A multi-behavior recommendation method and device based on structure-aware multi-view cascaded fusion

Abstract

A multi-behavior recommendation method and device based on structure-aware multi-view cascaded fusion, the method comprising: loading a data set and constructing a global interaction graph; global embedding initialization, global graph representation learning; decomposing behavior views for different behaviors; hierarchical fusion of different views and cascaded historical behavior information transmission; using a structure-aware alignment auxiliary task to improve robustness; multi-task collaborative optimization training, simultaneously optimizing the main task based on personalized weighted Bayesian loss and the structure-aware alignment auxiliary task; using the final embedding obtained by training to calculate the prediction score and generate a recommendation list. The application can effectively solve the semantic collapse and static bias problems of multi-view data, and improve the accuracy and robustness of the recommendation system.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a multi-behavior recommendation method and apparatus based on structure-aware multi-view cascade fusion. This method can be widely applied in various scenarios such as content distribution platforms and short video recommendations, and is used to model the inherent relationships between user behaviors at different stages, aiming to improve the accuracy and robustness of recommendation systems. Background Technology

[0002] With the explosive growth of internet content, user interactions across various platforms have become significantly more diverse and complex. Recommendation systems, as a crucial means of connecting users with information, aim to accurately capture user interests and preferences by analyzing historical behavioral data, providing personalized content or product recommendations. However, traditional single-behavior recommendation models (such as single-task learning based on clicks or purchases) often model only one type of interaction, neglecting the evolution of user behavior at different stages and failing to depict the behavioral transfer patterns from superficial interests to deeper intentions. This single-modeling approach can only capture preference signals in one dimension, making it difficult to fully utilize the potential correlations between multiple types of behavior. Therefore, it has limited performance in dealing with data sparsity, long-tail projects, and cold-start problems.

[0003] Therefore, Multi-Behavior Recommendation (MBR) has emerged. Its core idea is to enhance the modeling ability of the primary target behavior (such as purchasing) by introducing user auxiliary behaviors (such as browsing, saving, and adding to cart). Auxiliary behaviors often reflect users' potential interests or intention signals, thus providing richer contextual information for the target behavior. In recent years, with the rapid development of graph neural networks (GNNs), researchers have begun to introduce them into MBR scenarios. GNNs, by propagating neighborhood information on the user-item bipartite graph, can mine the implicit relationships between users and items at higher-order connection levels, capturing complex interaction structure features, thereby effectively improving the model's representation ability and recommendation accuracy.

[0004] Despite this, existing multi-action recommendation models still face several key challenges and bottlenecks in both theory and practice: First, there are the problems of view granularity conflict and semantic collapse. Multi-action recommendation systems typically require unified modeling across views of different granularities and sources. Some existing methods employ learnable attention-weighted strategies to fuse views, but because attention weights are susceptible to noise, this often leads to overfitting and instability. Other methods use simple averaging or linear fusion, lacking explicit modeling of the view hierarchy. These methods compress semantically different multi-source features into similar embedding spaces, causing the so-called semantic collapse phenomenon, thus weakening the model's representational power. Second, there is the problem of a lack of hierarchical constraints in structural alignment. In recent years, contrastive learning has been widely used to improve multi-view... Figure 1 While most methods achieve consistency, they directly align all views globally without considering the inherent hierarchical relationships and semantic role differences between views with different behaviors. Forcing all views to align pairwise during optimization can disrupt the original hierarchical structure, making it difficult for the model to maintain both consistency and diversity, and potentially exacerbating semantic collapse. Finally, there's the issue of behavioral imbalance and static bias. In real-world recommendation scenarios, the number of weak behavior samples far exceeds that of strong behaviors, and interaction data for long-tail items is extremely sparse. This inherent imbalance makes the model more susceptible to being dominated by high-frequency behaviors during training, leading to static bias and weakening the modeling accuracy for key target behaviors. Furthermore, excessive weak signals introduce noise, making it difficult for the model to effectively distinguish the contributions of different behaviors in multi-task learning, resulting in decreased recommendation performance.

[0005] In summary, existing multi-action recommendation methods still have significant shortcomings in handling multi-view fusion, structural alignment, and sample imbalance, making it difficult to achieve stable optimization while maintaining semantic diversity and structural consistency. Designing a method that can balance semantic differences and consistency in multi-level structures has become an important problem that urgently needs to be solved in the field of multi-action recommendation. Summary of the Invention

[0006] To overcome the aforementioned shortcomings of the prior art, this invention proposes a multi-behavior recommendation method and apparatus based on structure-aware multi-view cascade fusion, aiming to solve the problems of static bias in multi-behavior data and semantic collapse and insufficient retention of structural information in multi-view fusion in the prior art.

