Recommendation method and system for user interest aggregation representation based on large model enhancement

Abstract

The invention discloses a recommendation method and system for user interest aggregation representation based on large model enhancement, and relates to the technical field of personalized recommendation, and the method comprises the steps: obtaining an article static semantic vector, a user dynamic portrait vector and an article dynamic portrait vector through a large language model, carrying out the clustering of the article static portrait vector, generating an interest prototype center, and carrying out the clustering of the interest prototype center; for the interest prototype center of the article, calculating collaborative filtering specific attention, article dynamic attention and user dynamic attention, and generating user interest characterization; introducing a semantic difficult negative example to compare, learn and align a collaborative filtering embedded vector and an article static portrait vector, and offline caching a portrait vector, an interest prototype center and a user interest representation; and calling the pre-calculation vector online, and obtaining the user interest representation and the inner product of the article static portraits of the candidate articles for recommendation. The method improves the recommendation effect, reduces the reasoning complexity, accurately captures the dynamic interest preference of the user, and improves the matching degree between the recommendation result and the real demand of the user.

Description

Technical Field

[0001] This invention relates to the field of personalized recommendation technology, and more specifically, to a recommendation method and system based on a large model-enhanced user interest aggregation representation. Background Technology

[0002] Recommender systems play a crucial role in today's digital services, widely applied in e-commerce platforms, social media, digital advertising, and other fields. Accurately extracting and representing user interests in items is a key technical means to improve click-through rate prediction. Existing user interest representation methods mainly suffer from the following technical problems: First, traditional methods can only learn user interest features from closed training datasets, lacking a deep understanding of item attributes and user preferences, failing to acquire broad world knowledge beyond the training data, and struggling to capture the implicit semantic relationships between items and user interests. Second, existing refined interest modeling methods face serious efficiency problems in practical applications; complex attention mechanisms lead to high computational costs, making it difficult to meet the real-time requirements of online services and hindering efficient deployment in large-scale scenarios.

[0003] While existing methods have attempted to incorporate large language models to enhance interest representation, significant technical shortcomings remain. Firstly, the lack of an effective semantic knowledge alignment mechanism makes it difficult to balance the complementarity of collaborative filtering signals and semantic knowledge; the knowledge fusion methods are overly simplistic and fail to fully leverage the potential of large language models. Secondly, existing solutions suffer from fragmented online and offline feature computation processes, inefficient feature storage and retrieval, and poor service architecture scalability. Therefore, a new technical solution is urgently needed that can effectively utilize the world knowledge of large language models to enhance user interest representation while ensuring system computational efficiency and service performance. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention proposes a recommendation method and system based on large-model-enhanced user interest aggregation representation, which improves recommendation performance, reduces inference complexity, accurately captures users' dynamic interest preferences, and enhances the matching degree between recommendation results and users' actual needs.

[0005] The first aspect of this invention provides a recommendation method based on large model-enhanced user interest aggregation representation, comprising the following steps: The name, description, and category attributes of the item are input into a large language model to extract explicit features, and implicit attributes are inferred based on world knowledge to generate a static semantic vector of the item. Input user comment text and interaction behavior data into a large language model to generate user dynamic profile vectors and item dynamic profile vectors, and map them to the vector space through an embedding layer; K-means clustering is performed on the static image vector of the item to generate several interest prototype centers. For the interest prototype centers of the target item, collaborative filtering specific attention, item dynamic attention and user dynamic attention are calculated. Attention representations are integrated through multilayer perceptron to generate user interest representations. In offline pre-training, semantic hard negative example contrastive learning is introduced. The set of hard negative examples is selected by InfoNCE loss function to align and collaboratively filter the embedding vector and the static portrait vector of the item. The model parameters are updated by gradient descent based on contrastive loss, and the portrait vector, interest prototype center and user interest representation are cached. During the online phase, a pre-computed vector is invoked to obtain the inner product of the user interest representation and the static profile of the candidate items, generating a recommendation score, and then recommending items based on the recommendation score.

