A multi-modal recommendation method and system based on offline diversity rearrangement and itemCF boundary teacher distillation

Abstract

The application discloses a kind of based on offline diversity rearrangement and ItemCF boundary teacher distillation multimodal recommendation method and system, the matching degree of the final recommended result of existing multimodal recommendation method and user real demand is obviously reduced, directly cause the technical problem of the accuracy reduction of recommended result.Method includes obtaining user-item interaction matrix, item visual preprocessing feature matrix, item text preprocessing feature matrix, constructs the initial recommendation list of the article of user;Offline diversity rearrangement is carried out to the list, while dividing head protection area, boundary replaceable area, outside candidate area and constructing candidate exchange pair set, calculating ItemCF support and multi-source exchange evidence obtains candidate exchange pair with evidence;Combine head protection area to complete safe offline boundary exchange, output updated version recommendation list and effective safe exchange relationship;Based on effective safe exchange relationship, construct static sample, complete recommendation reasoning and output final recommended result.

Description

Technical Field

[0001] This invention relates to the field of recommender system technology, and in particular to a multimodal recommender method and system based on offline diversity rearrangement and ItemCF boundary teacher distillation. Background Technology

[0002] With the development of e-commerce, content distribution, and personalized information services, recommender systems have become an important technological means to connect user needs with candidate items. Traditional recommendation methods typically rely on historical interaction records between users and items to construct user preference representations and generate ranking results based on the degree of matching between user and item representations. However, in practical applications, user-item interaction data is often sparse, and simply relying on collaborative behavior signals is insufficient to fully characterize the content features of items, and it is also difficult to achieve stable recommendation results for niche items, long-tail items, or scenarios with few interactions.

[0003] To alleviate the aforementioned problems, multimodal recommendation methods have begun to incorporate auxiliary information such as visual features and textual descriptions of items. These methods typically construct a user-item interaction graph based on user-item interaction data and an item similarity graph based on visual or textual features. They then obtain multi-source representations of users and items through graph propagation, feature fusion, or attention mechanisms. Compared to methods that rely solely on the interaction matrix, multimodal recommendation enhances the semantic expressiveness of items and improves the model's ability to characterize user interests.

[0004] Existing multimodal recommendation methods, in improving recommendation diversity, typically calculate the similarity between items based solely on their visual, textual, and other multimodal content features. They then perform redundancy removal and reordering on the initial recommendation list based on similarity scores, suppressing items with excessively high content similarity to reduce the homogeneity of the recommendation results. This approach judges redundancy solely based on the item's content attributes, failing to differentiate based on the user's historical interaction behavior and genuine interests. It cannot accurately identify similar items that are content-redundant and highly relevant similar items that align with the user's core interests. This easily leads to the incorrect downgrading or even exclusion of similar items that highly match the user's true interests from the recommendation list. Consequently, many highly relevant and high-quality candidates fail to reach the top of the recommendation results, resulting in a significant decrease in the match between the final recommendation results and the user's actual needs, directly reducing the accuracy of the recommendation results. Summary of the Invention

[0005] This invention provides a multimodal recommendation method and system based on offline diversity rearrangement and ItemCF boundary teacher distillation, which solves the technical problem that the final recommendation result of existing multimodal recommendation methods has a significantly reduced degree of matching with the user's actual needs, which directly leads to a decrease in the accuracy of the recommendation result.

[0006] The first aspect of this invention provides a multimodal recommendation method based on offline diversity reordering and ItemCF boundary teacher distillation, comprising:

[0007] Obtain the user-item interaction matrix, the item visual preprocessing feature matrix, and the item text preprocessing feature matrix, and construct the user's initial item recommendation list based on the user-item interaction matrix, the item visual preprocessing feature matrix, and the item text preprocessing feature matrix;

[0008] Perform offline diversity rearrangement on the user's initial item recommendation list and output the offline diversity rearranged recommendation list;

[0009] The user's initial item recommendation list is divided into regions to obtain the head protection zone item set, the boundary replaceable zone item set, and the outer candidate zone item set. Exchange pairs are constructed and a candidate exchange pair set is output.

[0010] ItemCF support and multi-source exchange evidence are calculated for the candidate exchange pair set and the user-item interaction matrix, and candidate exchange pairs with evidence are output.

[0011] The offline boundary exchange filtering and replacement process is performed on the recommended list after offline diversity rearrangement using the candidate exchange pairs with evidence and the head protection zone item set, and an updated recommended list and effective secure exchange relationships are output.

[0012] Static samples are constructed based on the effective and secure exchange relationship, and recommendation inference is performed by combining the user's fusion ranking score for each item, the updated recommendation list, and the recommendation list after offline diversity re-ranking, and the final recommendation result is output.

[0013] Optionally, constructing the user's initial item recommendation list based on the user-item interaction matrix, the item visual preprocessing feature matrix, and the item text preprocessing feature matrix includes:

[0014] Construct a user-item interaction graph based on the user-item interaction matrix;

[0015] The K-nearest neighbor algorithm is used to construct visual modal item similarity maps and text modal item similarity maps based on the item visual preprocessing feature matrix and the item text preprocessing feature matrix;

[0016] Initialize the user embedding matrix and the item embedding matrix, and concatenate the user embedding matrix and the item embedding matrix to output the initial node representation of the graph;

[0017] Based on the initial node representation of the graph and the user-item interaction graph, multi-layer graph propagation is performed to obtain multiple propagation layer results;

[0018] The results from multiple propagation layers are aggregated to output a behavioral-side collaborative representation;

[0019] Linear mapping and gated fusion processing are performed on the visual preprocessing feature matrix of the object to output a basic representation of the visual object; graph propagation is performed by combining the basic representation of the visual object and the visual modal object similarity map to output a visual modal representation;

[0020] Linear mapping and gated fusion processing are performed on the preprocessed feature matrix of the item text to output the basic representation of the text item; graph propagation is performed by combining the basic representation of the text item and the text modal item similarity graph to output the text modal representation;

[0021] Attention-weighted calculations are performed on the visual modality representation and the text modality representation to output a shared modality representation;

[0022] Perform a difference operation on the visual modality representation and the shared modality representation to output a visual private representation;

[0023] Perform a difference operation on the text modal representation and the shared modal representation to output the text private representation;

[0024] Residual fusion is performed on the behavior-side collaborative representation, the shared modality representation, the visual private representation, and the text private representation to output the final user representation and the final item representation;

[0025] Calculate the basic ranking score based on the end-user representation and the final item representation;

[0026] Extract the set of items that the user has historically interacted with based on the user-item interaction matrix;

[0027] Based on the visual item representation, the text item representation, and the user's historical interaction item set, a shared view and a private view are constructed.

[0028] Based on the shared view and the private view, the shared score and the private score are calculated, and a weighted fusion calculation is completed by combining the preset shared score weight and the preset private score weight, and the user's fusion ranking score for each item is output.

[0029] Sort the items in descending order based on the user's combined ranking score for each item, and output the user's initial item recommendation list.

[0030] Optionally, the offline diversity re-ranking of the user's initial item recommendation list and the output of the offline diversity re-ranked recommendation list includes:

[0031] The user's initial item recommendation list is filtered by items based on the user's historical interactions, and an interaction-free filtered list is output.

[0032] Based on the fusion ranking score of each item in the user's initial item recommendation list, high-scoring items are extracted from the interaction-free filtering list, and the filtered rearranged candidate pool is output.

[0033] Cosine similarity is calculated for the visual modal representations of multiple items in the filtered rearranged candidate pool, and the visual similarity between each item in the filtered rearranged candidate pool is output.

[0034] Cosine similarity is calculated for the text modal representations of each item in the filtered rearranged candidate pool, and the text similarity between all items in the filtered rearranged candidate pool is output.

[0035] Cosine similarity is calculated for the behavioral modal representations of each item in the filtered rearranged candidate pool, and the behavioral similarity between each item in the filtered rearranged candidate pool is output.

[0036] A comprehensive similarity matrix is ​​constructed based on the visual similarity, text similarity, and behavioral similarity of each item in the filtered rearranged candidate pool;

[0037] Initialize an empty set as the selected recommendation set;

[0038] Multiple candidate items are selected from the filtered rearranged candidate pool;

[0039] Based on the user's fusion ranking score for each item and the comprehensive similarity matrix, calculate the re-ranking selection score for multiple candidate items;

[0040] The item with the highest rearrangement selection score is selected as the best candidate item in this round.

[0041] The best candidate items in this round are incorporated into the selected recommendation set to obtain the updated selected recommendation set;

[0042] When the number of items in the updated selected recommendation set reaches the preset recommendation length, the updated selected recommendation set is used as the recommendation list after diversity rearrangement.

[0043] Optionally, the step of calculating ItemCF support and multi-source exchange evidence on the candidate exchange pair set and the user-item interaction matrix, and outputting candidate exchange pairs with evidence, includes:

[0044] Based on the user-item interaction matrix, the set of interacting users and the popularity of each item are calculated.

[0045] Calculate the ItemCF similarity between each item based on the set of interacting users and the item popularity corresponding to each item.

[0046] Extract the set of items that the user has historically interacted with based on the user-item interaction matrix;

[0047] Based on the ItemCF similarity between the items and the set of items the user has interacted with in the past, for any user and its candidate items, calculate the maximum ItemCF similarity between the candidate items and the user's recent historical interaction items, and output the ItemCF support of each candidate item relative to the corresponding user.

[0048] Based on the set of candidate swap pairs and the ItemCF support of each candidate item relative to the corresponding user, calculate the ItemCF support difference for each set of candidate swap pairs;

[0049] Based on the candidate exchange pair set and the user's fusion ranking score for each item, calculate the multi-dimensional score difference for each candidate exchange pair.

[0050] The ItemCF support difference and multi-dimensional score difference of each candidate swap pair are bound to the corresponding candidate swap pair to complete the multi-source evidence association, forming auxiliary evidence fields for interpreting diagnosis, teacher pair export and dynamic teacher generation process parameter weight adjustment, and outputting candidate swap pairs with evidence.

[0051] Optionally, the step of using the candidate swap pairs with evidence and the head protection zone item set to perform secure offline boundary swap filtering and replacement processing on the recommended list after offline diversity rearrangement, and outputting an updated recommended list and valid secure swap relationships, includes:

[0052] The head protection zone item set is used to perform hard filtering on the candidate swap pairs with evidence, and multiple safe candidate swap pairs are output.

[0053] The ItemCF support difference carried by the multiple security candidate swap pairs is sorted in descending order from largest to smallest, and the truncation and filtering are completed in combination with the preset maximum number of swaps for the user, and the swap pairs to be executed are output.

[0054] The order of items in the offline diversity rearranged recommendation list is swapped using the swap pair to be executed. A two-level cascade swap is performed in combination with the preset boundary level to output an updated recommendation list.

[0055] Extract successfully executed item exchange combinations from the updated recommendation list, organize their relationships, and output valid and secure exchange relationships.

[0056] Optionally, the step of constructing static samples based on the effective secure exchange relationship, and combining the user's fusion ranking score for each item, the updated recommendation list, and the offline diversity-rearranged recommendation list to perform recommendation inference, outputting the final recommendation result, including:

[0057] A comparative analysis is performed using the updated recommendation list and the offline diversity rearranged recommendation list to extract the changes in the sorting positions of boundary replacement items and outer entry items, and a sorting difference table before and after the swap is output.

[0058] Based on the effective and secure exchange relationship and the sorting difference table before and after the exchange, items entering from the outside and items replaced at the boundary are distinguished. Positive and negative sample pairing is completed with user identifiers, and static teacher positive and negative sample pairs are output.

[0059] Based on the ItemCF support difference bound to the static teacher positive and negative sample pairs, the weight of the sample pairs is calculated through the truncation function, and auxiliary fields such as original sorting position, score difference, and exchange rule identifier are added to output weighted static teacher positive and negative sample pairs;

[0060] During model training, according to a preset round interval, based on the sorting results of the current optimal model state, items with low sorting discrimination at the preset cutoff boundary are selected to generate dynamic difficult example teacher pairs.

[0061] The static distillation loss is calculated based on the weighted static teacher positive and negative sample pairs, the dynamic distillation loss is calculated based on the dynamic hard example teacher pairs, and the total training loss is formed by combining the basic recommendation loss, the embedding regularization loss and the contrastive learning loss.

[0062] Backpropagation is performed based on the total training loss to iteratively update the model parameters and output the target model parameters after training is completed.

[0063] Based on the parameters of the trained target model, construct a target recommendation model;

[0064] The target recommendation model is used to calculate the user's target fusion ranking score for each item. After removing the items that have been interacted with, the first preset number of items are truncated in descending order of score, and the final recommendation result is output.

[0065] A second aspect of this invention provides a multimodal recommendation system based on offline diversity rearrangement and ItemCF boundary teacher distillation, comprising:

[0066] The acquisition module is used to acquire the user-item interaction matrix, the item visual preprocessing feature matrix, and the item text preprocessing feature matrix, and to construct the user's initial item recommendation list based on the user-item interaction matrix, the item visual preprocessing feature matrix, and the item text preprocessing feature matrix;

[0067] The reordering module is used to perform offline diversity reordering on the user's initial item recommendation list and output the recommendation list after offline diversity reordering.