[0007] Specifically, this invention proposes a multi-behavior recommendation method based on structure-aware multi-view fusion. The core technologies include: 1. Prior-Anchored Adaptive Fusion (PAF): This method uses prior-anchored adaptive fusion to achieve multi-level fusion within and between behaviors, adaptively adjusting the weights of each view based on a stable prior to balance the robustness and expressiveness of the model; 2. Structure-Aware Contrastive Alignment (SCA): By introducing a structure-aware InfoNCE contrastive alignment mechanism, and adhering to anchoring and exclusion principles, constraint alignment is performed at both cross-behavior and intra-behavior levels to effectively maintain the semantic differences and structural consistency of each view and prevent semantic collapse; 3. Personalized Weighted Collaborative Training: To address the static bias problem in multi-behavior data, personalized adaptive weights based on Jaccard similarity are introduced. Differentiated weights are assigned to each training sample, and multi-task collaborative training optimization is performed using the Bayesian ranking loss function (BPR) to balance the different contributions of each behavior. In summary, this method can effectively capture personalized user signals and significantly improve the accuracy of multi-behavior recommendations.

[0008] The first aspect of this invention relates to a multi-behavior recommendation method based on structure-aware multi-view fusion, comprising the following steps: S1: Loading the dataset and constructing the global interaction graph. First, load the original dataset containing various user-item interaction behaviors (including browsing, adding to cart, and purchasing). Then, construct a global user-item bipartite graph with users and items as nodes and interaction records as edges to capture the basic relationship structure independent of behavior. S2: Global embedding initialization uses a lightweight graph convolutional network (LightGCN) to aggregate neighbor information on the global graph, resulting in a behavior-independent global embedding vector set. ,in These represent the global representations of users and items, respectively. S3: Behavior view decomposition, for each behavior The relational signals are decomposed into three complementary structural views, namely: a behavioral interaction view that models the direct interaction between users and items; a hypergraph view that uses a parameterized hypergraph convolutional structure to model high-order correlations at the group level; and a semantic view that constructs semantic dependencies between items through point mutual information. S4: Multi-view fusion and behavior cascading. To balance the robustness and expressiveness of the model, Prior-Anchored Adaptive Fusion (PAF) is adopted in multi-view information fusion. The weights of each view are adaptively adjusted based on a stable prior to complete multi-level fusion. S5: Structure-Aware Contrastive Alignment. After multi-view fusion, in order to maintain the semantic differences and structural consistency of each view, a structure-aware InfoNCE contrast alignment mechanism is introduced to prevent semantic collapse. S6: Multi-task collaborative training and optimization. Considering the static bias problem in multi-behavioral data, that is, different types of user behaviors vary greatly in quantity and signal strength, in order to balance the different contributions of each behavior, this step recalculates and assigns weights to each training sample based on the similarity between the user's auxiliary behavior and the target behavior, so as to capture the user's personalized signal and remove some irrelevant noise interference.

[0009] S7: Utilize the trained and optimized user and item embeddings to calculate their inner product score, and select the top K items with the highest scores as the recommendation results output.

[0010] Further, step S4 includes: Traditional multi-view or multi-behavior fusion methods, such as simple averaging, concatenation, or fixed-weighted summation, generally result in poor model robustness and limited expressive power of recommendation performance, often failing to effectively model the inherent order dependencies or transformation relationships between different behaviors. Furthermore, using complex attention mechanisms can easily lead to model overfitting and interference from noise information in some views. Therefore, this method, to balance model robustness and expressiveness, employs Prior-Anchored Adaptive Fusion (PAF) in multi-view information fusion to achieve multi-level information fusion within and between behaviors. The hierarchical fusion mechanism of PAF is described from two levels: within and between behaviors. For view fusion within behaviors, for each behavior... Define an unbiased prior distribution And introduce learnable residuals. To calculate the fusion weight of three different views under the same behavior, the specific formula for calculating the fusion weight within a behavior is as follows:

[0011] This weight is then used to further fuse the behavior interaction view, hypergraph view, and semantic view to obtain the embedding vector under the current behavior. The specific formula is expressed as follows:

[0012] in Indicates the current action Embedding vectors for different views. After obtaining information from different views within the current behavior, further information fusion between behaviors is performed.

