Common preference enhanced collaborative recommendation method and device

Abstract

The invention discloses a common preference enhanced collaborative recommendation method and device, and the method comprises the steps: obtaining user implicit interaction data in a target domain and a source domain, and constructing corresponding user features and commodity feature representations; carrying out fusion coding on the target domain user features, the source domain user features and the target domain commodity features, and generating cross-domain migration features facing target domain recommendation tasks; performing fusion iteration de-noising processing on the cross-domain migration features, gradually inhibiting source domain noise interference and strengthening migratable collaborative preference information through multiple rounds of feature fusion and weight screening operation, and generating de-noised cross-domain user feature representation; after denoising is completed, user structure constraint is introduced, and consistency constraint is carried out on the similarity relation between the target domain users before and after feature enhancement, so that the overall distribution stability of the target domain user interest structure is kept; and finally, performing interactive modeling on the final user representation and the target domain commodity features to obtain a matching relationship between the user and the commodity. The device comprises a processor and a memory. Through a collaborative modeling mode of fusing iterative denoising and structural constraint, the problems of noise interference in cross-domain recommendation and user similarity structure change before and after feature migration can be relieved, and the stability and accuracy of a recommendation result are improved.

Description

Technical Field

[0001] This invention relates to the field of collaborative recommendation, and more particularly to a collaborative recommendation method and apparatus that enhances common preferences. Background Technology

[0002] Recommender systems effectively uncover users' potential interests and preferences, and are currently widely used in e-commerce platforms, content recommendation platforms, and online service systems. However, with the continuous expansion of platform scale and the explosive growth in the number of users and products, the interaction data between users and products is becoming increasingly sparse. This makes it difficult for traditional recommendation methods based on collaborative filtering and matrix factorization to accurately model user preferences, thus severely restricting further improvements in recommendation performance. To alleviate the problems of data sparsity and cold start, cross-domain recommendation technology has gradually gained attention. Its core objective is to transfer user preference knowledge contained in the auxiliary domain to the target domain to supplement the missing supervision signals in the target domain, thereby constructing a more complete and robust representation of user preferences and improving the recommendation effect in the target domain.

[0003] Existing cross-domain recommendation methods can be broadly categorized into two types: hybrid cross-domain recommendation methods and disentangled cross-domain recommendation methods. Hybrid cross-domain recommendation methods typically fuse user features from the source and target domains directly through feature concatenation, weighted fusion, or shared network structures to achieve cross-domain information transfer. However, in practical applications, not all user preferences learned in the source domain are applicable to the target domain. Disentangled cross-domain recommendation methods achieve more refined cross-domain modeling by decomposing user preferences into domain-invariant features and domain-specific features. Domain-invariant features are considered to be shareable across different domains, while domain-specific features only function within their respective domains. Existing research typically uses orthogonal constraints, probability distribution constraints, or information theory constraints to achieve feature disentanglement, which has improved cross-domain recommendation performance to some extent. However, existing disentangled cross-domain recommendation methods still have significant shortcomings in practical applications. First, most existing disentanglement mechanisms only operate at the user representation level, neglecting the crucial constraint role of target domain product semantics in cross-domain transfer. Existing methods typically decompose user preferences into domain-invariant and domain-specific features using loss functions such as orthogonality constraints, KL divergence, or contrastive consistency. However, this process mainly occurs in the user feature space, lacking direct interaction and alignment with the semantics of products in the target domain. Consequently, some cross-domain features considered "domain-invariant" are not truly applicable to the recommendation needs of specific products in the target domain, making it difficult for cross-domain transfer features to be effectively filtered for the target domain task, thus introducing irrelevant or even harmful transfer noise. Secondly, existing cross-domain transfer processes generally lack explicit constraints on the similarity structure of users in the target domain. During cross-domain knowledge transfer, user representations often undergo an overall shift. Users with similar interests in the target domain may be distanced, while users with significantly different interests may be incorrectly brought closer, thus disrupting the overall structural relationship of user interest distribution in the target domain. This leads to drift in user feature representation, affecting the recommendation performance and model stability based on collaborative relationships. Summary of the Invention

