Following the general approach used in existing recommendation methods, we initialize the u∈U and i∈I embedding vector as eu0∈Rd, ei0∈Rd, where d represents embedding size. Let P∈RU×d, Q∈RI×d represent the embedding matrices of users and items, respectively. Each user or item has a unique ID represented by a one-hot vector, where IDU, IDI are the one-hot vector matrix for all users and items. It is then transformed into low-dimensional embeddings through a shared embedding layer with P={pu10,pu20,…,pu|𝒰|0} and Q={qi10,qi20,…,qi|ℐ|0}. The embeddings of user u and item i are initialized as: (1)eu0=P⊤⋅IDuU,ei0=Q⊤⋅IDiI,where IDuU, IDiI denote the one-hot vectors of users and items, respectively.

Inspired by previous works [31], we have designed a user-item interaction graph encoder based on LightGCN with a residual design, where each LightGCN corresponds to a specific behavior, as shown in Fig. 4. This residual design uses residual connections to directly propagate the node representations from the previous layer to the next layer, thereby effectively preserving the node’s own features and mitigating the over-smoothing issue in networks. The output representations of all layers in the network maintain the same dimension d as the initial embedding, ensuring that representations from different layers can be directly fused. In this module, the residual fusion weights adopt a non-parametric and equal-weight setting, meaning the contribution weight of each layer is fixed at 1, enabling simple and efficient inter-layer information transfer. This constitutes a non-parametric residual connection. This constitutes a non-parametric residual connection. Through an iterative aggregation and update mechanism, it captures multi-order connections in the user-item interaction graph under different behaviors, thereby learning the user and item embeddings for each behavior. Given the representation euk, eik for user u and item i under behavior bk, the aggregation and update of the user-item interaction graph can be formulated as: (2)euk,(l+1)=∑i∈Nuk1|Nuk||Nik|eik,(l), (3)eik,(l+1)=∑u∈Nik1|Nik||Nuk|euk,(l),where euk,(l+1), eik,(l+1) represent the user u and item i embeddings after l + 1 layers of propagation and update under behavior K. Let Nuk, Nik represent the set of items interacted with by user u and the set of users who have interacted with item i.

After L layers of propagation, the user’s embedding is represented as {eu(b,0),eu(b,1),…,eu(b,L)} and item’s embedding is represented as {ei(b,0),ei(b,1),…,ei(b,L)}. We aggregate the user and item representations under behavior K as: (4)euk=∑l=0Leu(k,l),eik=∑l=0Lei(k,l),

After the feature transformation operation, euk, eik, as learned from the behavior bk, are used as the input embedding vectors for the LightGCN of the next behavior. Importantly, the representations learned from the previous LightGCN are used as the input features for the users and items in the next LightGCN. Importantly, the representation of the behavior bk is used as the initial user and item information for the LightGCN encoder under behavior bk, which can be expressed as: (5)euk,(0)=euk′,eik,(0)=eik′.

Different types of behaviors reveal varying degrees of user preferences. In a behavior chain, the subsequent behavior typically exhibits stronger signals or more accurate user preferences than the preceding behavior. However, the contributions of different auxiliary behaviors to the target behavior vary. Ignoring these differences may lead to the loss of the diverse information conveyed by the different behaviors. Therefore, we introduce the behavior transition strength matrix from behavior to behavior. Different types of behaviors reveal varying degrees of user preferences. In a behavior chain, the subsequent behavior typically exhibits stronger signals or more accurate user preferences than the preceding behavior. However, the contributions of different auxiliary behaviors to the target behavior vary. Ignoring these differences may lead to the loss of the diverse information conveyed by the different behaviors. Therefore, we introduce the behavior transition strength matrix from behavior bk′ to behavior bk.

Behavior Transition Strength Matrix. Let WUk∈RU×k be the behavior transition strength Matrix, which represents the user-adaptive distribution of all auxiliary behaviors {b1,b2,…,bk−1}. WUk′→k∈RU represents the behavior transition strength matrix from behavior bk′ to behavior bk for all users, which 0≤k′≤k−1. In the directed behavior path graph, a specific transition strength weight is assigned to each directed edge. We initialize the transition strength weights for each graph and update them through learning in subsequent modules.

