DSGRec: dual-path selection graph for multimodal recommendation

Multi-modal recommendation

With the growing prevalence of multimedia contents in the modern web era, many recommendation systems have evolved from single-modal approaches to multi-modal methods that integrate various data types and features. VBPR (He & McAuley, 2016) considers the visual appearance of the items by extracting visual features from pre-trained network. Recently, many works introduce GNNs into multimodal recommendation. Besides multi-modal item representations, MMGCN (Wei et al., 2019) enhances user representations from modalitiy specific user-item interactions. MICRO (Zhang et al., 2022a) injects a multi-modal representation of item-item into items and captures the relationships between different items with graph convolution. Attention-guided multi-step fusion network (TMFUN) (Zhou et al., 2023c) and Self-supervised graph disentangled network (SGDN) (Ren et al., 2023) are both trained on the interaction data between users and items to get the deep relationship between users and items. Self-supervised interest transfer network (SITN) (Sun et al., 2023) and MBSSL (Xu et al., 2023) operate the nodes to obtain a better recommendation effect. SITN (Sun et al., 2023) aggregates nodes with semantic invariance in different semantic Spaces. Bias constrained contrastive learning (BCCL) (Yang et al., 2023) utilizes data augmentation with constrained biases to enhance sample quality. However, existing recommendation methods fail to consider attractiveness and demand as the bidirectional characteristics of user-item interactions, leading to limited recommendation efficacy within specific item categories. DSGRec adopts two different methods to capture attraction and demand signals on user-item interaction graphs, allowing for more sophisticated modeling of complex user behaviors. MGNM (Mo et al., 2024) effectively filters out noise from the modality features, ensuring the refined information is more closely aligned with user preferences. Counterfactual knowledge distillation (CKD) (Zhang et al., 2024) enhances the information utilization rate of each modality under single-modality guidance, effectively capturing complementary information between different modalities.

Graph-based recommendation

In the field of recommender system, graph convolutional neural networks are favored by mainstream methods because of their excellent performance in high-order connectivity modeling. LightGCN (He et al., 2020) is a foundation stone for applying graph convolutional neural networks to multi-modal recommendation systems. Exploring user preferences from multiple modalities information of items (Zhou et al., 2023a; Zhang et al., 2021) has been a hot topic for MRS. For example, LATTICE (Zhang et al., 2021) enhances node representations in the graph by learning from multi-modal information. DRAGON (Zhou et al., 2023a) learns dual representations of users and items by constructing homogeneous and heterogeneous graphs. Multi-modal graph contrastive learning (MMGCL) (Yi et al., 2022) ensures the effective contribution of each modality through different data augmentation. FREEDOM (Zhou & Shen, 2023) freezes the large model and only trains the end of the model and graph denoising, achieving good results. Other methods such as DRAGON (Xia et al., 2023) learns dual representations of users and items by constructing both homogeneous and heterogeneous graphs. GraphCAR (Xu et al., 2018) uses graph convolutional network (GCN) to capture higher-order connectivity and enhance preference features, they have used GCN-based CF methods to explore higher-order relationships between users and items. Dynamic graph evolution learning (DGEL) (Tang et al., 2023) uses the joint training of interaction matching task and prediction task to ensure the rapidness and timeliness of the recommendation system for the interaction between users and items. Multimodal graph meta contrastive learning (MGMC) (Zhao & Wang, 2021) assigns meta-learning to contrastive learning to provide generalization capabilities for graph contrastive learning. In constructing nodes, current methods use a single pathway to learn a signal representing the method; however, due to the influence of noise, the more accurate the signal learned in a single pathway, the more severe the phenomenon of overfitting in the experiment will be. DSGRec uses two pathways and four different methods to build nodes to improve the performance and enhance the robustness of the recommender system. Disentangled Graph Variational Auto-Encoder (DGVAE) (Zhou & Miao, 2024) leverages GCNs to encode and disentangle ratings and multimodal information, enabling it to learn latent representations of items from their neighboring items.

