RETRACTED: A Global Structural Hypergraph Convolutional Model for Bundle Recommendation

Abstract

Bundle recommendations provide personalized suggestions to users by combining related items into bundles, aiming to enhance users’ shopping experiences and boost merchants’ sales revenue. Existing solutions based on graph neural networks (GNN) face several significant challenges: (1) it is demanding to explicitly model multiple complex associations using standard graph neural networks, (2) numerous additional nodes and edges are introduced to approximate higher-order associations, and (3) the user–bundle historical interaction data are highly sparse. In this work, we propose a global structural hypergraph convolutional model for bundle recommendation (SHCBR) to address the above problems. Specifically, we jointly incorporate multiple complex interactions between users, items, and bundles into a relational hypergraph without introducing additional nodes and edges. The hypergraph structure inherently incorporates higher-order associations, thereby alleviating the training burden on neural networks and the dilemma of scarce data effectively. In addition, we design a special matrix propagation rule that captures non-pairwise complex relationships between entities. Using item nodes as links, structural hypergraph convolutional networks learn representations of users and bundles on a relational hypergraph. Experiments conducted on two real-world datasets demonstrate that the SHCBR outperforms the state-of-the-art baselines by 11.07–25.66% on Recall and 16.81–33.53% on NDCG. Experimental results further indicate that the approach based on hypergraphs can offer new insights for addressing bundle recommendation challenges. The codes and datasets have been publicly released on GitHub.

1. Introduction

Recommender systems have emerged as a crucial tool in the e-commerce industry, contributing significantly to improving user experiences and driving product sales with their development [1,2]. Traditional recommender systems primarily emphasize individual item recommendations, which could not meet the growing personalized needs of users. To further enhance user satisfaction, the bundle recommendation as a marketing strategy has been proposed. Based on users’ purchasing behavior and the relevance of items, the bundle recommendation combines relevant items into bundles, such as music playlists [3], game bundles [4,5], and drug packages [6]. Recommending bundles containing related products not only provides users with more advantageous personalized combinations but also aids businesses in achieving profitability [7].

The essence of recommender systems lies in utilizing historical interaction information and similarity relationships between entities to predict the likelihood of interactions between users and items [8]. Traditional recommender systems only involve two types of entities: users and items. However, bundle recommendation tasks encompass three types of entities: users, items, and bundles. Therefore, the decision-making process for users to choose bundles will be more complex compared to selecting individual items. Specifically, users will simultaneously consider multiple items in the bundle and their combination discounts. Sometimes, even if users like all the items contained, they may not prefer the bundle because it is not a satisfactory well-matched combination. Only when users are highly satisfied with the combination of items within this bundle will they choose the bundle for consumption. Therefore, user–bundle interactions are usually sparser than user–item interactions. Moreover, the user–item-bundle relationship is much more complex than the pairwise relationship. For example, a user can interact with multiple items or bundles, and an item can exist in multiple bundles. In this case, the affinity relations are no longer dyadic (pairwise) but rather an integration of multiple binary relationships between users, items, and bundles (three types of entities). Accordingly, the aforementioned issues make bundle recommendation tasks highly challenging.

Over the past few years, graph neural networks (GNNs) [9] have become one of the research hotspots. GNNs are a type of neural network designed for processing and learning from graph-structured data. The core idea behind GNNs [10,11] is to perform message passing between neighboring nodes in the graph to learn node representations based on their local neighborhood information. This allows GNNs to capture complex and non-linear relationships between nodes in the graph. Due to their effective modeling of graph-structured data, GNNs have emerged as a cutting-edge algorithm in the field of recommender systems, finding wide applications. For example, GNNs have been applied in item recommendation [12,13], session-based recommendation [14,15], social recommendation [16,17], and bundle recommendation [18,19].

Despite the success achieved by existing solutions based on GNNs in bundle recommendation tasks, there are still some significant challenges that need to be addressed. Specifically, due to the interactions between users and items/bundles, as well as the affiliation relationships between items and bundles, it is difficult to explicitly model the multiple complex associations among three types of entities using standard graph neural networks. Furthermore, traditional graphs are limited in the sense that each edge can only connect two nodes. Consequently, a substantial number of additional nodes and edges need to be introduced to approximate higher-order associations between entities, which increases the training burden on neural networks. Finally, recommender systems generally encounter the challenge of sparse user–item interaction history data. Since bundles encompass multiple items, the user–bundle interaction history data is inherently scarcer than the user–item interaction history data. Therefore, the issue of data sparsity becomes more pronounced in bundle recommendation scenarios. The highly sparse historical interaction data make learning with neural networks difficult and unstable. This also makes it difficult for bundle recommendation models to accurately model user preferences.

Hypergraphs [20] provide a natural solution to address the above limitations. In comparison to traditional graphs, hypergraphs represent a more flexible type of graph structure. Figure 1 illustrates the differences between traditional graph structures and hypergraph structures. In traditional graphs, an edge can only connect two nodes, while a hyperedge in hypergraphs can connect multiple nodes simultaneously. Therefore, hypergraphs can more flexibly represent and capture complex relationships in real applications. Compared to traditional graph structures, hypergraph structures possess more intricate topologies and enhanced expressive capabilities. Consequently, some problems could be easier to solve with the more accurate representation provided by hypergraphs. For bundle recommendation tasks, higher-order complex associations exist among users, items, and bundles. Hypergraphs inherently possess advantages in handling such data.

Figure 1. The differences between a traditional graph and a hypergraph. In a traditional graph, every edge, represented as a line, solely connects two nodes. Conversely, in a hypergraph, every hyperedge, denoted by a colored ellipse, can connect more than two nodes simultaneously.

