Authors:
Huy-Son Nguyen、Tuan-Nghia Bui、Long-Hai Nguyen、Hoang Manh-Hung、Cam-Van Thi Nguyen、Hoang-Quynh Le、Duc-Trong Le
Paper:
https://arxiv.org/abs/2408.08906
Introduction
In the realm of e-commerce, recommendation systems play a pivotal role in enhancing user experiences and driving business strategies. Traditional recommendation systems often focus on suggesting individual items, but bundle recommendations, which involve suggesting sets of interconnected items, have emerged as a superior strategic marketing tactic. Bundle recommendations can significantly enhance business profitability and user convenience by grouping relevant items together, such as assortments of detective books or complementary items like a phone with a case.
However, recommending bundles based on user preferences poses greater challenges compared to recommending individual items due to data sparsity and the diverse composition of bundles. Existing state-of-the-art (SOTA) methods often overlook the intricate relationships between individual items within bundles, which can be asymmetric in nature. For instance, while blazers and cravats may be frequently purchased together, their influence on each other is not necessarily symmetrical. Understanding these dynamics requires examining item-item relations across diverse user preferences.
To address these challenges, the authors introduce BunCa, a novel bundle recommendation approach that employs item-level causation-enhanced multi-view learning. BunCa aims to provide comprehensive representations of users and bundles through two distinct views: the Coherent View and the Cohesive View. This approach ensures rigorous cohesion in direct user-bundle interactions and captures explicit intents through causation-sensitive relations among items.
Related Work
The field of bundle recommendation has seen various approaches aimed at accurately recommending pre-defined item sets to users. Early approaches, grounded in the Bayesian Personalized Ranking (BPR) framework, focused on users’ past interactions with lists of items and individual items. The DAM model recognized the importance of affiliated items within bundles and optimized both user-item and user-bundle interactions using attention mechanisms and multi-task learning.
Recent works have shifted towards recommending existing bundles, aiming to better align suggested bundles with users’ preferences. Multi-view learning architectures, such as BGCN and CrossCBR, have been employed to capture intricate relations between users, items, and bundles. These models leverage contrastive learning techniques to address challenges related to the inconsistency and integration of information concerning objects.
However, existing models often overlook the asymmetric relationships between items within bundles. MIDGN, for instance, tackles hidden intents inside user preferences and bundle construction but encounters performance and complexity challenges. BundleGT achieves SOTA performance with a Hierarchical Graph Transformer network but still relies on symmetric correlations. Our study introduces concrete and discrete contrastive learning to enhance the consistency and self-discrimination of representations from both views, addressing the limitations of previous works.
Research Methodology
Problem Formulation
The bundle recommendation problem is formalized as follows: Given a set of users ( U ), a set of bundles ( B ), and a set of items ( I ), the objective is to accurately predict unseen user-bundle interactions. The user-bundle interactions, user-item interactions, and bundle-item affiliations are represented as binary-valued matrices ( X ), ( Y ), and ( Z ), respectively.
Graph Construction
Cohesive View
In the Cohesive View, the user co-occurrence matrix ( C_U ) and the bundle co-occurrence matrix ( C_B ) are derived from user-bundle interactions. These matrices capture the hypothesis that users with mutual interactions share similar preferences and that bundles often bought collectively by numerous individuals indicate shared interests. The heterogeneous graph ( \bar{G} ) is constructed to integrate user-bundle interactions with homogeneous correlations, enhancing the propagation of high-order collaborative signals.
Coherent View
In the Coherent View, the bundle-item bipartite graph ( G_{BI} ) and the user-item bipartite graph ( G_{UI} ) are leveraged to aggregate information on item-level representations. The Multi-Prospect Causation Network (MPCNet) is employed to explicitly model asymmetric associations between items, capturing causation-sensitive relationships. The asymmetric causation matrix ( A ) is computed using attention mechanisms, and item representations are enhanced through the aggregation of information from connected items.
Representation Learning
Cohesive View Representation Learning
BunCa adopts the LightGCN operation to encode cohesive representations of users and bundles, excluding self-connections and non-linear transformations in the propagation function. The embeddings of users and bundles are learned through the unification of all embeddings from different layers, integrating information propagated from high-order neighbors.
Coherent View Representation Learning
In the Coherent View, BunCa leverages the asymmetric relations between items based on user preferences (UP sub-view) and bundle construction (BC sub-view). The Multi-Prospect Causation Network (MPCNet) models the causation-sensitive relationships among items, enhancing item representations. These representations are then aggregated to derive the final user and bundle representations through graph-based propagation and mean pooling operations.