[0068] The construction module is used to divide the user's initial item recommendation list into regions, obtain the head protection area item set, the boundary replaceable area item set, and the outer candidate area item set, and construct swap pairs, outputting a candidate swap pair set;

[0069] The calculation module is used to calculate ItemCF support and multi-source exchange evidence for the candidate exchange pair set and the user-item interaction matrix, and output candidate exchange pairs with evidence.

[0070] The replacement module is used to perform secure offline boundary exchange filtering and replacement processing on the recommended list after offline diversity rearrangement using the candidate exchange pairs with evidence and the head protection zone item set, and output an updated recommended list and valid secure exchange relationships;

[0071] The recommendation module is used to construct static samples based on the effective secure exchange relationship, and to perform recommendation inference by combining the user's fusion ranking score for each item, the updated recommendation list, and the recommendation list after offline diversity reordering, and output the final recommendation result.

[0072] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation as described above.

[0073] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation as described above.

[0074] The fifth aspect of the present invention provides a computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein, when the program instructions are executed by a computer, the computer performs the steps of the multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation as described above.

[0075] As can be seen from the above technical solutions, the present invention has the following advantages:

[0076] The present invention provides a multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation. It obtains a user-item interaction matrix, an item visual preprocessing feature matrix, and an item text preprocessing feature matrix. Based on these matrices, an initial item recommendation list is constructed for the user. Offline diversity rearrangement is performed on the initial recommendation list, outputting a revised recommendation list. The initial recommendation list is then divided into three regions: a head protection zone item set, a boundary replaceable zone item set, and an outer candidate zone item set. Exchange pairs are constructed, outputting a candidate exchange pair set. ItemCF support and multi-source exchange evidence are calculated on the candidate exchange pair set and the user-item interaction matrix, outputting candidate exchange pairs with evidence. The candidate exchange pairs with evidence and the head protection zone item set are used to perform secure offline boundary exchange filtering and replacement processing on the revised recommendation list, outputting an updated recommendation list and valid secure exchange relationships. Static samples are constructed based on the valid secure exchange relationships, and combined with the user's preferences for each... The invention uses the fusion ranking score of items, the updated recommendation list, and the offline diversity-rearranged recommendation list to perform recommendation inference and output the final recommendation result. Based on the above scheme, this invention divides the initial recommendation list into regions, placing high-confidence items at the top of the ranking into the head protection zone and not participating in the replacement operation. This mechanism can ensure the stability of highly relevant head recommendation results. At the same time, candidate exchange pairs are constructed for the boundary replaceable area and the outer candidate area. The ItemCF support is calculated by combining the user-item interaction matrix as the basis for exchange judgment. The degree of matching between candidate items and user interests can be quantified by using collaborative behavior signals. Only outer items with higher collaborative support are selected to replace boundary items. On this basis, the safe offline boundary exchange screening and replacement processing is carried out only in the boundary area with limited adjustments. This can optimize the diversity of the recommendation list without destroying the overall relevance of the recommendation. Finally, the effective safe exchange relationship is transformed into static samples and integrated into the recommendation inference process. The ranking optimization logic at the boundary can be transmitted to the generation stage of the final recommendation result, further improving the rationality of the candidate item ranking at the boundary position, thereby improving the accuracy of the final recommendation result. Attached Figure Description

[0077] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0078] Figure 1This is a flowchart illustrating the steps of a multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation, as provided in Embodiment 1 of the present invention.

[0079] Figure 2 This is an overall framework diagram of a multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation provided in Embodiment 1 of the present invention;

[0080] Figure 3 This is a structural block diagram of a multimodal recommendation system based on offline diversity rearrangement and ItemCF boundary teacher distillation, provided in Embodiment 2 of the present invention. Detailed Implementation

[0081] This invention provides a multimodal recommendation method and system based on offline diversity rearrangement and ItemCF boundary teacher distillation, which solves the technical problem that the final recommendation result of existing multimodal recommendation methods has a significantly reduced degree of matching with the user's actual needs, directly resulting in a decrease in the accuracy of the recommendation result.

[0082] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that in the optional embodiments of the present invention, the object information and other related data involved require the permission or consent of the object when the embodiments of the present invention are applied to specific products or technologies, and the collection, use, and processing of related data must comply with relevant laws, regulations, and standards. That is to say, if the embodiments of the present invention involve data related to the object, it needs to be obtained with the authorization and consent of the object, the authorization and consent of relevant departments, and in compliance with relevant laws, regulations, and standards. If personal information is involved in the embodiments, the acquisition of all personal information requires the consent of the individual. If sensitive information is involved, the separate consent of the information subject is required, and the embodiments also need to be implemented with the authorization and consent of the object.

[0083] Terminology Explanation:

[0084] Baby, Sports, and Clothing are the Amazon Baby Products dataset, the Amazon Sports & Outdoors dataset, and the Amazon Apparel, Footwear, and Jewelry dataset, respectively.

[0085] VBPR (Visual Bayesian Personalized Ranking from Implicit Feedback) refers to visual Bayesian personalized ranking based on implicit feedback.

[0086] MMGCN (Multi-modal Graph Convolution Network for Personalized Recommendation of Micro-video) represents a multimodal graph convolutional network for personalized recommendation of micro-videos.

[0087] SLMRec (Self-Supervised Learning for Multimedia Recommendation) represents self-supervised learning for multimedia recommendation.

[0088] BM3 (Bootstrap Latent Representations for Multi-modal Recommendation) represents bootstrap latent representation learning for multimodal recommendation.

[0089] MMSSL (Multi-Modal Self-Supervised Learning for Recommendation) represents multimodal self-supervised learning for recommendation.

[0090] MGCN (Multi-View Graph Convolutional Network for Multimedia Recommendation) represents a multi-view graph convolutional network for multimedia recommendation.

[0091] MMIL (Multimodal-aware Multi-intention Learning for Recommendation) represents multimodal-aware multi-intention learning for recommendation.

[0092] DA-MRS (Improving Multi-modal Recommender Systems by Denoising and Aligning Multi-modal Content and User Feedback) refers to improving the performance of multi-modal recommendation systems by denoising multi-modal content and aligning it with user feedback.

[0093] LGMRec (Local and Global Graph Learning for Multimodal Recommendation) represents local and global graph learning for multimodal recommendation.

[0094] MENTOR (Multi-level Self-supervised Learning for Multimodal Recommendation) represents multi-level self-supervised learning for multimodal recommendation.

[0095] MIG-GT (Modality-Independent Graph Neural Networks with GlobalTransformers for Multimodal Recommendation) represents a modality-independent graph neural network that incorporates global Transformers for multimodal recommendation.

[0096] Please see Figure 1 , Figure 1 The flowchart illustrates the steps of a multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation, as provided in Embodiment 1 of the present invention.

[0097] This invention provides a multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation, comprising:

[0098] Step 101: Obtain the user-item interaction matrix, the item visual preprocessing feature matrix, and the item text preprocessing feature matrix, and construct the user's initial item recommendation list based on the user-item interaction matrix, the item visual preprocessing feature matrix, and the item text preprocessing feature matrix.

[0099] The user-item interaction matrix is ​​a two-dimensional numerical matrix used to record the historical interaction behavior between all users and all items. The row dimension of the matrix corresponds to all users, and the column dimension corresponds to all items. The elements in the matrix are used to mark whether the user and the item at the corresponding position have interacted. It can include different types of interaction behavior records such as click, browse, favorite, add to cart, purchase, etc. It is the basic data source for constructing user-item interaction graphs and calculating collaborative filtering metrics.

[0100] The object visual preprocessing feature matrix is ​​a two-dimensional numerical matrix that stores the visual features of all objects. Each row of the matrix corresponds to the visual feature vector of an object, and each column corresponds to a visual feature dimension. This matrix is ​​generated by extracting features from the visual content of the corresponding image, cover, poster, etc. through a pre-trained visual model, and then performing normalization, dimensionality reduction and other standardized preprocessing. It contains numerical representations of visual attributes such as the appearance, color style, and layout of the object, and is the data foundation for constructing visual modal object similarity maps and calculating the visual similarity of objects.

[0101] The item text preprocessing feature matrix is ​​a two-dimensional numerical matrix that stores the semantic features of all item texts. Each row of the matrix corresponds to the text feature vector of an item, and each column corresponds to a text feature dimension. This matrix is ​​generated by extracting features from the text content such as the title, detailed description, attribute tags, and classification information of the item through a pre-trained text model, and then performing normalization, truncation and completion, and other standardized preprocessing. It contains numerical expressions of semantic attributes such as the item's category, function, attribute parameters, and scene location, and is the data foundation for constructing a text modal item similarity graph and calculating item text similarity.

[0102] It should be noted that, firstly, a user-item interaction graph is constructed based on the user-item interaction matrix. Simultaneously, the K-nearest neighbor algorithm is used to generate visual modal item similarity maps and text modal item similarity maps based on the item visual preprocessing feature matrix and the item text preprocessing feature matrix, respectively. The user embedding matrix and item embedding matrix are initialized and concatenated to obtain the initial node representation of the graph. This initial representation is input into the user-item interaction graph for multi-layer graph propagation. After aggregating the results of each layer, a behavior-side collaborative representation is obtained. At the same time, linear mapping and gating fusion are performed on the visual and text preprocessing features, respectively, and then input into the corresponding modal item similarity maps to complete propagation, obtaining visual modal representations and text modal representations. Attention weighting is applied to the two types of modal representations to obtain a shared modal representation, and then the visual private representation and text private representation are separated through difference operations. The behavior-side collaborative representation and the three types of modal representations are residually fused to output the final user representation and the final item representation. Based on the two final representations, multi-dimensional ranking scores are calculated and weighted to obtain the fused ranking score for each item. The items are then sorted in descending order of score to form the user's initial item recommendation list.

[0103] Further, step 101 may include the following sub-steps:

[0104] S11. Construct a user-item interaction graph based on the user-item interaction matrix;

[0105] S12. The K-nearest neighbor algorithm is used to construct visual modal item similarity maps and text modal item similarity maps based on the item visual preprocessing feature matrix and the item text preprocessing feature matrix.

[0106] S13. Initialize the user embedding matrix and the item embedding matrix, and concatenate the user embedding matrix and the item embedding matrix to output the initial node representation of the graph;

[0107] S14. Perform multi-layer graph propagation based on the initial node representation of the graph and the user-item interaction graph to obtain multiple propagation layer results;

[0108] S15. Aggregate the results from multiple propagation layers and output the behavioral-side collaborative representation;

[0109] S16. Perform linear mapping and gating fusion processing on the visual preprocessing feature matrix of the object to output the basic representation of the visual object; combine the basic representation of the visual object and the visual modal object similarity graph to perform graph propagation and output the visual modal representation.

[0110] S17. Perform linear mapping and gated fusion processing on the preprocessed feature matrix of the item text to output the basic representation of the text item; combine the basic representation of the text item and the text modal item similarity graph to perform graph propagation and output the text modal representation.

[0111] S18. Perform attention-weighted calculation on the visual modality representation and the text modality representation, and output the shared modality representation;

[0112] S19. Perform a difference operation on the visual modality representation and the shared modality representation, and output the visual private representation;

[0113] S110. Perform a difference operation on the text modal representation and the shared modal representation, and output the private text representation;

[0114] S111, Perform residual fusion on the behavior-side collaborative representation, shared modality representation, visual private representation, and text private representation, and output the final user representation and the final item representation;

[0115] S112. Calculate the basic sorting score based on the end-user representation and the final item representation;

[0116] S113. Extract the set of historical user-interacted items based on the user-item interaction matrix;

[0117] S114. Construct shared and private views based on the basic representation of visual items, the basic representation of text items, and the collection of items based on user history interactions;

[0118] S115. Based on the shared view and private view, calculate the shared score and private score, and combine the preset shared score weight and preset private score weight to complete the weighted fusion calculation, and output the user's fusion ranking score for each item;

[0119] S116. Sort the items in descending order based on the user's combined ranking scores for each item, and output the user's initial recommended list of items.

[0120] It should be noted that, firstly, the user set U, the item set I, the user-item interaction matrix R, and the item visual preprocessing feature matrix X are obtained. v and item text preprocessing feature matrix X t Construct a user-item interaction graph G based on the user-item interaction matrix. ui This is used to represent the collaborative behavioral relationship between users and items; a visual modal item similarity map G is constructed based on visual preprocessing features. v Based on the text preprocessing features, a text modal item similarity graph G is constructed. t .

[0121] In one implementation, a K-nearest neighbor approach can be used to construct item similarity graphs for visual and textual modalities, connecting each item to the few most similar items in its corresponding modal feature space. Initialize the user embedding matrix E. u and item embedding matrix E i The two are concatenated to form the initial node representation of the behavior graph, and multi-layer graph propagation is performed on the user-item interaction graph to obtain the behavior-side user representation and the behavior-side item representation.