In this work, we propose a novel global structural hypergraph convolutional model for bundle recommendation (SHCBR), which jointly incorporates multiple complex interactions between users, items, and bundles into a relational hypergraph. We directly connect user nodes, item nodes, and bundle nodes (three types of nodes) with a hyperedge, which explicitly models complex associations among the three types of entities without introducing additional nodes and edges. Further, we design a special matrix propagation rule that uses items as links to aggregate and update user embeddings and bundle embeddings on relational hypergraphs. We introduce the Laplacian matrix derivation in the matrix propagation rule, which can help one to better understand the evolution process of structural hypergraph convolutional neural networks. Additionally, we design a personalized weight operation to improve the accuracy of the final recommendation results. Meanwhile, inspired by LightGCN [13], we simplify the original hypergraph convolution by removing feature transformations and non-linear activation functions, making it more suitable for bundle recommendation scenarios. Hypergraph convolution can learn the hidden layer representation considering the high-order information, thereby acquiring more meaningful representations of users and bundles. The SHCBR further integrates the acquired representations to produce recommendations based on score rankings.

To summarize, the primary contributions of this study can be outlined as follows:

• We propose a novel model named the SHCBR, which introduces the hypergraph structure to explicitly model complex relationships between entities in bundle recommendation tasks. We directly connect three types of nodes with a hyperedge without introducing additional nodes and edges. By constructing a relational hypergraph containing three types of nodes, we can explore existing information from a global perspective, effectively alleviating the dilemma of data sparsity.
• We design a special matrix propagation rule and personalized weight operation in the proposed structural hypergraph convolutional neural network (SHCNN). Using items as links, we leverage efficient hypergraph convolution to learn node representations considering the high-order information. Since the relational hypergraph structure naturally incorporates higher-order associations, it is enough to generate node representations with one layer of SHCNN, further enhancing model efficiency;
• Experiments on two real-world datasets indicate that our SHCBR outperforms existing state-of-the-art baselines by 11.07–25.66% on Recall and 16.81–33.53% on NDCG. The experimental results further validate that hypergraphs provide a novel and effective method to tackle bundle recommendation tasks.

2. Related Work

In this section, we provide a concise review of related works concerning bundle recommendation and hypergraph learning.

2.1. Bundle Recommendation

Despite the extensive research on recommender systems, few efforts have been devoted to addressing the specific challenges of bundle recommendation tasks.

Initially, some works [21,22] modeled package recommendation as a linear knapsack problem [23]. These works used integer programming techniques, which overlooked the pairwise dependencies [24] among similar items. When cross-item dependencies were modeled as hard constraints, the computation became complex, and the system could not automatically recommend best matching results to users based on their preferences. During the same period, some studies [25,26] utilized association analysis techniques to address bundle recommendation problems. However, association rule methods primarily focused on the relationships between items, neglecting user and item feature information, which made it difficult to achieve personalized recommendations.

With the advancement of recommender systems, methods such as collaborative filtering [3,27], neural networks [28], and topic modeling [29] have been applied to bundle recommendations. For example, the embedding factor model (EFM) [3] extended the traditional factorization method to capture users’ preferences. The work [27] combined matrix factorization (MF) methods for collaborative filtering with learned fashion constraints to achieve clothing package recommendation. However, matrix factorization was primarily used to model user–item interactions. It was challenging to directly model higher-order associations, such as relationships between multiple items within bundles. Furthermore, methods based on collaborative filtering often utilized information from neighboring users or items for recommendations. However, they could be affected by noisy data and outliers, which might result in inaccurate recommendations. The DAM [28] model utilized a factorized attention network to learn bundle representations in a multi-task manner. However, neural network-based bundle recommendation methods could be affected by the issue of data sparsity. Neural network methods typically involved numerous parameters and complex model architectures, and sparse training data can lead to overfitting issues. Xiong et al. [29] proposed a personalized travel package recommendation model, which utilized topic analysis to obtain activities of interest to tourists. However, topic modeling methods might not accurately capture the true meaning and attributes of items when performing topic modeling on items. This could result in inaccurate topic representations, consequently affecting the effectiveness of recommendations.

In recent years, researchers have redirected their research focus towards graph neural networks. The BGCN [18] model unified three interactions between three types of entities into a heterogeneous graph. Then, the BGCN model used graph convolutional networks (GCN) [30] to acquire node representations for both users and bundles. The BundleNet [5] model formalized a recommendation task as a link prediction problem on the graph and addressed it by utilizing a neural network model capable of directly learning from graph-structured data. The IMBR [31] model used the graph neural network to extract multi-relation representations from various views. Additionally, an algorithm called BFTC was devised to enhance the precision of bundle representations.

Furthermore, some researchers have incorporated contrastive learning methods into GNN-based models. Contrastive learning can extract more meaningful and discriminative feature representations by learning the relationships between different entities in contrastive views. These feature representations can better capture the similarities and differences between entities, thereby enhancing the accuracy of bundle recommendation. The MIDGN [19] model utilized a graph neural network equipped with the neighbor routing mechanism to untangle the coupling between the user–item and bundle–item graphs, considering users’ intents. CrossCBR [32] utilized cross-view contrastive learning to capture cross-view cooperative associations. GPCL [33] innovatively embedded each entity as a Gaussian distribution and introduced a prototypical contrastive learning method to capture more refined representations.

The review of the mentioned works indicates that a significant portion of recent research has utilized GNN to model interactions between users, items, and bundles. GNN-based models possess the advantage of integrating node features and graph topology. However, bundle recommendation involves intricate relationships among users, items, and bundles. Graph neural networks may require the design of appropriate graph structures and connections to capture these intricate relationships. Accurately modeling higher-order associations could present a challenge. On the other hand, graph neural networks exhibit high computational complexity, especially for large-scale graph data. In bundle recommendations, there might be a substantial number of nodes and edges, approximating higher-order associations among the three types of entities. This can lead to an increased computational burden during training and inference. This brings forth the concept of the hypergraph [20]. Hypergraph is a special graph structure that can utilize hyperedges to simultaneously connect multiple nodes. In response to the challenges arising from the use of GNN, hypergraph neural networks present a natural and relative solution. The flexibility of hypergraphs makes them a novel approach for investigating bundle recommendation tasks.