Hypergraph learning for recommendation

The utilization of hyperedges connecting multiple nodes allows for the construction of hypergraphs, which can serve as a supplementary tool in CF for extracting unexplored user-item relationship information (Gao et al., 2020). The hypergraph neural networks (HGNN) (Feng et al., 2019) framework incorporates hyperedge convolution to manage data correlations during the representation learning process. A hardware-friendly recommendation algorithm based on hyperdimensional computing (HyperRec) (Wang et al., 2020) achieves more granular recommendation effects by stacking hypergraph convolution networks, residual gating layers, and fusion layers, Dual channel hypergraph collaborative filtering (DHCF) (Ji et al., 2020) learns user and item representations separately so that these two types of data can be interconnected while maintaining their specific attributes. Hypergraph Click-Through Rate prediction framework (HyperCTR) (He et al., 2021) leverages information interaction between users and micro-videos by considering different modalities, Multi-channel Hypergraph Convolutional Network (MHCN) (Yu et al., 2021) incorporates self-supervised learning to reduce aggregation loss through hierarchical mutual information maximization, dual channel hypergraph convolutional network (DHCN) (Xia et al., 2021) introduces a dual-channel hypergraph convolutional network and incorporates self-supervised tasks into network training. This approach enhances hypergraph modeling and improves the performance of recommendation tasks, Co-guided Heterogeneous Hypergraph Network (CoHHN) (Zhang et al., 2022b) employs a dual-channel aggregation mechanism within a heterogeneous hypergraph network to integrate diverse information from heterogeneous nodes and multiple relations. In the last couple of years, several research efforts have integrated hypergraphs with self-supervised learning techniques by employing self-supervised learning as a regularization method for hypergraph learning (Xia, Huang & Zhang, 2022). However, approaches like hypergraph contrastive collaborative filtering (HCCF) (Xia et al., 2022) and local and global graph learning for multimodal recommendation (lgmrec) (Guo et al., 2024) focus on self-supervised learning on hypergraph information. However, self-supervised learning without filtering information similarly leads to the enhancement of noise, which undermines the ability of the model to capture item characteristics and user preferences. In contrast, our approach not only leverages hypergraphs as supplementary signals but also incorporates multi-modal features of items and user preferences, thereby enhancing the precision of recommendations.

Heterogeneous networks-based information collaboration

The HNIC module serves as the foundational component, aiming to learn representations of user demand for items and the multi-modal preferences. These two parts are learned separately through the propagation of ID embeddings and multi-modal information on the $User-Item$ interaction graph, corresponding to behavior-oriented demand embedding and interest-guided preference embedding.

Behavior-aware multi-modal signal augmentation

Aiming to make the learned multi-modal feature complement with each other, we design a behavior-aware multi-modal signal augmentation module. Specifically, we first transform the original modality embeddings into shared feature space through a linear transform:

(10) $${\stackrel{..}{\mathbf{E}}}_i^m={\mathbf{W}}_1{\mathbf{E}}_i^m+b_1,$$where matrix ${\mathbf{W}}_1\in {\mathbb{R}}^{d\times d_m}$ and the bias vector $b_1\in {\mathbb{R}}^d$ in the transition here is trainable, unlike ${\mathbf{W}}_m$. Then, a behavior-guided purifier is utilized to select the preference-relevant modality features ${\stackrel{..}{\mathbf{E}}}_i^m$ from the item’s representation ${\mathbf{E}}_i^{id}$, which is proven to be effective in previous work (MGCN; Yu et al., 2023). The behavior-guided purifier $f_{gate}^m$ is defined as follow:

(11) $${\stackrel{..}{\mathbf{E}}}_i^m={\mathit{f}}_{gate}^m({\mathbf{E}}_i^{id},{\stackrel{.}{\mathbf{E}}}_i^m)={\mathbf{E}}_i^{id}\odot \sigma ({\mathbf{W}}_2({\stackrel{.}{\mathbf{E}}}_i^m+b_2)),$$where matrix ${\mathbf{W}}_2\in {\mathbb{R}}^{d\times d_m}$ and $b_2$ are learnable parameters, $\odot$ represents the element-wise product, and $\sigma$ is the sigmoid function.