[0122] ;

[0123] ;

[0124] in, This represents the user / item node features after aggregation of l-order neighbor information, i.e., the initial node representation of the graph. Let l represent the normalized adjacency matrix corresponding to the user-item interaction graph, and l represent the propagation layer index. Aggregating the representations from different propagation layers yields the behavior-side collaborative representation H. a .

[0125] ;

[0126] For the visual modality, preprocessed visual features are input into the visual feature mapping layer to obtain a visual feature representation. This representation is then fused with the item embedding through a visual modality gating function to form a basic visual item representation. Subsequently, graph propagation is performed on the visual modality item similarity map to obtain the visual modality representation. The text modality is processed in the same way to obtain the text modality representation.

[0127] ;

[0128] ;

[0129] in, The total propagation layer of the user-item interaction graph is a preset positive integer used to limit the number of iterations of graph propagation and the order of neighbor information aggregation; Represented by visual features; This is the weight matrix of the visual feature mapping layer, specifically a trainable two-dimensional parameter matrix; This is the bias term of the visual feature mapping layer, specifically a trainable one-dimensional parameter vector; For the basic representation of visual objects; is the weight matrix of the visual gating layer, and is a trainable two-dimensional parameter matrix; , where is the bias term of the visual gating layer, and is a trainable one-dimensional parameter vector; is the weight matrix of the text feature mapping layer, and is a trainable two-dimensional parameter matrix; , where is the bias term of the text feature mapping layer, and is a trainable one-dimensional parameter vector; The text feature representation is obtained by linear mapping from the preprocessed features of the item text. It is an intermediate feature result after the original text features are projected onto the target feature space. It carries the semantic information of the text after dimensional transformation and is the input data for subsequent gated fusion operations. is the weight matrix of the text gating layer, and is a trainable two-dimensional parameter matrix; , where is the bias term of the text gate layer, and is a trainable one-dimensional parameter vector; It serves as the basic representation of textual items, obtained by performing linear mapping and gated fusion processing on the preprocessed features of the item text. It integrates item embedding and text modal feature information, and is the basic input feature for graph propagation of text modal item similarity graphs.

[0130] Subsequently, in the visual modal item similarity graph G v The graph is propagated upwards to obtain the visual object modal representation:

[0131] ;

[0132] in, The visual object modal representation is obtained by combining the basic visual object representation with the visual modal object similarity graph and performing graph propagation. It incorporates the visual feature information of nearby similar objects. A v L represents the normalized adjacency matrix of the visual modality item similarity graph. m This indicates the number of modality propagation layers. Further, based on the user-item interaction matrix R, the visual item modality representation is aggregated to obtain the visual user modality representation:

[0133] ;

[0134] R represents the user-item interaction matrix, which represents the visual user modality. Visual object modal representation By concatenating the components, we obtain the joint representation of the visual modalities, i.e., the visual modal representation. :

[0135] ;

[0136] The same method can be used to obtain the text modality representation H. t .

[0137] After obtaining the visual modal representation H v and text modal representation H t Next, shared-private modality difference modeling is performed. First, shared attention scoring is performed on the joint representation of the visual modality and the joint representation of the text modality, respectively:

[0138] ;

[0139] in, For visual modality shared attention score, For text modality shared attention score, This represents the shared attention scoring function. The shared attention weights for the visual and textual modalities are obtained using Softmax.

[0140] ;

[0141] in, Indicates shared weights for visual modalities. This represents the shared attention weights of the text modality. A shared modality representation is generated based on the shared attention weights described above:

[0142] ;

[0143] in, This represents the shared representation between the visual and textual modalities. Further, private visual and private textual representations are obtained through a differential approach:

[0144] ;

[0145] ;

[0146] in, Representing shared modal representation, This represents a visual private representation. The text-private representation is used. To better match the private representation with user behavior preferences, visual preference gating and text preference gating are generated based on the behavior-side representation, and the visual private representation and text-private representation are recalibrated respectively. The recalibrated private representation and the shared representation together form the modality-side comprehensive representation, which is then residually fused with the behavior-side representation to obtain the final user representation and the final item representation.

[0147] ;

[0148] in, The visual preference gating coefficient matrix is ​​generated from the behavior-side collaborative representation and is used for element-wise recalibration of the visual private representation. The text preference gating coefficient matrix is ​​generated by the behavior-side collaborative representation and is used to recalibrate the text private representation element by element. The final representation, including the end-user representation and the final item representation, is obtained by residual fusion of the behavior-side collaborative representation, shared modality representation, visual private representation, and text private representation.

[0149] Finally, the base sort score is calculated based on the end-user representation and the final item representation.

[0150] Before performing diversity rearrangement, this invention first needs to calculate the shared score and the private score. Define the basic representation of visual and textual items:

[0151] ;

[0152] in, Let be the visual object basic representation of the i-th item, and let be the visual dimension feature vector of the corresponding item. It is obtained by fusing the item embedding and the reliability-weighted visual gating features, and is the visual dimension basic feature for constructing the shared representation of items and building the shared view and private view. Let be the textual item basic representation of the i-th item, and let be the textual dimension feature vector of the corresponding item. It is obtained by fusing item embedding with reliability-weighted textual gating features and is the textual dimension basic feature for constructing shared item representations and building shared and private views. Let be the item embedding vector of the i-th item, and let be the row vector of the item embedding matrix corresponding to item i. Let be the basic identity feature information of the item, and let be the row vector of the item embedding matrix corresponding to item i. Let be the basic representation of the item. Let i be the text feature representation of item i after text feature linear mapping and modal reliability routing. It is a d-dimensional feature vector, which is obtained by dimensional transformation and reliability filtering of the original text preprocessing features. It is the input feature of the text gating layer. , The reliability of visual and textual modalities are both scalars with values ​​ranging from 0 to 1. ∈R d Let i be the visual feature representation of item i after visual feature linear mapping and modal reliability routing processing. Construct an item-sharing representation based on reliability weighting:

[0153] ;

[0154] User-shared representation is equal to the mean of their historical interactive item-shared representation:

[0155] ;

[0156] in, This is the smoothing coefficient, which is a preset minimum positive value; The set of historical interaction items for user u contains all items for which the user has historical interaction behavior. It is a sample set for aggregate calculation of user shared representation and user private representation.

[0157] The code is actually implemented using a sparse interaction matrix:

[0158] ;

[0159] in, This is a shared representation matrix for all users. Each row of the matrix corresponds to a shared representation vector for a user. It is obtained by combining the shared representation of items through the user-item interaction matrix and then normalizing it element by element. It is a matrix form result of batch calculation of the shared score of all users. This is the shared representation matrix of all items. Each row of the matrix corresponds to the shared representation vector of an item. It is composed of the shared representations of all items arranged in the order of the items and is the input matrix for batch aggregation calculation of user shared representations. Hadamard division is an operation that divides corresponding elements of two matrices with identical dimensions one by one. It is used to achieve element-wise normalization of the matrix dimensions. This is a user interaction vector, where each element corresponds to the number of items a user has interacted with in the past. It is calculated by summing the rows of the user-item interaction matrix and is used to normalize the aggregated user-shared features.

[0160] Shared scores It is the inner product between normalized shared representations, i.e., cosine similarity:

[0161] ;

[0162] Similarly, we can conclude that... The three types of scores are weighted and merged to obtain the merged and ranked score.

[0163] ;

[0164] in, The shared representation is derived by averaging the shared representations of items interacted with by the user throughout their history, and is used to calculate the shared score. The shared item representation is constructed by weighting the visual and textual item representations and is used to calculate the sharing score. The base ranking score is calculated from the final user representation and the final item representation, representing the degree of matching between the user and the item in the comprehensive feature space. The sharing score, representing the cosine similarity between user-shared representations and item-shared representations, characterizes the degree of matching between the two in the common semantic dimension. The private score, calculated using the same logic as the shared score, represents the degree of matching between users and items in terms of differentiated features. This represents the merged sort score, where u represents the user and i represents the item. Indicates shared score weights. This represents the private score weight. The candidate items are sorted in descending order based on the fusion ranking score to obtain the user's initial recommendation list.

[0165] In this embodiment, a user-item interaction graph is first constructed based on the user-item interaction matrix. The graph uses all users and all items as nodes, and historical interactions between users and items as edges to represent collaborative behavior associations. The K-nearest neighbor algorithm is used to calculate the cosine similarity of visual feature vectors of each item in the item's visual preprocessing feature matrix and the cosine similarity of text feature vectors of each item in the item's text preprocessing feature matrix. Each item is then connected to its K nearest neighbors with the highest similarity in the corresponding modal feature space, thus constructing a visual modal item similarity graph and a text modal item similarity graph. A trainable user embedding matrix and item embedding matrix are initialized according to the formula... By concatenating the user embedding matrix and the item embedding matrix along the feature dimension, we obtain the initial node representation of the user-item interaction graph.

[0166] Subsequently, multi-level graph propagation is performed on the user-item interaction graph, and the propagation calculation satisfies the formula. The features of upper-level nodes are weighted and aggregated using the normalized adjacency matrix corresponding to the user-item interaction graph. Each layer of propagation yields node features aggregated with neighbor information of the corresponding order. This process is iterated until all preset propagation layers are completed, resulting in multiple propagation layer results. After all layers of propagation are completed, the results are processed according to the formula... Mean aggregation is performed on the node features of all propagation layers to output a behavioral-side collaborative representation. For the visual modality, the formula is used... The object's visual preprocessing feature matrix is ​​input into the visual feature mapping layer to perform a linear transformation. The original visual features are projected onto the target feature space through a learnable visual mapping weight matrix and bias terms to obtain a visual feature representation. Then, according to the formula... Element-wise gating coefficients are generated using a visual modality gating function consisting of a fully connected layer and a sigmoid activation function. These gating coefficients are then subjected to a Hadamard product operation with the item embedding matrix to achieve adaptive fusion, forming the basic visual item representation. Subsequently, the formula is followed... A graph propagation process with a preset number of layers is performed on the visual modal item similarity graph. The nearest neighbor similarity features of the items are aggregated through the normalized adjacency matrix of the visual modal item similarity graph to obtain the visual item modal representation. Further, the representation is obtained according to the formula... By weighted aggregation of the visual modal representations of all items that the corresponding user has interacted with in the past through the user-item interaction matrix, the visual user modal representation is obtained;

[0167] Finally, according to the formula The visual user modality representation and the visual object modality representation are concatenated along the feature dimension to obtain the visual modality representation. For the text modality, the processing flow is exactly the same as that for the visual modality, sequentially performing linear mapping, gated fusion, modality graph propagation, user-side feature aggregation, and feature concatenation operations to obtain the text modality representation.

[0168] After obtaining the visual modal representation and the text modal representation, shared-private modality difference modeling is performed, first according to the formula. Attention scores are assigned to both visual and textual modal representations using a shared attention scoring function to obtain the corresponding shared attention scores; then, according to the formula... Softmax normalization is applied to the two types of attention scores to obtain the visual modality shared weights and text modality shared weights; then, the formula is applied... A weighted summation is performed on the visual modality representation and the text modality representation to generate a shared modality representation that integrates the common semantics of the two modalities. After the shared modality representation is calculated, it is then processed according to the formula... Subtracting the shared modality representation from the visual modality representation and performing a difference operation, after stripping away common semantics, yields a visual private representation that carries only visually unique attributes; according to the formula... Subtracting the shared modality representation from the text modality representation and performing a difference operation yields a text-private representation that carries only the unique semantics of the text. After obtaining the behavior-side collaborative representation, shared modality representation, visual private representation, and text-private representation, residual fusion processing is performed according to the formula. Based on behavioral collaborative representation as the fundamental feature, the mean of shared modality representation and private representation after preference gating and recalibration is superimposed. Multimodal semantic features are incorporated while preserving behavioral collaborative information, outputting the final user representation and the final item representation. Subsequently, the inner product of the final user representation and the final item representation is calculated to obtain the basic ranking score. Simultaneously, the historical interaction item set of the corresponding user is extracted based on the user-item interaction matrix. Shared and private representations of the item dimension are constructed based on the visual item base representation and the text item base representation, respectively. The average of the shared representations of the user's historical interaction items is then taken to obtain the user's shared representation, and similarly, the user's private representation is obtained, thus constructing the shared view and the private view. According to the formula... Calculate the cosine similarity between the user's shared representation and the candidate item's shared representation, using this as the shared score. The same calculation logic is used to obtain the private score. Finally, follow the formula... The system combines preset shared score weights and private score weights to perform a weighted fusion of the basic ranking score, shared score, and private score, resulting in a user's fused ranking score for each item. All candidate items are then sorted in descending order according to their fused ranking scores, and the user's initial item recommendation list is output.

[0169] Step 102: Perform offline diversity rearrangement on the user's initial item recommendation list and output the offline diversity rearranged recommendation list.