2.2. Hypergraph Learning

Hypergraphs [20,34] are a specialized type of graph structure that extends traditional graphs. Hypergraphs allow a hyperedge to connect multiple nodes, forming a many-to-many relationship. Therefore, hypergraphs can more accurately represent complex relationships that traditional graphs cannot directly capture in real-world scenarios. For instance, in the knowledge graph [35], hypergraphs can directly represent relationships between one entity and multiple other entities using a single hyperedge. Such higher-order associations are useful for modeling complex semantic associations. In contrast, to approximate such higher-order associations, traditional graphs would require the introduction of a substantial number of nodes and edges.

Hypergraph learning [36] is a deep learning method based on hypergraph structures. In the past few years, hypergraph learning has gained significant attention due to its flexibility and ability to model complex data associations. Hypergraph learning was initially proposed as a label propagation method [37] for semi-supervised learning. Subsequently, Huang et al. [38] employed hypergraph learning for video object segmentation, delving into hypergraph construction methodologies. Weights had a great influence on the modeling of data correlation, and then learning the weights of hyperedges became a new research topic. Gao et al. [39] proposed a novel approach for evaluating the significance of hyperedges or subgraphs within a multi-hypergraph structure. The aim was to assign weightings that reflect their relative importance. Subsequently, the L2 regularization of weights to learn optimal hyperedge weights was proposed by [40]. Hypergraph neural networks (HGNN) have been introduced as a novel method on hypergraph structure, similar to graph neural networks [41]. Hypergraph neural networks employed the hypergraph Laplacian operator to represent the hypergraph from a spectral perspective. Hypergraph neural networks have an advantage over the current GNN methods in their ability to model and encode high-order relationships within data. The authors of [42] introduced two end-to-end trainable operators recently, namely, hypergraph convolution and hypergraph attention, which can handle non-pairwise relationships modeled in high-order hypergraphs.

Although research on hypergraph deep learning is still in its early stages, hypergraph neural networks have found widespread application across various fields [35,43,44,45,46] due to their exceptional representation capabilities. For example, in computer vision, Wang et al. [43] applied hypergraphs to visual classification tasks, using hypergraphs to describe the relations among visual features. In knowledge graph [35], hypergraph neural networks were utilized to capture higher-level correlations between entities and attributes. In the field of recommender systems, hypergraph-based recommendation models can capture complex interactions between entities and provide more accurate and diverse recommendation results. For example, the UHBR [44] model utilized hypergraphs to model complex user–item-bundle associations, thereby enhancing the efficiency and accuracy of the model. Hg-PDC [45] model proposed an algorithm based on dynamics clustering and similarity measurement in hypergraphs, thereby improving the recommendation performance. The GC–HGNN [46] model utilized the hypergraph convolutional neural network to obtain the global context information.

Hypergraph learning also comes with certain limitations in practical applications. Hypergraph learning is still a relatively new field, and there might not be enough well-established frameworks, libraries, or benchmarks available for researchers and practitioners. This can make it challenging to implement and compare different hypergraph-based models. It is worth noting that the application of hypergraph models may be more suitable for specific domains, whereas their applicability in other domains could be limited.

3. Preliminary

In this section, we briefly formulate the problem and introduce our bundle hypergraph definition.

3.1. Problem Formulation

In our research, we have a defined group of users $U=\{u_1,u_2,\dots ,u_N\}$, where the variable N is the total number of users, a defined group of items $I=\{i_1,i_2,\dots ,i_M\}$, the variable M is the total number of items, a defined group of bundles $B=\{b_1,b_2,\dots ,b_K\}$, and the variable K is the total number of bundles. Each individual bundle $b_s\in B$ is made up of a specific group of items, denoted as $b_s=\left\{i_{s_1},i_{s_2},\dots ,i_{s|b_s|}\right\}$, with the bundle size being greater than one. In order to gain a deeper understanding of the connections among users, items, and bundles, we present three matrices: user–item matrix $X_{N\times M}=\left\{x_{ui}\right|u\in U,i\in I\}$, user–bundle matrix $Y_{N\times K}=\left\{y_{ub}\right|u\in U$, $b\in B\}$, and bundle–item matrix $Z_{K\times M}=\left\{z_{bi}\right|b\in B,i\in I\}$. These matrices consist of binary values, where a value of 1 indicates an observed interaction or inclusion. In the user–item matrix, an entry $x_{ui}=1$ signifies that there is an interaction between user u and item i. Then, an entry $y_{ub}=1$ in the user–bundle matrix indicates that user u has interacted with bundle b. Similarly, an entry $z_{bi}=1$ in the bundle–item matrix implies that item i is a component of bundle b. From the definition provided earlier, we can formulate the issue of bundle recommendation in the following manner:

Input: Data on the interaction between users and items $X_{N\times M}$, data on the interaction between users and bundles $Y_{N\times K}$, and data on the dependency relationship between bundles and items $Z_{K\times M}$.

Output: A bundle recommendation model that calculates the likelihood of user u interacting with bundle b.

3.2. Bundle Hypergraph Definition

Hypergraphs provide an approach to obtain intricate higher-order connections. However, traditional hypergraphs struggle to represent multiple relationships in bundle recommendation tasks suitably. In response, we propose a novel data structure named the bundle hypergraph, which can more effectively capture the multiple associations among three types of entities in bundle recommendations.

