PaperlayerView exact revision on arXiv

Source-linked AI summary

CrossCBR: Cross-view Contrastive Learning for Bundle Recommendation

Yunshan Ma, Yingzhi He, An Zhang, Xiang Wang, Tat-Seng Chua

arXiv:2206.00242v5cs.IR

TL;DR

Bundle recommendation draws on complementary bundle and item views, but prior approaches merge the views or combine their predictions without adequately modeling cooperation between their representations. CrossCBR uses cross-view contrastive learning to align those views, and the authors report large-margin gains over SOTA baselines on three public datasets. The paper also identifies future directions for contrastive learning and cross-view cooperation.

  • Problem

    Bundle and item views capture complementary information, but prior methods merge them or combine view-specific predictions without adequately modeling cooperative association between their representations.

    Source: §1 · p. 2

  • Method

    CrossCBR learns bundle- and item-view representations and applies cross-view contrastive learning to model their cooperation.

    Source: §1 · p. 2, §2.2.2 · p. 3

  • Results

    CrossCBR outperforms SOTA baselines by a large margin on three public datasets.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The paper concludes that cross-view alignment and dispersion can mutually enhance view-aware representations and their discriminative power.

    Source: §2.5 · p. 5

    The authors describe contrastive learning for bundle or general recommendation as still in its infancy and identify model-based augmentation and alternative ways to model cross-view cooperation as future directions.

    Source: §5 · p. 9

Abstract

from arXiv · show

Bundle recommendation aims to recommend a bundle of related items to users, which can satisfy the users' various needs with one-stop convenience. Recent methods usually take advantage of both user-bundle and user-item interactions information to obtain informative representations for users and bundles, corresponding to bundle view and item view, respectively. However, they either use a unified view without differentiation or loosely combine the predictions of two separate views, while the crucial cooperative association between the two views' representations is overlooked. In this work, we propose to model the cooperative association between the two different views through cross-view contrastive learning. By encouraging the alignment of the two separately learned views, each view can distill complementary information from the other view, achieving mutual enhancement. Moreover, by enlarging the dispersion of different users/bundles, the self-discrimination of representations is enhanced. Extensive experiments on three public datasets demonstrate that our method outperforms SOTA baselines by a large margin. Meanwhile, our method requires minimal parameters of three set of embeddings (user, bundle, and item) and the computational costs are largely reduced due to more concise graph structure and graph learning module. In addition, various ablation and model studies demystify the working mechanism and justify our hypothesis. Codes and datasets are available at https://github.com/mysbupt/CrossCBR.

1 INTRODUCTION

CrossCBR addresses the overlooked cooperation between complementary bundle and item views by aligning their representations with cross-view contrastive learning. It reports large-margin gains over SOTA baselines on three public datasets, alongside reduced training time.

1 INTRODUCTION

  • The bundle view emphasizes behavioral similarity among users, while the item view captures bundle content relatedness and users’ item-level preferences.The paper identifies these views as complementary but different.Source: §1 · p. 2
  • Prior approaches either merge the views into one graph or combine view-specific predictions, leaving their cooperative association at the representation level insufficiently modeled.CrossCBR is proposed to encourage mutual enhancement between the views.Source: §1 · p. 2
  • CrossCBR uses cross-view contrastive learning to model cooperation between separately learned bundle- and item-view representations.The method jointly optimizes BPR and contrastive losses.Source: §1 · p. 2
  • CrossCBR outperforms SOTA baselines by a large margin on three public datasets.The authors also report largely reduced training time.Source: §1 · p. 2
  • The paper examines alignment and dispersion as mechanisms for mutual enhancement and stronger representation discrimination.It also presents the method as having minimal space complexity and low time complexity.Source: §1 · p. 2

2 METHODOLOGY

The methodology section introduces the bundle-recommendation task and CrossCBR framework, then previews discussion of its mechanism and computational complexity.

2 METHODOLOGY

  • The section first formulates bundle recommendation and presents CrossCBR’s overall framework.The framework has representation learning for two views and joint optimization of BPR and contrastive losses.Source: §2 · p. 2
  • CrossCBR’s framework comprises representation learning for the two user-and-bundle views.The cited framework caption identifies this as its first component.Source: §2 · p. 2
  • The section follows with discussion of CrossCBR’s working mechanism and computational complexity.Source: §2 · p. 2

2.1 Problem Formulation

The task uses historical user-bundle, user-item, and bundle-item data to predict unseen user-bundle interactions. User-bundle and user-item histories are binary and separately generated.

2.1 Problem Formulation

  • The task is to learn from historical user-bundle, user-item, and bundle-item data and predict unseen user-bundle interactions.Source: §2.1 · p. 2
  • User-bundle interactions, user-item interactions, and bundle-item affiliations are represented by binary matrices X, Y, and Z.A value of 1 indicates an interaction or that an item belongs to a bundle.Source: §2.1 · p. 2
  • The user-bundle and user-item matrices are separately generated and deduplicated, so their entries are binary and the two matrices contain different information.Source: §2.1 · p. 2

2.2 Learning of Two Views’ Representations

CrossCBR learns user and bundle representations from distinct bundle and item views using simplified graph-learning designs. The item view builds on user-item and bundle-item graphs, with item representations pooled to form bundle representations.