[0004] This invention provides a collaborative recommendation method and apparatus with enhanced common preferences. By introducing a fusion-first cross-domain modeling strategy and an iterative denoising and structure preservation constraint mechanism, this invention can gradually suppress source domain noise interference and maintain the stability of the target domain user preference structure during cross-domain knowledge transfer, thereby alleviating the negative transfer problem and improving the robustness and accuracy of cross-domain recommendation. See the description below for details: A collaborative recommendation method with enhanced common preferences, the method comprising: Obtain implicit user interaction data from the target domain and the source domain, construct the corresponding user-product interaction structures respectively, and initialize user embedding features and product embedding features; Based on the user-product interaction structure, information propagation and feature aggregation are performed on user features and product features in the target domain and source domain to obtain target domain user features, source domain user features, and target domain product features. The target domain user features, source domain user features, and target domain product features are jointly fused to generate cross-domain transfer features for target domain recommendation tasks. The cross-domain migration features are subjected to fusion iterative denoising processing. Through multiple rounds of feature updates and weight filtering operations, source domain noise features are gradually suppressed and transferable collaborative preferences are enhanced to generate a denoised cross-domain user feature representation. After the iterative denoising, a user structure preservation constraint is introduced to perform consistency correction on the similarity relationship between users in the target domain before and after migration, so as to maintain the overall distribution and local neighborhood stability of the user preference structure in the target domain. The predicted score is calculated based on the final user feature representation after structure preservation constraint correction and the target domain product features, and collaborative recommendation results are generated based on the predicted score.

[0005] Specifically, the information dissemination and feature aggregation based on the user-product interaction structure includes: A graph-based information propagation model is adopted to perform multi-level neighborhood feature aggregation on user nodes and product nodes in the user-product interaction structure. The node features are then updated by weighting the normalized adjacency relationship to obtain user feature and product feature representations that include high-order collaborative relationships.

[0006] in, and These represent the embedding features of users and products in domain A, respectively. This represents the number of convolutional layers in the graph.

[0007] Specifically, the cross-domain transfer features for generating target domain recommendation tasks are as follows: The target domain user features, source domain user features, and target domain product features are concatenated to construct a joint feature vector. The joint feature vector is then input into a mapping network for nonlinear transformation to generate a cross-domain transfer feature representation related to the semantics of the target domain products, which is used for subsequent fusion and iterative denoising processing.

[0008] in, This indicates a splicing operation. This represents user migration characteristics from the source domain to the target domain. and For learnable parameters, It is a non-linear activation function.

[0009] Specifically, the fusion iterative denoising process includes: Construct a denoising guidance target in each iteration. t In the middle, the current migration features will be Target domain user characteristics Perform fusion and construct fused feature representation Use it as the target for noise reduction. in, These are the preset fusion parameters; Based on the fusion feature representation Transfer characteristics of the current round The data is spliced ​​together, input into the meta-network, and generated through projection. Candidate transfer features representing different preference dimensions ; Based on the above candidate migration features, calculate its fusion features. The similarity between them is used to assign weight coefficients using the Softmax function. The candidate transfer features are weighted and aggregated to obtain the denoised transfer features for the current round. in, , Indicates the first Transfer features after rounds of iterations express The first round generated One candidate transfer feature, To calculate similarity for the inner product, This represents the corresponding weighting coefficient.

[0010] Specifically, the introduced user structure preservation constraint is as follows: Normalize the target domain user embedding vectors before and after migration fusion, and calculate the cosine similarity between users to construct the original user similarity matrix and the enhanced user similarity matrix respectively. in, and These represent the normalized user characteristics before and after migration, respectively. and These represent the original user similarity matrix and the enhanced user similarity matrix, respectively.

[0011] Based on the difference between the original user similarity matrix and the enhanced user similarity matrix, a global structural constraint term is constructed to maintain the consistency of the overall similarity distribution of users in the target domain; in, Represents the set of users in the target domain. The Frobenius norm is represented by the loss obtained by minimizing the MSE (minimum mean square error).

[0012] A collaborative recommendation apparatus with enhanced common preferences, the apparatus comprising: a processor and a memory, the memory storing program instructions, the processor being configured to invoke the program instructions to cause the apparatus to execute any of the collaborative recommendation methods with enhanced common preferences described above.

[0013] A computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform any of the common preference-enhanced collaborative recommendation methods described above.