A bundle hypergraph $\mathcal{G}=\left(\left.\mathcal{V},\mathcal{E}\right)\right.$ is a mathematical structure that consists of a set of hyperedges $\mathcal{E}$ and a set of nodes $\mathcal{V}$. The hyperedges are subsets of the nodes set and can contain multiple nodes. This definition allows for a more flexible representation of relationships between nodes compared to traditional graphs. The incidence matrix $H_{ve}\in {\left\{0,1\right\}}^{\left|\mathcal{V}\right|\times \left|\mathcal{E}\right|}$ represents the connection relationship between hyperedges and nodes. The entry $h\left(v,e\right)$ in the incidence matrix can be defined as:

$$h\left(v,e\right)=\left\{\begin{array}{c}1,ifv\in {\mathcal{E}}_e\hfill \\ 0,otherwise\hfill \end{array}\right.$$

(1)

Based on the hypergraph definition, node degrees and hyperedge degrees can be calculated. Next, we consider a node $v\in \mathcal{V}$, where $\mathcal{V}$ represents the set of nodes. In the bundle hypergraph, the degree of node v can be denoted as $d_v={\sum }_{e\in \mathcal{E}}h\left(v,e\right)$, which is calculated by summing up the contributions from all of the connected hyperedges in $\mathcal{E}$ (the set of hyperedges) using the function $h\left(v,e\right)$. Similarly, for a hyperedge $e\in \mathcal{E}$, its degree can be expressed as $d_e={\sum }_{v\in \mathcal{V}}h\left(v,e\right)$. It is determined by adding up the values calculated by applying the function $h\left(v,e\right)$ for each node v in $\mathcal{V}$.

4. Method

The construction of our proposed relational hypergraph is presented in Section 4.1. Following that, we elaborate on the structural hypergraph convolutional neural networks in Section 4.2. Moving on to Section 4.3, we introduce the matrix propagation rule in our research. Lastly, a detailed explanation of the model’s prediction and training process is given in Section 4.4.

4.1. Relational Hypergraph Construction

As shown in Figure 2, we construct a relational hypergraph matrix with items as links. To obtain more meaningful bundle representations, we first construct three types of relation graphs around users and bundles. The graphs that are included in this study are the interaction graph between users and items, the interaction graph between users and bundles, and the affiliation graph between bundles and items. Based on these three relation graphs, we can obtain the three interaction matrices mentioned in Section 3.1, namely, X, Y, and Z. In order to encode the input, we perform one-hot encoding for each node on the hypergraph and compress each node into a dense real-value vector. $E_u$, $E_i$, and $E_b$ represent the embeddings of user, item, and bundle in the initial matrix representation, respectively. The embedding size d remains consistent for all three matrices $E_u$, $E_i$, and $E_b$. We define the feature vectors of user u, item i, and bundle b in the following manner:

$$e_u=E_u^T{v}_u,e_i=E_i^T{v}_i,e_b=E_b^T{v}_b$$

(2)

where $E_u\in {\mathbb{R}}^{N\times d}$, $E_i\in {\mathbb{R}}^{M\times d}$, and $E_b\in {\mathbb{R}}^{K\times d}$. $v_u\in {\mathbb{R}}^N$, $v_i\in {\mathbb{R}}^M$, and $v_b\in {\mathbb{R}}^K$ are one-hot vectors of user, item, and bundle, respectively.

Figure 2. The SHCBR framework can be divided into two main components: in the relational hypergraph construction, we constructs a relational hypergraph matrix with items as links and obtains user/bundle initial embeddings; the structural hypergraph convolutional neural network is proposed for capturing high-order relationships of users and bundles on a relational hypergraph.

According to the definition in Section 3.1, in the application scenario of bundle recommendation, we acquire the interaction matrix X between users and items, the interaction matrix Y between users and bundles, and the affiliation matrix Z between bundles and items, in a natural manner. Given the complexity of relationships among the three types of nodes, the relevant items can serve as links between users and bundles. We construct the item association matrix as follows:

$$H_I={\left[\begin{array}{cc}X& Z\end{array}\right]}^T$$

(3)

where $X\in {\mathbb{R}}^{\left|U\right|\times \left|I\right|}$, $Z\in {\mathbb{R}}^{\left|B\right|\times \left|I\right|}$, and $H_I\in {\mathbb{R}}^{\left(\left|U\right|+\left|B\right|\right)\times \left|I\right|}$. Similarly, using items as connectors and leveraging second-order interaction data, we can infer the similarity between users. Likewise, we infer the similarity between bundles. Highly overlapping users and bundles aid in predicting user interests in bundles. Therefore, we further define the user–bundle adjacency matrix as follows:

$$H_{UB}=\left[\begin{array}{cc}{S}_U& Y\\ Y^T& S_B\end{array}\right]$$

(4)

$$S_U=X·X^T,S_B=Z·Z^T$$

(5)

where $H_{UB}\in {\mathbb{R}}^{\left(\left|U\right|+\left|B\right|\right)\times \left(\left|U\right|+\left|B\right|\right)}$. $S_U\in {\mathbb{R}}^{\left|U\right|\times \left|U\right|}$ and $S_B\in {\mathbb{R}}^{\left|B\right|\times \left|B\right|}$ denote the user similarity-overlap matrix and the bundle similarity-overlap matrix constructed based on second-order interactions, respectively.

Based on the flexibility of hypergraphs in representing and capturing complex relationships in practical problems, the relational hypergraph matrix we constructed can be defined as follows:

$$H_{SC}=\left[\begin{array}{cc}{H}_{UB}& H_I\end{array}\right]$$

(6)

where $H_{UB}\in {\mathbb{R}}^{\left(\left|U\right|+\left|B\right|\right)\times \left(\left|U\right|+\left|B\right|\right)}$, $H_I\in {\mathbb{R}}^{\left(\left|U\right|+\left|B\right|\right)\times \left|I\right|}$, and $H_{SC}\in {\mathbb{R}}^{\left(\left|U\right|+\left|B\right|\right)\times \left(\left|U\right|+\left|B\right|+\left|I\right|\right)}$. The three types of nodes are jointly incorporated in a relational hypergraph $H_{SC}$ from a global perspective, naturally capturing higher-order connections within it.

