A multi-behavior recommendation method based on differential distillation condition diffusion

Abstract

The application relates to the technical field of deep learning, and proposes a multi-behavior recommendation method based on differential distillation conditional diffusion. The method first distinguishes between conversion behaviors and non-conversion behaviors and models the same, and on this basis, differential distillation conditional diffusion is used to guide auxiliary behavior embedding to be close to conversion behaviors, while keeping the distinction from non-conversion behaviors. In order to reduce the interference of redundant structures in non-conversion behaviors, a target-guided attention mask graph convolutional network is used for structural denoising, and a hierarchical semantic enhancement fusion module is designed in the fusion stage, and the weights of each behavior are determined by combining the structure score and the semantic similarity with the target behavior. The application realizes deeper and more refined modeling of user behaviors in the multi-behavior recommendation scene. The experimental results on the public benchmark dataset show that the recommendation performance of the method is significantly better than that of existing mainstream methods, and the application provides an effective technical solution to solve the challenges of data sparsity and behavior semantic difference.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and more specifically, to a multi-behavior recommendation method based on differential distillation conditional diffusion. Background Technology

[0002] Recommender systems are an indispensable technology for solving the problem of information overload. Over the past few decades, they have made significant progress, enhancing personalized user experiences across various fields such as streaming services, e-commerce, and social media. Collaborative filtering (CF) recommends items to users based on interactions between similar users. Most traditional CF-based methods utilize matrix factorization (MF) or graph-based models. However, these methods rely on single behaviors of items, leading to data sparsity and limited performance. In many real-world recommendation scenarios, there are multiple types of user-item interaction data, such as browsing, saving, purchasing, and adding to cart. These behaviors are typically categorized into auxiliary behaviors (browsing, saving, etc.) and target behaviors (purchasing). Methods that leverage auxiliary behaviors to facilitate target behavior recommendations are called multi-behavior recommendation (MBR). Many researchers have proposed innovative methods that utilize advanced deep learning techniques, including Graph Convolutional Networks (GCN) and Transformers, to effectively recommend items to users based on multi-behavior interactions. Since multi-behavioral interactions can be modeled as a unified interaction graph, GCN-based methods learn node embeddings from this unified graph. Some methods further employ multi-task learning (MTL) to predict assistive behaviors. Inspired by cascading sequences (e.g., browsing, adding to cart, and finally purchasing), some studies have used these sequences to model embeddings within the GCN framework. Recently, some research has indicated a multifaceted relationship between assistive behaviors and purchasing behaviors, such as browsing without purchasing or browsing and purchasing simultaneously.

[0003] However, the above methods all overlook two key issues: 1. Significant semantic differences exist between auxiliary behaviors and target behaviors. Existing methods typically integrate auxiliary behaviors simply as reinforcing signals into target behaviors or use multi-task learning to predict them separately; however, these methods fail to explicitly model the fine-grained impact of different modes of auxiliary behaviors on target behaviors. For example, there are fundamental differences in preferences between users who only browse and those who both browse and purchase. While Mule models the multifaceted relationships between auxiliary behaviors and target behaviors separately to capture richer semantic information, it lacks a mechanism to indirectly guide target behavior learning from auxiliary behaviors, resulting in limited semantic transfer effects. 2. Structural redundancy and weak semantic correlation exist in auxiliary behaviors, making them difficult to model effectively. Existing methods typically rely on global structures or simple attention mechanisms to fuse auxiliary behavior embeddings, failing to distinguish redundant and weakly correlated connections in the unconverted behavior graph, and failing to fully utilize the semantic similarity between different behaviors and target behaviors. As a result, irrelevant signals are over-propagated, masking potential conversion cues and affecting recommendation performance. Summary of the Invention

[0004] In view of this, the present invention proposes a multi-behavior recommendation method based on differential distillation conditional diffusion to solve the problems existing in the prior art.

[0005] To achieve the above objectives, this invention proposes a multi-behavior recommendation method based on differential distillation conditional diffusion, comprising the following steps: Construct a global interaction graph of users and items, and learn initial embedding representations of users and items based on a graph convolutional network; Based on the initial embedding representation, a cascaded single-behavior graph learning module is used to learn the corresponding behavior embedding representation for each single behavior. Based on the auxiliary behavior, the transformation behavior and the non-transformation behavior are defined, and on the basis of the behavior embedding representation, the transformation behavior embedding and the non-transformation behavior embedding are learned through the transformation behavior graph learning module and the target-guided attention mask graph convolutional network. Based on the aforementioned auxiliary behavior embedding, a differential distillation conditional diffusion module is used to perform forward diffusion and reverse conditional denoising with the unconverted behavior embedding as a condition, generating a denoised embedding, and updating the converted behavior embedding based on the denoised embedding. A hierarchical semantic enhancement fusion module is used to perform hierarchical fusion of the behavior embedding representation corresponding to the single behavior learning, the converted behavior embedding and the unconverted behavior embedding to obtain the user and item fusion embedding. Based on the user and item fusion embedding, the target behavior prediction score between the user and the item is calculated, and the model is trained end-to-end using a total loss function that includes Bayesian personalized ranking loss and diffusion loss.

[0006] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention uses the differential distillation conditional diffusion (DDCDiff) module to take unconverted behavior as a condition and guide auxiliary behavior to be embedded closer to the semantic space of conversion behavior. It can accurately extract potential conversion signals in auxiliary behavior and suppress noise, thereby directly reducing the semantic gap between auxiliary behavior and final purchase goal at the representation level, and significantly enhancing the model's ability to capture users' potential conversion preferences.