[0013] A comprehensive representation of the current behavior and the output of the previous action Cascaded fusion is equivalent to combining historical context and behavior transformation information to obtain the final embedded representation of the current behavior after two levels of fusion:

[0014] in These represent the inter-behavior fusion weights for historical and current behaviors, respectively. Their calculation method is consistent with that of the intra-behavior fusion weights, in order to capture the dependencies between behaviors.

[0015] Further, step S5 includes: Although the PAF in step S4 effectively fuses multi-view representations, the fused different views (especially derived views, such as hypergraphs and semantic views) still suffer from redundancy and insufficient correlation in the feature space. Traditional contrastive learning alignment, if it simply forces all views to align with each other, will lead to the loss of their uniqueness and semantic differences, causing the so-called "semantic collapse" problem, where the representations learned by each view become highly similar and redundant, failing to capture the unique structural signals contained in the original view. To prevent this semantic collapse caused by view redundancy and to ensure that the fused view representation is meaningful and distinguishable, this invention introduces the Structure-Aware Contrastive Alignment (SCA) mechanism, which follows two core principles: 1. Anchoring principle: derived views must be aligned based on behavioral interaction views. This is because behavioral interaction views capture the most basic and direct user-item interaction signals and are the most basic and stable anchoring views. Through anchoring alignment, it is ensured that the representations of all derived views have consistency with the basic interaction views; 2. Exclusion principle: different derived views should not be forced to align with each other. This is done to avoid their representations becoming redundant, and to deliberately allow them to retain and capture their own unique, complementary structural signals.

[0016] To illustrate this more clearly, the method is explained from two perspectives: cross-behavior alignment and intra-behavior alignment. Cross-behavior alignment ensures that the embeddings of users and items across different behavior graphs remain consistent with the global graph embedding, thus achieving semantic consistency across multiple behaviors. The loss function for cross-behavior alignment is defined on both the user and item sides as follows:

[0017] in These represent the user and the item in the global embedding graph, respectively. For behavior Bottom view Embedding; For similarity functions, cosine similarity is usually used; It is a temperature coefficient used to adjust the difficulty of the model in distinguishing negative samples; Represents the complete set of users and items. This represents a negative sample. Similarly, intra-behavior alignment refers to the consistent alignment of interactive views and their derived views (hypergraph views, semantic views) within the same behavior, thereby enhancing the semantic coupling within the structure. The loss function for intra-behavior alignment is defined on the user side and the item side respectively:

[0018] The final loss for overall alignment is defined as:

[0019] Compared to traditional methods, this method effectively captures structural information from the original view, improves recommendation performance, and alleviates the semantic collapse problem caused by traditional methods.

[0020] Furthermore, step S6 includes: Traditional multi-task learning methods typically optimize all behaviors or all training samples using fixed or uniform weights, failing to effectively balance the different contributions of each behavior. This leads to models favoring high-frequency behaviors while sacrificing the performance of low-frequency but crucial behaviors, ultimately resulting in insufficient robustness and fairness in the recommendation results. To overcome the contribution imbalance problem caused by the aforementioned static bias, this invention proposes a personalized weighted collaborative training mechanism. This mechanism assigns adaptive weights to each training sample based on the similarity between different user behaviors, thereby balancing the contributions of each behavior and effectively utilizing signals with higher information content. The personalized weighted collaborative training mechanism is as follows: First, personalized adaptive weights are calculated for different users. These weights aim to measure the intensity of behavior conversion or behavior similarity, i.e., the probability that a user will convert from an auxiliary behavior to a target behavior. Specifically, users are defined... In behavior With target behavior The Jaccard similarity weights between them are:

[0021] in Indicates user In auxiliary behaviors The collection of items that can be interacted with below. Indicates user In target behavior The collection of items that can be interacted with below; It indicates the degree of overlap between two user behaviors and can approximately reflect the user's behavior conversion strength signal.

[0022] Based on this adaptive weighting, the personalized weighted Bayesian ranking loss function (BPR) is used for optimization, enabling the model to prioritize behavioral signals with higher information content. The specific recommendation loss formula is defined as follows:

[0023] in The training triplet set representing behavior k, They are positive samples and negative samples, respectively. This represents the predicted score for user u on item i under behavior k. This loss function reduces the dominance of high-frequency, low-information behaviors by reweighting each sample, further improving the robustness and accuracy of the recommendation results.

