Collaborative filtering recommendation method fusing large language model semantic information and contrast learning alignment mechanism

Abstract

This invention discloses a collaborative filtering recommendation method that integrates semantic information from a large language model with a contrastive learning alignment mechanism, comprising the following steps: Data preprocessing: constructing a user-item bipartite graph based on e-commerce interaction data; Collaborative representation learning: obtaining user-item collaborative embeddings using LightGCN; Semantic representation generation: obtaining semantic embeddings by offline encoding of item text using MiniLM; Top-p Gate semantic gating fusion: performing dimensional filtering on the semantic embeddings and fusing them with the collaborative embeddings; Contrastive alignment training: constructing semantic-collaborative alignment and contrastive learning constraints; Joint optimization and outputting Top-K recommendation results. This invention employs approximate singular value decomposition and embedding space structure correlation for data augmentation, effectively mitigating noise problems. Furthermore, the data augmentation strategy of this invention only needs to be executed once, greatly reducing time overhead and ultimately improving the accuracy and robustness of the recommendations.

Description

Technical Field

[0001] This invention belongs to the field of recommendation systems and artificial intelligence technology, and more specifically, relates to a collaborative filtering recommendation method that integrates semantic information from a large language model with a contrastive learning alignment mechanism. Background Technology

[0002] With the rapid development of e-commerce, digital content platforms, and online service systems, the scale of goods and information accessible to users continues to expand. Traditional information distribution methods based on manual screening or fixed rule matching are no longer sufficient to meet personalized needs. To alleviate information overload and improve user experience, personalized recommendation systems have gradually become a core infrastructure of internet platforms, playing a crucial role in various scenarios such as e-commerce shopping, content distribution, advertising, information retrieval, and social media.

[0003] Among numerous recommendation technologies, collaborative filtering is widely used due to its weak dependence on domain features, strong transferability, and relatively simple modeling approach. Traditional collaborative filtering mainly infers potential preferences by mining historical user-item interactions, such as using matrix factorization or neighborhood similarity methods for rating prediction. However, in real e-commerce environments, user behavior typically exhibits a significant long-tail distribution: a few popular products have a large number of interaction records, while many long-tail products have only a few or even zero interactions, resulting in a highly sparse interaction matrix. This sparsity problem makes it difficult for traditional collaborative filtering to learn stable and reliable user and item representations, thus affecting the recall coverage and ranking quality of recommendations, especially in cold start and new product scenarios.

[0004] To enhance the ability to model higher-order relationships, graph neural networks have been introduced into recommender systems in recent years. These networks capture more complex structural information by constructing user-item bipartite graphs and performing multi-layer neighbor propagation. Among them, LightGCN, as a lightweight graph collaborative filtering model, has received widespread attention. By simplifying the traditional graph convolutional structure and retaining only neighbor information propagation and inter-layer aggregation, it maintains good recommender performance while reducing computational complexity, making it one of the current mainstream baselines.

[0005] However, existing LightGCN-based methods primarily rely on interaction structures for representation learning, essentially remaining a "behavior-driven" recommendation paradigm that infers preferences solely from user-item interaction patterns. When interaction data is sparse or the differences between items are mainly reflected in textual descriptions, simply relying on structural signals is insufficient to fully characterize the fine-grained semantic features of items. For example, two items may belong to the same category but differ significantly in function, style, or applicable scenarios. This information is usually contained in the title and description text, and traditional collaborative filtering struggles to effectively utilize this content information, thus limiting further improvements in recommendation performance.

[0006] Meanwhile, pre-trained language models have made significant progress in text understanding and semantic representation, providing a new technical path for introducing interpretable semantic information into recommender systems. Some studies have attempted to use language models to encode product titles, descriptions, or reviews, and then fuse the resulting semantic embeddings with collaborative representations to alleviate the expressive limitations caused by sparse interactions. However, e-commerce text often contains marketing noise, templated expressions, or information weakly related to user preferences. Simply fusing semantic features can easily lead to an over-reliance on textual information, thereby weakening the original collaborative structure signal and even causing performance fluctuations.

[0007] Furthermore, although contrastive learning has been used to enhance recommendation representations, existing methods mostly focus on contrastive constraints on the co-embedding itself, lacking an effective mechanism to align the two different representation spaces of "semantic embedding" and "co-embedding," making it difficult for textual semantics and behavioral structure to form a stable complementarity. Therefore, it is necessary to propose a novel recommendation method that introduces semantic information while maintaining the stability of the co-embedding structure, and achieves cross-space consistency optimization through gating fusion and contrastive alignment mechanisms. Summary of the Invention

[0008] To address the aforementioned issues, this invention proposes a collaborative filtering recommendation method that integrates semantic information from a large language model with a contrastive learning alignment mechanism. It employs approximate singular value decomposition and embedding spatial structure correlation for data augmentation, effectively mitigating noise problems. Furthermore, the data augmentation strategy of this invention only needs to be executed once, significantly reducing time overhead and ultimately improving the accuracy and robustness of the recommendations.