[0006] In this approach, the name, description, and category attributes of an item are input into a large language model to extract explicit features. Based on world knowledge, implicit attributes are inferred to generate a static semantic vector for the item. Specifically: Collect structured metadata of items, including name, description and category attributes; perform data cleaning on the structured metadata; and design prompting strategies from the perspectives of explicit feature extraction and implicit attribute inference. In the explicit feature extraction perspective, a structured prompt template is constructed, and the large language model is required to extract explicit attributes based on the structured prompt template. The large language model returns key-value pairs of the explicit attributes, and the integrity is verified by the rule engine. In the implicit attribute reasoning perspective, the world knowledge of the large language model is stimulated through thought chain prompts, and the large language model outputs implicit attributes, which are then combined with a preset domain knowledge base for semantic calibration. The extracted attribute key-value pairs are concatenated into a text sequence, which is then input into a large language model to obtain a dense vector representation, which serves as the explicit attribute vector. The inferred implicit attribute text is then input separately into the same large language model to generate an implicit attribute vector. The explicit attribute vector and the implicit attribute vector are weighted and summed, then mapped to a unified dimension through a linear layer, and the high-dimensional vector is compressed into a low-dimensional space through a multilayer perceptron to generate the static semantic vector of the item.

[0007] In this solution, user comment text and interaction behavior data are input into a large language model to generate user dynamic profile vectors and item dynamic profile vectors. These vectors are then mapped to the vector space through an embedding layer, specifically: Collect user comments on items, standardize the comments, segment the standardized comments into sentences, extract structured semantic units, integrate user behavior logs, divide behavior sessions by time windows, normalize discrete behaviors, and perform one-hot encoding on categorical behaviors. We construct time-series correlated text templates, combine behavioral data with comment text, and use a large language model to analyze the consistency between behavior and text to identify potential preferences. All user interaction data are input into the large language model. Based on the potential preference prompting large language model, interest tags are generated and a probabilistic interest distribution vector is output. The vector is then mapped to the vector space through the embedding layer to generate a user dynamic semantic vector. In addition, the global data of the item being interacted with is used to prompt the large language model to extract group perception features, which are then mapped to the vector space through the embedding layer to generate an item dynamic semantic vector. The user dynamic semantic vector and the item dynamic semantic vector are transformed using a multilayer perceptron to obtain the user dynamic profile vector and the item dynamic profile vector.

[0008] In this scheme, K-means clustering is performed on the static image vectors of items to generate several interest prototype centers, specifically: Collect static image vectors of all items and perform data verification. Introduce particle swarm optimization algorithm to optimize the number of interest prototypes and obtain the optimal number of interest prototypes. Use cosine similarity as distance metric. Based on the optimal number of interest prototypes, the K-means++ algorithm is used to initialize the interest prototype centers. The cosine distance from the static portrait vector of each item to all prototype centers is calculated and assigned to the cluster to which the nearest center belongs. The mean vector of each cluster is recalculated as the new interest prototype center. Clustering stops when the change in the center of interest is less than the threshold or the maximum number of iterations is reached. Clusters with fewer than the preset threshold of samples are removed and reassigned to the nearest valid cluster to obtain the final center of interest. Interpretable labels are then generated through a large language model.

[0009] In this scheme, for the interest prototype center of the target item, collaborative filtering specific attention, item dynamic attention, and user dynamic attention are calculated. Attention weights are then fused using a multilayer perceptron to generate a user interest representation. Specifically: Based on the static profile vector of the target item, the nearest interest prototype center is retrieved from several pre-calculated interest prototype centers and used as the target interest prototype center. The user's historical interaction item set is extracted, which includes three types of features: collaborative filtering embedding vector, item dynamic profile vector, and user dynamic profile vector. For each type of feature, the attention weights with the target interest prototype center are calculated separately. In the specific attention calculation of collaborative filtering, the target interest prototype center is used as the query vector and the collaborative filtering embedding vector is used as the key value. The weights are calculated by scaling the dot product attention and then summed to obtain the collaborative filtering interest representation. In the dynamic attention calculation of items, the target interest prototype center is used as the query vector and the dynamic profile vector of the item is used as the key value. When calculating similarity, a time decay factor is introduced to generate attention weights and output the dynamic interest representation of the item in a weighted manner. In the calculation of user dynamic attention, the target interest prototype center is used as the query vector, the user dynamic profile vector is used as the key value, and the attention weight is calculated by adding the behavior type weight, and the user dynamic interest representation is output in a weighted manner. By integrating collaborative filtering interest representations, dynamic interest representations of items, and dynamic interest representations of users through a multilayer perceptron, a user interest representation is output.