To obtain an accurate item modality vector and avoid contamination from user preferences, we only propagate modality features in the I − I graph. First, we conduct KNN sparsification (Chen, Fang & Saad, 2009) on the I − I graph. KNN sparsification can effectively reduce training costs and remove unnecessary noise as much as possible. The similarity between item $a$ and $b$ on modality $m$ is denoted as $s_{a,b}^m$:

(12) $$s_{a,b}^m=\frac{{(e_a^m)}^T{e}_b^m}{||e_a^m||||e_b^m||},$$where $e_a^m$ and $e_b^m$ represent the feature of item $a$ and $b$ in modality $m$. To ensure that the edges retained in the graph are truly effective, we retain the edges with the top $k$ highest similarity in each row and remove the other edges. All the similarity values construct a affinity matrix ${\mathbf{S}}^m$ in Eq. (13), where the element in the row $a$ and column $b$ of it is computed with Eq. (12).

(13) $${\stackrel{.}{\mathbf{S}}}^m=\{\begin{array}{cc}{s}_{a,b}^m,\hfill & s_{a,b}^m\in \mathrm{t}\mathrm{o}\mathrm{p}-\mathrm{K}(s_{a,c},c\in \mathit{I}),\hfill \\ 0,\hfill & \mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}.\hfill \end{array}$$

We also normalize the item-item affinity matrix to prevent gradient explosion, which has been proven effective in prior work (Chen, Fang & Saad, 2009). The normalizeditem-item affinity matrix is defined as:

(14) $${\stackrel{..}{\mathbf{S}}}^m={\mathbf{D}}^{m\frac{1}{2}}{\stackrel{.}{\mathbf{S}}}^m{\mathbf{D}}^{m-\frac{1}{2}}.$$

Here, ${\mathbf{D}}^m$ is the diagonal matrix of ${\stackrel{.}{\mathbf{S}}}^m$. Then, we propagate all item modal features ${\stackrel{..}{\mathbf{E}}}_i^m$ through the corresponding $Item-Item$ affinity matrix ${\stackrel{.}{\mathbf{S}}}^m$ by the LightGCN (He et al., 2020):

(15) $${\overline{\mathbf{E}}}_i^m={\stackrel{..}{\mathbf{S}}}^m{\stackrel{..}{\mathbf{E}}}_i^m.$$

It can enrich features by capturing the common characteristics of similar items. However, in the I − I view, as the propagation path increases, the semantic similarity of the node modality features significantly decreases. Stacking multiple graph convolution layers not only leads to the over-smoothing problem of nodes but also easily captures noise features. Finally, we obtain the second user modality ${\overline{\mathbf{E}}}_u^m$ under the I − I graph, user $u$’s modality feature ${\overline{e}}_u^m$ is expressed as:

(16) $${\overline{e}}_u^m=\sum _{i\in N_u}\frac{1}{\sqrt{|N_i||N_u|}}{\overline{e}}_i^m,$$where ${\overline{e}}_i^m$ is the vector of ${\overline{\mathbf{E}}}_i^m$. By concatenating ${\overline{\mathbf{E}}}_u^m$ with ${\overline{\mathbf{E}}}_i^m$, we obtain the behavior-related modal aggregation auxiliary feature ${\overline{\mathbf{E}}}^m\in {\mathbb{R}}^{d\times (|U|+|I|)}$ for each modality. To model the multi-modal preference signals with a fine-grained fusion mechanism, we design a behavior-aware fusion $P_m$ for each modality $m$. This module learns and adjusts parameters to determine the weight of each item’s attraction to users based on the behavior information matrix. It is defined as:

(17) $${\mathbf{P}}_m=\sigma ({\mathbf{W}}_3{\mathbf{E}}_{BD}+b_3),$$where ${\mathbf{W}}_3\in {\mathbb{R}}^{d\times d_m}$ and $b_3\in {\mathbb{R}}^d$ are learnable parameters, and $\sigma$ is the sigmoid function. $P_m$ describe the importance of each modality for different users. For flexible fusion weight allocation based on user modality preferences, we decompose the modal features into common features and unique features. First, we assume that the common features, derived from interaction information, are highly similar to the modal features of the user’s target needs. Then, we capture these common features (Vaswani et al., 2017; Wang, Wu & Hoashi, 2019) using an attention mechanism, resulting in the modal weight matrix $\alpha$:

(18) $${\alpha }_m=\mathrm{s}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{m}\mathrm{a}\mathrm{x}(q_1^{\mathrm{\top }}\tanh ({\mathbf{W}}_4{\mathbf{E}}^m+b_4)),$$where $q_1\in {\mathbb{R}}^d$ is the attention vector, ${\mathbf{W}}_4\in {\mathbb{R}}^{d\times d_m}$ and $b_4\in {\mathbb{R}}^d$ are the weight matrix and bias vector, respectively. These parameters are shared across all modalities. The shared modal features ${\mathbf{E}}_s$ are obtained by summing the weighted modality features:

(19) $${\mathbf{E}}_s=\sum _{m\in M}{a}_m{\overline{\mathbf{E}}}_m.$$

The unique modal features ${\tilde{\mathbf{E}}}_m$ are then derived by subtracting the shared features ${\mathbf{E}}_s$:

(20) $${\tilde{\mathbf{E}}}_m={\overline{\mathbf{E}}}_m-{\mathbf{E}}_s.$$

Finally, we adaptively fuse the modal-specific features ${\tilde{\mathbf{E}}}_m$ with the shared features ${\mathbf{E}}_s$ to obtain the behavior-related modal aggregation auxiliary embeddings ${\mathbf{E}}_f$:

(21) $${\mathbf{E}}_f={\mathbf{E}}_s+\frac{1}{|M|}\sum _{m\in M}{\tilde{\mathbf{E}}}_m\odot {\mathbf{P}}_m.$$

To ensure that the auxiliary information ${\mathbf{E}}_f$ is learned effectively under sparse data conditions and to better explore the relationship between CF signals and MF signals, we design a self-supervised auxiliary task to optimize ${\mathbf{E}}_f$ with the loss ${\mathbf{L}}_f$:

(22) $${\mathbf{L}}_f=\sum _{i\in I}-\log \frac{\exp (e_{i,f}\cdot e_{i,id}/{\tau }_1)}{{\sum }_{m\in I}\exp (e_{m,f}\cdot e_{m,id}/{\tau }_1)}+\sum _{u\in U}-\log \frac{\exp (e_{u,f}\cdot e_{u,id}/{\tau }_1)}{{\sum }_{n\in U}\exp (e_{n,f}\cdot e_{n,id}/{\tau }_1)}.$$

Here, $L_f$ can learn meaningful feature representations by maximizing the similarity of positive sample pairs while minimizing the similarity of negative sample pairs, and ${\tau }_1$ is a temperature hyper-parameter, $e_{u,f}$ and $e_{i,f}$ denote the features of $u$ and $i$ in ${\mathbf{E}}_f$, $e_{u,id}$ and $e_{i,id}$ denote the features of $u$ and $i$ in ${\overline{\mathbf{E}}}_{BD}$.

Hypergraph-guided cooperative signal enhancement

The over-smoothing effect in deeper graph-based CF architectures can lead to indistinguishable user representations and a degradation in recommendation quality. Inspired by the ability of hyper-graphs to capture complex high-order relationships globally, we introduce hyper-graph to capture high-order global hybrid representations of CF signal and MF signal missed that are missed in local learning due to data sparsity.

To make the model to learn the hypergraph structure adaptively, we define learnable vectors ${\mathbf{V}}_k^m\in {\mathbb{R}}^d$ to represent a set of hyper-edges, where $k\in K$ and K is the number of hyper-edges, $m\in M$ and M represents the set of modalities. As illustrated in Fig. 4, the process include hyperedge learning, matrix calculation and hyper-graph embedding learning.