[0009] To address at least one of the aforementioned technical problems, according to one aspect of the present invention, a collaborative filtering recommendation method that integrates semantic information from a large language model with a contrastive learning alignment mechanism is provided, comprising the following steps:

[0010] Data preprocessing: Constructing a user-item bipartite graph based on e-commerce interaction data;

[0011] Collaborative representation learning: LightGCN is used to obtain user-item collaborative embeddings;

[0012] Semantic representation generation: Semantic embeddings are obtained by offline encoding of item text using MiniLM;

[0013] Top-p Gate semantic gating fusion: dimensional filtering of semantic embeddings and fusion with co-embeddings;

[0014] Contrastive Alignment Training: Construct semantic-cooperative alignment and contrastive learning constraints; jointly optimize and output Top-K recommendation results.

[0015] Furthermore, the data preprocessing specifically includes:

[0016] We selected the Amazon Books dataset, cleaned it, and retained user IDs, item IDs, and interaction records. We then divided the dataset into training and testing sets.

[0017] Construct a user-item bipartite graph G=(U,V,E) based on the training set, where U is the set of users, V is the set of items, and E is the set of historical interactions.

[0018] Furthermore, collaborative representation learning specifically involves:

[0019] Using LightGCN for multi-level neighbor propagation in a bipartite graph, the update of node v at layer k+1 is:

[0020]

[0021] in, Represents the target node The embedding vector at layer k+1 represents the target node. The feature representation of a node after k+1 rounds of neighbor information propagation; For nodes The set of first-order neighbors, These are the neighboring nodes in the set; They are nodes and neighboring nodes Degree, normalization coefficient Used to mitigate deviations caused by differences in node degree; For neighboring nodes The embedding vector at the k-th layer, as a node Update the source of neighbor features aggregated during embedding;

[0022] The embeddings obtained from multi-layer propagation are averaged and aggregated across layers to obtain the final collaborative embeddings of users and items. , The formula is:

[0023]

[0024] in , Each item node and user nodes The final collaborative embedding; K is the total number of propagation layers of LightGCN, which is 3 to 5 layers in this invention; , Each item node and user nodes The embedding vector at the k-th layer; The average weight is used to fuse structural information at different propagation depths.

[0025] Furthermore, the specific steps for semantic representation generation are as follows: after cleaning the item title and description, input them into the pre-trained language model MiniLM for offline encoding to obtain the item semantic embedding. .

[0026] Furthermore, the specific steps of Top-p Gate semantic gating fusion are as follows:

[0027] Calculate semantic dimension importance and construct mask vector ;

[0028] Dimensions whose cumulative importance reaches a threshold p are retained;

[0029] Fusion item representation:

[0030]

[0031] in, For item nodes The final fusion embedding vector obtained after semantic gating fusion integrates collaborative structural features and high-value text semantic features; For item nodes The collaborative embedding vectors, learned by LightGCN from the user-item interaction bipartite graph, represent the behavioral structural attributes of the item; These are semantic embedding weighting coefficients, used to adjust the contribution weight of semantic information in the fusion process; It is a dynamic mask vector used to filter high-importance dimensions in semantic embedding and suppress text noise; For item nodes The semantic embedding vector represents the fine-grained textual semantic features of the item.

[0032] Furthermore, the contrast alignment training specifically involves:

[0033] The same item and These are considered positive samples, while different items constitute negative samples.

[0034] Using InfoNCE to compare losses:

[0035]

[0036] Simultaneously, L2 alignment loss is introduced to reduce the distance between the same item in two representation spaces, achieving cross-space consistency optimization. The alignment loss formula is as follows:

[0037]

[0038] in, For InfoNCE, compare the loss values. This represents the L2 alignment loss value. This is the cosine similarity function, used to calculate the similarity between two embedding vectors; This is a temperature coefficient used to adjust the smoothness of the similarity distribution; For a collection of items, For the set excluding the target item Other items; The square operation of the L2 norm is used to measure the Euclidean distance between two embedding vectors; For items The collaborative embedding vector, For items semantic embedding vector, For items The semantic embedding vector.