[0014] The beneficial effects of the technical solution provided by this invention are: (1) This invention integrates target domain user features, source domain user features and target domain product features into a model, and introduces an iterative denoising mechanism after integration. This enables the gradual identification and suppression of source domain noise features that do not match the target domain recommendation task during cross-domain migration. Thus, without the need for complex reconstruction or explicit filtering rules, adaptive filtering of cross-domain migration noise is achieved, significantly improving the stability and accuracy of cross-domain recommendation. (2) The present invention introduces user structure preservation constraints during cross-domain migration, and maintains the stability of the user preference structure in the target domain from the global similarity distribution, which effectively alleviates the problem of user similarity relationship drift caused by cross-domain migration and avoids the damage of negative migration to collaborative recommendation performance; (3) By combining fusion before denoising with structure preservation, this invention breaks the dependence of existing technologies on static cross-domain mapping and one-time fusion strategies. In complex cross-domain scenarios, it can make fuller use of the effective information of the source domain and provide robust and reliable recommendation results for the target domain. Attached Figure Description

[0015] Figure 1 This is a flowchart of a collaborative recommendation method for enhancing common preferences according to an embodiment of the present invention; Figure 2 This is a flowchart of the collaborative recommendation algorithm with enhanced common preferences, as described in an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0017] Example 1 This invention provides a collaborative recommendation method with enhanced common preferences, see [link to relevant documentation]. Figure 1 The method includes the following steps: 101: Construct the user-product interaction structure based on the implicit feedback data from the source and target domains respectively; Obtain implicit user interaction data from the source and target domains, construct corresponding user-product interaction structures, and initialize embedded feature representations for user nodes and product nodes for subsequent cross-domain feature propagation and modeling.

[0018] 102: Based on the interaction structure, feature propagation is performed to obtain multi-domain user features and product features; Information propagation operations are performed on the user-product interaction structures in the source and target domains, respectively, and neighborhood feature aggregation is performed on user nodes and product nodes to obtain user features in the target domain, user features in the source domain, and product features in the target domain.

[0019] 103: Introduce target domain product semantics, fuse and model the features of the target domain and source domain, and generate initial cross-domain transfer features; By jointly fusing target domain user features, source domain user features, and target domain product features, an initial cross-domain transfer feature related to the target domain recommendation task is generated through a mapping network, thereby achieving cross-domain feature transfer guided by the semantics of target domain products.

[0020] 104: Construct a fusion-guided denoising target, perform fusion iterative denoising processing, and gradually suppress cross-domain migration noise; Using the fused representation of target domain user features and initial cross-domain transfer features as a denoising reference, candidate transfer features are generated in multiple iterations. These candidate features are then weighted and updated based on their similarity to the fused representation, gradually reducing noisy features in the source domain that are mismatched with the target domain recommendation task, thus obtaining a denoised cross-domain user feature representation. 105: Introduce user structure preservation constraints to stabilize the target domain user preference structure; By applying consistency constraints to the similarity relationships between users in the target domain before and after migration, the overall structure of the target domain user interest distribution and the stability of local neighborhood relationships are maintained, thus avoiding damage to the original user similarity structure during the cross-domain migration process.

[0021] 106: Calculate the final interaction prediction score and perform ranking, outputting the collaborative recommendation results for the target domain; The user-product interaction prediction score is calculated using the final obtained user feature representation and target domain product features. The target domain products are then ranked according to the prediction score, and the recommendation result is output as the collaborative recommendation result.

[0022] Example 2 The solution in Example 1 will be further described below: 201: Obtain implicit interaction data between the source and target domains and construct the user-product interaction structure; Given source domain With the target domain The implicit feedback data are represented as user-product interaction sets. and ,in Indicates user index, Product indexes belonging to the target domain / source domain. Based on the implicit interaction set, construct user-product interaction matrices for the target domain and source domain respectively: Among them, when user With goods When there is interaction, Otherwise, initialize the user and product embedding matrices of the target domain and source domain respectively: in, For embedded feature dimensions.

[0023] 202: Based on the interaction structure, information dissemination and feature aggregation are performed to obtain multi-domain user features and target domain product features; To capture higher-order cooperative signals and complex user-item relationships, the interaction matrix is ​​normalized based on the user-item interaction structure of the target and source domains, resulting in a normalized adjacency matrix for the target domain. Normalized adjacency matrix of source domain In the During layer propagation, the user feature matrix in the target domain With product feature matrix The update method is defined as follows: Similarly, in the source domain, the user feature matrix With product feature matrix The mode of transmission is as follows: in, and This represents the normalized adjacency matrix.