[0007] This invention innovatively proposes a target-guided attention-masked graph convolutional network, TamGCN, to address the weakly correlated interactions present in unconverted behaviors. This module actively identifies and weakens redundant and noisy connections in the graph through a target behavior-guided attention mechanism and a quantile pruning strategy, while retaining strongly correlated conversion cues. This results in learning a purer and more discriminative representation of unconverted behaviors, fundamentally alleviating the problem of excessive propagation of irrelevant signals.

[0008] This invention utilizes a hierarchical semantic enhancement fusion module (HSEF) to hierarchically fuse single behaviors with target-related behaviors. It innovatively combines two complementary scoring mechanisms: contextual structure relationships and semantic similarity. This ensures that the fusion process not only considers the structural dependencies between behaviors but also emphasizes semantic closeness to the target behavior. This allows for more accurate aggregation of semantic information from different sources, generating high-quality final representations and directly improving the accuracy of recommendation results. Attached Figure Description

[0009] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a schematic diagram of the overall process of a multi-behavior recommendation method based on differential distillation conditional diffusion proposed in this invention; Figure 2 This is a schematic diagram of the overall structure of the DDCDRec model in an embodiment of the present invention; Figure 3 This illustrates the impact of the number of LightGCN and TamGCN layers on different datasets in this embodiment of the invention. Figure 4 This illustrates the impact of the number of LightGCN and TamGCN layers on different datasets in this embodiment of the invention. Figure 5 This illustrates the impact of different diffusion loss coefficients on the model in this embodiment of the invention. Figure 6 This illustrates the effect of different pruning quantiles on the model in this embodiment of the invention. Figure 7This invention illustrates the impact of different data sparsity on the Taobao dataset in an embodiment of the invention. Detailed Implementation

[0010] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0011] This embodiment proposes a multi-behavior recommendation method based on differential distillation conditional diffusion, such as... Figure 1 As shown, it includes the following steps: Construct a global interaction graph of users and items, and learn initial embedding representations of users and items based on a graph convolutional network; Based on the initial embedding representation, a cascaded single-behavior graph learning module is used to learn the corresponding behavior embedding representation for each single behavior. Based on the auxiliary behavior, the conversion behavior and the non-converted behavior are defined, and the conversion behavior embedding and the non-converted behavior embedding are learned by the conversion behavior graph learning module and the goal-guided attention mask graph convolutional network, respectively, on the basis of the behavior embedding representation. Based on the auxiliary behavior embedding, a differential distillation conditional diffusion module is used to perform forward diffusion and reverse conditional denoising with the unconverted behavior embedding as a condition to generate the denoised embedding, and the converted behavior embedding is updated based on the denoised embedding. A hierarchical semantic enhancement fusion module is used to perform hierarchical fusion of the behavior embedding representation, the converted behavior embedding and the unconverted behavior embedding corresponding to a single behavior learning, to obtain the user and item fusion embedding. Based on user and item fusion embedding, the target behavior prediction score between users and items is calculated, and the model is trained end-to-end using a total loss function that includes Bayesian personalized ranking loss and diffusion loss.

[0012] Specifically, this embodiment first draws on the definition of target-related behaviors, subdividing auxiliary behaviors into conversion behaviors (such as browsing and purchasing) and non-conversion behaviors (such as browsing without purchasing), and models them separately. Based on this, a differential distillation conditional diffusion module is proposed. Using non-conversion behaviors as conditions, it guides the embeddings of auxiliary behaviors towards conversion behaviors, thereby bridging the semantic gap between auxiliary behaviors and target behaviors and improving the representation of potential user conversion preferences. It is worth noting that the input to the diffusion operation is the embedding of the auxiliary behaviors, with the aim of extracting potential conversion signals and suppressing noise. Since the goal is to better characterize the semantics of conversion behaviors, the embeddings obtained after diffusion are updated in the representation corresponding to the conversion behaviors. This design avoids directly altering the diversity of the auxiliary behaviors themselves while improving the representation quality of conversion behaviors, thus better serving the target task. Overall, differential distillation conditional diffusion does not simply align all auxiliary behaviors to the target behavior, but rather uses conversion behaviors as a bridge to guide the representation of the potential conversion portion of the auxiliary behaviors towards the conversion behaviors, while maintaining distinction from non-conversion behaviors. This is because in multi-behavior recommendation scenarios, the target behavior is the user's final conversion result, while auxiliary behaviors include both converted and non-converted components. These two components are semantically different. Directly aligning auxiliary behaviors with the target behavior may confuse the interference signals from non-converted behaviors, thereby reducing the model's performance.

[0013] Secondly, to address the structural redundancy and weak semantic relevance issues in auxiliary behaviors, this embodiment designs a Target-Guided Attention Masked Graph Convolutional Network (TamGCN) and a Hierarchical Semantic Enhancement Fusion (HSEF) module in the unconverted behavior modeling and fusion stages, respectively. TamGCN introduces the target behavior as a guiding signal, masking the interaction graph of the unconverted behavior to effectively weaken weakly related or irrelevant nodes and edges while retaining potential conversion cues, thereby mitigating the interference of redundant structures in the unconverted behavior graph on representation learning. In the fusion stage, the HSEF module assigns weights to different behavior embeddings based on two complementary scoring mechanisms: one is a contextual relationship score, which captures the structural relationships between different embeddings through an attention mechanism; the other is a similarity-guided score, which directly measures the semantic similarity between each behavior embedding and the target behavior embedding. The two scores are added together and normalized to obtain the final weight, ensuring that the fusion process considers not only structural relationships but also semantic proximity, effectively highlighting target-related behavioral signals. Simultaneously, HSEF employs a hierarchical fusion strategy, fusing single behaviors first and then target-related behaviors, fully capturing semantic complementarity at different levels and improving the final representation quality.