4.2. Structural Hypergraph Convolutional Neural Networks

According to the depicted in Figure 2, we propose a structural hypergraph convolutional neural network (SHCNN), which captures the higher-order associations between entities on the hypergraph structure. Hypergraphs [20] are an extension of graphs where hyperedges can connect multiple nodes. Compared with traditional graph structures, hypergraphs have more complex connectivity, making them more suitable for bundle recommendation tasks. Hypergraph convolution, mentioned in the work by authors [42], can be perceived as a form of message passing. Knowing how to define the convolution operation and efficiently propagate information between adjacent nodes is very crucial in bundle recommendation systems. It is a natural approach to introduce more convolutional layers in GNN to acquire higher-order associations between entities. However, this approach accompanies the downside of imposing a substantial escalation in computational requirements. Since the relational hypergraph structure naturally incorporates higher-order associations, it enables the efficient learning of node representations using only one layer of SHCNN. We further simplify the formula for hypergraph convolution operation as follows:

$$\widehat{H}=D_v^{-\frac{1}{2}}H{D}_e^{-\frac{1}{2}}{H}^T{D}_v^{-\frac{1}{2}}$$

(7)

$$E^{\left(l+1\right)}=\widehat{H}{E}^{\left(l\right)}$$

(8)

where $\widehat{H}$ denotes relational hypergraph’s Laplacian spectral normalization matrix. The output of layer l denoted as $E^{\left(l+1\right)}$.

Furthermore, we design a personalized weight operation to enhance the accuracy of learning user and bundle representations. Inspired by the hypergraph attention operator [42], we make improvements to the similarity-overlap matrix in Equation (5). Taking $S_U$ as an example, the similarity between user $u_i$ and user $u_j$ can be computed in following manner:

$$S_{ij=}\frac{exp\left(\sigma \left(sim\left(u_i,u_j\right)\right)\right)}{{\sum }_{k\in N_i}exp\left(\sigma \left(sim\left(u_i,u_k\right)\right)\right)}$$

(9)

where $S_{ij}\in S_U$. $N_i$ is the neighborhood set of user $u_i$. $\sigma \left(·\right)$ is the nonlinear activation function like LeakyReLU [47] and eLU [48], enhancing the learning ability of neural networks. The similarity function $sim\left(·,·\right)$ is used to calculate the similarity between two vertices. This function is presented as follows:

$$sim\left(u_i,u_j\right)=a^T\left[u_i\left|\right|u_j\right]$$

(10)

Here, a represents a weight vector that is utilized to generate a scalar value indicating similarity. Likewise, we compute the similarity between bundles. Through the above computations, we can obtain the user similarity-overlap matrix $S_U$ and the bundle similarity-overlap matrix $S_B$.

4.3. Matrix Propagation Rule

Combining the application scenario of the bundle recommendation and the SHCBR model, we propose a specialized matrix propagation rule to achieve the embedding propagation of the entire relational hypergraph. In order to simulate the complex interaction logic between three types of entities, we define the relational hyperedge adjacency matrix as follows:

$$A=H_{SC}{D}_e^{-\frac{1}{2}}{H}_{SC}^T$$

(11)

where $H_{SC}$ is the relational hypergraph matrix we constructed in Equation (6). The Laplacian matrix L based on the relational hypergraph matrix $H_{SC}$ can be defined as:

$$D_v=\left[\begin{array}{cc}{D}_U& 0\\ 0& D_B\end{array}\right]$$

(12)

$$L=D_v^{-1/2}A{D}_v^{-1/2}+\left[\begin{array}{cc}{I}_U\hfill & 0\hfill \\ 0\hfill & I_B\hfill \end{array}\right]=\left[\begin{array}{cc}{I}_U& D_U^{-1/2}{A}_U{D}_B^{-1/2}\\ D_B^{-1/2}{A}_B{D}_U^{-1/2}& I_B\end{array}\right]$$

(13)

where $A_U$ and $A_B$ are the relational hyperedge adjacency matrix of users and bundles, respectively. $A_U$ and $A_B$ together constitute the relational hyperedge adjacency matrix A. $D_U$ and $D_B$ are the node degree matrices of users and bundles, respectively. $I_U\in {\mathbb{R}}^{\left|U\right|\times \left|U\right|}$ and $I_B\in {\mathbb{R}}^{\left|B\right|\times \left|B\right|}$ are the identity matrix. Inspired by LightGCN, the purpose of adding the identity matrix is to connect its own nodes.

4.4. Model Prediction and Training

In the realm of hypergraph convolution, traditional methods like hypergraph neural networks [41] use the embeddings from the final layer as their ultimate representation. However, when the quantity of layers augments, the embeddings incline towards excessive smoothing, making it unreasonable to use only the last layer embeddings as the ultimate representation. In existing bundle recommendation models, BGCN [18] concatenates embeddings from all layers to merge information obtained from neighbors at different depths for prediction, potentially intensifying training challenges. Inspired by LightGCN [13], we choose to incorporate embeddings from different layers as the ultimate representation. This approach captures the information from different layers to enrich the semantics of the final representation. Next, we combine the embeddings from each layer. By merging the embeddings of each layer, we are able to generate comprehensive representations of each user and bundle:

$$e_u=\sum _{k=0}^K{\alpha }_k{e}_u^{\left(k\right)},e_b=\sum _{k=0}^K{\alpha }_k{e}_b^{\left(k\right)}$$

(14)

where ${\alpha }_k\ge 0$ represents the significance of the embedding of the l-th layer in forming the ultimate embedding. To ensure fairness in contribution, we assign ${\alpha }_k$ as $1/(K+1)$.