[0039] Furthermore, the joint optimization and output of the Top-K recommendation results are as follows:

[0040] Predict user u's rating for item v:

[0041] ;

[0042] Recommended loss using BPR:

[0043] ;

[0044] Joint optimization objective:

[0045]

[0046] in, For users For items Predicted score; This is a matrix transpose operation; For users ; For items fused embedding vector; Recommend a loss value for BPR; Use the Sigmoid activation function; For users For negative sample items Predicted score Items that user u has not interacted with; The total loss value for the joint optimization objective; Recommended loss for BPR; This represents the L2 alignment loss. For InfoNCE contrast loss; To align the loss weights; The loss weights are used to compare the constraint strength of each loss.

[0047] According to another aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the collaborative filtering recommendation method of the present invention, which integrates semantic information of a large language model with a contrastive learning alignment mechanism.

[0048] According to another aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the collaborative filtering recommendation method of the present invention, which integrates semantic information of a large language model with a contrastive learning alignment mechanism.

[0049] Compared with existing technologies, the beneficial effects of the above-described method of the present invention are as follows:

[0050] This invention introduces semantic embedding of a large language model on the basis of the LightGCN collaborative filtering framework, and performs adaptive filtering of semantic dimensions through the Top-p Gate gating mechanism, effectively suppressing text noise interference and improving the semantic expression ability of items while preserving the stability of the collaborative structure.

[0051] The semantic encoding process of this invention uses offline MiniLM encoding, which eliminates the need to call large language models during recommendation model training or online inference, significantly reducing computational overhead and system latency.

[0052] This invention introduces semantic-cooperative alignment loss and contrastive learning loss for joint training, sharing a cooperative representation space. This not only enhances consistency across representation spaces but also improves the stability and robustness of the model in deep Top-K ranking. Attached Figure Description

[0053] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below. Obviously, the drawings described below only relate to some embodiments of the present invention and are not intended to limit the present invention.

[0054] Figure 1 This is a flowchart of a preferred embodiment of the present invention;

[0055] Figure 2 This is a schematic diagram of a preferred embodiment of the present invention. Detailed Implementation

[0056] 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, not all, of the embodiments of the present invention.

[0057] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0058] Example 1:

[0059] like Figure 1-2 As shown, this invention provides a collaborative filtering recommendation method that integrates semantic information from a large language model with a contrastive learning alignment mechanism, comprising the following steps:

[0060] S1. Data preprocessing: Constructing a user-item bipartite graph based on e-commerce interaction data;

[0061] S2. Collaborative Representation Learning: LightGCN is used to obtain collaborative embeddings between users and items;

[0062] S3. Semantic Representation Generation: Semantic embedding is obtained by offline encoding of item text using MiniLM;

[0063] S4, Top-p Gate Semantic Gated Fusion: Dimensional filtering of semantic embeddings and fusion with collaborative embeddings;

[0064] S5. Contrastive Alignment Training: Construct semantic-cooperative alignment and contrastive learning constraints; S6. Joint Optimization and Output of Top-K Recommendation Results.

[0065] Step S1 includes the following steps:

[0066] S101. Select the Amazon-Books dataset, clean it, retain user IDs, item IDs, and interaction records, and divide it into training and test sets;

[0067] S102. Construct a user-item bipartite graph G=(U,V,E) based on the training set, where U is the user set, V is the item set, and E is the historical interaction set.

[0068] Step S2 includes the following steps:

[0069] S201. Using LightGCN for multi-level neighbor propagation on a bipartite graph, the update of node v at layer k+1 is:

[0070]

[0071] in, Represents the target node The embedding vector at layer k+1 represents the feature representation obtained by the node after k+1 rounds of neighbor information propagation; For nodes The set of first-order neighbors, These are the neighboring nodes in the set; They are nodes and neighboring nodes Degree, normalization coefficient Used to mitigate deviations caused by differences in node degree; For neighboring nodes The embedding vector at the k-th layer, as a node Update the source of neighbor features aggregated during embedding.

[0072] S202. Perform cross-layer average aggregation on the embeddings obtained from multi-layer propagation to obtain the final collaborative embeddings of users and items. , The formula is:

[0073]

[0074] in , Each item node and user nodes The final collaborative embedding; K is the total number of propagation layers of LightGCN, which is 3 to 5 layers in this invention; , Each item node and user nodes The embedding vector at the k-th layer; The average weight is used to fuse structural information at different propagation depths.

[0075] Step S3 includes the following steps:

[0076] After cleaning the item titles and descriptions, they are input into the pre-trained language model MiniLM for offline encoding to obtain the item semantic embeddings. .