[0170] It should be noted that the initial recommendation list of items for the user is first filtered based on the user's historical interaction item set, removing items that the user has already interacted with, and outputting a filtered list. Then, a preset number of high-scoring items are selected from high to low according to the fusion ranking score to form a filtered rearrangement candidate pool. Subsequently, the visual modal representation, text modal representation, and behavioral modal representation of each item in the rearrangement candidate pool are extracted respectively, and the cosine similarity between each pair of items is calculated to obtain the corresponding visual similarity, text similarity, and behavioral similarity. The comprehensive similarity matrix of the rearrangement candidate pool is constructed by combining the three types of similarity. An empty set is initialized as the selected recommendation set. The rearrangement selection score of the remaining candidate items is calculated round by round, and the item with the highest score is selected as the best candidate item in this round and added to the selected recommendation set. This process is continued until the number of items in the selected recommendation set reaches the preset recommendation length, and finally, the offline diversity rearranged recommendation list is output.

[0171] Furthermore, step 102 may include the following sub-steps:

[0172] S21. Filter the user's initial item recommendation list by the user's historical interaction items and output the no-interaction filter list.

[0173] S22. Based on the fusion ranking scores of each item in the user's initial item recommendation list, extract high-scoring items from the interactive filter list and output the filtered rearranged candidate pool.

[0174] S23. Calculate the cosine similarity of the visual modal representations of multiple items in the filtered rearranged candidate pool, and output the visual similarity between each item in the filtered rearranged candidate pool.

[0175] S24. Calculate the cosine similarity of the text modal representations of each item in the filtered rearranged candidate pool, and output the text similarity between all items in the filtered rearranged candidate pool.

[0176] S25. Calculate the cosine similarity of the behavioral modal representations of each item in the filtered rearranged candidate pool, and output the behavioral similarity between each item in the filtered rearranged candidate pool.

[0177] S26. Construct a comprehensive similarity matrix based on the visual similarity, text similarity, and behavioral similarity of each item in the filtered rearranged candidate pool;

[0178] S27. Initialize an empty set as the selected recommendation set;

[0179] S28. Select multiple candidate items from the filtered rearranged candidate pool;

[0180] S29. Calculate the re-selection score of multiple candidate items based on the user's fusion ranking score and comprehensive similarity matrix for each item.

[0181] S210. Select the item with the highest rearrangement score as the best candidate item for this round.

[0182] S211. Merge the best candidate item in this round into the selected recommendation set to obtain the updated selected recommendation set;

[0183] S212. When the number of items in the updated selected recommendation set reaches the preset recommendation length, the updated selected recommendation set will be used as the recommendation list after diversity rearrangement.

[0184] It should be noted that after obtaining the fused ranking score output by the multimodal base ranking model, offline diversity re-ranking is further performed to reduce content redundancy in the recommendation list while maintaining recommendation accuracy, thereby improving the diversity and coverage of the recommendation results.

[0185] For any user u, first obtain the score set of user u for all candidate items based on the fused ranking score output by the multimodal basic ranking model. :

[0186] ;

[0187] right Sort by score from highest to lowest to obtain user u's basic sorted list, which is the user's initial recommended list of items:

[0188] ;

[0189] in, This represents a descending sort function, used to sort a set of scores from largest to smallest, and output the sorted results for the corresponding items. This represents the initial candidate ranking result directly output by the multimodal base ranking model. To reduce computational overhead, the top K0 items can be extracted from the base ranking list as an offline re-ranking candidate pool, i.e., the recommended list after offline diversity re-ranking. :

[0190] ;

[0191] in, This function extracts the top K0 elements from a descendingly sorted base list, generating a rearrangement candidate pool where K0 represents the size of the candidate pool, and typically satisfies the following conditions:

[0192] K0>K;

[0193] K represents the length of the final recommendation list. In other words, the base ranking model first recalls a large set of highly relevant candidates, and the offline re-ranking module then generates an offline re-ranking list from this candidate set. In this embodiment, this offline re-ranking list primarily serves as input for subsequent boundary-safe swapping and the construction of teacher positive-negative sample pairs.

[0194] In another implementation, to avoid repeatedly recommending items that the user has already interacted with, the candidate pool can be filtered based on the user's historical interactions. That is, the initial item recommendation list for the user is filtered for items based on their historical interactions, and a filtered list of items to remove interactions is output. Let the set of historical interaction items for user u be... for:

[0195] ;

[0196] Then the filtered candidate pool It can be represented as:

[0197] ;

[0198] After obtaining the candidate pool Next, the similarity between candidate items is calculated to measure the redundancy within the recommendation list. The similarity between candidate items can be obtained by weighting visual similarity, textual similarity, and behavioral similarity:

[0199] ;

[0200] in, This represents the overall similarity between item i and item j. This represents the similarity between item i and item j in the visual modality. This indicates the similarity between the two in the text modality. This indicates the similarity between the two in the cooperative behavior representation space; , and These represent the weights of visual similarity, textual similarity, and behavioral similarity, respectively.

[0201] In one specific implementation, cosine similarity can be used to calculate the similarity of items across different modalities. For example, visual similarity can be expressed as:

[0202] ;

[0203] Text similarity can be represented as:

[0204] ;

[0205] Behavioral similarity can be represented as:

[0206] ;

[0207] in, The visual item representation of item i, i.e., the item-side feature vector corresponding to the visual modality representation, is the feature basis for calculating visual similarity. Visual feature representation is obtained by performing a linear mapping on the visual preprocessing features of the item. This visual feature representation is then fused with the item embedding through a visual modality gating function to obtain the basic visual item representation. This basic representation is then input into the visual modality item similarity graph and subjected to multi-layer graph propagation. After aggregating the visual similarity features of neighboring items, the visual item modality representation is obtained. The feature vector corresponding to the i-th item is then extracted from this representation. ; The textual item representation of item i is obtained by performing a linear mapping on the preprocessed text features of the item to obtain the text feature representation. This text feature representation is then fused with the item embedding using a text modality gating function to obtain the basic textual item representation. This basic textual item representation is then input into a text modality item similarity graph and subjected to multi-layer graph propagation. After aggregating the textual similarity features of nearest-neighbor items, the textual item modality representation is obtained. The feature vector corresponding to the i-th item is then extracted from this representation. ; The final behavior fusion representation (i.e., behavior modality representation) of item i is obtained by initially embedding the concatenated user and item into the input user-item interaction graph and performing multi-layer graph propagation to obtain node features at different propagation layers. The mean aggregation of the multi-layer propagation results yields the behavior-side collaborative representation, from which the node feature vector corresponding to the i-th item is extracted. .

[0208] The offline diversity re-ranking module constructs the final recommendation list using a stepwise selection method. Let the set of recommendations selected in the r-th round be:

[0209] ;

[0210] Initially:

[0211] ;

[0212] In round r+1, select the set of items from the candidate pool that have not yet been selected. \ Choose item j. The reordering score for this item is determined by both the relevance term and the diversity penalty term:

[0213] ;

[0214] in, Let represent the set of recommended items selected by user u after the r-th iteration, containing all items selected in the previous r iterations, and serve as the reference set for calculating the redundancy penalty term. This represents the initial set of selected recommendations, which can be an empty set, corresponding to the starting state of the diversity rearrangement iteration. This represents the reordering selection score of candidate item j corresponding to user u. It is a scalar value calculated by combining the relevance term and the diversity penalty term, and serves as the criterion for selecting the optimal candidate item in each round. This represents the diversity adjustment coefficient, used to control the trade-off between recommendation accuracy and recommendation diversity. This represents the redundancy penalty term between candidate item j and the currently selected recommendation set. The redundancy penalty term can be represented by the maximum similarity between the candidate item and the selected items:

[0215] ;

[0216] when season:

[0217] ;

[0218] Therefore, the first round of selection is mainly based on the fusion score of the multimodal basic ranking model, while subsequent rounds gradually suppress candidate items that are too similar to the selected items based on the relevance score.

[0219] The item selected in round (r+1) is the best candidate item selected in round (r+1). :

[0220] ;

[0221] Then update the recommendation set:

[0222] ;

[0223] Repeat the above process until the recommendation set contains K items:

[0224] ;

[0225] Finally, the offline diversity rearrangement results for user u are obtained. :

[0226] ;

[0227] In this embodiment, when performing offline diversity rearrangement of the user's initial item recommendation list, the initial item recommendation list is first filtered based on the user's historical interaction item set, removing all items in the list that the user has interacted with, and outputting a de-interaction filtered list. Based on the fusion ranking score of each item in the user's initial item recommendation list, the de-interaction filtered list is sorted in descending order of score, and a preset number of high-scoring items are truncated, outputting a filtered rearrangement candidate pool. The size of the rearrangement candidate pool is larger than the length of the final recommendation list to reduce the computational cost of rearrangement while ensuring the basic relevance of the recommendations. Subsequently, the multi-dimensional similarity between each pair of items in the filtered rearrangement candidate pool is calculated, the visual modal representation of each item in the rearrangement candidate pool is extracted, and the cosine similarity is used to calculate the similarity between each pair of items, outputting the visual similarity between each item in the filtered rearrangement candidate pool. The corresponding calculation formula is as follows: Extract the text modal representation of each item in the rearrangement candidate pool, calculate the pairwise similarity between items using cosine similarity, and output the text similarity between all items in the filtered rearrangement candidate pool. The corresponding calculation formula is as follows: Extract the behavioral modal representation of each item in the rearranged candidate pool, calculate the pairwise similarity between items using cosine similarity, and output the behavioral similarity between each item in the filtered rearranged candidate pool. The corresponding calculation formula is as follows: Based on the visual similarity, text similarity, and behavioral similarity of each item in the filtered and rearranged candidate pool, a weighted sum is performed using preset visual similarity weights, text similarity weights, and behavioral similarity weights to obtain the comprehensive similarity between each pair of items. The corresponding calculation formula is as follows: A comprehensive similarity matrix is ​​constructed based on the combined similarity of all item pairs. After similarity calculation, a greedy, stepwise optimization strategy is used to generate the reordering results. An empty set is initialized as the selected recommendation set. Initially, the selected recommendation set is empty, and the redundancy penalty term for the corresponding candidate items is set to 0. The item with the highest fusion ranking score is directly selected as the first-round optimal candidate item and added to the selected recommendation set. In each subsequent iteration, all unselected items are selected as candidate items from the filtered reordering candidate pool. Based on the user's fusion ranking score for each item and the comprehensive similarity matrix, the reordering selection score for each candidate item is calculated. The reordering selection score is jointly determined by the relevance term and the diversity penalty term, and the corresponding calculation formula is as follows: After calculation, the item with the highest rearrangement selection score is selected as the optimal candidate item for this round, expressed by the following formula: The best candidate item in this round is merged into the selected recommendation set to obtain the updated selected recommendation set. The above iterative process of item selection and set update is repeated. When the number of items in the updated selected recommendation set reaches the preset recommendation length, the iteration is terminated and the updated selected recommendation set is used as the recommendation list after diversity rearrangement.

[0228] Step 103: Divide the user's initial item recommendation list into regions to obtain the head protection zone item set, the boundary replaceable zone item set, and the outer candidate zone item set, and construct swap pairs to output the candidate swap pair set.

[0229] It should be noted that, after obtaining the initial recommendation list R for user u... u Next, the recommendation list is divided into regions. The purpose of region division is to distinguish between high-confidence head recommendations, boundary results that can be safely adjusted, and outer candidate results that may enter the Top-5.

[0230] ;

[0231] in, Indicates the head protection zone. Indicates the boundary replacement area. denoted as the outer candidate region, and r represents the item's ranking position in the recommendation list. Items in the head protection zone are not replaced to ensure the stability of high-confidence recommendation results; items in the boundary replaceable region are candidates to be replaced; items in the outer candidate region are candidates that may enter the boundary replaceable list. , , , This indicates the preset position threshold for each region; This indicates the item whose sorting position is r.

[0232] For user u's initial recommendation list R u First, establish the Top-5 protected areas:

[0233] ;

[0234] Items in the Top-5 protected areas are not replaced to maintain the stability of high-confidence head recommendation results.

[0235] For the Top-10 boundary, set the Top-10 boundary replaceable area:

[0236] ;

[0237] And the outer candidate region of Top-10:

[0238] ;

[0239] in, The items in the text are boundary items that may be replaced. The items in the list are the outer candidate items that may enter the Top-10 boundary position.

[0240] For the Top-20 boundary, set the Top-20 boundary replaceable area:

[0241] ;

[0242] And the outer candidate region of Top-20:

[0243] ;

[0244] in, The items in the list are the Top-20 boundary items that may be replaced. The items in the list are the outer candidate items that may enter the Top-20 boundary position.

[0245] Therefore, the candidate exchange set can be represented as:

[0246] ;

[0247] in, This represents the set of Top-10 candidate swap pairs for user u, generated by pairing candidate items from the outer edge of that level with replaceable items from the boundary, with each pair consisting of a triplet. Represents a group of independent switching units. This represents the set of Top-20 candidate swap pairs for user u, generated by pairing candidate items from the outer edge of that level with replaceable items from the boundary, with each pair consisting of a triplet. Represents a group of independent switching units, i eIndicates the outer candidate entry item, i r Indicates replaceable items at the boundary.