Ultimately, we define the inner production of final user representations and bundle representations as our model prediction:

$${\widehat{y}}_{ub}=e_u^T{e}_b$$

(15)

The ranking scores for generating recommendations rely on the final results. Observations indicate that when there is a interaction (e.g., purchase, click) between the user and the bundle, it can be inferred that this user has an interest in the bundle as a whole or in a majority of its individual items. Conversely, if a user does not interact with a bundle, it can be assumed that the user is not aware of that bundle. In order to predict bundles, we classify bundles with interactions as positive samples and randomly choose an unobserved bundle as a negative sample. In conclusion, a pairwise framework for learning is established, and the Bayesian personalized ranking loss is applied to predict bundles:

$$L_{bundle}=(1-\alpha )\sum _{\left(j,e\right)\in R^{+}}ln\sigma \left({\widehat{y}}_{je}-y_{je}\right)+\alpha \sum _{\left(j,f\right)\in R^{-}}ln\sigma \left({\widehat{y}}_{jf}-y_{jf}\right)$$

(16)

where $R=\left\{\left(j,e,f\right)|\left(j,e\right)\in R^{+},\left(j,f\right)\in R^{-}\right\}$. $R^{+}$ denotes bundles that has interactions with user $u_j$, while $R^{-}$ denotes bundles that have no interactions with user $u_j$. We regard $(j,e)$ as a positive sample and $(j,f)$ as a negative sample. $\sigma \left(·\right)$ denotes a sigmoid function. To prevent the model from overfitting, we add $L_2$ regularization into the loss function.

Subsequently, the AdamW optimizer [49] is employed to minimize the loss function, which iteratively updates neural networks’ weights using the provided training data.

5. Experiment

In this section, we perform experiments on two public datasets to assess the SHCBR, with the aim of answering following research questions:

• RQ1: How does our SHCBR compare in performance to previous models?
• RQ2: How do key components influence the SHCBR’s performance?
• RQ3: How do different parameter settings affect the SHCBR’s results?

5.1. Experiment Settings

5.1.1. Datasets and Metrics

We evaluate SHCBR and other baseline methods over two real-world datasets. The training/test/validation sets are randomly divided in a ratio of 70%/20%/10%, and the statistical data is indicated in Table 1.

Table 1. Dataset statistics.

Dataset	NetEase	Youshu
User	18,528	8039
Item	22,864	4771
Bundle	123,628	32,770
User–item	1,128,065	138,515
User–bundle	302,303	51,337
Bundle–item	1,778,065	176,667
Avg. bundle interactions	16.32	6.39
Avg. item interactions	60.88	17.23
Avg. bundle size	77.80	37.03
User–item density	0.05%	0.05%
User–bundle density	0.07%	0.13%

• NetEase: This is a dataset constructed using data provided by a Chinese music platform, Netease Cloud Music (http://music.163.com, accessed on 7 August 2017). As a social music software, it allows users to freely choose their favorite songs and add them to their favorites. Users can also choose to listen to playlists bundled with different songs.
• Youshu: This is a dataset constructed by the famous Chinese book review website Youshu (https://www.yousuu.com/, accessed on 7 August 2017). Youshu allows users to create their own booklists with different styles and types. Additionally, it can provide users with a bundle of related books.

To evaluate models’ recommendation performance, two commonly-adopted ranking metrics are employed in experiments: $\mathit{Recall}@K$ and $\mathit{NDCG}@K$. $\mathit{Recall}@K$ and $\mathit{NDCG}@K$ are metrics used to assess the performance of information retrieval and recommendation systems. $\mathit{Recall}@K$ measures how many of the actual target items are present among the top-K recommended items. $\mathit{NDCG}@K$ aims to measure the ranking quality and quantity of truly relevant items within the top-K recommended items. It is worth mentioning that we have set the values of K as: {20, 40, 80}. Recall and NDCG can be calculated as below:

$$\mathit{Recall}@K=\frac{\left|D\cap B\right|}{D}$$

(17)

where K signifies that we are employing a top-K ranking approach. For every user, we produce a bundle list B containing K items. D represents a bundle set that the user has previously interacted with. The symbol $b_k$ refers to the bundle located at the k-th position in the generated bundle list B, while d denotes a bundle in the bundle set D that the user has indeed interacted with.

$$\mathit{DCG}@K=\sum _{k=1}^K\frac{2^{{hit}_k}-1}{{log}_2\left(k+1\right)}$$

(18)

$$hi{t}_k=\left\{\begin{array}{c}1,if{b}_k\in D\hfill \\ 0,otherwise\hfill \end{array}\right.$$

(19)

$$\mathit{NDCG}@K=\frac{\mathit{DCG}@K}{\mathit{IDCG}@K}$$

(20)

where $\mathit{IDCG}@K$ is typically used as the denominator in the calculation of $\mathit{NDCG}@K$ to compute the standardized NDCG value. $\mathit{IDCG}@K$ is computed by arranging the truly relevant items in the ideal ranking order and then calculating their cumulative discounted gains.

5.1.2. Baselines

To demonstrate the superior performance of SHCBR, we conduct a comparison between it and the below models:

• MFBPR [24]: This is a popular MF model based on BPR loss optimization, which is widely used for implicit feedback.
• RGCN [50]: RGCN is a method based on graph convolutional networks that is specifically designed to handle multi-relational graphs.
• LightGCN [13]: This is an efficient and lightweight model that incorporates both graph neural networks and collaborative filtering ideas.
• BundleNet [5]: BundleNet constructs a tripartite graph consisting of users, bundles, and items, which utilizes GNN to learn node representation of entities.
• DAM [28]: DAM is a deep learning model that incorporates attention mechanisms to facilitate the acquisition of comprehensive bundle representations.
• BGCN [18]: BGCN leverages the powerful ability of GNN in learning from higher-order connections on graphs, modeling the complex relations between users, items, and bundles.
• MIDGN [19]: MIDGN designed a multi-view intent resolution graph network, using a graph neural network to separate user intent from different perspectives.