2.2 Learning of Two Views’ Representations

  • CrossCBR learns representations from two views using a simpler graph-learning approach than the cited BGCN design.The framework represents users and bundles in both views and jointly optimizes BPR and contrastive losses.Source: §2.2 · p. 2, Figure 2 · p. 3

2.2.1 Bundle-view Representation Learning.

  • The bundle view applies LightGCN to a user-bundle bipartite graph, removing self-connections and nonlinear transformations.It also omits bundle-bundle connections, whose overlap information can be distilled from the item view, reducing graph-learning costs.Source: §2.2.1 · p. 3

2.2.2 Item-view Representation Learning.

  • The item view builds user-item and bundle-item bipartite graphs, learns user and item representations with LightGCN, and averages item representations to obtain bundle representations.The bundle pooling follows the bundle-item graph.Source: §2.2.2 · p. 3
  • The item-view design shares initial user parameters across views and aggregates information after K propagation layers.The authors report that sharing does not affect performance and reduces parameter count.Source: §2.2.2 · p. 3

2.3 Cross-view Contrastive Learning

Cross-view contrastive learning models cooperation between bundle and item views by aligning representations of the same users and bundles while separating different ones. The section also describes graph- and embedding-based augmentation alongside an original-preservation setting.

2.3 Cross-view Contrastive Learning

  • Contrastive learning uses naturally available views when possible; augmentation can supply views otherwise and may improve robustness to noise.The paper keeps original preservation as its default and also presents graph- and embedding-based augmentation methods.Source: §2.3.1 · p. 4, §2.3 · p. 4

2.3.1 Data Augmentation.

  • Graph-based edge dropout randomly removes a proportion of graph edges, while embedding-based message dropout masks propagated-embedding elements.Both methods use dropout ratio ρ.Source: §2.3.1 · p. 4
  • Original preservation uses no augmentation or added randomness, relying on the distinct data sources of the two views.The method retains only the original representations.Source: §2.3.1 · p. 4

2.3.2 Cross-view Contrastive Loss.

  • InfoNCE cross-view loss aligns the same user or bundle across views and separates representations of different users or bundles.The user and bundle losses are averaged; cosine similarity and a temperature parameter are used, with in-batch negative sampling.Source: §2.3.2 · p. 4

2.4 Prediction and Optimization

The method combines item-view and bundle-view predictions additively. It trains with BPR loss, weighted together with contrastive loss and L2 regularization.

2.4 Prediction and Optimization

  • Item-view and bundle-view predictions are calculated by inner products and added to produce the final recommendation prediction.The prediction combines the two views additively.Source: §2.4 · p. 4
  • The conventional Bayesian Personalized Ranking (BPR) loss is used as the main loss.The objective also includes contrastive loss and L2 regularization.Source: §2.4 · p. 4
  • The BPR, contrastive, and L2 regularization terms are weighted using hyperparameters λ1 and λ2.Θ denotes the model parameters.Source: §2.4 · p. 4

2.5 Model Discussion

The cross-view contrastive loss aligns representations of the same user or bundle and separates different pairs. The discussion links alignment to mutual enhancement and to greater within-view separation.

2.5 Model Discussion

  • Aligning the same user or bundle across views lets each view distill distinctive information from the other, enhancing both representations and the combined prediction.The discussion describes the contrastive loss as an additional regularizer for representation learning.Source: §2.5 · pp. 4–5
  • The loss directly pulls same-user cross-view pairs together and pushes cross-user pairs apart, indirectly enlarging distances between different users within each view.The text describes the direct and indirect effects for users; Figure 3 illustrates these effects.Source: §2.5 · p. 5, Figure 3 · p. 5
  • Cross-view dispersion alone, without alignment, cannot produce ego-view dispersion or enhance representation self-discrimination.The discussion connects the loss's alignment and dispersion effects to improved bundle recommendation quality.Source: §2.5 · p. 5

2.6 Complexity Analysis

CrossCBR uses three embedding sets and has lower graph-learning time complexity than BGCN. Its contrastive loss has cubic complexity in batch size, though its practical training time is described as limited.

2.6 Complexity Analysis

  • CrossCBR uses three embedding sets and has space complexity O((M+N+O)d), with fewer parameters than BGCN because it removes feature transformation matrices.The embedding sets are for users, bundles, and items.Source: §2.6 · p. 5
  • CrossCBR's graph-learning time complexity is lower than BGCN's because it removes feature transformations and uses fewer edges.The edge reduction comes from removing self-connections and bundle-bundle connections.Source: §2.6 · p. 5
  • The contrastive loss has cubic complexity O(dT^2), but its practical training time is described as very limited; inference has the same time complexity as BGCN.The practical-time statement is attributed to GPU hardware and optimized libraries.Source: §2.6 · p. 5

3 EXPERIMENTS

Experiments across Youshu, NetEase, and iFashion compare CrossCBR with recommendation baselines and probe its components, view behavior, hyperparameters, and efficiency. The results report gains over baselines and examine how contrastive learning affects performance and representations.