[0077] This process is completed before training and does not increase online inference overhead.

[0078] Step S4 includes the following steps:

[0079] S401. Calculate the semantic dimension importance and construct the mask vector. :

[0080] S402. Only retain dimensions whose cumulative importance reaches the threshold p;

[0081] S403, Fusion Item Representation:

[0082]

[0083] in, For item nodes The final fusion embedding vector obtained after semantic gating fusion integrates collaborative structural features and high-value text semantic features; For item nodes The collaborative embedding vectors, learned by LightGCN from the user-item interaction bipartite graph, represent the behavioral structural attributes of the item; These are semantic embedding weighting coefficients, used to adjust the contribution weight of semantic information in the fusion process; It is a dynamic mask vector used to filter high-importance dimensions in semantic embedding and suppress text noise; For item nodes The semantic embedding vector represents the fine-grained textual semantic features of the item.

[0084] Step S5 includes the following steps:

[0085] S501, The same item and These are considered positive samples, while different items constitute negative samples.

[0086] S502, using InfoNCE to compare losses:

[0087]

[0088] S503. Introducing Alignment Loss: Simultaneously introducing L2 alignment loss reduces the distance between the same item in two representation spaces, achieving cross-space consistency optimization. The alignment loss formula is as follows:

[0089]

[0090] in, For InfoNCE, compare the loss values. This represents the L2 alignment loss value. This is the cosine similarity function, used to calculate the similarity between two embedding vectors; This is a temperature coefficient used to adjust the smoothness of the similarity distribution; For a collection of items, For the set excluding the target item Other items; The square operation of the L2 norm is used to measure the Euclidean distance between two embedding vectors; For items The collaborative embedding vector, For items semantic embedding vector, For items The semantic embedding vector.

[0091] Step S6 includes the following steps:

[0092] S601, Predict user u's rating for item v:

[0093]

[0094] S602, using BPR recommended loss:

[0095]

[0096] S603, Joint Optimization Objective:

[0097] .

[0098] in, For users For items Predicted score; This is a matrix transpose operation; For users ; For items fused embedding vector; Recommend a loss value for BPR; Use the Sigmoid activation function; For users For negative sample items Predicted score Items that user u has not interacted with; The total loss value for the joint optimization objective; Recommended loss for BPR; This represents the L2 alignment loss. For InfoNCE contrast loss; To align the loss weights; The loss weights are used to compare the constraint strength of each loss.

[0099] This invention proposes a collaborative filtering recommendation method that integrates semantic information and contrastive learning. By introducing semantic representation of a large language model, Top-p Gate gating fusion, and contrastive alignment learning, it effectively alleviates the interaction sparsity problem, suppresses text noise interference, and improves the consistency between shallow Top-K ranking and deep ranking structure. This method improves recommendation accuracy and robustness while ensuring training efficiency.

[0100] Example 2:

[0101] The computer-readable storage medium of this embodiment stores a computer program that, when executed by a processor, implements the steps in the collaborative filtering recommendation method of Embodiment 1, which integrates semantic information of a large language model with a contrastive learning alignment mechanism.

[0102] The computer-readable storage medium in this embodiment can be an internal storage unit of the terminal, such as the terminal's hard disk or memory; the computer-readable storage medium in this embodiment can also be an external storage device of the terminal, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc. equipped on the terminal; furthermore, the computer-readable storage medium can include both the terminal's internal storage unit and external storage devices.

[0103] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0104] Example 3:

[0105] The computer device of this embodiment includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the collaborative filtering recommendation method of Embodiment 1 that integrates semantic information of a large language model with a contrastive learning alignment mechanism.

[0106] In this embodiment, the processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The memory can include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0107] Those skilled in the art will understand that the content disclosed in the embodiments can be provided as a method, system, or computer program product. Therefore, this solution can take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this solution can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage) containing computer-usable program code.

[0108] This solution is described with reference to flowchart illustrations and / or block diagrams of methods and computer program products according to embodiments of this solution. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0109] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0110] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0111] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0112] The examples described herein are merely preferred embodiments of the invention and are not intended to limit the concept and scope of the invention. Any modifications and improvements made by those skilled in the art to the technical solutions of the invention without departing from the design concept of the invention should fall within the protection scope of the invention.

[0113] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the specific embodiments described above. The specific embodiments and descriptions in the specification are merely for further illustrating the principles of the invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the claims and their equivalents.