[0248] This candidate swap set serves as the common input for subsequent ItemCF support calculations, secure offline rearrangement, and the construction of teacher positive and negative sample pairs.

[0249] In this embodiment, after obtaining the user's initial item recommendation list, a three-stage region division is performed based on the item's sort position r in the list. The general division rule corresponds to three sets of set calculation formulas, and the item set in the head protection zone satisfies... The set of items in the boundary replaceable zone satisfies The set of items in the outer candidate area satisfies In this embodiment, the head protection zone is set to a fixed Top-5 interval. After substituting the threshold, the formula for calculating the set of items in the head protection zone is as follows: All items within this interval are fixedly included in the head protection zone item set; then, according to multi-level truncation boundaries, corresponding levels of boundary replacement zones and outer candidate zones are delineated. For the Top-10 boundaries, the formula for calculating the boundary replacement zone item set is as follows: The formula for calculating the set of items in the outer candidate area is: For the Top-20 boundary, the formula for calculating the set of replaceable items in the boundary area is as follows: The formula for calculating the set of items in the outer candidate area is: After completing the division of all areas, items from the outer candidate areas of each level are used as outer candidate entry items. Items in the same level boundary replaceable zone are considered boundary replaceable items. The two types of items are paired up in pairs to construct exchange units. The top-10 candidate exchange pairs follow the formula. The generation of the Top-20 tier candidate swap pair set follows the formula After generating and integrating all hierarchical pairing units, a complete set of candidate swap pairs is output. This set serves as the unified input data for subsequent ItemCF support calculation, secure offline boundary swap screening, and construction of teacher positive and negative sample pairs.

[0250] Step 104: Calculate ItemCF support and multi-source exchange evidence for the candidate swap pair set and user-item interaction matrix, and output candidate swap pairs with evidence.

[0251] It should be noted that, firstly, based on the user-item interaction matrix, the set of interacting users and the popularity of each item are statistically analyzed. Then, based on the overlap of interacting users, the ItemCF similarity (Item-based Collaborative Filtering Similarity) between any two items is calculated. Next, the set of items with the user's historical interactions is extracted, and the maximum ItemCF similarity between each candidate item and the user's historical interacting items is calculated, outputting the ItemCF support of each candidate item relative to the corresponding user. Subsequently, all exchange units within the candidate exchange pair set are traversed, and the ItemCF support difference between the outer entering item and the boundary replacing item in each exchange pair is calculated, yielding the ItemCF support difference for the corresponding exchange pair. Simultaneously, combined with the user's fusion ranking score for each item, the multi-dimensional score difference for each candidate exchange pair is calculated. Finally, the ItemCF support difference and the multi-dimensional score difference are used as auxiliary evidence fields and bound to the corresponding candidate exchange pair, completing the multi-source exchange evidence association and outputting candidate exchange pairs with evidence.

[0252] Furthermore, step 104 may include the following sub-steps:

[0253] S41. Calculate the set of interacting users and the popularity of each item based on the user-item interaction matrix;

[0254] S42. Calculate the ItemCF similarity between items based on the set of interacting users corresponding to each item and the item popularity.

[0255] S43. Extract the set of items that the user has interacted with in the past based on the user-item interaction matrix;

[0256] S44. Based on the ItemCF similarity between items and the set of items that the user has interacted with in the past, for any user and its candidate items, calculate the maximum ItemCF similarity between the candidate items and the user's recent historical interaction items, and output the ItemCF support of each candidate item relative to the corresponding user.

[0257] S45. Based on the candidate swap pair set and the ItemCF support of each candidate item relative to the corresponding user, calculate the ItemCF support difference for each candidate swap pair.

[0258] S46. Based on the candidate swap pair set and the user's fusion ranking score for each item, calculate the multi-dimensional score difference for each candidate swap pair.

[0259] S47. Bind the ItemCF support difference and multi-dimensional score difference of each candidate swap pair to the corresponding candidate swap pair to complete the multi-source evidence association, form auxiliary evidence fields for interpreting diagnosis, teacher export and dynamic teacher generation process parameter weight adjustment, and output candidate swap pairs with evidence.

[0260] It should be noted that for any candidate swap pair in the candidate swap set, the ItemCF support is calculated first. Based on the training interaction data, the set of interacting users corresponding to each item is counted. Let U be the set of interacting users for item a. a The set of users interacting with item b is U. b Then the ItemCF similarity between item a and item b is... for:

[0261] ;

[0262] in, This indicates the popularity of item A. Let represent the popularity of item b. For user u, let the set of items they have recently interacted with be . Then, the ItemCF support of candidate item i for user u is defined as the maximum collaborative similarity between the candidate item and the user's recent historical items.

[0263] ;

[0264] in, This represents the set of items that user u has recently interacted with. It includes all items that the user has interacted with within a preset recent time window and serves as the reference sample set for calculating ItemCF support. represents the ItemCF support of candidate item i relative to user u, which is a scalar value. The value is the maximum ItemCF similarity between the candidate item and the user's recent historical interaction items. c represents the unified identifier of a single candidate exchange pair, which represents a complete boundary exchange unit containing three core elements: target user, externally entering candidate item, and boundary item to be replaced.

[0265] Furthermore, for a candidate swap pair c=(u,entry,replace), the ItemCF support difference between the candidate item and the item to be replaced is calculated.

[0266] ;

[0267] Here, `entry` represents the outermost candidate item in the candidate swap pair, originating from the outer candidate area item set, and is the item to be moved into the boundary region during the boundary swap operation. `replace` represents the boundary item to be replaced in the candidate swap pair, originating from the boundary replaceable area item set, and is the item to be moved out of the boundary region during the boundary swap operation. This represents the difference in ItemCF support for candidate swap pair c. It is a scalar value, and its value is the difference between the ItemCF support of the incoming item and the ItemCF support of the item to be replaced. It is used to quantify the net change in collaborative support brought about by the swap operation.

[0268] when Greater than or equal to the preset threshold At that time, it was assumed that the candidate item had stronger collaborative filtering support than the item to be replaced. (Except for...) In addition, it can also calculate multi-source evidence such as teacher score difference, basic ranking score difference, diversity score difference, integration score difference, and historical interaction support difference. It is mainly used to record the changes in multi-source scores before and after the exchange, and plays a role in interpretation, diagnosis, and teacher supervision derivation, and affects parameter weights in the dynamic teacher process.

[0269] ;

[0270] in, This represents the difference in teacher scores corresponding to the candidate swap pairs. It is a scalar value, taking the value as the difference between the teacher ranking scores of the incoming item and the item to be replaced. It is used to measure the ranking gain from the teacher's perspective before and after the swap. This represents the difference in base ranking scores for the candidate swap pairs. It is a scalar value, taking the value as the difference between the base ranking scores of the incoming item and the item to be swapped. It is used to measure the change in scores in the base ranking dimension before and after the swap. This represents the difference in fusion scores corresponding to candidate swap pairs. It is a scalar value, taking the value as the difference in fusion ranking scores between the incoming item and the item to be replaced. It is used to measure the change in scores across the overall ranking dimension before and after the swap. This represents the teacher ranking score for the boundary item to be replaced corresponding to user u. It is obtained from the teacher recommendation model and is used in conjunction with the teacher score of the entering item to calculate the change in teacher-dimensional score before and after the swap operation. This represents the teacher ranking score for the boundary item to be replaced corresponding to user u. It is obtained from the teacher recommendation model and is used in conjunction with the teacher score of the entering item to calculate the change in teacher-dimensional score before and after the swap operation. The base ranking score for user u entering candidate item entries is calculated from the matching degree between the final user representation and the final item representation. It represents the basic matching degree between the two in the comprehensive feature space and is the core input data for calculating the difference in base ranking scores. This represents the base sort score of the boundary item to be replaced corresponding to user u. It is used in conjunction with the base score of the incoming item to calculate the change in the base sort dimension score before and after the swap operation. This represents the fusion ranking score of the candidate item entry corresponding to user u. It is obtained by weighted fusion of the base ranking score, shared score, and private score. It is the core indicator for measuring the overall recommendation priority of items and also the core input data for calculating the fusion score difference. This represents the fusion sort score of the boundary item to be replaced corresponding to user u, which is used in conjunction with the fusion score of the incoming item to calculate the change in the overall sort dimension score before and after the swap operation.

[0271] By supporting difference and multi-source evidence calculation through ItemCF, candidate swap pairs not only have ranking score information, but also cooperative behavior support, boundary position information and security judgment basis, providing an evidentiary basis for subsequent secure offline reordering.

[0272] After obtaining the candidate swap set and its multi-source evidence, a secure offline rearrangement is performed on the candidate swap pairs. Secure offline rearrangement includes hard filtering, weight calculation, sorting truncation, and secure swapping. Hard filtering is used to eliminate candidate swap pairs that do not meet the security conditions, preventing weak evidence candidate items from entering the boundary positions.

[0273] Hard filtering conditions include: the difference in teacher scores between candidate items and items to be replaced must reach a preset threshold; the decrease in basic ranking scores cannot exceed a preset range; items to be replaced must not be located in the head protection zone; the ItemCF support difference must not be less than a preset threshold; and candidate swap pairs must not violate the constraint on the maximum number of swaps per user.

[0274] ;

[0275] in, This represents the collaborative filtering support margin of candidate swap pair c. Its value corresponds to the difference in ItemCF support for the swap pair. It is a core criterion in the hard filtering process for determining whether a swap operation meets the requirements for collaborative behavior support, and is used to screen swap pairs with sufficient collaborative benefits. This represents the number of boundary swaps performed by user u. It is a non-negative integer used to count the total number of swap operations completed by a single user and to verify whether the maximum number of swaps is met. This represents the maximum number of exchanges a user can make, and is a preset positive integer parameter.

[0276] In this embodiment, the set of interacting users and the popularity of each item are first calculated based on the user-item interaction matrix, where the set of interacting users for item a is denoted as U. a Let U be the set of users interacting with item b. b Item popularity is the total number of users interacting with the corresponding item. Let the popularity of item a be denoted as . The popularity of item b is Based on the set of interacting users and the popularity of each item, calculate the ItemCF similarity between items using the following formula: The numerator is the number of elements in the intersection of the user sets interacting with the two items, i.e., the total number of users who interact with both items simultaneously. The denominator is the square root of the product of the popularity of the two items. Popularity normalization reduces the bias of popular items on similarity calculation, resulting in the corrected collaborative similarity between items. Based on the unique identifier of the target user, the row data of the corresponding user is located in the user-item interaction matrix; all columns in the row corresponding to the item dimension are traversed to filter out items with valid interaction records, and the interaction behavior timestamps corresponding to each item are retrieved; based on the pre-set recent time window length, all items whose interaction time falls within the window range are filtered out, and after deduplication, the set of recent historical interaction items of the user is obtained. ;

[0277] Subsequently, based on the ItemCF similarity between items and the user's recent historical interaction item set, for any user and its candidate items, iterate through all of the user's recent historical interaction items and calculate the ItemCF similarity with each candidate item. The maximum similarity among all similarities is taken as the ItemCF support of the candidate item relative to the corresponding user. The calculation formula is as follows: This invention outputs the ItemCF support of each candidate item relative to the corresponding user. Specifically, it deduplicates and merges all boundary replaceable items and outer candidate items across all levels to obtain all candidate items for the user, i.e., all items that could potentially participate in boundary swaps and items to be swapped or entered. Based on the candidate swap pair set and the ItemCF support of each candidate item relative to the corresponding user, each candidate swap pair is denoted as c=(u,entry,replace). The difference in ItemCF support between the outer entering item and the item to be swapped is calculated to obtain the ItemCF support difference for each candidate swap pair. The calculation formula is as follows: Based on the candidate exchange pair set and the user's fusion ranking score for each item, the multi-dimensional score difference for each candidate exchange pair is calculated, including the teacher score difference, the basic ranking score difference, and the fusion score difference, fully covering the score changes in different ranking dimensions.

[0278] Finally, a unique primary key is assigned to each exchange unit in the candidate exchange pair set. The triple "user identifier + external entry item identifier + replacement boundary item identifier" is used as the unique index for a single exchange pair to ensure that each exchange pair can be accurately located and matched. Subsequently, the ItemCF support, teacher ranking score, basic ranking score, and fusion ranking score corresponding to all candidate items are pre-calculated in batches. These are stored as key-value mapping tables with "user identifier + item identifier" as the key and the corresponding score value as the value, for subsequent batch query calls. After the pre-calculation is completed, the candidate exchange pair set is traversed one by one. Based on the unique index of each exchange pair, the scores of the corresponding entry item and the item to be replaced are extracted. The ItemCF support difference, teacher score difference, basic ranking score difference, and fusion score difference of the exchange pair are calculated in sequence. Then, the data structure of the original candidate exchange pair is expanded by adding two auxiliary evidence fields, ItemCF support difference and multi-dimensional score difference, in addition to the original basic identifier field. The calculated difference values ​​are written into the corresponding fields to complete the multi-source evidence association of a single exchange pair. After all exchange pairs are bound, an integrity check is performed. The evidence fields of each exchange pair are checked one by one to see if they are complete and if the values ​​are within a reasonable range. Exchange units with calculation errors are eliminated. Finally, all exchange units that pass the check and have complete evidence fields are summarized and organized to output a set of candidate exchange pairs with evidence.