[0024] The final overall optimization objective function is defined as a joint optimization of the recommendation loss and the structure-aware alignment loss, to achieve dual constraints on recommendation performance and preservation of structural information. The formula for the total loss of the entire training framework is as follows:

[0025] in The balancing coefficient is used to adjust the weight allocation between recommendation accuracy and structural consistency during joint training. This method assigns differentiated weights to different users and behaviors, enabling the model to focus on learning highly relevant behavioral signals and reducing the dominance effect of high-frequency behaviors, thereby improving the robustness and fairness of the recommendation results.

[0026] A second aspect of the present invention relates to a multi-behavior recommendation method and apparatus based on structure-aware multi-view cascade fusion, comprising a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the multi-behavior intelligent recommendation method of the present invention, which exhibits collaborative robustness and accuracy.

[0027] A third aspect of the invention relates to a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the multi-behavioral intelligent recommendation method of the present invention, which exhibits synergistic robustness and accuracy.

[0028] The innovations of this invention are as follows: First, it proposes Prior-Anchored Adaptive Fusion (PAF), which combines a stable prior distribution with learnable adaptive residuals through a hierarchical cascading mechanism to achieve dynamic fusion of multi-view information and modeling of sequence dependencies between behaviors. Second, it introduces a Structure-Aware Contrastive Alignment (SCA) mechanism, which follows anchoring and exclusion principles to establish selective constraints at both cross-behavior and intra-behavior levels, preventing semantic collapse in multi-view feature learning. Third, it designs a personalized weighted collaborative training mechanism based on Jaccard similarity, which assigns adaptive weights to the behavioral conversion intensity of different users, achieving joint optimization of multiple tasks.

[0029] The advantages of this invention are: by employing a priori anchoring adaptive fusion strategy, a structure-aware contrast alignment mechanism, and personalized weighted collaborative training, the balance between robustness and expressiveness in multi-behavioral recommendation systems is optimized, effectively addressing the shortcomings of traditional static fusion methods in capturing complex behavioral dependencies; through selective contrast alignment constraints, the unique semantic differences of each view are preserved while maintaining structural consistency, avoiding semantic collapse caused by redundant information; simultaneously, the adaptive weight allocation strategy effectively overcomes the static bias problem in multi-behavioral data, balancing the contributions of high-frequency and low-frequency behaviors, thereby significantly improving the accuracy of recommendation results while ensuring recommendation robustness, demonstrating high technical innovation and application value. This invention can effectively solve the semantic collapse and static bias problems of multi-view data, improving the accuracy and robustness of recommendation systems. Attached Figure Description

[0030] Figure 1 This is a diagram of the model training framework of the present invention; Figure 2 This is a schematic diagram of the device of the present invention. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0032] Example 1

[0033] like Figure 1 As shown, a multi-behavior recommendation method based on structure-aware multi-view cascading fusion includes the following specific steps: Step S1: Load the dataset and build the global interaction graph In some embodiments, publicly available datasets can typically be obtained from websites such as Taobao Tianchi. These datasets generally contain information such as user IDs, item IDs, and behavior types. Then, using users and items as nodes and interaction records as edges, a global user-item bipartite graph is constructed to capture the underlying structure unrelated to behavior.

[0034] Step S2: Global embedding initialization, i.e., global graph learning

[0035] Based on the loaded dataset, initial embedding vectors are generated for each user and item, with a vector dimension of d. These embedding vectors represent the features of users and items and will be dynamically updated based on training. A lightweight graph convolutional network, LightGCN, is used to aggregate neighbor information on the global graph to obtain the initial global embeddings. ,in These represent the global representations of users and items, respectively.

[0036] Step S3: View Decomposition and Feature Construction

[0037] For each different behavior, this method decomposes its relationship signal into three complementary structural views, specifically: using LightGCN to model the behavioral interaction view between users and items to capture the most direct interaction relationship; using a parameterized hypergraph convolutional network HGCN to model the hypergraph view and mine the potential high-order group relationships between items; and constructing a semantic view based on point mutual information to mine the potential semantic associations between items from a global statistical perspective. After obtaining the above three views, they are used for subsequent view representation fusion.