Inspired by the Markov process, we only consider the nearest preceding behavior in each behavior path, assuming that it has the most direct influence on the terminating behavior. We investigated the correlation from the given starting behavior bk′ to the terminating behavior bk, comprehensively considering all paths leading to the terminating behavior and fully accounting for the influence of preceding behaviors in its formation. Taking the user’s directed behavior path graph as an example, we calculate the probability of items that appeared only in the starting behavior bk′ appearing in the terminating behavior bk. First, we remove the items from the terminating behavior bk that appeared in the behaviors between bk′ and bk. Then, we take the intersection with Nuk′ to get the result, which represents the set of items directly transitioning from behavior bk′ to bk in the user’s behavior path. Finally, we use the posterior probability of the terminating behavior bk to normalize its correlation with all possible starting behaviors, obtaining the behavior transition strength from behavior resents the set of items directly transitioning from behavior bk′ to bk for the user and item. This is formally represented as follows: (6)wuk′→k=|(Nuk−∪j=kv+1k−1Nuj)∩Nuk′||Nuk|, (7)wik′→k=|(Nik−∪j=k′+1k−1Nij)∩Nik′||Nik|,where Nuk, Nik represent the set of items interacted with by user u and the set of users interacting with item i under behavior bk∈ℬ. To mitigate scaling differences, we employ weight normalization to account for message variations across different behaviors, using the Softmax function to normalize the behavior distribution: (8)w~uk′→k=wuk′→k∑k′=0k−1wuk′→k, (9)w~ik′→k=wik′→k∑k′=0k−1wik′→k.

We propose a graph attention-based information aggregation encoder, which effectively aggregates user and item feature embeddings across different behaviors based on varying behavior transition strengths. It updates user preferences and item characteristics through the behavior chain. To preserve the features of previous behaviors, our model employs a residual operation to connect the feature embeddings learned from prior behaviors with those learned from current behavior. This method connects the feature embeddings learned from previous behaviors with those learned from the current behavior, enabling the model to effectively learn high-order interaction patterns while retaining crucial information from the original behavioral features. All embeddings maintain the same feature dimension d, including the embedding of the current behavior bk, the embedding of its direct predecessor behavior bk−1, and the embeddings corresponding to all historical behaviors. This ensures that features from each stage can be directly fused via weighted aggregation. In this module, the residual fusion weights are set as non-learnable parameters, with the weight for the residual term corresponding to the direct predecessor behavior bk−1 fixed at 1. This enables our model to benefit from higher-order GCN operations while retaining the original behavior features. We update user and item embeddings using a graph attention information aggregation encoder: (10)e~uk=euk−1+∑k′=0k−1w~uk′→keuk′, (11)e~ik=eik−1+∑k′=0k−1w~ik′→keik′,where e~uk, e~ik represent the user preference representation and item characteristic representation under behavior bk, respectively. Similarly, euk′, eik′ represent the user preference representation and item characteristic representation under behavior bk′.

In the model optimization process, we introduce the Multi-Task Learning (MTL) framework [32,33] to quantify the contribution of each auxiliary task to the target task prediction, enabling collaborative optimization across different tasks. Therefore, to ensure that the target task can effectively benefit from the predictions of auxiliary interactions, it is essential to accurately measure the correlation between the target behavior and auxiliary behaviors. We calculate the item similarity under the target and auxiliary behaviors as a metric for evaluating the correlation between the two behaviors. We calculate the correlation score cuk→K from the auxiliary behavior {b1,b2,…,bK−1} to the target behavior bK using the Cosine similarity for user u, which can be defined as: (13)cuk→K=|Nuk∩NuK||Nuk|⋅|NuK|,k=1,2,…,K,where Nuk is the set of items interacted with by user u under behavior bk. For a specific user, a higher correlation score indicates a stronger interaction between the given auxiliary behavior and the target behavior, suggesting that the auxiliary behavior has a significant impact on the prediction of the target behavior.

Similar to other ranking-based recommendation approaches, a pairwise learning strategy is employed for model optimization. In the multi-behavior dataset, Bayesian Personalized Ranking (BPR) is used as the optimization criterion during the training phase. The BPR loss function is widely used in recommendation systems [10], and by emphasizing the relative order of observed and unobserved user-item interactions, it guides the model to capture user preferences and item characteristics more accurately. Based on this, a loss function is constructed to ensure that the model assigns higher predicted scores to observed interactions. The loss function is formulated as follows: (14)ℒk=∑(u,i,j)∈Ok−cuk→K(ln⁡σ(y^uik−y^ujk)),where Ok={(u,i,j)|(u,i)∈Yk,+,(u,j)∈Yk,−} is the training set for behavior bk. Yk,+ (Yk,−) represent the observed (or unobserved) samples in the target behavior. σ denotes the sigmoid function.