[0014] In this embodiment, the DDCDRec model proposed for multi-behavior recommendation includes the following three parts: (1) multi-behavior graph learning module; (2) differential distillation conditional diffusion (DDCDiff); and (3) hierarchical and semantic-enhanced fusion (HSEF). Figure 2 The overall architecture of DDCDRec is shown.

[0015] Problem Definition set up and Let them represent the user set and the item set, respectively. and These represent the total number of users and items, respectively. Let the set of behavior types be represented. Indicates the target behavior (e.g., purchasing behavior) User-item behavior The interaction diagram above is represented as ,in , Representation and Behavior The corresponding set of edges. Represents the user-item interaction matrix, when... hour, Otherwise, it is 0. set up and They represent The adjacency matrix and its symmetric normalized matrix are defined as follows: (1) in It is a diagonal matrix.

[0016] Given the effectiveness of LightGCN in graph-based collaborative filtering and its wide application in multi-behavior recommendation, this embodiment employs a variant of LightGCN as a graph encoder to learn user and item embeddings. The LightGCN layer is defined as follows: (2) in Initial embedding matrix, Indicates the first The embedding matrix obtained from the layer. In LightGCN, graph convolution operations on the user-item interaction graph are defined as follows: , It is the same as the symmetric normalized matrix defined in equation (1).

[0017] The goal of multi-behavior recommendation is to use user sets Item collection and multi-behavior interaction sets As input. The goal is to output the user's... With items A sorted score between , indicating user In items Execute target behavior The probability of.

[0018] Multi-behavior graph learning In multi-behavior recommendation, users typically interact with items through different types of behaviors (e.g., browsing, adding to cart, purchasing), each carrying different semantic and conversion implications. To effectively model these diverse and interconnected behaviors, this paper designs a hierarchical, progressive embedding learning module.

[0019] Initialization Embedding Learning Based on Global Interaction Graph Drawing on existing deep learning-based recommendation methods, this paper uses ID embedding to represent users. and items They are initialized to... and ,in Represents the embedding dimension. Let... This represents the initial learnable embedding matrix. As proposed in previous methods, embeddings learned from the global graph representing integrated interactions can serve as good initial embeddings when learning embeddings for specific behaviors. Building on this, this paper also employs the same strategy to obtain effective warm-start embeddings.

[0020] First, construct a global interaction graph to integrate all types of user-item interactions. Let... This represents the global interaction graph. So, ,in This refers to the integrated interaction of all behaviors. This embodiment uses LightGCN for learning. As shown below: (3) in This represents the user and item embeddings learned from the global interaction graph. express The adjacency matrix.

[0021] Cascaded Single Behavior Interaction Graph Learning After obtaining the representation of the global user-item interaction graph, fine-grained representations are further learned for each individual behavior. Inspired by existing methods, this embodiment further introduces dependencies between behaviors and global information during the learning process of each specific behavior embedding, thereby improving the cascaded method. To this end, this embodiment proposes a cascaded single-behavior graph learning module, which effectively models the sequential dependencies between behaviors by using the embeddings of each behavior before cascading and the global embedding as inputs to the current behavior.

[0022] For a single action sequence ， The first action directly uses the global embedding as the initial input: (4) For each subsequent action Its input embedding is computed as a weighted sum of the global embedding and all previous behavior embeddings: (5) Among them, weight Softmax normalization is used to adaptively adjust the contributions of each part. Then, a residual gating module is introduced to enhance global information and avoid overfitting to the previous behavior. (6) Finally, regarding a single behavior The embeddings are learned using LightGCN to represent their embeddings: (7) This module allows us to obtain a representation of each individual behavior, for example... , , (or These technologies, such as [examples of technologies], can effectively capture the sequential dependencies between global information and individual behaviors.

[0023] Learning from target-related behavioral interaction graphs To model the relationship between auxiliary behaviors and target behaviors in recommendation tasks, existing work has formally proposed the concept of target-related behaviors. Based on previous research, this embodiment defines two categories of target-related behaviors: transformed behaviors and untransformed behaviors. Assumptions... It is a collection of auxiliary behaviors. It is the target behavior. Therefore, This represents a set of conversion behaviors used to capture the auxiliary behaviors (such as browsing and purchasing items) that successfully convert to the target behavior. Similarly, This represents the set of unconverted behaviors, used to represent auxiliary behaviors that have not been converted into the target behavior (e.g., browsing items that were not purchased).

[0024] After defining conversion and non-conversion behaviors, their embeddings are further learned on the corresponding interaction graphs. Both types of behaviors use the embeddings of their auxiliary behaviors (e.g., browsing, adding to cart) as initial input, but DDCD-Rec employs different graph learning strategies for these two different behaviors. Let... This represents a conversion behavior interaction graph. For each... ， The following embeddings are learned on this graph: (8) in It is a picture The adjacency matrix, It is an auxiliary behavior Embedded.