[0038] Step S4: Multi-view fusion and cascading transfer of historical behavior information

[0039] Traditional multi-view or multi-behavior fusion methods generally suffer from poor model robustness and limited expressive power in recommendation performance, often failing to effectively model the inherent order dependencies or transformation relationships between different behaviors. Therefore, this method, to balance model robustness and expressiveness, designs a Prior-Anchored Adaptive Fusion (PAF) in multi-view information fusion to achieve multi-level information fusion within and between behaviors. Specifically, the hierarchical fusion mechanism of the PAF is described from two levels: within and between behaviors. For view fusion within behaviors, for each behavior... Define an unbiased prior distribution And introduce learnable residuals. To calculate the fusion weight of three different views under the same behavior, the specific formula for calculating the fusion weight within a behavior is as follows:

[0040] This weight is then used to further fuse the behavior interaction view, hypergraph view, and semantic view to obtain the embedding vector under the current behavior. The specific formula is expressed as follows:

[0041] in Indicates the current action Embedding vectors for different views. After obtaining information from different views within the current behavior, further information fusion between behaviors is performed.

[0042] A comprehensive representation of the current behavior and the output of the previous action Cascaded fusion is equivalent to combining historical context and behavior transformation information to obtain the final embedded representation of the current behavior after two levels of fusion:

[0043] in These represent the inter-behavior fusion weights for historical and current behaviors, respectively. Their calculation method is consistent with the intra-behavior fusion weights, thus capturing the dependencies between behaviors. It's important to note that these weights can also be learned using attention or gating mechanisms. However, through specific experimental testing and analysis, it was found that under sparse data conditions (a common problem in e-commerce datasets), attention and gating mechanisms are prone to learning noise, further demonstrating the effectiveness of this method.

[0044] S5: Structure-Aware Contrastive Learning

[0045] While the aforementioned methods can effectively fuse multi-view representations, the fused different views (especially derived views, such as hypergraphs and semantic views) still suffer from redundancy and insufficient correlation in the feature space. Using traditional contrastive learning alignment methods leads to the loss of their uniqueness and semantic differences, causing the so-called "semantic collapse" problem. This means that the representations learned by each view become highly similar and redundant, failing to capture the unique structural signals contained in the original views. To prevent this semantic collapse caused by view redundancy and ensure that the fused view representation is meaningful and discriminative, this method designs a Structure-Aware Contrastive Alignment (SCA) mechanism. This mechanism follows two core principles: 1. Anchoring principle: derived views must be aligned based on behavioral interaction views. This is because behavioral interaction views capture the most basic and direct user-item interaction signals, making them the most basic and stable anchoring views. Anchoring alignment ensures that the representations of all derived views are consistent with the basic interaction views; 2. Exclusion principle: different derived views should not be forced to align with each other. This is done to avoid their representations becoming redundant, and to deliberately allow them to retain and capture their own unique, complementary structural signals.

[0046] To illustrate this method more clearly, it is explained from two perspectives: cross-behavior alignment and intra-behavior alignment. Cross-behavior alignment ensures that the embeddings of users and items in different behavior graphs remain consistent with the global graph embedding, thus achieving semantic consistency across multiple behaviors. The loss function for cross-behavior alignment is defined on the user side and the item side respectively:

[0047] in These represent the user and the item in the global embedding graph, respectively. For behavior Bottom view Embedding; For similarity functions, cosine similarity is usually used; It is a temperature coefficient used to adjust the difficulty of the model in distinguishing negative samples; Represents the complete set of users and items. This represents a negative sample. Similarly, intra-behavior alignment refers to the consistent alignment of interactive views and their derived views (hypergraph views, semantic views) within the same behavior, thereby enhancing the semantic coupling within the structure. The loss function for intra-behavior alignment is defined on the user side and the item side respectively:

[0048] The final loss for overall alignment is defined as:

[0049] Compared to traditional methods, this method effectively ensures the structural and semantic consistency of multi-view embedding, mitigating the semantic collapse problem.

[0050] Step S6: Multi-task collaborative training optimization, combining the main task's Bayesian personalized loss with the auxiliary task's structure-aware contrastive learning for training optimization.