To achieve the joint modeling of behavioral patterns and item semantics, we propose an item semantic alignment constraint. This constraint aligns the feature representations from two distinct perspectives within a latent space by minimizing the discrepancy between item features learned from behavioral path graphs and item attribute features obtained via an attribute encoder, thereby enhancing the representational capacity of item features. First, we use the pre-trained language model Recformer [34] to learn item-attribute embeddings. This model transforms all textual attribute information for an item (e.g., title, color, brand) into key-value pairs, allowing different items to have different sets of attributes. Subsequently, all key-value pairs of an item are flattened into an item “sentence”. Finally, the pre-trained language model is utilized to derive the feature representation of this item sentence. Formally, given an item i with n attributes, its key-value pairs {(x1i,v1i),(x2i,v2i),⋯,(xni,vni)} are flattened into a sequence {x1i,v1i,x2i,v2i,⋯,xni,vni}. Prepending the special [CLS] token, the input is fed into the Recformer model: (15)e~i=Recformer(Flatten({(x1i,v1i),(x2i,v2i),…(xni,vni)}))where (x1i,v1i),(x2i,v2i),…(xni,vni) denotes the n attribute key-value pairs of item i. Flatten(⋅) represents flattening the key-value pair sequence into a continuous text sequence and adding a special [CLS] token at the beginning. Recformer(⋅) denotes a pretrained language model, which takes the flattened text sequence as input and outputs the hidden state corresponding to the [CLS] token as the attribute semantic feature e~i of item i.

To establish synergy between behavioral patterns and item semantics, the proposed alignment constraint reduces the distance in the vector space between the behavior-learned item representation and the semantically-learned attribute representation. Specifically, for item i, its behavior-based feature e~iK and its attribute feature ei~ form a positive pair, while e~iK and the attribute features of other items in the batch form negative pairs. The contrastive loss is defined as: (16)ℒcontra=−1|ℐbatch|∑i∈ℐbatchlog⁡exp⁡(sim(e~iK,ei~)/τ)∑j∈ℐbatchexp⁡(sim(e~iK,ej~)/τ),where ℐbatch is the set of items in the current training batch, sim(⋅) denotes the cosine similarity function, and τ>0 is the temperature parameter.

This contrastive loss effectively promotes the deep integration of behavioral patterns and item semantics by maximizing the agreement between the behavioral-view and semantic-view representations of the same item while minimizing the similarity between representations of different items, thereby enhancing the discriminative power of the learned item features. Finally, the objective function of the DBMR model proposed in this paper can be expressed as: (17)ℒ=ℒk+α∑k=0K−1ℒk+λℒcontra+β||Θ||22,where α is the discounting parameter between the auxiliary task and the target task, λ is the coefficient of the semantic alignment constraint function for the item, and β controls the L2-regularization to prevent overfitting. Θ={P,Q,WU,WI} is the representation of all trainable parameters. The detailed training process of DBMR is shown in the Algorithm 1.

To evaluate the performance of DBMR, we conducted experiments on two real-world publicly available datasets: Tmall and Taobao. The statistics of these two datasets are summarized in Table 1.

Tmall¹1

https://www.tmall.com

Taobao²2

https://tianchi.aliyun.com/dataset/dataDetail?dataId=649

To ensure fairness and facilitate comparison with existing methods, we merged duplicate user-item interaction data and filtered out users and items with fewer than 5 purchase interactions to ensure the representativeness and validity of the data. Subsequently, the last purchase record of each user was taken as the test set, while the remaining records were used as the training set. In the performance evaluation phase, we employed two commonly used Top-K recommendation evaluation metrics: Normalized Discounted Cumulative Gain (N@K) and Hit ratio (H@K), where K represents the length of the recommendation list and K = 10, 20, 50, 80.

To thoroughly evaluate the recommendation performance of DBMR, this study selected two types of baseline algorithms (single-behavior recommendation models and multi-behavior recommendation models) for comparative experiments.

(1) Single-Behavior Recommendation Model

NeuMF [35]: It is a typical deep learning-based recommendation algorithm that enhances nonlinear matrix factorization through a multi-layer perceptron module and a generalized matrix factorization module, allowing it to simultaneously capture both low-dimensional and high-dimensional user-item interaction signals.

MF-BPR [10]: It is a widely used matrix factorization model with BPR as its objective function, optimized using Bayesian Personalized Ranking (BPR) loss for matrix factorization, applied to Top-K recommendation tasks.

LightGCN [3]: It is a lightweight graph convolutional network recommendation model that eliminates feature transformations and nonlinear activations from NGCF, making it simpler and more suitable for recommendation tasks.

(2) Multi-Behavior Recommendation Model

NMTR [23]: This approach integrates Neural Collaborative Filtering (NCF) with multi-task learning to model the dependencies of various types of user interaction behaviors, mining user preferences from multi-behavior data for recommendation.