[0025] For unconverted behaviors, this embodiment proposes a target-guided attention mask graph convolutional network (Target Attention Network) based on graph attention networks. Guided Attention Mask GCN (TamGCN) is aimed at Perform graph learning. Specifically, in graph... In this work, we aim to enhance the influence of interactions that, while remaining relevant to the target behavior despite not being transformed, while suppressing noisy or irrelevant interactions. Existing work mitigates the noise problem in untransformed behaviors by introducing a soft target-guided attention mechanism that reweights all edges in the graph according to their relevance to the target behavior. However, this purely soft attention approach still retains a large number of weakly relevant or irrelevant interactions because softmax assigns non-zero weights to all neighbors. This residual noise may mask the truly meaningful signals, limiting the discriminative power of the learned embeddings. To overcome this limitation, this embodiment introduces a masking strategy on top of the soft attention mechanism, proposing a Target-Guided Attention Masked Graph Convolutional Network (TamGCN). This module first computes the target-guided attention edge weights as described in [previous work], and then prunes edges with weights below a certain quantile threshold. This not only suppresses irrelevant interactions more effectively but also allows the model to focus on the most significant and target-relevant signals.

[0026] Given a graph and target behavior embedding TamGCN Layer learning to obtain embeddings : (9) in It is the initial embedding matrix, at the th Layer-updated embedding matrix The definition is as follows: (10) in It is the first The adjacency attention weight matrix of the layer. For each interaction ,from arrive Attention weights are calculated as follows: (11) in It is a node In the figure The set of neighboring nodes in, and This represents the relevance score, here This represents the inner product operation. Unlike existing soft attention methods that preserve all edges, TamGCN introduces a masking mechanism to further suppress irrelevant interactions. Specifically, when computing all... Normalized attention weights Then, in this embodiment, the quantile threshold is first calculated based on the distribution of all edge weights. Then, the weight of each edge is checked; if it is greater than... Either retain the weight of the edge, or reduce its weight to 0. This process means: (12) This method effectively preserves edges with larger weights and strong relevance to the target, while removing edges that are embedded with the target. The low-correlation noise edges result in a sparser and cleaner graph structure.

[0027] from arrive attention weights Calculated in a similar way, using and .for ， set up To ensure that the attention matrix is ​​preserved The sparsity and structure of the data. This embodiment employs embeddings learned from auxiliary behavior graphs. initialization This provides a good warm start on rich interactions with coarser granularity.

[0028] Through this module, we ultimately arrive at the embedding of two behaviors: conversion behavior embedding. Embedded unconverted behaviors .

[0029] Differential distillation conditional diffusion In multi-behavior recommendation tasks, there is a potential conversion relationship between auxiliary behaviors and target behaviors. Previously, TamGCN was introduced to denoise the unconverted portion (unconverted behavior) of auxiliary behaviors, effectively removing most noise irrelevant to the target behavior at the graph structure level. However, a representational gap still exists between auxiliary and target behaviors. To further narrow this gap, leveraging the excellent denoising and generation properties of diffusion models, this embodiment proposes a Differential Distillation Conditional Diffusion (DDCDiff) module. This module no longer directly optimizes the representation of auxiliary behaviors. Instead, it uses their representation as input, combined with the embedding of unconverted behaviors as diffusion conditions, to generate a representation with a stronger conversion intent. This representation is then used to optimize the representation of the target behavior, indirectly guiding the auxiliary behavior embedding to approximate the conversion behavior embedding. Specifically, the representation generated through the conditional diffusion process serves as a pseudo-label for the conversion behavior, distilling and optimizing the target behavior embedding while maintaining its distinguishability from the unconverted behavior embedding. In this way, DDCDiff not only fully utilizes the implicit conversion preferences in auxiliary behaviors but also further alleviates the semantic gap between auxiliary and target behaviors, enhancing the ability to express users' potential conversion intent.

[0030] Forward diffusion process Embedded data based on assistive behavior (For example , This paper designs a forward process for differential distillation conditional diffusion (DDCDiff). In this process, Gaussian noise is progressively added to the original data. Middle. (Number) The data of the step is recorded as Step 0 is the original data, i.e. . No. The data for step one is from the first step. The data from the step is calculated and defined as follows: (13) in Control the level of noise. This represents the Gaussian distribution used to generate the noise. With... As the noise level increases, the data... It eventually converges to pure Gaussian noise. Thanks to the additivity of the Gaussian distribution, the first... The data from each step is directly written as about the original data. and a series of pre-calculated The expression is obtained by pre-compiling the following quantities to avoid recursive computation: (14) To determine each step This embodiment introduces two hyperparameters. and Generate sequences through linear interpolation Using these pre-calculated values, It can be efficiently obtained through the following calculations: (15) in This represents independent Gaussian noise. Utilizing these additive properties, efficient inference can be performed without recursive computation. .

[0031] Reverse conditional diffusion process In the forward pass of DDCDiff, auxiliary behavior representation Gaussian noise is gradually added on, eventually resulting in pure noise. To recover a meaningful representation from a noisy input, this paper designs a backdiffusion process, in which the untransformed behavior... Under the guidance of the conditions, Gradual denoising to clean auxiliary behavior representation .