[0051] Traditional multi-task learning methods typically optimize all behaviors or all training samples using fixed or uniform weights, failing to effectively balance the different contributions of each behavior. This leads to models favoring high-frequency behaviors while sacrificing the performance of low-frequency but crucial behaviors, ultimately resulting in insufficient robustness and fairness in the recommendation results. To overcome the imbalance of contributions caused by the aforementioned static bias, this invention introduces Jaccard similarity weights to perform personalized user weighting, thereby balancing the contributions of each behavior. It is important to note that the personalized weighted Bayesian weighting mechanism is not an original algorithmic idea, but this invention integrates it into a multi-view cascaded hierarchical fusion network architecture and further integrates auxiliary loss from structure-aware contrastive learning, further optimizing recommendation performance; it should be viewed as a whole. The specific process is as follows: First, calculate the personalized adaptive weights for different users. These weights aim to measure the intensity of behavior conversion or behavior similarity, i.e., the probability of a user converting from an auxiliary behavior to a target behavior. Specifically, define users... In behavior With target behavior The Jaccard similarity weights between them are:

[0052] in Indicates user In auxiliary behavior The collection of items that can be interacted with below. Indicates user In target behavior The collection of items that can be interacted with below; This indicates the degree of overlap between two user behaviors, and can approximately reflect the user's behavior conversion strength signal. Based on this adaptive weight, this method uses a personalized weighted Bayesian ranking loss function (BPR) for optimization, enabling the model to prioritize behavioral signals with higher information content. The specific recommendation loss formula is defined as follows:

[0053] in The training triplet set representing behavior k. They are positive samples and negative samples, respectively. Let represent the predicted score of user u for item i under behavior k. This loss function reduces the dominant effect of high-frequency, low-information behaviors by reweighting each sample, thereby further improving the robustness and accuracy of the recommendation results. The final overall optimization objective function is defined as a joint optimization of the recommendation loss and the structure-aware alignment loss, achieving a dual constraint on recommendation performance and structural information preservation. The formula for the total loss of the entire training framework is as follows:

[0054] This method assigns differentiated weights to different users and behaviors, enabling the model to focus on learning highly relevant behavioral signals and reduce the dominance effect of high-frequency behaviors, thereby improving the robustness and fairness of the recommendation results.

[0055] Step S7: Based on the final trained user embeddings and item embeddings, calculate their inner product to obtain the user's rating of the items, and recommend the K highest-rated items to the user.

[0056] Finally, retain the trained embedding representation. and The interaction probability between items can be obtained by performing inner product calculation; the inner product result represents the user's interaction probability. With the project The relevance score indicates the user's level of interest in the project; a higher score signifies greater user interest. Specifically, for each user... By calculating the relevance score between the target item and all candidate items, a set of matching scores is obtained. The matching scores are then sorted from largest to smallest and the recommended results are output. Typically, the top K items with the highest scores are selected as the final recommended results.

[0057] Example 2

[0058] Reference Figure 2 This embodiment relates to a multi-behavior recommendation device based on structure-aware multi-view cascade fusion, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the intelligent recommendation method of collaborative diversity and accuracy in Embodiment 1.

[0059] Example 3

[0060] This embodiment relates to a computer-readable storage medium storing a program that, when executed by a processor, implements the multi-behavior intelligent recommendation method of Embodiment 1, which offers collaborative robustness and accuracy.

[0061] It should be noted that although the embodiments described above are illustrative, they are not intended to limit the invention. Therefore, the invention is not limited to the specific embodiments described above. Any other embodiments obtained by those skilled in the art under the guidance of this invention without departing from its principles are considered to be within the protection scope of this invention.