5.1.3. Implementation Details and Environment

For SHCBR, we utilize the PyTorch framework (https://pytorch.org/, accessed on January 2017) and optimize the model using the AdamW optimizer. In our experiments, we set the embedding size to 128. A batch size of 2048 is used to process both datasets. The search for the learning rate is conducted within the value set {$1\times {10}^{-4}$, $5\times {10}^{-4}$, $1\times {10}^{-3}$, $5\times {10}^{-3}$, $1\times {10}^{-2}$}. Based on the experiments, it is determined that the optimal choice is $5\times {10}^{-3}$. To address the issue of overfitting, we have implemented the use of L2-norm with a specific value of 0.2. Moreover, we have also incorporated a dropout rate of 0.2 into our methodology. These measures aim to mitigate the potential negative impact of overfitting on our outcome.

To ensure experimental rigor, all experiments were performed on the same server. Our server is equipped with Nvidia GeForce RTX 2080Ti GPUs and runs on Ubuntu 14.04.2.

5.2. Performance Comparison and Analysis

We evaluate SHCBR’s overall recommendation performance against existing baseline methods on two datasets. From Table 2, we can observe that SHCBR outperforms the baseline methods in terms of recommendation performance. The best-performing method, which is SHCBR, is highlighted in bold, and the strongest performing baseline is underlined. Referring to Table 2, we can make the observations and analysis as below:

Table 2. The performance of SHCBR and other baseline methods on two datasets. Bold represents the best results, while underline represents the second-best results.

Method	NetEase				Youshu
	Recall @20	NDCG @20	Recall @40	NDCG @40	Recall @20	NDCG @20	Recall @40	NDCG @40
MFBPR	0.0355	0.0181	0.0600	0.0246	0.1959	0.1117	0.2735	0.1320
RGCN	0.0407	0.0210	0.0670	0.0280	0.2040	0.1069	0.3017	0.1330
LightGCN	0.0496	0.0254	0.0795	0.0334	0.2286	0.1344	0.3190	0.1592
BundleNet	0.0391	0.0201	0.0661	0.0271	0.1895	0.1125	0.2675	0.1335
DAM	0.0411	0.0210	0.0690	0.0281	0.2082	0.1198	0.2890	0.1418
BGCN	0.0491	0.0258	0.0829	0.0346	0.2347	0.1345	0.3248	0.1593
MIDGN	0.0678	0.0343	0.1085	0.0451	0.2682	0.1527	0.3712	0.1808
Ours	0.0852	0.0458	0.1304	0.0561	0.3049	0.1838	0.4123	0.2112
Imp%	25.66%	33.53%	20.18%	24.39%	13.68%	20.37%	11.07%	16.81%

Compared with the traditional machine learning method BPR, the graph-based methods LightGCN and RGCN exhibit more powerful learning capabilities. This is attributed to the advantages of graph neural networks in learning and aggregating multi-hop collaborative information. The performance of RGCN demonstrates the importance of modeling relationship between entities in personalized bundle recommendation tasks. Among the three methods, LightGCN performs the best, thanks to its simplified design of GCN, making it more suitable for bundle recommendation scenarios.

Over the past few years, there has been notable advancement in bundle recommendation research. It is worth mentioning that BundleNet constructs a user–item-bundle tripartite graph. Then, BundleNet formalizes the task of bundle recommendation as a graph-based link prediction problem. However, its performance is not satisfactory due to its overly simplistic neural network model. It has also been observed that some excellent graph neural network-based models (such as RGCN and BundleNet) perform even worse than non-graph machine learning methods like DAM, which leverages deep attention mechanisms and a multi-task manner to predict user preferences jointly. In addition, BGCN adopts graph neural networks to explicitly model complex relationships among entities. MIDGN decouples user–item and bundle–item interactions to obtain potential intentions of users and bundles from diverse perspectives. At the same time, it is also the best model among the bundle recommendation methods in the baseline. While graph-based methods are effective, hypergraphs offer a more flexible structure. Capturing complex associations between users, items, and bundles is crucial in bundle recommendation tasks, and hypergraphs have a natural advantage in handling higher-order associations. Therefore, SHCBR achieves the best results. Specifically, the performance of SHCBR on NetEase dataset is 20.18–33.53% higher than the best baseline. The performance of SHCBR on Youshu dataset is 11.07–20.37% higher than the best baseline.

5.3. Ablation Study of SHCBR

Next, ablation studies are conducted to analysis the effectiveness of several key designs in SHCBR. We sequentially remove some key components of SHCBR to create various derivative models. Referring to the findings displayed in Table 3, the conclusions can be made:

Table 3. Ablation study. Bold represents the best results.