[0032] In each reverse step A learnable neural network under certain conditions and time step Predict Noise components in. Reverse from arrive The calculation method is as follows: (16) in and They represent the first time. The mean and variance of the reverse process predicted in the first step are defined as: (17) It consists of two consecutive fully connected layers. The time steps... Embedded Unconverted behavior embedding and time step embedding The concatenated data is then used as input to the neural network. Conditional guidance is provided to ensure that the denoised auxiliary behavior representation remains distinguishable from the unconverted behavior during the diffusion process.

[0033] This reverse conditional diffusion aims to suppress noise while maintaining harmony with... The semantic distinctiveness allows for subsequent use of independent loss terms to represent auxiliary behaviors and transform behaviors. Align and get ready.

[0034] Optimization target and training loss To optimize the proposed differential distillation conditional diffusion (DDCDiff) module, this paper employs transformation behavior embedding. As an evidence lower bound (ELBO), this paper performs diffusion and denoising on the embeddings of users and items, respectively, based on their respective embedding data. Specifically, for the two types of embeddings, user and item embeddings are used with their respective conditional diffusion models, and the evidence lower bound is determined by the embedding of the corresponding transformation behavior. and Give it. Remember. The paradigm for maximizing ELBO is:

[0035] Here, the first term represents the reconstruction term, and the second term represents the denoising term. The goal of the denoising term is to minimize the true distribution. and the learned inverse transition distribution The KL divergence between them. Based on the conclusions of existing work, the standard deviation can be simplified to... (19) The corresponding noise reduction term It can be written as: (20) in, Indicates by , and Predicted by a multilayer perceptron (MLP) As shown in Equation (17). Based on the derivation of Equation (17), the reconstruction term can also be approximated as the prediction embedding and the transformation behavior embedding. Squared loss between: (twenty one) It is important to note the semantic information of auxiliary behaviors. As input to the DDCDiff module, in the untransformed behavior As a conditional guideline, the auxiliary behavior embedding is noise-removed and optimized through distillation to align the transformation behavior embedding. Furthermore, the denoised embeddings are directly used to update the representation of conversion behavior, replacing the previous embeddings with cleaner ones that better reflect the semantics of conversion. Finally, the denoised auxiliary behavior embedding is replaced by the conversion behavior embedding, expressed by the formula: (twenty two) in This represents the output of the conditional diffusion process. This design explicitly distinguishes between auxiliary and unconverted behaviors, while improving the representation quality of the converting behavior by utilizing useful portions of the auxiliary behaviors. To determine the diffusion steps... This paper employs a uniform sampling strategy. Diffusion loss Defined as: (twenty three) Hierarchical semantic enhancement fusion After the preceding modules, we have obtained embeddings of various behaviors. The goal of this section is to effectively fuse these embeddings to capture the hierarchical relationships and semantic complementarity between different behaviors. To this end, this embodiment designs a Hierarchical and Semantic-Enhanced Fusion (HSEF) module to effectively aggregate embeddings of multiple semantics.

[0036] This module assigns weights to each behavior embedding based on two complementary scoring mechanisms: a context-driven score, which captures structural relationships through an attention mechanism; and a similarity-driven score, which directly calculates the semantic similarity between each behavior embedding and the target embedding. The two scores are summed and normalized using a softmax function to obtain the final behavior weight, thus integrating structural and semantic information. Simultaneously, HSEF hierarchically fuses single behaviors and target-related behaviors, effectively enhancing the semantic information of each behavior embedding to improve recommendation performance.

[0037] First, HSEF embeds the target Embedded with each individual behavior The data is concatenated and a behavioral context score is calculated using a linear transformation: (twenty four) To incorporate semantic information, further calculations are performed. and Cosine similarity between them: (25) Then, the semantic score is passed through a learnable parameter. The attention score is incorporated, and the final weights are obtained through softmax normalization. (26) Reuse Weighted aggregation of the embeddings of each behavior yields the intermediate embeddings. : (27) In the second stage, this embodiment will use a similar approach. Embedding of target-related conversion behaviors and non-conversion behaviors Fusion: (28) HSEF outputs the final merged embedding of all users and items. and Compared to methods that rely solely on simple averaging or single-structure attention scores, the method in this embodiment achieves better performance experimentally by jointly modeling the contextual relationships and semantic relevance between behaviors.

[0038] Using final embedding This embodiment calculates the user and items Scores between As shown below: (29) This embodiment employs Bayesian Personalized Ranking (BPR) loss, which assumes that interacted items should receive a higher score than uninteracted items. Given a set of target behavior interactions... BPR loss function Defined as: (30) in It is the sigmoid function. It is a set of triples ,in Indicates a positive sample. This represents a negative sample randomly selected in each epoch. The final total loss function is: (31) Weights that control diffusion loss It is a regularization hyperparameter. This represents the set of all model parameters.

[0039] Dataset This experiment was conducted on two publicly available real-world benchmark datasets: Taobao and Tmall. Detailed statistics for each dataset are shown in Table 1.

[0040] Taobao: The dataset contains 15,449 users and 11,953 products, with three user behavior types: browsing, adding to cart, and purchasing.

[0041] Tmall: The dataset contains 41,738 users and 11,953 products. Unlike Taobao, this dataset includes four user behavior types: browsing, collecting, adding to cart, and purchasing.

[0042] Table 1 ; Evaluation indicators To evaluate the model's performance, this experiment uses hit ratio (HR@K) and normalized discounted cumulative gain (NDCG@K) as evaluation metrics.