MB-GMN [36]: It is a recommendation method capable of extracting effective representations of users and items from complex multi-behavior data. This model combines meta-learning with graph neural networks, utilizing a graph meta-network to capture meta-knowledge about users, items, and behavior categories, thereby capturing behavior-based user and item feature information for accurate recommendations.

S-MBRec [37]: This model learns the importance of different behaviors in a supervised manner, while capturing the commonalities and differences in user behaviors, thereby improving the accuracy of recommendations.

CRGCN [24]: This method utilizes a cascaded GCN model with residual connections to learn user preferences and optimizes the model under a multi-task recommendation framework.

MB-CGCN [25]: This method effectively models the dependencies of behaviors within a sequential behavior chain, using user and item embeddings learned from one behavior as input features for the embedding learning process of the next behavior. Additionally, a feature transformation module is designed before each behavior’s information propagation stage in the model, effectively filtering and transforming noisy information.

DA-GCN [26]: This method utilizes the partial order relationship in a monotonic behavior chain and extends it to a personalized directed acyclic behavior graph to leverage behavior dependencies for recommendation.

In the aforementioned multi-behavior recommendation baselines, NMTR, CRCCN, and MB-CGCN are categorized as cascaded models, primarily relying on the determined behavior chains. In contrast, MB-GMN, S-MBRec, and DA-GCN are classified as parallel models, typically using higher-order operations to process the semantic and contextual information of each behavior separately, thereby forming complex and diverse model architectures.

For all models, unless explicitly stated in the literature, this experiment uses the Adam optimizer and stochastic gradient descent strategy for model optimization and training, with the Adam optimizer’s learning rate set to 1e−3. For the comparison methods, we first initialize the baseline algorithms with the parameters specified in the literature, and then tune the parameters based on the dataset and optimization strategy. To ensure a fair comparison, the embedding dimensions of both the models and the baseline comparison methods are set to 64. For the user-item interaction graph, we explore the number of Light-GCN layers for each behavior within the set 1, 2, 3, 4. For multitasking, the goal reference [26] recommend varying the trade-off hyperparameter between the target task and auxiliary tasks over the range 0.2 to 3.0, with a step size of 0.2, and setting the regularization coefficient to 0.001.

The proposed DBMR employs attention mechanisms in two distinct modules: behavioral transition intensities for message aggregation in directed behavioral path graphs and behavioral correlation for multi-task learning. To analyze the impact of different attention mechanisms on recommendation performance, this section compares DBMR with the following two method variants under specific configurations:

w/o BTS: This variant removes the Behavior Transition Strengths (BTS) module from the DBMR method, which learns contributions of auxiliary behaviors. By setting the behavioral transition strength w in Eqs. (8) and (9) to the same value, this variant treats all behaviors’ influence on the target equally during modeling.

w/o BRE: This variant removes the Behavioral Relevance (BRE) module from the multi-task learning part of the DBMR method. The BRE module measures the correlation between target and auxiliary behaviors. By setting behavioral correlation scores c in Eq. (13) to identical values, this variant gives equal importance to recommendation tasks across different behaviors during multi-task learning.

Fig. 5 presents the performance comparison between the DBMR method and its two variants on the Tmall and Taobao datasets. The experimental results demonstrate that the removal of either attention mechanism leads to a decline in DBMR’s performance on both H@K and N@K evaluation metrics, validating the critical role of attention mechanisms in the model. Specifically, the maximum performance reductions observed on the Tmall dataset reach 2.07% and 3.32%, while on the Taobao dataset they reach 6.44% and 11.21%, respectively. These experimental results confirm that using uniform weights fails to achieve optimal performance, whereas explicitly learning the distinct contributions of different auxiliary behaviors to target behavior prediction through differentiated weighting strategies significantly enhances the model’s effectiveness.

To validate the effectiveness of introducing the Semantic Alignment Constraint (SAC), a model variant designated as w/o SAC was generated by eliminating this constraint. This section compares the DBMR method with the w/o SAC variant.

Fig. 6 presents a comparative analysis of the recommendation performance between DBMR and the w/o SAC variant across both datasets. The experimental results indicate that the w/o SAC variant exhibits varying degrees of performance degradation in predictive capability on both datasets. These findings demonstrate the critical role of semantic alignment between attribute features and behavioral paths in enhancing recommendation accuracy. By integrating static item attributes with dynamic behavioral patterns within the representation space, the model learns more discriminative feature representations, effectively mitigates data sparsity, and enhances the ability to capture users’ genuine intents, thereby significantly improving the precision and robustness of recommendations.