Method	NetEase		Youshu
	Recall@20	NDCG@20	Recall@20	NDCG@20
$SHCB{R}_{w/oHG}$	0.0801	0.0397	0.2879	0.1743
$SHCB{R}_{w/oHC}$	0.0797	0.0411	0.2838	0.1716
$SHCB{R}_{w/oMP}$	0.0832	0.0441	0.2978	0.1806
$SHCBR$	0.0852	0.0458	0.3049	0.1838

• $SHCB{R}_{w/oHG}$. This model removes the module of the relational hypergraph construction. In this part, we exclude the similarity overlap matrix form the user–bundle adjacency matrix. At the same time, we also eliminate the construction of the structural hypergraph matrix. We discover that SHCBR outperforms significantly $SHCB{R}_{w/oHG}$, demonstrating the effectiveness of the relational hypergraph construction module. Moreover, the experimental results also highlight the importance of the hypergraph structure in capturing node feature.
• $SHCB{R}_{w/oHC}$. This model removes the part of the structural hypergraph convolutional neural networks. Here, we replace the structural hypergraph convolutional neural networks with a simple graph convolutional neural network. It is apparent that SHCBR outperforms $SHCB{R}_{w/oHC}$. This is due to simple graph convolutional neural networks have limited aggregation capabilities compared to hypergraph convolutional neural networks. This demonstrates the superiority of our proposed structural hypergraph convolutional neural networks.
• $SHCB{R}_{w/oMP}$. This model removes the special matrix propagation rule module but retains other designs of SHCBR. It can be observed that SHCBR is only slightly superior than $SHCB{R}_{w/oMP}$. Although SHCBR is not highly competitive compared to $SHCB{R}_{w/oHG}$, we can still see that the special matrix propagation rule is helpful for improving model performance.

5.4. Hyper-Parameters Analysis

We further explore how hyper-parameters $\alpha$ and batch size influence the performance of SHCBR.

Research on hyper-parameters $\alpha$. As shown in Figure 3, we analyze the loss function ratio of positive to negative samples in Equation (16) to investigate its influence on the performance of SHCBR. We compare SHCBR’s $\mathit{Recall}@20$ and $\mathit{NDCG}@20$ when setting different $\alpha$. We can observe that for both datasets, the highest $\mathit{NDCG}@20$ is achieved when the value of $\alpha$ is 0.4, and the highest $\mathit{Recall}@20$ is achieved when the value of $\alpha$ is 0.3. Therefore, we can draw the conclusion that the impact of negative samples is somewhat smaller compared to that of positive samples. This situation arises mainly due to the random selection strategy for negative samples.

Figure 3. Impact of hyper-parameters $\alpha$.

Research on batch size. To study the impact of batch size on the model, we incrementally raise it from 128 to 8192. As shown in Figure 4, it can be observed that with the increase in batch size, the model’s performance experiences an initial rapid improvement and then reaches a stable state. Based on the observed results, we adopt a 2048-size in our experiments on both datasets.

Figure 4. Impact of batch size.

5.5. Experimental Efficiencies

The running cost is an important factor that needs to be considered. Lower running costs allow for faster iterations and reduce expenses. In this section, we analyze the runtime of the SHCBR model for each epoch and compared it with other models as indicated in Table 4. All experiments are performed utilizing the NetEase dataset. We choose the NetEase dataset because it has a significantly larger data interactions compared to the Youshu dataset.

Table 4. Execution cost of different method.

Method	Time for Train per Epoch	Time for Test
BundleNet	1610 s	5 s
DAM	1805 s	5507 s
BGCN	92 s	3 s
MIDGN	—	—
SHCBR	32 s	3 s

Table 4 demonstrates that our SHCBR is the most cost-effective approach, considering both the training and testing phases. The DAM model involves a large number of parameters, leading to a high optimization cost. This is primarily because the DAM model not only includes item-based factorized attention parameters but also incorporates extensive representations for users, bundles, and items. Additionally, DAM utilizes the multi-task learning method, which jointly models interactions between entities. This heavily relies on the powerful parallel computing capabilities of GPUs. However, due to limitations in GPU performance, the DAM model is less cost-effective. For BundleNet, the construction of the user–item-bundle triple graph significantly contributes to the increased runtime. Furthermore, the complexity of the graph propagation layer also results in negative effects on time efficiency. On the other hand, BGCN, MIDGN, and SHCBR employ lightweight graph convolutional networks, which significantly improves training efficiency. For MIDGN, this model requires a substantial amount of GPU memory due to the generation of numerous representation vectors in its graph decomposition process. As indicated in Table 4, since the Nvidia RTX 2080Ti GPU ran out of memory, the running time of MIDGN model is not displayed. In contrast, our proposed SHCBR model utilizes structural hypergraph convolutional networks, which reduces aggregation layers and eliminates non-linear transformations. Additionally, we design a denser hypergraph Laplacian matrix, facilitating more efficient message propagation for convolution processing. Consequently, the SHCBR model exhibits an outstanding training time advantage.

6. Conclusions and Future Work

In this work, we investigate bundle recommendation tasks. Unlike traditional single-item recommendation tasks, bundle recommendation involves recommending a group of related items, i.e., bundles, to users. We propose a novel model named SHCBR, which jointly incorporates nodes of users, items, and bundles into a relational hypergraph from a global perspective. We utilize the flexible hypergraph structure to model multiple complex associations among three types of entities. With item nodes as links, we leverage efficient hypergraph convolution to learn representations of users and bundles while considering the high-order information, improving the quality of node representations. This modeling approach allows for better exploration of the underlying interests and associations behind user behavior, alleviating the dilemma of data scarcity. Our experiments conducted on two real-world datasets significantly illustrate the superior performance of SHCBR compared to baseline models. The experimental results further affirm that hypergraphs can offer a novel and effective approach to address bundle recommendation tasks.

This work is our initial attempt to explore the application of hypergraphs in the field of bundle recommendation. There are a number of challenging directions based on hypergraphs that warrant further exploration to address the difficulties in bundle recommendation. For example, in our experiments, hyperedges are equally weighted. When additional knowledge of the data distributions is available, theoretically, utilizing appropriate weighting mechanisms can enhance the accuracy of recommendation results. It is even possible to incorporate learnable modules within neural networks and optimize the weights using gradient descent. Combining hypergraphs with other deep learning methods is also promising. For example, the users and bundles on both sides of our relational hypergraph construction module can also serve as a suitable contrastive learning paradigm. In the future, we plan to study the integration of contrastive learning frameworks into our existing work to enhance recommendation efficiency while also alleviating the dilemma of data sparsity.