First, we learn the hyperedges of item nodes from the original modality embeddings in the low-dimensional embedding space:

(23) $${\mathbf{H}}_i^m={\tilde{\mathbf{E}}}_i^m\cdot {{\mathbf{V}}^m}^{\mathrm{\top }},$$where ${\mathbf{H}}_i^m\in {\mathbb{R}}^{d\times |K|}$ represents the item hyperedge dependency matrices. ${\tilde{\mathbf{E}}}_i^m$ is the original item modality feature matrix computed with Eq. (5), and ${\mathbf{V}}_m$ = $[V_1^m,..,V_K^m]\in R^{K\times d_m}$ is the hyperedge vector matrix. Based on ${\mathbf{H}}_i^m$ and user behavior, the hyperedge ${\mathbf{H}}_u$ for users can be obtained as follows:

(24) $${\mathbf{H}}_u^m={\mathbf{A}}_u\cdot {{\mathbf{H}}_i^m}^{\mathrm{\top }},$$where ${\mathbf{H}}_u^m$ represents user-hyperedge dependency matrices, and ${\mathbf{A}}_u\in {\mathbb{R}}^{|U|\times |I|}$ is the user-related adjacency matrix extracted from adjacency matrix $\mathbf{A}$. This ensures that the hyperedges connect nodes with similar modalality features, allowing the hyperedge embedding to globally correct the embedding vector previously obtained by CF through user behavior information. To ensure fairness in each node’s contribution to the hypergraph and avoid the same node being captured by multiple hyperedges, we apply Gumbel-Softmax reparameterization (Jang, Gu & Poole, 2017) for each node:

(25) $${\tilde{h}}_{i,\ast }^m=\mathrm{S}\mathrm{O}\mathrm{F}\mathrm{T}\mathrm{M}\mathrm{A}\mathrm{X}\left(\frac{\log \delta -\log (1-\delta )+h_{i,\ast }^m}{\tau }\right),$$where ${\tilde{h}}_{i,\ast }^m\in {\mathbb{R}}^K$ is the $i$-th row vector of ${\mathbf{H}}_i^m$. $\delta \in {\mathbb{R}}^K$ is a noise vector, where each value ${\delta }_k\sim \mathrm{U}\mathrm{n}\mathrm{i}\mathrm{f}\mathrm{o}\mathrm{r}\mathrm{m}(0,1)$, and $\tau$ is the temperature hyperparameter. The SOFTMAX represents the softmax function, which ensures differentiable smapling. Subsequently, we obtain the enhanced item-attribute hypergraph dependency matrix ${\hat{\mathbf{H}}}_i^m$. By performing a similar operation on ${\mathbf{H}}_u^m$, we derive the enhanced user-attribute relationship matrix ${\tilde{\mathbf{H}}}_u^m$. Similar with traditional CF approaches, we use hyperedges as bridges for the message-passing mechanism. Due to the characteristics of hyperedges, modality information is no longer limited by hop distance during transmission. The information of the entire graph can be propagated to each user $u$ and item $i$ using the following formulas:

(26) $${\mathbf{E}}_i^{m(l+1)}=S{H}_i\cdot {\mathbf{E}}_i^{m(l)},{\mathbf{E}}_u^{m(l+1)}=S{H}_u\cdot {\mathbf{E}}_u^{m(l)}.$$where ${\mathbf{E}}_i^{m(l)}$ and ${\mathbf{E}}_u^{m(l)}$ represents the hypergraph embedding matrices of items and users under modality $m$ at the $l$-th layer, respectively. We use ${\mathbf{E}}_{BD}$ as the initial embedding when $l$ = 0. $S{H}_i$ and $S{H}_u$ are matrices describing the global relationship between nodes with learned hyperedges, computed as follows based on prior work (Guo et al., 2024):

(27) $$S{H}_i=Drop({\tilde{\mathbf{H}}}_i^m)\cdot Drop({\tilde{\mathbf{H}}}_i^{m^{\mathrm{\top }}}),S{H}_u=Drop({\tilde{\mathbf{H}}}_u^m)\cdot Drop({\tilde{\mathbf{H}}}_i^{m^{\mathrm{\top }}})$$where $Drop()$ represents the dropout function.