[0043] Experimental details DDCDRec is implemented using PyTorch. To ensure fair comparison, following existing methods, the embedding dimension is set to 64, the batch size to 1024, and the number of training epochs to 200. This paper uses the Adam optimizer, with the learning rate and regularization coefficients tuned in {0.001, 0.0001, 0.00001} and {0, 0.00001}, respectively. The number of LightGCN and TGAP layers is tuned in {1, 2, 3} and {1, 2, 3, 4, 5}, respectively. Furthermore, the two hyperparameters of this model—the diffusion loss coefficient—are optimized. Quantile threshold in {0.05, 0.1, 0.3, 0.5, 0.8} The parameters were optimized within the range of {0.5, 0.6, 0.7, 0.8, 0.9}. For other methods, this embodiment sets the hyperparameters according to the range described in the corresponding paper and employs an early stopping strategy during the training phase.

[0044] Baseline model To verify the effectiveness of DDCDRec, this experiment compares it with many baseline models in recent years and divides them into two categories: (1) Single-behavior methods: MF-BPR, NeuMF, LightGCN; (2) Multi-behavior methods: RGCN, GNMR, NMTR, MBGCN, CRGCN, MB-CGCN, MB-HGCN, Mule, COPF.

[0045] Overall experimental results This experiment presents a detailed comparative analysis of the performance of the DDCD-Rec model and all baseline methods. Experimental results for the Taobao and Tmall datasets are shown in Tables 2 and 3, respectively. The best results are indicated in bold, and the second best results are indicated by underlining. Impr.% represents the percentage improvement of the best performance relative to the second best.

[0046] Table 2 ; Table 3 ; Overall, multi-behavior methods outperform single-behavior methods, highlighting the effectiveness of modeling multiple behaviors. Among multi-behavior methods, DDDCDRec significantly outperforms other methods. Compared to the best baseline model, this model achieves significant improvements in recommendation performance. Specifically, on the Taobao dataset, HR@10 and NDCG@10 performance are improved by 7.12% and 8.37%, respectively; on the Tmall dataset, HR@10 and NDCG@10 performance are improved by 3.88% and 7.01%, respectively.

[0047] For single-behavior recommendation models, the LightGCN method outperforms MF-BPR and NeuMF on all metrics across both datasets. This is because LightGCN further improves performance by leveraging higher-order information in the user-item interaction graph, highlighting the advantages of graph convolutional network-based models.

[0048] For multi-behavior recommendation models, while RGCN models different types of behaviors, it fails to effectively distinguish the importance of these behaviors, resulting in relatively poor performance. GNMR and MBGCN, on the other hand, more effectively consider the contributions of different behaviors, outperforming RGCN, indicating that different behaviors contribute differently to the target behavior. NMTR, CRGCN, and MB-CGCN all consider the cascading relationships between multiple behaviors, but they all lag behind COPF because COPF considers the cascading relationships of multiple behaviors while also taking into account the constraints of user behavior patterns from a combinatorial optimization perspective. COPF treats multi-behavior fusion as a combinatorial optimization problem, imposing different constraints at different stages of each behavior to limit the solution space, thus significantly improving fusion efficiency. This sufficiently demonstrates that adding good constraints can effectively capture better behavioral information when learning the embedding representation of a single behavior. Although cascading methods achieve good performance, they are slightly inferior to parallel methods like MB-HGCN and MuLe due to noise accumulation issues. As a parallel method, MuLe outperforms other methods such as MB-HGCN, which is attributed to MuLe directly improving the uncertainty of behaviors unrelated to the target behavior through a graph attention network guided by the target behavior. While MB-HGCN also uses uniform or behavior-specific interactions, they neglect graph learning of goal-related behaviors, resulting in poor performance. This further demonstrates the importance of learning goal-related behaviors.

[0049] ablation experiment This embodiment carefully analyzes the effectiveness of each module proposed in this method through ablation experiments.

[0050] Effectiveness analysis of differential distillation conditional diffusion module The proposed DDCDiff method introduces unconverted behavior as a condition. Through a diffusion process, it guides the representation of the auxiliary behavior to move closer to the representation of the converted behavior and updates the converted behavior representation, while maintaining its distinction from the unconverted behavior. To verify the effectiveness of the DDCDiff module, this experiment designed three comparative variants: w / o.Diff, which removes the diffusion process and updates the converted behavior embedding solely through simple MLP denoising of the auxiliary behavior embedding; and w / o.Replace, which retains the diffusion process but directly replaces the auxiliary behavior embedding with the diffused embedding instead of using it to update the converted behavior embedding. The code snippet `w / o.ALL` removes the entire DDCDiff module, performs no operations on auxiliary behaviors, and uses the original embedding. The experimental results are shown in Table 4.

[0051] Table 4 ; Observations show that the complete DDDCDRec achieves the best HR@10 and NDCG@10 on Taobao and Tmall, reaching 0.2091 / 0.1200 and 0.2207 / 0.1256 respectively, significantly outperforming all variants. Compared to w / o.ALL, performance drops significantly without any diffusion or conditional guidance, indicating that the diffusion module as a whole makes a significant contribution to the model's performance. Compared to w / o.Diff, the complete DDDCDRec effectively performs progressive noise reduction on auxiliary behaviors through the diffusion process, guiding the embeddings closer to the target behavior, thereby improving model performance. Finally, compared to w / o.Replace, its performance is lower than DDDCDRec, indicating that directly covering auxiliary behaviors may confuse untransformed signals, while updating the design of the transformation behavior can better maintain semantic discriminability and improve recommendation performance.

