Spectral-Based Graph Neural Networks for Complementary Item Recommendation

Abstract: Modeling complementary relationships greatly helps recommender systems to accurately and promptly recommend the subsequent items when one item is purchased. Unlike traditional similar relationships, items with complementary relationships may be purchased successively (such as iPhone and Airpods Pro), and they not only share relevance but also exhibit dissimilarity. Since the two attributes are opposites, modeling complementary relationships is challenging. Previous attempts to exploit these relationships have either ignored or oversimplified the dissimilarity attribute, resulting in ineffective modeling and an inability to balance the two attributes. Since Graph Neural Networks (GNNs) can capture the relevance and dissimilarity between nodes in the spectral domain, we can leverage spectral-based GNNs to effectively understand and model complementary relationships. In this study, we present a novel approach called Spectral-based Complementary Graph Neural Networks (SComGNN) that utilizes the spectral properties of complementary item graphs. We make the first observation that complementary relationships consist of low-frequency and mid-frequency components, corresponding to the relevance and dissimilarity attributes, respectively. Based on this spectral observation, we design spectral graph convolutional networks with low-pass and mid-pass filters to capture the low-frequency and mid-frequency components. Additionally, we propose a two-stage attention mechanism to adaptively integrate and balance the two attributes. Experimental results on four e-commerce datasets demonstrate the effectiveness of our model, with SComGNN significantly outperforming existing baseline models.

• The paper introduces SComGNN, a novel method that isolates low-frequency (relevance) and mid-frequency (dissimilarity) components for complementary item recommendation.
• It employs customized spectral filters and a two-stage attention mechanism to effectively transform spectral data into actionable spatial convolutions.
• Contrastive learning experiments on e-commerce datasets demonstrate significant improvements in Hit Rate and NDCG compared to traditional recommendation models.

Spectral-Based Graph Neural Networks for Complementary Item Recommendation

The paper "Spectral-Based Graph Neural Networks for Complementary Item Recommendation" introduces a novel approach for complementary item recommendation by leveraging spectral-based Graph Neural Networks (GNNs). This approach, termed Spectral-based Complementary Graph Neural Networks (SComGNN), aims to effectively capture and model the distinct attributes of complementarity between items: relevance and dissimilarity. The key innovation is the use of spectral properties to distinguish between low-frequency and mid-frequency components, associating these with relevance and dissimilarity, respectively.

Introduction to Complementary Item Recommendation

Complementary item recommendation is critical for e-commerce platforms, where the objective is to suggest items that enhance or are compatible with a previously purchased item. Conventional recommendations often focus on similarity, leading to substitutable item recommendations. However, complementary items share relevance while also exhibiting dissimilarity, such as an iPhone and AirPods Pro. Thus, capturing both attributes is pivotal to improving recommendation quality.

Figure 1: Item relationships in recommender systems.

Spectral Perspective on Complementary Relationships

The paper identifies that complementary item graphs predominantly consist of low-frequency and mid-frequency components in the spectral domain. Low-frequency components represent relevance, indicating node similarities within the graph. On the contrary, mid-frequency components depict dissimilarity, reflecting the inherent differences between complementary items. This dual spectral characteristic allows SComGNN to selectively process and enhance these attributes using specialized filtering techniques.

Figure 2: Spectrums of low-pass and mid-pass GCNs.

Methodology

Spectral-based GCN Filters

SComGNN employs customized spectral filters to isolate and extract low-frequency (relevance) and mid-frequency (dissimilarity) components via dedicated low-pass and mid-pass filters. These filters transform the graph's spectral information back into spatial executions, allowing for efficient node representation and convolution operations.

Attention Mechanism

The model integrates a two-stage attention mechanism. Initially, a pairwise attention mechanism adjusts the focus on relevance and dissimilarity for items in a pair, followed by a self-attention mechanism that refines these integrations independently.

Contrastive Learning

SComGNN leverages contrastive learning to optimize the model by distinguishing between correct (positive) and incorrect (negative) item pairs. This ensures that complementary items exhibit strong signal alignment, whereas non-complementary items are distanced in the representation space.

Figure 3: The overall framework of our proposed model SComGNN.

Experimental Evaluation

Empirical results on multiple e-commerce datasets demonstrate that SComGNN outperforms existing approaches significantly, especially in diverse marketplaces. Performance metrics such as Hit Rate (HR) and NDCG consistently show high scores across datasets, highlighting the advantage of spectral-based representation in capturing complementary relationships.

Implications and Future Directions

The application of spectral analysis within GNNs for item recommendation opens avenues for further refinement of GNN architectures to capture complex relationships. Future work could explore more sophisticated spectral filtering and attention mechanisms or extend this framework to support real-time recommendation systems with adaptive capabilities.

Figure 4: Hyperparameter sensitivity evaluation.

Conclusion

The paper presents a sound methodology for leveraging spectral properties of GNNs to enhance complementary item recommendations. By decoupling and balancing the dual attributes of relevance and dissimilarity, SComGNN achieves superior performance in identifying actionable and contextually complementary items, thereby enriching user interaction and satisfaction on e-commerce platforms.