[0024] To simultaneously capture local neighborhood information and global graph structure features, this invention aggregates all... The embedding vectors of the layers are used to generate the final in-domain user and product feature representations: in, and Representing users respectively End-user feature representation in the target and source domains Indicates target domain products The final product characteristics are represented.

[0025] 203: Integrate source domain user features, target domain user features, and target domain product features to generate cross-domain transfer features perceived by the target domain; The target domain user features, source domain user features, and target domain product features are concatenated to construct a joint feature vector. The joint feature vector is then input into a mapping network for nonlinear transformation to generate a cross-domain transfer feature representation related to the semantics of the target domain products, which is used for subsequent fusion and iterative denoising processing.

[0026] in, This indicates a splicing operation. This represents user migration characteristics from the source domain to the target domain. and For learnable parameters, It is a non-linear activation function.

[0027] 204: Construct a fusion-guided denoising target and perform progressive iterative denoising through multiple rounds of feature updates and weight selection; To further suppress noise features introduced during cross-domain migration, this embodiment introduces a progressive iterative denoising mechanism based on the fused features, in each iteration... t In the middle, the current migration features will be Target domain user characteristics Perform fusion and construct fused feature representation Use it as the target for noise reduction. in, These are the preset fusion parameters; Subsequently, based on the fusion feature representation Transfer characteristics of the current round The data is spliced ​​together, input into the meta-network, and generated through projection. Candidate transfer features representing different preference dimensions ; Based on the aforementioned candidate migration features, calculate the sum of each candidate migration feature and the fusion feature. The similarity between them is calculated, and weight coefficients are assigned based on the similarity using the Softmax function. Then, the candidate transfer features are weighted and aggregated to obtain the denoised transfer features for the current round.

[0028] in, , Indicates the first Transfer features after rounds of iterations express The first round generated One candidate transfer feature, To calculate similarity for the inner product, This represents the corresponding weighting coefficient.

[0029] After completing the multiple rounds of iterative updates, the fused representation obtained in the final round of iteration is used as the user's enhanced feature representation, defined as: in, Indicates the process T The final feature representation after round-fusion guided iterative denoising and fusion.

[0030] 205: Calculate the similarity relationship between users in the target domain before and after migration, and introduce structure preservation constraints to stabilize the user preference representation; Normalize the target domain user embedding vectors before and after migration fusion, and calculate the cosine similarity between users to construct the original user similarity matrix and the enhanced user similarity matrix respectively. in, and These represent the normalized user characteristics before and after migration, respectively. and These represent the original user similarity matrix and the enhanced user similarity matrix, respectively.

[0031] Based on the difference between the original user similarity matrix and the enhanced user similarity matrix, a global structural constraint term is constructed to maintain the consistency of the overall similarity distribution of users in the target domain; in, Represents the set of users in the target domain. The Frobenius norm is represented by the loss obtained by minimizing the MSE (minimum mean square error).

[0032] 206: Calculate interaction prediction scores based on end-user and product characteristics, and output the target domain collaborative recommendation results; In this embodiment of the invention, for the recommendation task of the target domain, the Bayesian Personalized Ranking (BPR) loss is used as the final matching relationship optimization objective. The BPR loss, by modeling positive and negative sample pairs, encourages the model to assign higher prediction scores to items that the user has already interacted with. Its calculation method is as follows: in, This represents the set of user-positive / negative item triples constructed in the training set; Indicates user u Products that have interacted within the target domain. Indicates user u Products that have not been interacted with in the target domain; It is the Sigmoid non-linear activation function.

[0033] In this embodiment, the user For target domain products The predicted score is calculated from the final target domain user features and target domain product features, and its specific form is as follows: in, This represents the final user feature representation obtained after fusion-guided iterative denoising and structure-preserving constraints. Indicates target domain products The final product characteristics are represented.

[0034] Furthermore, considering both the recommendation loss and the structure preservation constraint, the overall optimization objective function of the model is obtained as follows: in, This represents the user structure preservation constraint loss term; Representing model parameters Regularization term; and These are hyperparameters used to balance the effects of various loss terms.

[0035] After the model training is complete, the prediction score is based on the user-item prediction score. The products in the target domain are sorted, and the products with higher scores are selected as the recommended results to obtain the collaborative recommendation results under the target domain.