[0052] Effectiveness Analysis of Goal-Driven Attention Mask Networks To verify the effectiveness of the proposed target-guided attention mask graph convolutional network TamGCN, this experiment replaced TGAP with TDA proposed by LightGCN and Mule to learn each untransformed behavior. The results are shown in Table 5.

[0053] Table 5 ; Table 5 shows that on both the Taobao and Tmall datasets, TamGCN significantly outperforms LightGCN and TDA, demonstrating that TamGCN can more effectively remove irrelevant or redundant connections in the non-converted behavior graph, retaining more useful conversion cues and thus improving the recommendation performance of the target behavior. The fact that LightGCN performs the worst indicates that the non-converted behavior graph contains a large amount of noise, and treating each node in the interaction graph equally will reduce the recommendation performance.

[0054] Effectiveness analysis of hierarchical semantic enhancement fusion module This embodiment considers three variants of HSEF: w / o.SA, removing the structured score; w / o.SE, removing the similarity semantic score; and w / o.SS, removing both, with the embeddings of each behavior directly added together. The experimental results are shown in Table 6.

[0055] Table 6 ; Specifically, removing the structured score (w / o.SA) significantly reduced both HR@10 and NDCG@10 (e.g., Taobao's score dropped to 0.1352 / 0.0867), indicating that the structured score effectively depicted the hierarchical relationships between different behaviors, contributing to improved recommendation quality. Removing the semantic similarity score (w / o.SE) also resulted in a performance decrease, but to a relatively minor degree (e.g., Taobao's score dropped to 0.2045 / 0.1185), suggesting that the semantic similarity score could further supplement the structured score and positively impact the fusion effect. The performance drop was greatest when simply adding the embeddings of each behavior (w / o.SS), indicating that simple linear fusion is insufficient to fully explore the relationships between the embeddings of each behavior, validating the irreplaceable role of the HSEF module in behavior fusion.

[0056] GCN Layer Analysis DDCDRec uses LightGCN for graph learning for single and transformation behaviors, while employing TamGCN for convolutional operations for non-transformation behaviors. This experiment compares the performance impact of different numbers of LightGCN and TamGCN layers on different datasets, and the results are as follows: Figure 3 , Figure 4 As shown.

[0057] Through the Figure 3 , Figure 4Observations show that when the number of LightGCN layers is 1, the model achieves optimal results on both HR@10 and NDCG@10 datasets; however, as the number of layers increases, the model performance gradually decreases. This indicates that additional LightGCN layers do not benefit the model, and 1 layer is sufficient to obtain a warm-start embedding representation. Conversely, as the number of TamGCN layers increases, the model performance gradually improves, peaking at 4 layers. This demonstrates that TamGCN can effectively aggregate rich high-order information, thus significantly improving recommendation performance.

[0058] DDCDRec has two hyperparameters, one of which is the coefficient of diffusion loss. Quantile thresholds of TGAP This experiment statistically analyzed the performance of DDCD-Rec on two datasets with different hyperparameter values. Specific results are as follows: Figure 5 and Figure 6 As shown.

[0059] from Figure 4 It can be observed that on the Taobao dataset, HR@10 and NDCG@10 are... The value reaches its highest point when it equals 0.5. Overall, The value of has a certain impact on performance. For the Tmall dataset, HR@10 and NDCG@10 have different performance characteristics. The coefficient peaks around 0.1, and both metrics decrease when the coefficient is increased further to 0.8. Overall, they show a trend of first rising and then falling, indicating that appropriately introducing diffusion loss constraints can help enhance representation learning and improve recommendation performance. However, when the diffusion loss accounts for too high a proportion of the total objective, the optimization of the main task is weakened, resulting in a performance decline.

[0060] Figure 5 The experimental results show hyperparameters The impact on model performance. With Increasing from 0.5 to 0.9, HR@10 and NDCG@10 on the Taobao dataset... The highest values ​​were reached when the value was 0.7, specifically 0.2091 and 0.1200. The Tmall dataset, on the other hand, showed... It performs best when the value is 0.5. The overall trend is relatively stable, but... Performance drops significantly when the pruning ratio is 0.9. This indicates that moderate pruning can remove noisy edges and improve structural quality, thereby enhancing recommendation performance. However, excessive pruning can damage the graph structure, lose useful information, and lead to performance degradation. Therefore, an appropriate pruning ratio is crucial for model performance.

[0061] Data sparsity analysis To evaluate the performance of this method under different user interaction volumes, this experiment divided users in the test set into four groups based on the number of interactions in the Taobao dataset. Specifically, users were divided into Q1 (the top 25% of users with the highest interaction volume), Q2 (25–50%), Q3 (50–75%), and Q4 (the bottom 25% of users with the lowest interaction volume) according to the number of interactions. Figure 7 The data shows the model's recommendation performance across different groups.

[0062] As the number of user interactions increases, the performance of both our proposed method and Mule shows a continuous improvement across all groups. Users in Q1, with the highest number of interactions, achieved the best results on both HR@10 and NDCG@10 metrics, while users in Q4, with the fewest interactions, performed the worst. Notably, the relative performance improvement of our proposed model compared to Mule also increases with interaction density, indicating that our method can more effectively utilize rich user interaction data to generate more accurate representations. These results demonstrate that introducing a conditional diffusion alignment mechanism with differential distillation can effectively improve recommendation performance, especially when dealing with highly active users.

