correlated with other attributes, making their use noisy for inference.Source: §5.3.2 · p. 9
  • The comparison includes LP, GR, Semi-GCN, and BLA; BLA is generally the strongest baseline, and AGCN improves further.The paper describes BLA as jointly modeling the two tasks; AGCN is reported to better model complex graph structure.Source: §5.3.1 · p. 8, §5.3.2 · p. 9

5.4 Detailed Model Analysis

The analysis examines propagation depth and the balance between recommendation and attribute-inference losses. Best depth and task balance vary across datasets and tasks.

5.4.1 The Propagation Layer Depth

  • Best propagation depth is K = 3 on Amazon-Video Games and K = 2 on Movielens-1M and Movielens-20M.Performance initially improves as higher-order neighbors are included; K = 4 decreases performance on all datasets, which the authors attribute to overfitting.Source: §5.4.1 · p. 9
  • The authors attribute early gains to first-order neighbors alleviating sparsity, while excessive depth can cause over-smoothing.Amazon-Video Games has 0.028% rating density; the paper links its greater sparsity to a benefit from more neighbor aggregation.Source: §5.4.1 · p. 9

5.4.2 Task Balance Weight

  • Joint modeling improves both tasks over the independent-loss setting as γ rises from 0 to 0.001, but performance often drops at γ = 0.1 or 1.Different tasks and attributes reach their best performance at different γ values, so the paper selects γ separately for each task.Source: §5.4.2 · p. 9

6 CONCLUSION

The conclusion presents AGCN for joint item recommendation and attribute inference on attributed user-item graphs with missing attribute values. Its iterative attribute updates provide weak supervision, and experiments on three real-world datasets show effectiveness.

6 CONCLUSION

  • AGCN jointly addresses item recommendation and attribute inference in an attributed user-item bipartite graph with missing attributes.Source: §6 · p. 9
  • AGCN iteratively learns graph embeddings using previously learned attribute values, then updates the graph input with newly approximated attributes.Source: §6 · p. 10
  • The authors report that experiments on three real-world datasets show AGCN's effectiveness.Source: §6 · p. 10

Related papers

Loading 2005.12021v1…