[0036] Example 3 The effectiveness of the schemes in Examples 1 and 2 is evaluated below based on specific experimental data, as detailed in the following description: 301: Data Composition In this embodiment, several publicly available Amazon cross-domain datasets, widely used in recommender system research, are selected to verify the method of the present invention. These Amazon datasets record user rating behaviors across different product domains, covering information such as users, products, and their interactions, and are characterized by strong representativeness and high sparsity.

[0037] In this embodiment, only user rating information is retained, and the data is preprocessed according to preset rules: first, users and products with fewer than a preset threshold of interactions are filtered out; second, only the common set of users with interaction records in both related domains is retained, thereby constructing a cross-domain recommendation dataset. The four cross-domain datasets finally selected include: Sport&Phone, Sport&Cloth, Elec&Phone, and Elec&Cloth, and their statistical information is shown in Table 1.

[0038] Table 1. Information on four Amazon cross-domain datasets 302: Evaluation Criteria To comprehensively evaluate recommendation performance, this embodiment uses hit rate HR@K and normalized cumulative gain NDCG@K as evaluation metrics. HR@K measures whether the target product appears in the top K recommendation results, and NDCG@K measures the quality of the target product's ranking position in the recommendation list. In this embodiment, K is set to 10.

[0039] 303: Comparison Algorithm In this embodiment, to verify the effectiveness and advancement of the method of the present invention, it is compared with a variety of existing recommendation methods, including two main categories: single-domain recommendation methods and cross-domain recommendation methods. Single-domain recommendation methods include Bayesian Personalized Ranking (BPR), Neural Collaborative Filtering (NCF), LightGCN based on graph structure modeling, and Disentangled Contrastive Learning (DCCF), used to characterize user preference modeling capabilities within a single domain. Cross-domain recommendation methods include CoNet based on feature fusion, DDTCDR based on bidirectional transfer, DML based on metric learning, MADD based on multiple attention, ETL based on preference equivalence transformation, and various disentangled or graph structure-driven cross-domain recommendation methods, such as DisenCDR, GDCCDR, DCCDR, and BITGCF. The above comparison methods cover the mainstream technical routes in the current cross-domain recommendation field, and the specific comparison results are listed in Tables 2 and 3, respectively.

[0040] Table 2. Comparison results (%) between the method and state-of-the-art baseline methods on the Sport&Phone and Sport&Cloth datasets. Table 3. Comparison results (%) between the method and state-of-the-art baseline methods on the Elec&Phone and Elec&Cloth datasets. Tables 2 and 3 present the performance comparison results of the method of this invention on four Amazon cross-domain datasets (Sport & Phone, Sport & Cloth, Elec & Phone, and Elec & Cloth). The experimental results show that as data sparsity and cross-domain differences increase, the recommendation performance of existing single-domain recommendation methods and some traditional cross-domain recommendation methods are affected to varying degrees. However, after introducing the technical solution of this invention, the model achieves further improvements in HR@10 and NDCG@10 metrics on each dataset.

[0041] Furthermore, compared with the existing best-performing cross-domain recommendation methods, the method of this invention achieves performance improvements of approximately 0.5%–1.8% in both HR@10 and NDCG@10 metrics across four cross-domain datasets, and exhibits a more significant advantage in scenarios with higher data sparsity and more significant cross-domain differences. These results demonstrate that this invention, through target-domain-aware cross-domain feature fusion, a fusion-priority iterative denoising mechanism, and user structure preservation constraints, can effectively suppress cross-domain migration noise and stabilize user preference structures, thereby achieving superior recommendation performance in various cross-domain recommendation scenarios, verifying the effectiveness and versatility of the method.

[0042] Furthermore, it should be noted that the method of this invention did not exhibit performance fluctuations or degradation under different cross-domain combinations and data scales, indicating that the proposed fusion iterative denoising and structural constraint strategy has good stability and universality. In summary, the experimental results fully verify the effectiveness of the method of this invention in collaborative recommendation tasks, which can stably improve recommendation performance in complex cross-domain scenarios and has high practical application value.