[0010] In this scheme, semantic hard negative example contrastive learning is introduced into offline pre-training. The set of hard negative examples is selected by aligning and collaboratively filtering the embedding vectors and the static image vectors of the items using the InfoNCE loss function. Specifically: For each target item, candidate negative samples are screened from the set of uninterrupted items, items that the user is clearly not interested in are excluded, and the cosine similarity between the target item and the candidate negative samples is calculated based on the static profile vector of the item. Candidate negative samples with cosine similarity higher than a preset threshold are screened, and negative examples with similar behavior patterns to positive samples are excluded, thus forming a set of difficult negative examples. For the target item, extract the collaborative filtering embedding vector. The collaborative filtering embedding vector and the static image vector of the item are mapped to the same contrast space through the projection head to eliminate distribution differences. Calculate positive sample pairs Similarity score Calculate the target item and the set of difficult examples. Each negative sample Similarity score The InfoNCE loss function is used for optimization, and contrastive learning is used to align the collaborative filtering embedding vectors with the item static image vectors. The contrastive loss... Represented as: ; in This represents the temperature parameter.

[0011] In this scheme, during the online phase, a pre-computed vector is invoked to obtain the inner product of the user interest representation and the static profile of candidate items, generating a recommendation score. Item recommendations are then made based on this recommendation score. Specifically: Based on the user ID, pre-computed user interest representations are read from the vector database in real time, and static profile vectors of items are loaded in batches from the cache based on the candidate item ID list. For each candidate item, the inner product of the item's static profile vector and the user's interest representation is calculated. The result of the inner product is used as the recommendation score. The candidate item list is sorted in descending order of the recommendation score. Category diversity control is introduced into the sorting results to avoid excessive homogenization. A preset number of candidate items are selected for recommendation.

[0012] The second aspect of this invention provides a recommendation system based on large-model-enhanced user interest aggregation representation, the system comprising: a basic data module, a large language model service module, a semantic hard negative example contrastive learning module, a user interest aggregation representation module, and an inference optimization module; The basic data module stores the name, description, and category attribute information of items, as well as user comment text and interaction behavior data; The large language model service module extracts explicit features based on the name, description, and category attributes of the item using the large language model, and infers implicit attributes based on world knowledge to generate static semantic vectors for the item. User comment text and interaction behavior data are input into the large language model to generate dynamic user profile vectors and dynamic item profile vectors. These vectors are then mapped to a unified representation space through an embedding layer and subjected to dimensional adjustment and normalization. The semantic hard negative example contrastive learning module introduces semantic hard negative example contrastive learning, selects the hard negative example set by using the InfoNCE loss function to align and collaboratively filter the embedded vector and the static image vector of the item, and updates the model parameters based on the contrastive loss through gradient descent. The user interest aggregation and representation module performs K-means clustering on the static profile vector of the item to generate several interest prototype centers. For the interest prototype centers of the target item, it calculates collaborative filtering specific attention, item dynamic attention, and user dynamic attention, and integrates the attention representation through a multilayer perceptron to generate user interest representation. The inference optimization module offline caches the profile vector, interest prototype center, and user interest representation. In the online stage, it calls the pre-computed vector to obtain the inner product of the user interest representation and the static profile of the candidate item, generates a recommendation score, and recommends items based on the recommendation score.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a user interest aggregation representation recommendation method based on a large language model, achieving significant improvements in recommendation effectiveness and system performance optimization. Regarding recommendation quality, the multi-dimensional semantic profiles generated by the large language model enable the system to deeply understand the explicit features and implicit attributes of items, while accurately capturing the dynamic interest preferences exhibited by users during interactions, thus greatly improving the matching degree between recommendation results and users' actual needs. In terms of computational efficiency, the innovative interest aggregation representation mechanism significantly reduces the complexity of online inference, and combined with offline pre-computation strategies, shortens system response time, enabling stable support for tens of thousands of concurrent requests per second. Furthermore, the application of semantic hard negative example contrastive learning technology effectively solves the problem of the separation between semantic information and behavioral data in traditional recommendations. Through refined negative sample selection and cross-space alignment, the model fully integrates the world knowledge of the large language model while maintaining the advantages of collaborative filtering. After the system is deployed in actual business scenarios, it improves the diversity of recommendation results and, through an interpretable interest prototype mechanism, makes the generation of recommendation reasons more intuitive and credible, significantly improving user satisfaction. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples 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 according to these drawings without creative effort.

[0015] Figure 1 A flowchart of a recommendation method based on large model-enhanced user interest aggregation representation is shown; Figure 2 A flowchart for generating user interest representations is shown; Figure 3 The architecture diagram of a recommender system based on user interest aggregation representations enhanced by a large model is shown. Detailed Implementation

[0016] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0017] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0018] Figure 1A flowchart of a recommendation method based on user interest aggregation representations enhanced by a large model is shown.