Claims

1. A multi-behavior recommendation method based on structure-aware multi-view cascade fusion, characterized in that, Includes the following steps: S1: Loading the dataset and building the global interaction graph; First, load the original dataset containing various user and item interaction behaviors; then, with users and items as nodes and interaction records as edges, build a global user-item bipartite graph to capture the basic relationship structure that is independent of behavior; S2: Global Embedding Initialization; LightGCN, a lightweight graph convolutional network, is used to aggregate neighbor information on the global graph to obtain a behavior-independent global embedding vector set. ,in These represent the global representations of users and items, respectively. S3: Behavior view decomposition; for each behavior The relationship signal is decomposed into three complementary structural views, specifically including: Behavior Interaction View, which models the direct interaction between users and items; and Hypergraph View, which uses a parameterized hypergraph convolutional structure to model high-order correlations at the group level. A semantic view of semantic dependencies between learning projects is constructed using point mutual information. S4: Multi-view fusion and behavior cascading; To balance the robustness and expressiveness of the model, a priori anchored adaptive fusion PAF is adopted in multi-view information fusion. The weights of each view are adaptively adjusted based on a stable prior to complete multi-level fusion. S5: Structure-Aware Alignment (SCA); After multi-view fusion, in order to maintain the semantic differences and structural consistency of each view, a structure-aware InfoNCE contrast alignment mechanism is introduced to prevent semantic collapse. S6: Multi-task collaborative training and optimization; Considering the static bias problem in multi-behavioral data, that is, different types of user behaviors vary greatly in quantity and signal strength, in order to balance the different contributions of each behavior, the weights are recalculated and assigned to each training sample based on the similarity between the user's auxiliary behavior and the target behavior, so as to capture the user's personalized signals and remove noise interference. S7: Utilize the trained and optimized user and item embeddings to calculate their inner product score, and select the top K items with the highest scores as the recommendation results output; Step S4 specifically includes: Prior anchoring-based adaptive fusion is divided into intra-behavior multi-view information fusion and inter-behavior contextual information fusion; for each behavior, intra-behavior view fusion is performed first, starting with each behavior... Define an unbiased prior distribution And introduce learnable residuals. The formula for calculating the fusion weight of three different views under the same behavior is as follows: This weight is then used to further fuse the behavior interaction view, hypergraph view, and semantic view to obtain the embedding vector under the current behavior. The specific formula is expressed as follows: in Indicates the current action Embedding vectors for different views; further, a comprehensive representation of the current behavior. and the output of the previous action Performing cascading fusion between behaviors is equivalent to combining historical context information to ultimately obtain a comprehensive representation of the current behavior after hierarchical fusion: in These represent the weights between historical and current behaviors, respectively. Their calculation method is consistent with the intra-behavior fusion weights, in order to capture the dependencies between behaviors. Step S5 specifically includes: SCA follows the anchoring principle and the exclusion principle, and aligns different views through contrast loss. Specifically, the anchoring principle means that derived views should be aligned with behavioral interaction views, that is, to maintain consistency with the base interaction view. The exclusion principle means that different derived views should not be forced to align, so as to preserve their unique semantic information.

2. The method according to claim 1, characterized in that, Cross-behavior alignment ensures that the embeddings of users and items across different behavior graphs remain consistent with the global graph embedding, thereby achieving semantic consistency across multiple behaviors. The loss function for cross-behavior alignment is defined on the user side and the item side, respectively: in These represent the user and the item in the global embedding graph, respectively. For behavior Bottom view Embedding; The similarity function is cosine similarity. It is a temperature coefficient used to adjust the difficulty of the model in distinguishing negative samples; Represents the complete set of users and items. Represents negative samples; In-behavior alignment refers to consistent alignment of interactive views and their derived views within the same behavior, thereby enhancing the semantic coupling within the structure. The loss function for in-behavior alignment is defined on the user side and the item side respectively: The final loss for overall alignment is defined as: 。 3. The method according to claim 1, characterized in that, Step S6 specifically includes: To measure the conversion relationship between different user assistive behaviors and target behaviors, a Jaccard similarity weight is introduced to approximate the intensity signal of user behavior conversion. The specific definition of its formula is as follows: in Indicates user In auxiliary behaviors The collection of items that can be interacted with below. Indicates user In target behavior A collection of items for interaction; This represents the degree of overlap between two user behaviors, approximately reflecting the user's behavior conversion strength signal. Based on this, a personalized weighted Bayesian ranking loss function is used for optimization. The specific loss function is as follows: in The training triplet set representing behavior k, They are positive samples and negative samples, respectively. Indicates user For items In behavior Based on the predicted scores, the final overall optimization objective is defined as: in The balancing coefficient is used to adjust the weight distribution of recommendation accuracy and structural consistency in joint training. By assigning differentiated weights to different users and behaviors, the model can focus on highly relevant behavioral signals, reduce the dominant effect of high-frequency behaviors, and thus improve the robustness and fairness of the recommendation results.

4. A multi-behavior recommendation device based on structure-aware multi-view cascade fusion, characterized in that, The method includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the multi-behavior recommendation method based on structure-aware multi-view cascade fusion as described in any one of claims 1-3.

5. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the multi-behavior recommendation method based on structure-aware multi-view cascading fusion as described in any one of claims 1-3.

Images