[0043] Example 4 A collaborative recommendation device with enhanced common preferences includes a processor and a memory, wherein the memory stores program instructions, and the processor invokes the program instructions stored in the memory to cause the device to perform the following method steps in Embodiment 1: Obtain implicit user interaction data from the target domain and the source domain, construct corresponding user-product interaction structures respectively, and initialize user embedding features and product embedding features; Based on the user-product interaction structure, information propagation and feature aggregation are performed on user features and product features in the target domain and source domain to obtain target domain user features, source domain user features, and target domain product features. The target domain user features, source domain user features, and target domain product features are jointly fused to generate cross-domain transfer features for target domain recommendation tasks. The cross-domain migration features are subjected to fusion iterative denoising processing. Through multiple rounds of feature updates, source domain noise features are gradually suppressed and transferable collaborative preferences are enhanced to generate a denoised cross-domain user feature representation. After iterative denoising, user structure constraints are introduced to constrain the similarity relationship between users in the target domain before and after migration, thereby maintaining the stability of the user preference structure in the target domain. The predicted scores of the final user representation after structure preservation constraint correction and the target domain product features are calculated to obtain the matching relationship between users and products.

[0044] A graph-based information propagation model is adopted to perform multi-layer neighborhood feature aggregation on user nodes and product nodes in the user-product interaction structure. In each propagation layer, node features are updated by weighting normalized adjacency relationships to obtain user features and product features that include higher-order collaborative relationships.

[0045] The target domain user features, source domain user features, and target domain product features are concatenated to construct a joint feature vector, which is specifically represented as follows: in, Indicates user User characteristics in the target domain Indicates user User characteristics in the source domain, Indicates the products in the target domain Product characteristics, Indicates feature concatenation operation; The joint feature vector is input into a mapping network for nonlinear transformation to generate a cross-domain transfer feature representation oriented towards the semantics of the target domain product, specifically as follows: in, This indicates a splicing operation. This represents user migration characteristics from the source domain to the target domain. and For learnable parameters, It is a non-linear activation function.

[0046] Construct a denoising guidance target in each iteration. t In the middle, the current migration features will be Target domain user characteristics Perform fusion and construct fused feature representation Use it as the target for noise reduction. in, These are the preset fusion parameters; Based on the fusion feature representation Transfer characteristics of the current round The data is spliced ​​together, input into the meta-network, and generated through projection. Candidate transfer features representing different preference dimensions ; Based on the above candidate migration features, calculate its fusion features. The similarity between them is used to assign weight coefficients using the Softmax function. The candidate transfer features are weighted and aggregated to obtain the denoised transfer features for the current round. in, , Indicates the first Transfer features after rounds of iterations express The first round generated One candidate transfer feature, To calculate similarity for the inner product, This represents the corresponding weighting coefficient.

[0047] Normalize the target domain user embedding vectors before and after migration fusion, and calculate the cosine similarity between users to construct the original user similarity matrix and the enhanced user similarity matrix respectively. Based on the difference between the original user similarity matrix and the enhanced user similarity matrix, a global structural constraint term is constructed to maintain the consistency of the overall similarity distribution of users in the target domain. The structural constraint is calculated as follows: in, Represents the set of users in the target domain. Denotes the Frobenius norm. This represents the user similarity matrix before migration. This represents the user similarity matrix after migration, calculated by minimizing the MSE (Minimum Mean Square Error) loss.

[0048] The interaction prediction score is calculated based on the end-user feature representation and the target domain product features. The target domain products are then sorted according to the interaction prediction score, and a recommendation list is output as the collaborative recommendation result.

[0049] It should be noted that the device descriptions in the above embodiments correspond to the method descriptions in the embodiments, and the embodiments of the present invention will not be repeated here.

[0050] The execution entities of the processor 1 and memory 2 mentioned above can be devices with computing functions such as computers, microcontrollers, and microcontrollers. In specific implementation, the embodiments of the present invention do not limit the execution entity and can select it according to the needs of actual application.

[0051] Data signals are transmitted between memory 2 and processor 1 via bus 3, which will not be described in detail in this embodiment of the invention.

[0052] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium, the storage medium including a stored program, which, when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiments.

[0053] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0054] It should be noted that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not be repeated here.

[0055] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0056] A computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted through a computer-readable storage medium. A computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic or semiconductor, etc.