In hypergraph transmission, by using the hypergraph dependency perceived through user behavior as input, we achieve global CF and MF information propagation, supplementing or denoising the information missed in the local information transmission of previous work. Subsequently, the embeddings of users and items are stacked to obtain the hybrid feature embedding matrix ${\mathbf{E}}_h$:

(28) $${\mathbf{E}}_h=\sum _{m\in M}Concat({\mathbf{E}}_u^{m(l)},{\mathbf{E}}_i^{m(l)}),$$where ${\mathbf{E}}_u^{m(l)}\in {\mathbb{R}}^{|U|\times d}$ and ${\mathbf{E}}_i^{m(l)}\in {\mathbb{R}}^{|I|\times d}$ are the embedding matrices representing the global hybrid features of user $u$ and item $i$ under modality $m$ at the $l$-th layer, respectively. $Concat()$ represents the concatenate operation.

To explore the collaborative relationship between demand signals and modality embedding signals, we optimize each signal independently before fusion. For the HCSE module, due to data scarcity and the complexity of comparing cross-modal information, we utilize contrastive learning to optimize $E_h$ in a self-supervised manner (Gutmann & Hyvärinen, 2010). Specifically, we treat embeddings of the same user/item under the different modalities as postive pairs and embeddings of different users/items under different modalities as negative pairs:

(29) $$L_h=\sum _{u\in U}-\log \frac{\exp ({\mathbf{e}}_u^{v(l)}\cdot {\mathbf{e}}_u^{t(l)}/{\tau }_2)}{{\sum }_{u^{\mathrm{\prime }}\in U}\exp ({\mathbf{e}}_u^{v(l)}\cdot {\mathbf{e}}_{u^{\mathrm{\prime }}}^{t(l)}/{\tau }_2)}+\sum _{i\in I}-\log \frac{\exp ({\mathbf{e}}_i^{v(l)}\cdot {\mathbf{e}}_i^{t(l)}/{\tau }_2)}{{\sum }_{i^{\mathrm{\prime }}\in I}\exp ({\mathbf{e}}_i^{v(l)}\cdot {\mathbf{e}}_{i^{\mathrm{\prime }}}^{t(l)}/{\tau }_2)},$$where ${\tau }_2$ is the temperature factor for this loss function, and $u^{\mathrm{\prime }}$ and $i^{\mathrm{\prime }}$ are randomly sampled negative samples for user $u$ and item $i$, respectively.

Adaptive fusion and prediction

We obtain the final embedding $e$ of users and items by aggregating the collaborative embedding ${\mathbf{E}}_{HNIC}$ from the heterogeneous network, the augmented multi-modal features ${\mathbf{E}}_f$, and the global hybrid features ${\mathbf{E}}_h$ from the hyper-graph network:

(30) $$e={\mathbf{E}}_{HNIC}+\alpha \cdot NORM({\mathbf{E}}_h)+\beta \cdot NORM({\mathbf{E}}_f),$$where $NORM()$ is a normalization function to alleviate the value scale difference among embeddings, and $\alpha$ and $\beta$ are weighting factors. Following He et al. (2020), we adopt the inner product to calculate the prediction score between user $u$ and item $i$.

(31) $${\hat{\mathbf{r}}}_{u,i}=e_u^T{e}_i.$$

After obtaining the ${\hat{\mathbf{r}}}_{u,i}$, we use Bayesian personalized ranking (BPR) loss (Rendle et al., 2012) to optimize the parameters of DSGRec:

(32) $${\mathbf{L}}_{BPR}=-\frac{1}{|D|}\sum _{(u,i^{+},i^{-})\in D}\ln \sigma ({\hat{\mathbf{r}}}_{u,i^{+}}-{\hat{\mathbf{r}}}_{u,i^{-}}),$$where $(u,i^{+},i^{-})$ is a set of triples for training. Here, $u$ is the user embedding vector, $i^{+}$ is an item that user $u$ has interacted with, and $i^{-}$ is a randomly sampled negative item from the dataset. Finally, we integrate the loss functions of each component as follows:

(33) $$\mathbf{L}={\mathbf{L}}_{BPR}+{\lambda }_f{\mathbf{L}}_f++{\lambda }_h{\mathbf{L}}_h+{\lambda }_E||\varphi |{|}_2,$$where ${\lambda }_f$, ${\lambda }_h$, and ${\lambda }_E$ are hyperparameters for weighting the loss terms. $\varphi$ represents the model parameters, which are regularized using $L_2$ regularization to prevent overfitting.