[0279] Step 105: Perform secure offline boundary swap filtering and replacement processing on the recommended list after offline diversity rearrangement using candidate swap pairs with evidence and head protection zone item sets, and output an updated recommended list and valid secure swap relationships.

[0280] It should be noted that, firstly, the set of items in the head protection zone is used as a fixed and unadjustable core interval. Items in the corresponding positions in the recommended list after offline diversity re-ranking are locked, ensuring they do not participate in any swapping or replacement operations. Then, all candidate swap pairs with evidence are traversed, and multi-dimensional hard filtering checks are performed sequentially. The check rules include: the item to be replaced does not belong to the head protection zone; the ItemCF support difference is not lower than a preset threshold; the decrease in the basic ranking score does not exceed a preset range; and the corresponding user's cumulative swap count has not reached the maximum swap count limit. Weak evidence swap pairs that do not meet all the check conditions are eliminated, resulting in a set of qualified swap pairs. Next, qualified exchange pairs are sorted in descending order according to their overall benefit. Boundary replacements are carried out group by group from high to low priority. Items to be replaced in the corresponding boundary replaceable area are removed from the recommended list, and items entering from the outer candidate area are added to the corresponding positions. After each replacement is completed, the user's cumulative number of exchanges and the item positions in the recommended list are updated simultaneously until there are no qualified exchange pairs or the maximum number of exchanges per user is reached. All successfully executed exchange units and their position and score changes before and after replacement are recorded throughout the process to form valid and safe exchange relationships. Finally, an updated recommended list and valid and safe exchange relationships are output.

[0281] Furthermore, step 105 may include the following sub-steps:

[0282] S51. Use the head protection zone item set to perform hard filtering of candidate swap pairs with evidence to output multiple safe candidate swap pairs;

[0283] S52. Sort the ItemCF support differences carried by multiple security candidate swap pairs in descending order from largest to smallest, and perform truncation filtering in combination with the preset maximum number of swaps for the user, and output the swap pairs to be executed.

[0284] S53. Use the swap pair to perform the offline diversity rearrangement of the recommended list to swap the order of items, and combine the preset boundary level to perform two-level cascade swaps to output an updated recommended list.

[0285] S54. Extract successfully executed item exchange combinations from the updated recommendation list, organize their relationships, and output valid and safe exchange relationships.

[0286] It should be noted that, for the Top-10 boundary stage, from the candidate exchange set Candidate swaps that meet security conditions are selected. Security conditions include: item i to be replaced... r Not part of the Top-5 protected area P u ; Enter candidate item i e With the item to be replaced i r ItemCF supports difference Not less than the threshold Each user's number of exchanges in the current boundary phase does not exceed a preset limit. Candidate admission and ranking in the static phase are primarily based on... and the aforementioned boundary security rules.

[0287] Each user performs at most one swap during the Top-10 boundary phase. For candidate swap relationships that meet security conditions, they are processed according to the ItemCF support difference. Sort the candidates from largest to smallest, and select the candidate swaps with the highest scores to perform boundary replacement. This selection process can be represented as:

[0288] ;

[0289] in, This represents the optimal swap pair to be executed in the Top-10 boundary phase. This represents the set of candidate swaps for user u that passed the security filtering in the Top-10 boundary stage.

[0290] For the Top-20 boundary stage, from the candidate exchange set The same filtering and sorting method is used, with each user performing at most one swap during the Top-20 boundary phase. The selection process can be represented as:

[0291] ;

[0292] in, This represents the optimal swap pair to be executed in the Top-20 boundary phase. This represents the set of candidate swaps that user u passed the safety filtering during the Top-20 boundary stage. Therefore, a user can generate at most one interpretable swap each at the Top-10 and Top-20 boundaries.

[0293] For any selected candidate swap relationship Its exchange weight can be expressed as:

[0294] ;

[0295] in, This represents the swap weight corresponding to candidate swap pair c, and is a scalar value. Support for differential amplification factor in ItemCF This represents a numerical clipping function. and These represent the lower and upper bounds of the weighted clipping, respectively. By introducing weights, teacher pairs with stronger ItemCF support, more explicit boundary exchange evidence, or higher confidence in safe exchange can have a greater impact on distillation training.

[0296] When performing the replacement, the outer candidate item ie Place the item to be replaced. r The location of the boundary, and the item to be replaced i r Remove or with i e Swap positions. If using... Let r represent the items in user u's sorted list at rank r. Then the swap can be represented as:

[0297] ;

[0298] in, Indicate i r The original sort position, Indicate i e The original sorting position. Since the replacement only occurs within the preset boundary area, and the Top-5 protected area does not participate in the replacement, this process can improve the sorting quality near the Top-10 and Top-20 truncation boundaries while maintaining the stability of the head recommendation.

[0299] In this embodiment, when performing secure offline boundary exchange filtering and replacement processing, a two-level cascaded processing flow of Top-10 and Top-20 is adopted to sequentially complete hard filtering, sorting filtering, position replacement, and relationship organization. First, hard filtering of exchange pairs is performed on candidate exchange pairs with evidence, using the head protection zone item set as an insurmountable constraint boundary. For the Top-10 boundary stage, three security access conditions are verified group by group from the candidate exchange pair subsets of the corresponding level: matching the regional affiliation of the item to be replaced one by one to confirm that it does not belong to the Top-5 head protection zone item set; retrieving the Item bound to the exchange pair. The ItemCF support difference is verified, and its value is confirmed to be no less than the preset ItemCF support difference threshold. The cumulative number of swaps executed by the corresponding user in the current Top-10 boundary stage is counted, and it is confirmed that it does not exceed the preset maximum number of swaps per level. Swap pairs that meet all three conditions are judged as safe candidate swap pairs and added to the corresponding user's Top-10 safe candidate swap set. The candidate swap pair subsets of the Top-20 level are processed using completely consistent verification rules to obtain the corresponding user's Top-20 safe candidate swap set. After hard filtering, all safe candidate swap pairs for both levels are output. After filtering, the safe candidate swap pairs for both levels are sorted in descending order according to their ItemCF support difference, and truncation filtering is performed based on the preset maximum number of swaps per user per level. Only the pair with the largest ItemCF support difference is selected as the swap pair to be executed for each level. The optimal swap pair to be executed for the Top-10 level satisfies the formula... The optimal swap pairs to be executed at the Top-20 level satisfy the formula This ensures that each user generates at most one pair of swaps to be executed at each of the two boundary levels. Then, two-level cascaded swaps are executed in the order of Top-10 boundary first, then Top-20 boundary. For each pair of swaps to be executed, the original ranking positions of the items to be replaced and the original ranking positions of the externally entering items are first located in the recommendation list after offline diversity re-ranking. Then, the swaps are executed according to the formula... Complete the item swap operation between the two levels, moving the outer candidate item to its corresponding boundary position and moving the item to be replaced to the original position of the outer item. The swaps at both levels are performed sequentially without interference. After all replacements are complete, an updated recommendation list is obtained. Simultaneously, the swap weight is calculated for each selected swap pair. The weight calculation formula is... The weights are constrained within a preset upper and lower limit range by a pruning function. ItemCF supports swap pairs with positive differences, and the larger the difference, the higher the corresponding swap weight. Finally, all successfully executed item swap combinations at both boundaries are extracted from the updated recommendation list. Complete information such as the user identifier, the identifiers of the entering and replacing items, the order before and after the swap, and the difference and swap weight supported by ItemCF are compiled for each swap, and a structured, valid, and safe swap relationship is output.

[0300] Step 106: Construct static samples based on valid and secure exchange relationships, and combine the user's fusion ranking score for each item, the updated recommendation list, and the recommendation list after offline diversity re-ranking to perform recommendation inference and output the final recommendation result.

[0301] It should be noted that, firstly, based on the effective and secure exchange relationships, all successfully executed and valid boundary exchange units are extracted. Positive and negative pairing samples are constructed with the incoming item and the item to be replaced in each exchange as the core, and corresponding exchange weights, ItemCF support difference and other attribute labels are bound simultaneously to complete the static sample construction. Then, based on the user's fusion ranking score for each item as the ranking benchmark, combined with the diversity optimization position of the recommendation list after offline diversity re-ranking and the boundary exchange adjustment results of the updated recommendation list, the ranking position of all candidate items is uniformly aligned and consistent. The original ranking of items in the head protection zone, the ranking of items after boundary exchange adjustment, and the relative ranking of items in non-adjusted areas are fixed, completing the complete recommendation reasoning process and outputting the final recommendation result.

[0302] Furthermore, step 106 may include the following sub-steps:

[0303] S61. Perform a comparative analysis using the updated recommendation list and the recommendation list after offline diversity rearrangement, extract the changes in the sorting positions of boundary replacement items and outer entry items, and output a table of sorting differences before and after the swap.

[0304] S62. Based on the effective and secure exchange relationship and the sorting difference table before and after the exchange, distinguish between items entering from the outside and items replaced at the boundary. Combine this with user identifiers to complete positive and negative sample pairing and output static teacher positive and negative sample pairs.

[0305] S63. Based on the ItemCF support difference bound to the static teacher positive and negative sample pairs, calculate the sample pair weights through the truncation function, and add auxiliary fields such as original sorting position, score difference, and exchange rule identifier to output weighted static teacher positive and negative sample pairs.

[0306] S64. During the model training process, according to the preset round interval, based on the sorting results of the current stage's optimal model state, select items with low sorting discrimination at the preset cutoff boundary to generate dynamic difficult example teacher pairs.

[0307] S65. Calculate the static distillation loss based on weighted static teacher positive and negative sample pairs, calculate the dynamic distillation loss based on dynamic difficult example teacher pairs, and combine the basic recommendation loss, embedding regularization loss and contrastive learning loss to form the total training loss.

[0308] S66. Perform backpropagation based on the total training loss, iteratively update the model parameters, and output the target model parameters after training is completed.

[0309] S67. Based on the parameters of the trained target model, construct a target recommendation model;

[0310] S68. Use a target recommendation model to calculate the user's target fusion ranking score for each item, remove the items that have been interacted with, and then extract the first preset number of items in descending order of score, and output the final recommendation result.

[0311] It should be noted that after the boundary interpretable replacement filters out candidate swap pairs that meet the safe swap conditions, the swap process is further transformed into a boundary ranking supervision signal that can be used to train the recommendation model. For any safe swap relationship c=(u,entry,replace), the item entry that enters the boundary position from the outer candidate area is taken as the teacher positive sample, and the item replace that was originally in the boundary replaceable area and was replaced is taken as the teacher negative sample.

[0312] ;

[0313] The teacher's interpretation is: Under the judgment of the secure offline rearrangement module, for user u, the outer candidate items... Compared to items within the boundary This should be recommended for higher rankings. Unlike random negative sampling or ordinary pairwise training samples, the teacher positive and negative sample pairs in this invention have a clear Top-10 boundary source, which can more directly express the local ranking and error correction relationship at the boundary.

[0314] Furthermore, after obtaining the static set of teacher positive and negative sample pairs derived from the offline safe rearrangement module, it is used for recommender model training. For each teacher pair:

[0315] ;

[0316] The student recommendation model calculates the number of positive sample items for user u. and negative sample items Shared-private fusion sorting score:

[0317] ;

[0318] in, The positive sample item fusion ranking score output by the student model; The negative sample item fusion ranking score output by the student model;

[0319] The student model's ranking interval on the teacher pair for:

[0320] ;

[0321] To enable the student model to learn the boundary ordering relationships in the teacher pair, pairwise distillation loss is employed. :

[0322] ;

[0323] in, This indicates the teacher's weighting. This represents the softplus activation function. When the score of positive sample items output by the student model is already higher than the score of negative sample items, the distillation loss is small; when the student model still ranks positive sample items after negative sample items, the distillation loss is large, thus prompting the model to correct boundary sorting errors.

[0324] For static teachers, set P static Its distillation loss is:

[0325] ;

[0326] In addition to static teacher pairs, a dynamic set P of teacher pairs is also generated during the training process based on the current student model state. dynamic The dynamic teacher first sorts candidate items according to their current merged teacher scores, then constructs candidate pairs from the vicinity of the Top-10 truncation boundary and within the Top-20 relevance range. The teacher fusion interval is specified in the original text. It can be used for candidate interval filtering and candidate utility calculation, base interval The difference between the teacher integration interval and the student integration interval can be used to identify teacher-student difficulty cases that students currently find difficult to distinguish.