[0019] like Figure 1 As shown, this embodiment provides a recommendation method based on large model-enhanced user interest aggregation representation, including: S102, input the name, description and category attributes of the item into the large language model to extract explicit features, and infer implicit attributes based on world knowledge to generate a static semantic vector of the item; S104, Input user comment text and interaction behavior data into the large language model to generate user dynamic profile vectors and item dynamic profile vectors, and map them to the vector space through the embedding layer; S106, perform K-means clustering on the static image vector of the item to generate several interest prototype centers. For the interest prototype centers of the target item, calculate the collaborative filtering specific attention, the item dynamic attention, and the user dynamic attention. Integrate the attention representations through a multilayer perceptron to generate user interest representations. S108 introduces semantic hard negative example contrastive learning in offline pre-training, selects hard negative example set by InfoNCE loss function to align and collaboratively filter embedded vectors and item static portrait vectors, updates model parameters by gradient descent based on contrastive loss, and caches portrait vectors, interest prototype centers and user interest representations. S110, during the online phase, the pre-computed vector is invoked to obtain the inner product of the user interest representation and the static profile of the candidate items, generate a recommendation score, and recommend items based on the recommendation score.

[0020] It should be noted that the structured metadata of the items is collected. This structured metadata includes name, description, and category attributes. For example, for e-commerce products, the input includes the product title (wireless Bluetooth headphones), detailed description (supports active noise cancellation, 30-hour battery life), and category (electronic products, headphones). The structured metadata is then cleaned, including removing special symbols, filtering stop words, and standardizing abbreviations.

[0021] A prompting strategy is designed from the perspectives of explicit feature extraction and implicit attribute inference. From the explicit feature extraction perspective, a structured prompt template is constructed, for example: "Please extract explicit attributes from the following product description, listed in the format of attribute name: attribute value." Based on this structured prompt template, a large language model is required to extract the explicit attributes. The large language model returns key-value pairs of the displayed attributes, such as "Connection method: Wireless Bluetooth" and "Battery life: 30 hours," and the integrity is verified by a rule engine. In the implicit attribute reasoning perspective, the world knowledge of the large language model is stimulated through thought chain prompts. For example, based on product descriptions and common sense, potential usage scenarios and user groups are inferred. The large language model outputs implicit attributes, such as "applicable scenarios: commuting, sports" and "target users: music enthusiasts," and performs semantic calibration in conjunction with a preset domain knowledge base. The extracted attribute key-value pairs are concatenated into a text sequence and input into the large language model to obtain a dense vector representation, which serves as the explicit attribute vector. The implied attribute text obtained through reasoning is input separately into the same large language model to generate an implicit attribute vector. The explicit attribute vector and the implicit attribute vector are weighted and summed, and then mapped to a unified dimension through a linear layer. The high-dimensional vector is compressed into a low-dimensional space through a multilayer perceptron, preserving core semantic features while reducing computational overhead. After normalization, a static semantic vector of the item is generated. The generated static semantic vector of the item is stored in an offline database, and an inverted index is established to support efficient retrieval.

[0022] It should be noted that the process involves collecting user reviews of items and standardizing these reviews by removing irrelevant symbols, correcting spelling, and standardizing sentiment words. The standardized reviews are then segmented into sentences to extract structured semantic units. Furthermore, user behavior logs, including clicks, purchases, and dwell time, are integrated, and behavioral sessions are grouped by time windows, such as grouping behaviors within a 7-day period. Discrete behaviors such as click counts are numerically normalized, and categorical behaviors are encoded using one-hot encoding.

[0023] A temporally correlated text template is constructed, combining behavioral data with comment text. For example, a user clicked on and purchased item A twice in the past three days, commenting on its excellent sound quality; they only browsed item B for 10 seconds, finding it unsuitable. A large language model is used to analyze the consistency between behavior and text, identifying potential preferences. All user interaction data is input into the large language model, and based on these potential preferences, the model generates interest tags, such as "prefers high-quality audio devices and is sensitive to battery life," and outputs a probabilistic interest distribution vector. This vector is then mapped to a vector space through an embedding layer to generate a user dynamic semantic vector. Additionally, global data on item interactions, such as all user comments and behavior statistics, is used to extract group perception features from the large language model. These features are then mapped to a vector space through an embedding layer to generate an item dynamic semantic vector, such as "most users believe the battery life is insufficient, but the sound quality rating is high." A multilayer perceptron (MLM) is used to perform dimensionality transformation on the user and item dynamic semantic vectors, resulting in user dynamic profile vectors and item dynamic profile vectors. These are stored offline in a vector database, indexed by user ID and item ID. Weekly LLM adjustments are triggered based on new behavioral data to update the dynamic vectors.