[0063] In summary, this embodiment proposes a novel multi-behavior recommendation framework, DDDCDRec, which aims to effectively bridge the semantic gap between auxiliary behaviors and target behaviors and improve the modeling ability of user conversion preferences. Addressing the significant semantic differences and structural noise issues in auxiliary behaviors, this embodiment designs three core modules: (1) a multi-behavior graph learning module, which distinguishes and learns differentiated representations between conversion behaviors and non-converted behaviors; simultaneously, for non-converted behaviors, this paper also proposes a target-guided attention-pruned graph convolutional network (TamGCN) to prune irrelevant or weakly related nodes and edges on the non-converted behavior graph, reducing the interference of redundant structures on representation learning. (2) a differential distillation conditional diffusion (DDCDiff) module, which utilizes the good noise reduction capability of the diffusion model and introduces non-converted behaviors as conditions to guide the embedding of auxiliary behaviors towards conversion behaviors while maintaining the distinction from non-converted behaviors; (3) a hierarchical semantic enhancement fusion (HSEF) module, which combines structural relationships and semantic similarity to aggregate multiple behavior embeddings in a hierarchical manner, effectively improving representation quality.

[0064] Extensive experimental results on two real-world datasets demonstrate that DDDCDRec can fully utilize auxiliary behavioral information to effectively improve recommendation performance and outperform existing multi-behavior recommendation methods.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A multi-behavior recommendation method based on differential distillation conditional diffusion, characterized in that, Includes the following steps: Construct a global interaction graph of users and items, and learn initial embedding representations of users and items based on a graph convolutional network; Based on the initial embedding representation, a cascaded single-behavior graph learning module is used to learn the corresponding behavior embedding representation for each single behavior. Based on the auxiliary behavior, the transformation behavior and the non-transformation behavior are defined, and on the basis of the behavior embedding representation, the transformation behavior embedding and the non-transformation behavior embedding are learned through the transformation behavior graph learning module and the target-guided attention mask graph convolutional network. Based on the auxiliary behavior embedding, a differential distillation conditional diffusion module is used to perform forward diffusion and reverse conditional denoising with the unconverted behavior embedding as a condition to generate a denoised embedding, and the converted behavior embedding is updated based on the denoised embedding. A hierarchical semantic enhancement fusion module is used to perform hierarchical fusion of the behavior embedding representation corresponding to the single behavior learning, the converted behavior embedding and the unconverted behavior embedding to obtain the user and item fusion embedding. Based on the user and item fusion embedding, the target behavior prediction score between the user and the item is calculated, and the model is trained end-to-end using a total loss function that includes Bayesian personalized ranking loss and diffusion loss.

2. The method according to claim 1, characterized in that, In the cascaded single-behavior graph learning module: The input for the first single action is the global embedding; The input for subsequent single actions is a weighted sum of the global embedding and all previous action embeddings, and the global information is enhanced and overfitting is suppressed by the residual gating module. The embedding of each individual behavior is learned on the graph structure using LightGCN.

3. The method according to claim 1, characterized in that, The target-guided attention mask map convolutional network is used to learn the embeddings of untransformed behaviors, specifically including: Attention weights for each edge in the computation graph are embedded based on the target behavior. Based on the distribution of all edge weights, a quantile threshold is set, and edges with weights below the quantile threshold are pruned and their weights reduced to zero. Graph convolution is performed based on the pruned attention weight matrix to iteratively update node embeddings.

4. The method according to claim 1, characterized in that, The differential distillation conditional diffusion module includes a forward diffusion process and a reverse conditional diffusion process. The forward diffusion process includes: gradually adding Gaussian noise into the auxiliary behavior embedding until it becomes pure noise; The reverse conditional diffusion process includes: using the unconverted behavior embedding and the time step embedding as conditions, predicting noise through a neural network and gradually denoising to restore the auxiliary behavior embedding; The original transformation behavior is embedded instead of the denoised auxiliary behavior, thereby achieving semantic alignment and representation enhancement.

5. The method according to claim 4, characterized in that, The optimization objective of the differential distillation conditional diffusion module is constructed based on an evidence lower bound, and includes a reconstruction term and a denoising term, wherein: The reconstruction term is approximately the squared loss between the prediction embedding and the transformation behavior embedding; The denoising term is the mean square error between the actual noise and the predicted noise.

6. The method according to claim 1, characterized in that, The hierarchical semantic enhancement fusion module employs a two-stage fusion process: In the first stage, the target behavior embedding and each individual behavior embedding are merged. The weights are calculated by combining the context relationship score and the similarity guidance score, and then weighted aggregation is performed to obtain the intermediate embedding. In the second stage, the intermediate embedding, conversion behavior embedding, and non-conversion behavior embedding are merged to obtain the user and item fused embedding.

7. The method according to claim 6, characterized in that, The context relationship score is calculated using linear transformation and attention mechanism; the similarity guidance score is calculated using the cosine similarity between the target behavior embedding and each behavior embedding.

8. The method according to claim 1, characterized in that, The total loss function is expressed as: ， in, For BPR loss, For the spread of loss, Weights that control diffusion loss It is a regularization hyperparameter. This represents the set of all model parameters.

Images