[0057] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0058] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A collaborative recommendation method with enhanced common preferences, characterized in that, The method includes: Obtain implicit user interaction data from the target domain and the source domain, construct corresponding user-product interaction structures respectively, and initialize user embedding features and product embedding features; Based on the user-product interaction structure, information propagation and feature aggregation are performed on user features and product features in the target domain and source domain to obtain target domain user features, source domain user features, and target domain product features. The target domain user features, source domain user features, and target domain product features are jointly fused to generate cross-domain transfer features for target domain recommendation tasks. The cross-domain migration features are subjected to fusion iterative denoising processing. Through multiple rounds of feature updates, source domain noise features are gradually suppressed and transferable collaborative preferences are enhanced to generate a denoised cross-domain user feature representation. After iterative denoising, user structure constraints are introduced to constrain the similarity relationship between users in the target domain before and after migration, thereby maintaining the stability of the user preference structure in the target domain. The predicted scores of the final user representation after structure preservation constraint correction and the target domain product features are calculated to obtain the matching relationship between users and products.

2. The collaborative recommendation method with enhanced common preferences according to claim 1, characterized in that, The specific steps for information dissemination and feature aggregation of user characteristics and product characteristics are as follows: A graph-based information propagation method is used to perform multi-layer neighborhood feature aggregation on user nodes and product nodes in the user-product interaction structure. In each propagation layer, node features are updated by weighting normalized adjacency relationships to obtain user features and product features that include higher-order collaborative relationships.

3. The collaborative recommendation method for enhancing common preferences according to claim 1, characterized in that, The specific method for generating cross-domain transfer features is as follows: The target domain user features, source domain user features, and target domain product features are concatenated to construct a joint feature vector, which is specifically represented as follows: in, Indicates user User characteristics in the target domain Indicates user i User characteristics in the source domain, Indicates the products in the target domain j Product characteristics, Indicates feature concatenation operation; The joint feature vector is input into a mapping network for nonlinear transformation to generate a cross-domain transfer feature representation oriented towards the semantics of the target domain product, specifically as follows: in, This indicates a splicing operation. This represents user migration characteristics from the source domain to the target domain. and For learnable parameters, It is a non-linear activation function.

4. The collaborative recommendation method with enhanced common preferences according to claim 1, characterized in that, The fusion iterative denoising process specifically includes: Construct a denoising guidance target in each iteration. t In the middle, the current migration features will be Target domain user characteristics Perform fusion and construct fused feature representation Use it as the target for noise reduction. in, These are the preset fusion parameters; Based on the fusion feature representation Transfer characteristics of the current round The data is spliced ​​together, input into the meta-network, and generated through projection. Candidate transfer features representing different preference dimensions ; Based on the above candidate migration features, calculate its fusion features. The similarity between them is used to assign weight coefficients using the Softmax function. The candidate transfer features are weighted and aggregated to obtain the denoised transfer features for the current round. in, , Indicates the first Transfer features after rounds of iterations express The first round generated One candidate transfer feature, To calculate similarity for the inner product, This represents the corresponding weighting coefficient.

5. The collaborative recommendation method for enhancing common preferences according to claim 1, characterized in that, The specific meaning of introducing user structure constraints is as follows: Normalize the target domain user embedding vectors before and after migration fusion, and calculate the cosine similarity between users to construct the original user similarity matrix and the enhanced user similarity matrix respectively. Based on the difference between the original user similarity matrix and the enhanced user similarity matrix, a global structural constraint term is constructed to maintain the consistency of the overall similarity distribution of users in the target domain.

6. The collaborative recommendation method for enhancing common preferences according to claim 5, characterized in that, The structural constraints are calculated as follows: in, Represents the set of users in the target domain. Denotes the Frobenius norm. This represents the user similarity matrix before migration. This represents the user similarity matrix after migration, calculated by minimizing the MSE (Minimum Mean Square Error) loss.

7. The collaborative recommendation method for enhancing common preferences according to claim 1, characterized in that, Specifically, the process of generating recommendation results involves calculating an interaction prediction score between the final user feature representation and the target domain product features, sorting the target domain products based on the interaction prediction score, and outputting a recommendation list as the collaborative recommendation result.

8. A collaborative recommendation device with enhanced common preferences, characterized in that, The device includes: A processor and a memory, wherein the memory stores program instructions, and the processor is configured to invoke the program instructions to cause the apparatus to perform the method of any one of claims 1–7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method of any one of claims 1–7.

Images