[0327] In the dynamic teacher pair generation, let candidate positive items be i and candidate negative items be j, then we can calculate:

[0328] ;

[0329] in, This represents the teacher ranking score for user u on item i, as output by the teacher model. This represents the base ranking score of user u for item i, as output by the base model. Indicates the teacher sorting interval. Indicates the sorting interval of the basic model. This represents the teacher-student interval difference. Furthermore, the discount payoff can be calculated based on the difference in the ranking position of candidate item pairs. :

[0330] ;

[0331] in, and These represent the positions of item i and item j in the current candidate sorting list, respectively. Dynamic teacher pairs are selected preferentially from candidate item pairs that satisfy the minimum teacher interval, minimum sorting interval, and boundary window constraints.

[0332] For dynamic teachers, set P dynamic Its distillation loss is:

[0333] ;

[0334] The static-dynamic mixed distillation loss is:

[0335] ;

[0336] in, This represents the static teacher's weight on the distillation process. This represents the dynamic teacher's weight for distillation.

[0337] The total training loss of the recommendation model can be expressed as:

[0338] ;

[0339] in, This represents the basic recommendation ranking loss. This represents the embedding regularization loss. Indicates the contrast learning loss. and These represent the weights of the regularization loss and the contrastive learning loss, respectively.

[0340] Through the aforementioned static-dynamic hybrid distillation loss, the teacher's ranking knowledge is fed back into the student recommendation model in the form of gradient constraints. Specifically, the student model is initialized with the parameters of the multimodal base recommendation model, and during training, the base ranking loss, regularization loss, contrastive learning loss, and teacher-applied distillation loss are simultaneously optimized. The teacher-applied distillation loss constrains the ranking interval between positive and negative samples, allowing the student model to gradually approximate the teacher model's ranking preferences on boundary samples. After training, the offline safe re-ranking module does not participate in online inference; ultimately, only the distillation-enhanced student recommendation model is retained for recommendation ranking.

[0341] In this embodiment, a position-by-position comparison analysis is performed using the updated recommendation list and the offline diversity-rearranged recommendation list. It iterates through all item positions in the Top-10 and Top-20 boundary regions, matching the item identifiers at each position. It extracts boundary replacement items and external entry items that have undergone position swaps, records the ranking position and corresponding fusion ranking score for each item before and after the swap, and generates a ranking difference table before and after the swap. Based on the effective safe swap relationships and the ranking difference table before and after the swap, external entry items in each effective swap are used as teacher positive samples, and replaced boundary items are used as teacher negative samples. These are paired with corresponding user identifiers to complete the positive and negative sample pairing. The structure of the static teacher positive and negative sample pairs satisfies the formula... Based on the ItemCF support difference bound to each static teacher positive and negative sample pair, the corresponding sample pair weights are calculated using a pruning function. Simultaneously, auxiliary fields such as original sorting position, multi-dimensional score difference, and exchange rule identifier are added to each sample pair, outputting weighted static teacher positive and negative sample pairs. During model training, at preset intervals, based on the sorting output of the current optimal student model, the item sorting intervals for the teacher model and the base model are calculated separately, where the teacher sorting interval satisfies... The basic model sorting interval satisfies Further calculation of teacher-student interval differences The discount is calculated by combining the ranking of the two items. From the vicinity of the cutoff boundary between Top-10 and Top-20, low-discriminative items that satisfy the minimum teacher interval, minimum sorting interval, and are within the boundary window constraints are preferentially selected for pairing, generating a dynamic set of difficult example teacher pairs P. dynamic .

[0342] After constructing the two types of samples, the static distillation loss is calculated based on the weighted static teacher positive and negative sample pairs. First, the student model calculates the shared-private fusion ranking score for positive and negative samples in each group, using the following formula: The ranking interval of the student model on this sample pair is obtained. Then, the pairwise distillation loss of a single sample is calculated using the softplus function. The static distillation loss is obtained by taking the average value of all static teacher samples and processing the dynamic difficult teacher pairs using the same loss calculation logic to obtain the dynamic distillation loss. Subsequently, the two types of distillation losses are weighted and fused to obtain the static-dynamic hybrid distillation loss. This loss is then combined with the basic recommendation ranking loss, embedding regularization loss, and contrastive learning loss to form the total training loss.

[0343] Finally, backpropagation is performed based on the total training loss. The gradients of all trainable parameters are calculated layer by layer using chained differentiation. After gradient clipping to avoid gradient explosion, the optimizer iteratively updates the parameters. After each training round, a ranking evaluation metric is calculated on the validation set. When the improvement of the validation metric is lower than the preset convergence threshold for multiple consecutive rounds, or when the number of training rounds reaches the set upper limit, the training process is terminated, and all trainable parameters of the current model are exported and saved, i.e., the target model parameters after training. In the online inference stage, the target model parameters are loaded to complete the construction of the target recommendation model (student model). For the user to be recommended, the user-side features and the multimodal features of all candidate items are input. The target recommendation model completes the multimodal feature encoding, shared-private representation alignment, and score fusion calculation, and outputs the target fusion ranking score of the user for each candidate item. The historical interaction item set of the corresponding user is retrieved, and all items that have generated interaction behavior are removed from the candidate items. The remaining items are sorted in descending order according to the target fusion ranking score, and a preset number of items with the highest ranking are selected to generate the final recommendation result and output it.

[0344] It is worth mentioning that this invention addresses the Top-10 boundary optimization problem in multimodal recommendation scenarios where user interests are more concentrated and the diversity space is small. It proposes a multimodal recommendation scheme based on offline diversity reordering and ItemCF boundary teacher distillation. This scheme first generates a fused ranking score using a multimodal graph recommendation model. Then, it stabilizes and optimizes the Top-10 boundary through head protection, boundary candidate mining, ItemCF support calculation, multi-source evidence filtering, and restricted secure swapping. Subsequently, the secure swapping relationships are precipitated as teacher positive and negative sample pairs, and then fed back into the recommendation model training process through static-dynamic hybrid knowledge distillation.

[0345] Through the above design, this invention can discover candidate items with strong collaborative support outside the boundary without destroying the high-confidence head recommendation results, thereby improving the local ranking quality at the boundary of the recommendation list. At the same time, offline re-ranking knowledge is no longer just a post-processing result in the inference stage, but can be transformed into training supervision signals that the model can learn, thereby improving the Top-10 boundary ranking ability, stability and interpretability of the multimodal recommendation model itself.

[0346] The experimental results are shown in Table 1. Experimental verification confirms that this invention has achieved certain results. Here, Ours represents the improved method based on the MGCN-based ranking model, incorporating the offline multimodal rearrangement, Top-K boundary interpretability replacement, and static-dynamic hybrid distillation module proposed in this invention.

[0347] Table 1 Experimental Results

[0348]

[0349] For comparison of technical effectiveness, existing technologies can be referenced. Multimodal recommendation methods are beginning to incorporate auxiliary information such as visual features and textual descriptions of items. These methods typically construct user-item interaction graphs based on user-item interaction data and item similarity graphs based on visual or textual features. They then obtain multi-source representations of users and items through graph propagation, feature fusion, or attention mechanisms. Compared to methods relying solely on interaction matrices, multimodal recommendation enhances the semantic expressiveness of items and improves the model's ability to characterize user interests. However, existing multimodal recommendation methods still have several shortcomings:

[0350] Existing diversity ranking methods cannot solve the problem of poor recall in Top-10 recommendations in domains where user interests are more concentrated and diversity space is small. Traditional diversity ranking usually only performs redundancy removal based on the multimodal similarity between items, which can easily lead to the accidental deletion or downgrading of similar items that are highly relevant to the user's true interests, resulting in limited or even decreased recall improvement.

[0351] Existing offline re-ranking methods typically adjust the initial recommendation list directly based on diversity, similarity, rule scores, or manual strategies, lacking a safe swapping mechanism for the Top-10 boundaries. If the re-ranking process does not distinguish between high-confidence head positions and replaceable boundary positions, it may disrupt the already relatively stable head items in the original recommendation results, causing unnecessary fluctuations in the recommendation results.

[0352] Existing offline re-ranking methods often exist only as a post-processing module in the inference stage, and the effective boundary exchange information generated during the re-ranking process is difficult to feed back to the recommendation model itself. Even if offline re-ranking can discover that some candidate items outside the boundary are better than items inside the boundary, if this exchange relationship cannot be transformed into positive and negative sample supervision during the training process, the recommendation model's own Top-10 boundary ranking ability will still be difficult to continuously improve.

[0353] Traditional knowledge distillation methods typically require the student model to fit the score, soft label, or overall ranking distribution of the teacher model. For Top-10 boundary scenarios, simply fitting scores may not be sufficient to fully express the pairwise substitution relationship of "which outer candidate should exceed which inner candidate," especially when the score difference between candidate items near the boundary is small. The lack of teacher positive and negative sample pairs for boundary swapping limits the model's ability to correct local ranking errors.

[0354] To address the aforementioned issues, this invention proposes a multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation. This method is suitable for multimodal recommendation scenarios where user interests are more concentrated and the diversity space is small. It aims to solve problems in existing multimodal recommendation methods, such as high homogeneity of recommendation results, insufficient utilization of boundary candidate items, and difficulty in effectively feeding offline rearrangement signals back to the model training process.

[0355] like Figure 2 As shown, the core idea of ​​this invention is as follows: First, a multimodal basic ranking model is constructed, which uses user-item interaction information, item visual information, and item text information to learn a comprehensive representation of users and items, and generates a fusion ranking score for users on candidate items. Then, offline diversity re-ranking is performed on the basic ranking result, so that the recommended candidate list reduces content redundancy while maintaining relevance. Furthermore, an ItemCF boundary teacher is constructed on the list after diversity re-ranking, and a safe exchange mechanism is used to determine whether outer candidate items should enter the Top-K boundary region, thereby generating teacher positive and negative sample pairs. Finally, the teacher signal is used for static-dynamic hybrid knowledge distillation, so that the target recommendation model can learn the preference information provided by offline re-ranking and ItemCF boundary judgment, thereby improving the accuracy, diversity, and security of the final recommendation result.

[0356] The offline diversity reordering module performs post-processing optimization on the candidate ranking list output by the multimodal basic ranking module, enabling the recommendation list to maintain high relevance while possessing better diversity. The ItemCF boundary teacher construction module constructs boundary teacher signals on the candidate list after offline diversity reordering, determining whether certain candidate items outside the Top-K but with strong collaborative behavior support should enter the recommendation boundary region. In this way, the invention introduces the relatively stable and interpretable collaborative behavior signals from the traditional ItemCF method into the boundary judgment of the deep multimodal recommendation model, thereby improving the reliability of boundary candidate item selection. The secure offline exchange module controls the exchange process of ItemCF boundary teachers, avoiding unreliable replacements introduced by offline reordering or ItemCF signals that could affect the quality of the recommendation list. This module makes the offline reordering process more controllable, avoiding a significant decrease in relevance due to simply pursuing diversity. The teacher sample construction module transforms the secure offline exchange results into teacher positive and negative sample pairs that can be used for model training. Through this module, the offline reordering signal, which originally only existed as a post-processing result, is transformed into explicit training samples, enabling the recommendation model to learn the preference relationships reflected in offline reordering and ItemCF boundary judgment during training. The static-dynamic hybrid knowledge distillation module feeds back offline-constructed teacher signals to the target recommendation model, enhancing its ranking capabilities. Through this module, the target recommendation model learns not only user preference information from historical interactions but also boundary ranking knowledge provided by offline diversity re-ranking and ItemCF boundary teachers. The trained target recommendation model calculates the user's fused ranking score for candidate items based on user and item representations, filters items the user has already interacted with, and finally outputs the recommendation result.

[0357] In this embodiment of the invention, a multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation is provided. The method obtains a user-item interaction matrix, an item visual preprocessing feature matrix, and an item text preprocessing feature matrix. Based on these matrices, an initial item recommendation list for the user is constructed. Offline diversity rearrangement is performed on the initial item recommendation list, outputting a revised recommendation list. The initial item recommendation list is then divided into regions: a head protection zone item set, a boundary replaceable zone item set, and an outer candidate zone item set. Exchange pairs are constructed, outputting a candidate exchange pair set. ItemCF support and multi-source exchange evidence are calculated on the candidate exchange pair set and the user-item interaction matrix, outputting candidate exchange pairs with evidence. The candidate exchange pairs with evidence and the head protection zone item set are used to perform secure offline boundary exchange filtering and replacement processing on the revised recommendation list, outputting an updated recommendation list and valid secure exchange relationships. Static samples are constructed based on the valid secure exchange relationships, combined with user feedback. The invention performs recommendation inference based on the fusion ranking scores of various items, the updated recommendation list, and the recommendation list after offline diversity re-ranking, and outputs the final recommendation result. Based on the above scheme, this invention divides the initial recommendation list into regions, placing high-confidence items at the top of the ranking into the head protection zone and not participating in the replacement operation. This mechanism can ensure the stability of highly relevant head recommendation results. At the same time, candidate exchange pairs are constructed for the boundary replaceable area and the outer candidate area. The ItemCF support is calculated by combining the user-item interaction matrix as the basis for exchange judgment. The degree of matching between candidate items and user interests can be quantified by using collaborative behavior signals. Only outer items with higher collaborative support are selected to replace boundary items. On this basis, the safe offline boundary exchange screening and replacement processing is carried out only in the boundary area with limited adjustments. This can optimize the diversity performance of the recommendation list without destroying the overall relevance of the recommendation. Finally, the effective safe exchange relationship is transformed into static samples and integrated into the recommendation inference process. The ranking optimization logic at the boundary can be transmitted to the generation stage of the final recommendation result, further improving the rationality of the ranking of candidate items at the boundary position, thereby improving the accuracy of the final recommendation result.