Contrastive Learning Module
BunCa employs both discrete and concrete contrastive learning to enhance the consistency and self-discrimination of representations from different views. Discrete contrastive learning minimizes inconsistencies in representations across views, while concrete contrastive learning distinguishes the unique characteristics of each user and bundle.
Prediction and Joint Optimization
The comprehensive representations of users and bundles are derived by capturing augmented information from both views. The interaction probability between a user and a bundle is calculated using the inner-product, and the Bayesian Personalized Ranking (BPR) loss is utilized to enhance probabilities for interacted user-bundle pairs while reducing probabilities for non-interacted pairs. The final loss is a weighted summation of the BPR loss, contrastive loss, and L2 regularization.
Experimental Design
Datasets
The authors conducted extensive experiments on three benchmark datasets: Youshu, NetEase, and iFashion. These datasets capture user interactions with bundles of books, song lists, and outfits, respectively. The statistics of these datasets are shown in Table 1.
Evaluation Metrics
The performance of the methods was evaluated using Recall (R@K) and Normalized Discounted Cumulative Gain (N@K) metrics with ( K \in {10, 20} ). These metrics measure the proportion of test bundles within the top-K ranking list and the normalized discounted cumulative gain scores, respectively.
Baselines
BunCa was compared to three groups of state-of-the-art models: Traditional Bundle Recommendation (BPR, LightGCN, DAM), Multi-view Learning Bundle Recommendation (CrossCBR, BGCN, MIDGN, BundleGT, EBRec), and Distillation Bundle Recommendation (DGMAE).
Implementation Details
BunCa was implemented using PyTorch and evaluated on NVIDIA P100 and T4 GPUs. The initial embedding size was set to 64, and the learning rate was set to ( 1e^{-3} ). Xavier initialization and Adam optimizer were used for trainable parameters. The hyper-parameters were tuned using grid search.
Results and Analysis
Performance Comparison
Table 2 illustrates the performance comparison between BunCa and related baselines in terms of R@K and N@K metrics. The results show that multi-view learning methods consistently outperform traditional techniques. Among the multi-view learning models, BunCa achieves the best performance across all benchmark datasets, demonstrating the effectiveness of integrating discrete and concrete contrastive learning.
Ablation Study
Effects of Learning from Different Views
Table 3 shows the impact of selectively omitting different views (Cohesive View, Coherent View, UP sub-view, BC sub-view) on BunCa’s performance. The results indicate that learning representations from each view encodes valuable information, and excluding any view significantly diminishes performance.
Effect of Contrastive Learning Module
Table 4 examines the effectiveness of the contrastive learning module by removing the entire module, discrete contrastive learning, and concrete contrastive learning. The results emphasize the pivotal role of modeling multi-view information integrated with contrastive learning.
Importance of Asymmetric Causation Matrix
Figure 4 compares the performance of BunCa with different alternatives for the causation matrix. The results validate the significance of asymmetric relationships in real-world scenarios, affirming the hypothesis discussed in the introduction.
Effects of Key Hyper-parameters
Figure 5 depicts the impact of key hyper-parameters ( \beta ) and ( L ) on BunCa’s performance for iFashion. The results show that the optimal settings for iFashion are ( \beta = 0.8 ) and ( L = 5 ).
Qualitative Showcase
Figure 6 illustrates some qualitative examples from the iFashion dataset, highlighting the asymmetric weights that indicate the different effects among items within bundles. These examples emphasize the ability to model key components within bundles via asymmetric relationships, validating the hypothesis of leveraging item-level causation effects for bundle recommendation.
Overall Conclusion
This study presents BunCa, a novel approach for bundle recommendation that emphasizes the significance of asymmetric relationships between items. BunCa leverages item-level causation-enhanced multi-view learning, showcasing significant improvements over existing methods. The architecture of BunCa lies in learning two distinct views: the Coherent View, employing the Multi-Prospect Causation Network for causation-enhanced representation learning, and the Cohesive View, utilizing high-order collaborative signals of user-bundle interaction. Additionally, BunCa integrates concrete and discrete contrastive learning, enhancing the consistency and self-discrimination of representations from both views. The extensive experiments conducted on three benchmark datasets validate the efficacy of BunCa and emphasize the importance of considering asymmetric item relationships for bundle-related tasks.