3.1 Experimental Settings

  • Experiments use Youshu, NetEase, and iFashion with Recall@K and NDCG@K for K=20 or 40, comparing general user-item recommenders with bundle-specific baselines.The datasets are randomly split 70%/10%/20% for training, validation, and testing; compared methods include MFBPR, LightGCN, SGL, DAM, BundleNet, and BGCN.Source: §3 · p. 5, §3.1 · p. 5, §3.1.1 · pp. 5–6

3.2 Performance Comparison (RQ1)

  • CrossCBR significantly outperforms the strongest baselines across three datasets, including SGL and bundle-specific methods, and performs consistently across book, music, and fashion scenarios.The overall comparison is reported in Table 2.Source: Table 2 · p. 6, §3.2 · p. 6

3.3.1 Effectiveness of Cross-view Contrastive Learning.

  • Removing the cross-view contrastive loss causes a severe performance decline, while alignment and dispersion together outperform either signal alone; dispersion alone can yield little benefit or collapse.CrossCBR-CL still outperforms BGCN, which the authors attribute to CrossCBR’s LightGCN kernel.Source: §3.3.1 · pp. 6–7, Table 3 · p. 6
  • Performance differences among original preservation, edge dropout, and message dropout are negligible relative to CrossCBR’s gains over baselines.The authors interpret this as evidence that the two views’ original data already provide sufficient variance for cross-view contrastive learning.Source: §3.3.2 · p. 7

3.3.2 Effectiveness of Data Augmentations.

  • Adding self-connections or bundle-bundle links contributes little or no performance benefit, while those links add training cost.The authors state that self-connections add no information with LightGCN, and bundle overlap can be distilled from the item view through cross-view alignment.Source: §3.3.3 · p. 7

3.4.1 Mutual Enhancement Effect.

  • Contrastive learning significantly improves bundle-, item-, and both-view predictions, while bundle-view predictions outperform item-view predictions.Bundle-view prediction is even slightly better than both-view prediction on NetEase.Source: Figure 4 · p. 7, §3.4.1 · p. 7

3.4.2 Cross-View Alignment and Dispersion Analysis.

  • Contrastive loss increases cross-view alignment and reduces user-representation dispersion, improving user discrimination; bundle-item structure also informs item-view bundle dispersion that is distilled to the bundle view.Table 4 reports the alignment and dispersion measures for users and bundles across views.Source: Table 4 · p. 7, §3.4.2 · pp. 7–8

3.5 Hyper-parameter and Computational Efficiency Analysis (RQ4)

  • Performance is sensitive to temperature, while increasing batch size from 128 to 8192 yields quick gains followed by a plateau; batch size 2048 performs sufficiently well.CrossCBR’s original-preservation variant is more efficient than BGCN, the contrastive loss adds negligible cost, and bundle-bundle links cost about three times Cr_OP on NetEase.Source: Figure 5 · p. 8, §3.5.1 · p. 8, Table 5 · p. 8

4 RELATED WORK

The related work reviews graph-based and bundle recommendation alongside contrastive learning in recommendation. It positions CrossCBR within two-view modeling and cross-view learning.

Graph-based and Bundle Recommendation

  • LightGCN removes redundant modules from NGCF and serves as the backbone for CrossCBR.Graph-based recommendation models use graph neural networks to model higher-order interactions.Source: §4 · p. 8
  • Bundle recommendation models range from bundle IDs to methods that use user-item interactions and bundle-item affiliations.BundleNet mixes user, bundle, and item relations, whereas BGCN decomposes preferences into item and bundle views; set, basket, and bundle-generation tasks differ from recommending pre-defined themed bundles.Source: §4 · p. 8

Contrastive Learning in Recommendation

  • Contrastive recommendation methods construct pairs through data augmentation or multiple data views, and this work combines both approaches.Examples include graph and sequence augmentations, and methods that contrast item/session or modality/user/item views.Source: §4 · pp. 8–9

5 CONCLUSION AND FUTURE WORK

CrossCBR applies cross-view contrastive learning to model cooperative association in bundle recommendation. Its authors report improved performance on three public datasets and identify future directions for the field.

5 CONCLUSION AND FUTURE WORK

  • CrossCBR significantly enhanced state-of-the-art bundle recommendation performance on three public datasets.The authors also report that ablation and model studies elucidate the mechanism behind the performance gain.Source: §5 · p. 9
  • Future work could use model-based augmentations for hard negatives and diverse positives, explore other cross-view association methods, and generalize the paradigm to tasks with two distinctive views.The authors describe contrastive learning for bundle or general recommendation as still in its infancy.Source: §5 · p. 9
Loading 2206.00242v5…