[0358] Please see Figure 3 , Figure 3 This is a structural block diagram of a multimodal recommendation system based on offline diversity rearrangement and ItemCF boundary teacher distillation, provided in Embodiment 2 of the present invention.

[0359] This invention provides a multimodal recommender system based on offline diversity rearrangement and ItemCF boundary teacher distillation, comprising:

[0360] The acquisition module 301 is used to acquire the user-item interaction matrix, the item visual preprocessing feature matrix, and the item text preprocessing feature matrix, and to construct the user's initial item recommendation list based on the user-item interaction matrix, the item visual preprocessing feature matrix, and the item text preprocessing feature matrix.

[0361] The rearrangement module 302 is used to perform offline diversity rearrangement on the user's initial item recommendation list and output the recommendation list after offline diversity rearrangement.

[0362] Module 303 is used to divide the user's initial item recommendation list into regions, obtain the head protection zone item set, the boundary replaceable zone item set, and the outer candidate zone item set, and construct swap pairs, outputting a candidate swap pair set;

[0363] Calculation module 304 is used to calculate ItemCF support and multi-source exchange evidence for the candidate exchange pair set and user-item interaction matrix, and output candidate exchange pairs with evidence.

[0364] Replacement module 305 is used to perform secure offline boundary exchange filtering and replacement processing on the recommended list after offline diversity rearrangement using candidate exchange pairs with evidence and head protection zone item sets, and outputs an updated recommended list and valid secure exchange relationships.

[0365] The recommendation module 306 is used to construct static samples based on valid and secure exchange relationships, and to perform recommendation inference by combining the user's fusion ranking score for each item, the updated recommendation list, and the recommendation list after offline diversity re-ranking, and output the final recommendation result.

[0366] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the system and its modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0367] This invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program; when the computer program is executed by the processor, the processor performs the steps of the multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation as described in the above embodiments.

[0368] This invention also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation as described in the above embodiments.

[0369] This invention also provides a computer program product, including a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, wherein when the program instructions are executed by a computer, the computer performs the steps of the multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation as described in the above embodiments.

[0370] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0371] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0372] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0373] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0374] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation, characterized in that, include: Obtain the user-item interaction matrix, the item visual preprocessing feature matrix, and the item text preprocessing feature matrix. Based on these matrixes, construct an initial item recommendation list for the user, including: Construct a user-item interaction graph based on the user-item interaction matrix; The K-nearest neighbor algorithm is used to construct visual modal item similarity maps and text modal item similarity maps based on the item visual preprocessing feature matrix and the item text preprocessing feature matrix; Initialize the user embedding matrix and the item embedding matrix, and concatenate the user embedding matrix and the item embedding matrix to output the initial node representation of the graph; Based on the initial node representation of the graph and the user-item interaction graph, multi-layer graph propagation is performed to obtain multiple propagation layer results; The results from multiple propagation layers are aggregated to output a behavioral-side collaborative representation; Linear mapping and gated fusion processing are performed on the visual preprocessing feature matrix of the object to output a basic representation of the visual object; graph propagation is performed by combining the basic representation of the visual object and the visual modal object similarity map to output a visual modal representation; Linear mapping and gated fusion processing are performed on the preprocessed feature matrix of the item text to output the basic representation of the text item; graph propagation is performed by combining the basic representation of the text item and the text modal item similarity graph to output the text modal representation; Attention-weighted calculations are performed on the visual modality representation and the text modality representation to output a shared modality representation; Perform a difference operation on the visual modality representation and the shared modality representation to output a visual private representation; Perform a difference operation on the text modal representation and the shared modal representation to output the text private representation; Residual fusion is performed on the behavior-side collaborative representation, the shared modality representation, the visual private representation, and the text private representation to output the final user representation and the final item representation; Calculate the basic ranking score based on the end-user representation and the final item representation; Extract the set of items that the user has historically interacted with based on the user-item interaction matrix; Based on the visual item representation, the text item representation, and the user's historical interaction item set, a shared view and a private view are constructed. Based on the shared view and the private view, the shared score and the private score are calculated, and a weighted fusion calculation is completed by combining the preset shared score weight and the preset private score weight, and the user's fusion ranking score for each item is output. Sort the items in descending order based on the user's combined ranking scores for each item, and output the user's initial item recommendation list; Perform offline diversity rearrangement on the user's initial item recommendation list and output the offline diversity rearranged recommendation list; The user's initial item recommendation list is divided into regions to obtain the head protection zone item set, the boundary replaceable zone item set, and the outer candidate zone item set. Exchange pairs are constructed and a candidate exchange pair set is output. ItemCF support and multi-source exchange evidence are calculated for the candidate exchange pair set and the user-item interaction matrix, and candidate exchange pairs with evidence are output. The offline boundary exchange filtering and replacement process is performed on the recommended list after offline diversity rearrangement using the candidate exchange pairs with evidence and the head protection zone item set, and an updated recommended list and effective secure exchange relationships are output. Static samples are constructed based on the effective and secure exchange relationship, and recommendation inference is performed by combining the user's fusion ranking score for each item, the updated recommendation list, and the recommendation list after offline diversity re-ranking, and the final recommendation result is output.

2. The multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation according to claim 1, characterized in that, The process of performing offline diversity re-ranking on the user's initial item recommendation list and outputting the offline diversity re-ranked recommendation list includes: The user's initial item recommendation list is filtered by items based on the user's historical interactions, and an interaction-free filtered list is output. Based on the fusion ranking score of each item in the user's initial item recommendation list, high-scoring items are extracted from the interaction-free filtering list, and the filtered rearranged candidate pool is output. Cosine similarity is calculated for the visual modal representations of multiple items in the filtered rearranged candidate pool, and the visual similarity between each item in the filtered rearranged candidate pool is output. Cosine similarity is calculated for the text modal representations of each item in the filtered rearranged candidate pool, and the text similarity between all items in the filtered rearranged candidate pool is output. Cosine similarity is calculated for the behavioral modal representations of each item in the filtered rearranged candidate pool, and the behavioral similarity between each item in the filtered rearranged candidate pool is output. A comprehensive similarity matrix is ​​constructed based on the visual similarity, text similarity, and behavioral similarity of each item in the filtered rearranged candidate pool; Initialize an empty set as the selected recommendation set; Multiple candidate items are selected from the filtered rearranged candidate pool; Based on the user's fusion ranking score for each item and the comprehensive similarity matrix, calculate the re-ranking selection score for multiple candidate items; The item with the highest rearrangement selection score is selected as the best candidate item in this round. The best candidate items in this round are incorporated into the selected recommendation set to obtain the updated selected recommendation set; When the number of items in the updated selected recommendation set reaches the preset recommendation length, the updated selected recommendation set is used as the recommendation list after diversity rearrangement.

3. The multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation according to claim 1, characterized in that, The step involves calculating ItemCF support and multi-source exchange evidence for the candidate exchange pair set and the user-item interaction matrix, outputting candidate exchange pairs with evidence, including: Based on the user-item interaction matrix, the set of interacting users and the popularity of each item are calculated. Calculate the ItemCF similarity between each item based on the set of interacting users and the item popularity corresponding to each item. Extract the set of items that the user has historically interacted with based on the user-item interaction matrix; Based on the ItemCF similarity between the items and the set of items the user has interacted with in the past, for any user and its candidate items, calculate the maximum ItemCF similarity between the candidate items and the user's recent historical interaction items, and output the ItemCF support of each candidate item relative to the corresponding user. Based on the set of candidate swap pairs and the ItemCF support of each candidate item relative to the corresponding user, calculate the ItemCF support difference for each set of candidate swap pairs; Based on the candidate exchange pair set and the user's fusion ranking score for each item, calculate the multi-dimensional score difference for each candidate exchange pair. The ItemCF support difference and multi-dimensional score difference of each candidate swap pair are bound to the corresponding candidate swap pair to complete the multi-source evidence association, forming auxiliary evidence fields for interpreting diagnosis, teacher pair export and dynamic teacher generation process parameter weight adjustment, and outputting candidate swap pairs with evidence.

4. The multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation according to claim 1, characterized in that, The process involves using the candidate swap pairs with evidence and the head protection zone item set to perform secure offline boundary swap filtering and replacement on the recommended list after offline diversity rearrangement, outputting an updated recommended list and valid secure swap relationships, including: The head protection zone item set is used to perform hard filtering on the candidate swap pairs with evidence, and multiple safe candidate swap pairs are output. The ItemCF support difference carried by the multiple security candidate swap pairs is sorted in descending order from largest to smallest, and the truncation and filtering are completed in combination with the preset maximum number of swaps for the user, and the swap pairs to be executed are output. The order of items in the offline diversity rearranged recommendation list is swapped using the swap pair to be executed. A two-level cascade swap is performed in combination with the preset boundary level to output an updated recommendation list. Extract successfully executed item exchange combinations from the updated recommendation list, organize their relationships, and output valid and secure exchange relationships.

5. The multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation according to claim 1, characterized in that, The process involves constructing static samples based on the effective secure exchange relationship, combining the user's fusion ranking score for each item, the updated recommendation list, and the offline diversity-rearranged recommendation list to perform recommendation inference, and outputting the final recommendation result, including: A comparative analysis is performed using the updated recommendation list and the offline diversity rearranged recommendation list to extract the changes in the sorting positions of boundary replacement items and outer entry items, and a sorting difference table before and after the swap is output. Based on the effective and secure exchange relationship and the sorting difference table before and after the exchange, items entering from the outside and items replaced at the boundary are distinguished. Positive and negative sample pairing is completed with user identifiers, and static teacher positive and negative sample pairs are output. Based on the ItemCF support difference bound to the static teacher positive and negative sample pairs, the weight of the sample pairs is calculated through the truncation function, and auxiliary fields such as original sorting position, score difference, and exchange rule identifier are added to output weighted static teacher positive and negative sample pairs; During model training, according to a preset round interval, based on the sorting results of the current optimal model state, items with low sorting discrimination at the preset cutoff boundary are selected to generate dynamic difficult example teacher pairs. The static distillation loss is calculated based on the weighted static teacher positive and negative sample pairs, the dynamic distillation loss is calculated based on the dynamic hard example teacher pairs, and the total training loss is formed by combining the basic recommendation loss, the embedding regularization loss and the contrastive learning loss. Backpropagation is performed based on the total training loss to iteratively update the model parameters and output the target model parameters after training is completed. Based on the parameters of the trained target model, construct a target recommendation model; The target recommendation model is used to calculate the user's target fusion ranking score for each item. After removing the items that have been interacted with, the first preset number of items are truncated in descending order of score, and the final recommendation result is output.

6. A multimodal recommendation system based on offline diversity rearrangement and ItemCF boundary teacher distillation, applied to the multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation as described in claim 1, characterized in that, include: The acquisition module is used to acquire the user-item interaction matrix, the item visual preprocessing feature matrix, and the item text preprocessing feature matrix, and to construct the user's initial item recommendation list based on the user-item interaction matrix, the item visual preprocessing feature matrix, and the item text preprocessing feature matrix; The reordering module is used to perform offline diversity reordering on the user's initial item recommendation list and output the recommendation list after offline diversity reordering. The construction module is used to divide the user's initial item recommendation list into regions, obtain the head protection area item set, the boundary replaceable area item set, and the outer candidate area item set, and construct swap pairs, outputting a candidate swap pair set; The calculation module is used to calculate ItemCF support and multi-source exchange evidence for the candidate exchange pair set and the user-item interaction matrix, and output candidate exchange pairs with evidence. The replacement module is used to perform secure offline boundary exchange filtering and replacement processing on the recommended list after offline diversity rearrangement using the candidate exchange pairs with evidence and the head protection zone item set, and output an updated recommended list and valid secure exchange relationships; The recommendation module is used to construct static samples based on the effective secure exchange relationship, and to perform recommendation inference by combining the user's fusion ranking score for each item, the updated recommendation list, and the recommendation list after offline diversity reordering, and output the final recommendation result.

7. An electronic device, characterized in that, The system includes a memory and a processor, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of the multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation as described in any one of claims 1-5.

9. A computer program product, characterized in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, wherein, when the program instructions are executed by a computer, the computer performs the steps of the multimodal recommendation method based on offline diversity rearrangement and ItemCF boundary teacher distillation as described in any one of claims 1-5.

Images