Claims

1. A collaborative filtering recommendation method that integrates semantic information from a large language model with a contrastive learning alignment mechanism, characterized in that, Includes the following steps: Data preprocessing: Constructing a user-item bipartite graph based on e-commerce interaction data; Collaborative representation learning: LightGCN is used to obtain user-item collaborative embeddings; Semantic representation generation: Semantic embeddings are obtained by offline encoding of item text using MiniLM; Top-p Gate semantic gating fusion: dimensional filtering of semantic embeddings and fusion with co-embeddings; Contrastive Alignment Training: Constructing Semantic-Co-alignment and Contrastive Learning Constraints; Jointly optimize and output Top-K recommendation results.

2. The method as described in claim 1, characterized in that, Data preprocessing specifically includes: Select the Amazon-Books dataset, clean it, retain user IDs, item IDs, and interaction records, and divide it into training and test sets; Construct a user-item bipartite graph G=(U,V,E) based on the training set, where U is the set of users, V is the set of items, and E is the set of historical interactions.

3. The method as described in claim 2, characterized in that, Collaborative representation learning specifically includes: Using LightGCN for multi-level neighbor propagation in a bipartite graph, the update of node v at layer k+1 is: ； in, Represents the target node The embedding vector at layer k+1 represents the target node. The feature representation of a node after k+1 rounds of neighbor information propagation; For nodes The set of first-order neighbors, These are the neighboring nodes in the set; They are nodes and neighboring nodes Degree, normalization coefficient Used to mitigate deviations caused by differences in node degree; Neighboring nodes The embedding vector at the k-th layer, as a node Update the source of neighbor features aggregated during embedding; The embeddings obtained from multi-layer propagation are averaged and aggregated across layers to obtain the final collaborative embeddings of users and items. , The formula is: ； in , Each item node and user nodes The final collaborative embedding; K is the total number of propagation layers of LightGCN, which is 3 to 5 layers in this invention; , Each item node and user nodes The embedding vector at the k-th layer; The average weight is used to fuse structural information at different propagation depths.

4. The method as described in claim 3, characterized in that, The specific steps for generating semantic representations are as follows: after cleaning the item titles and descriptions, they are input into the pre-trained language model MiniLM for offline encoding to obtain the item semantic embeddings. .

5. The method as described in claim 4, characterized in that, The specific steps of Top-p Gate semantic gating fusion are as follows: Calculate semantic dimension importance and construct mask vector ; Dimensions whose cumulative importance reaches a threshold p are retained; Fusion item representation: ； in, For item nodes The final fusion embedding vector obtained after semantic gating fusion integrates collaborative structural features and high-value text semantic features; For item nodes The collaborative embedding vectors, learned by LightGCN from the user-item interaction bipartite graph, represent the behavioral structural attributes of the item; These are semantic embedding weighting coefficients, used to adjust the contribution weight of semantic information in the fusion process; It is a dynamic mask vector used to filter high-importance dimensions in semantic embedding and suppress text noise; For item nodes The semantic embedding vector represents the fine-grained textual semantic features of the item.

6. The method as described in claim 5, characterized in that, The specific steps of the contrast alignment training are as follows: The same item and These are considered positive samples, while different items constitute negative samples. Using InfoNCE to compare losses: ； Simultaneously, L2 alignment loss is introduced to reduce the distance between the same item in two representation spaces, achieving cross-space consistency optimization. The alignment loss formula is: ； in, For InfoNCE, compare the loss values. This represents the L2 alignment loss value. This is the cosine similarity function, used to calculate the similarity between two embedding vectors; This is a temperature coefficient used to adjust the smoothness of the similarity distribution; For a collection of items, For the set excluding the target item Other items; The square operation of the L2 norm is used to measure the Euclidean distance between two embedding vectors; For items The collaborative embedding vector, For items semantic embedding vector, For items The semantic embedding vector.

7. The method as described in claim 6, characterized in that, The joint optimization and output of the Top-K recommendation results are as follows: Predict user u's rating for item v: ； Recommended loss using BPR: ； Joint optimization objective: ； in, For users For items Predicted score; This is a matrix transpose operation; For users ; For items fused embedding vector; Recommend a loss value for BPR; Use the Sigmoid activation function; For users For negative sample items Predicted score Items that user u has not interacted with; The total loss value for the joint optimization objective; Recommended loss for BPR; This represents the L2 alignment loss. For InfoNCE contrast loss; Alignment loss weights; The loss weights are used to compare the constraint strength of each loss.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the steps in the collaborative filtering recommendation method that integrates semantic information of a large language model with a contrastive learning alignment mechanism as described in any one of claims 1 to 7.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the collaborative filtering recommendation method that integrates semantic information of a large language model with a contrastive learning alignment mechanism as described in any one of claims 1 to 7.

Images