Experimental setup

Overall performance comparison (RQ1)

Table 2 presents the performance of all methods on three datasets. Specifically, our method outperforms baseline models and shows strong competitiveness against the latest models on two datasets, which improved by $2.84\mathrm{\%}$, and $2.17\mathrm{\%}$ in terms of Recall@ $20$ for the baby and sports, respectively. In addition, the performance on the Clothing dataset is superior to most baseline method. Further analysis of the two types of models reveals: (1) Comparison with CF-type models: Our model performs better on the Sports and Clothing datasets, which we attribute to its emphasis on modality information. Some CF models overestimate the role of collaborative signals in capturing user interests, leading to suboptimal performance in interest-dominated environments. (2) Comparison with multi-modal models: Our model’s performance on the demand-guided baby dataset indicates that some models overly rely on modal information learning while neglecting demand information exploration. DSGRec adaptively integrates high-order modality information with demand signals, ensuring robust performance.

Ablation study (RQ2)

To investigate the complexity of user behavior and explore the specific contributions of different modules, we conduct an ablation study by comparing DSGRec with four variants:

• w/o BDE: Disable the behavior-oriented demand embedding module but retain its participate in auxiliary tasks.

• w/o IPE: Remove the interest-guided preference embedding module, and set ${\mathbf{E}}_{HNIC}={\overline{\mathbf{E}}}_{BD}$.

• w/o BMSA: Remove the behavior-aware modal signal augmentation module.

• w/o HCSE: Remove the hypergraph-guided cooperative signal enhancement module.

Table 3 reports Recall@ $20$ and NDCG@ $20$ of these variants on the three datasets, leading to the following findings:

(1) Among these four models, the variants w/o BDE and w/o IPE exhibit the worst performance, demonstrating that user behavior cannot be comprehensively modeled by methods using a single signal. DSGRec’s decomposition of behavioral signals into demand and perefernce components enables independent learning and effective collaboration.

(2) The modules w/o BMSA and w/o HCSE indeed play a role in optimization. The performance gap varies across datasets, indicating that interest and demand signals are not conflicting and can sometimes be optimized by a single module. This highlights the limitations of single-path modeling and the benefits of multi-module collaboration.

Hyperparameter discussion (RQ3)

In this section, we discuss the hyperparameters settings for the three datasets in our experiments. Table 4 details the optimal parameter settings. Additionally, Fig. 5 presents the Recall@ $20$ and NDCG@ $20$ for different $\alpha$ and $\beta$ in detail.

As the Table 4 demonstrates, the weight $\alpha$ of Behavior-aware Modal Signal Augmentation is positive for the Baby and Sports datasets but negative for the Clothing datasets. Conversely, the weight $\beta$ of Hypergraph-guided Cooperative Signal Enhancement is opposite to $\alpha$. This pattern, illustrated in Fig. 5, aligns with our hypothetical model: the interpretation of interactive behavior should be context-dependent.

In the demand-driven Baby dataset, modal information may interfere with the recommendation accuracy. On the contrary, due to the GCN limitations, the demand signal requires hypergraph enhancement to achieve optimal performance. In the Clothing dataset, the optimal $\alpha$ value is $-0.9$, indicating that the demand signal struggles to capture user preferences effectively. The preference signal demonstrates a significant advantage, explaining why DSGRec underperforms compared to multi-modal methods on this dataset.

Visualization

We utilize a visualization module to create $2$D graphs showing the representation of nodes from initial view, demand view, preference view and DSGRec view. We randomly sample $5\text{,}000$ items from the dataset and map their embedding vectors into a $2$-dimensional space using t-SNE (Van der Maaten & Hinton, 2008). In Fig. 6, the first row visualizes initial embedding vectors, and the second row shows results after training. Each column corresponds to different method variants, as in the ablation study. Node colors represent different clusters. From the figure, we observe that the initial view embeddings displays a random distribution of users and items, while other views result in more distinguishable distributions.