[0024] It should be noted that the static profile vectors of all items are collected and data validation is performed to check for missing or outlier values. The vectors are filled with the mean or outlier samples are removed. A particle swarm optimization algorithm is introduced to optimize the number of interest prototypes. A fitness function is constructed based on the silhouette coefficient. The silhouette coefficient is calculated iteratively for different values ​​of the interest prototype number C to obtain the optimal number of interest prototypes. Cosine similarity is used as the distance metric. Based on the optimal number of interest prototypes, the K-means++ algorithm is used to initialize the interest prototype centers. The cosine distance from each item's static profile vector to all prototype centers is calculated, and the item is assigned to the cluster of the nearest center. The mean vector of each cluster is recalculated as the new interest prototype center. Clustering stops when the change in the interest prototype center is less than a threshold or the maximum number of iterations is reached. Clusters with fewer than a preset threshold are removed, and the items are reassigned to the nearest valid cluster to obtain the final interest prototype centers. Interpretable labels are generated using a large language model and stored in an offline database for real-time access by the online service. When a new item is added, it is assigned to an existing prototype through nearest neighbor, without the need for re-clustering.

[0025] Figure 2 A flowchart for generating user interest representations is shown.

[0026] According to an embodiment of the present invention, for the interest prototype center of the target item, collaborative filtering specific attention, item dynamic attention, and user dynamic attention are calculated, and attention weights are fused through a multilayer perceptron to generate a user interest representation, specifically: S202, based on the static profile vector of the target item, retrieve the nearest interest prototype center from a number of pre-calculated interest prototype centers, use it as the target interest prototype center, and extract the user's historical interaction item set, which includes three types of features: collaborative filtering embedding vector, item dynamic profile vector, and user dynamic profile vector. S204. For each type of feature, calculate the attention weight with the target interest prototype center. In the collaborative filtering specific attention calculation, the target interest prototype center is used as the query vector and the collaborative filtering embedding vector is used as the key value. The weight is calculated by scaling the dot product attention and the weighted sum is used to obtain the collaborative filtering interest representation. S206, In the dynamic attention calculation of items, the target interest prototype center is used as the query vector and the dynamic profile vector of the items is used as the key value. When calculating similarity, a time decay factor is introduced to generate attention weights and output the dynamic interest representation of items in a weighted manner. S208, in the calculation of user dynamic attention, the target interest prototype center is used as the query vector, the user dynamic profile vector is used as the key value, the behavior type weight is added to calculate the attention weight, and the user dynamic interest representation is output in a weighted manner. S210 integrates collaborative filtering interest representations, dynamic interest representations of items, and dynamic interest representations of users through a multilayer perceptron, and outputs user interest representations.

[0027] It should be noted that the collaborative filtering embedding vector in the set of historical interaction items... ID class representation from collaborative filtering models such as matrix factorization, and dynamic profile vectors of items. To analyze the timeliness characteristics based on large language models, user dynamic profile vectors Reflecting users' interactions with items The momentary interest during interaction.

[0028] Collaborative filtering with specific attention aims to capture implicit collaborative signals in users' historical behavior. Overscaling the dot product attention for each user, it calculates... The dot product similarity between the item and the target interest prototype center is normalized using Softmax to obtain the attention score. Item dynamic attention aims to integrate the dynamic attributes of the item itself, introducing a time decay factor when calculating similarity. For example, the closer the interaction time, the higher the weight. For each item in the interaction time series, its time decay weight is calculated using an exponential decay method. This time decay weight is incorporated into the standard dot product attention to generate the attention score. The time decay factor allows the model to adaptively track interest drift. User dynamic attention aims to capture changes in user interest under different interaction scenarios, incorporating behavior type weights. For example, purchasing behavior has a 30% higher weight than clicking behavior. For each interactive item, behavior type weights are obtained based on its behavior type. The key vector is weighted using these behavior type weights, and the attention score is calculated using the weighted key vector. Explicit behavior weighting avoids low-value interactions (such as accidental clicks) from polluting the interest representation.

[0029] By integrating three types of attention representations through a multilayer perceptron, the vectors of these three types of representations are concatenated into a joint vector, outputting a fused interest representation. Collaborative filtering, item dynamics, and user dynamics complement each other's attention, taking into account both long-term preferences and real-time interests. In addition, attention weights can be back-mapped to specific historically interacted items, supporting the generation of recommendation reasons.

[0030] It should be noted that semantically difficult negative example contrastive learning is introduced in offline pre-training. For each target item, candidate negative samples are screened from the set of uninterrupted items to exclude items that the user is clearly not interested in, based on the item's static profile vector. Calculate the cosine similarity between the target item and the candidate negative samples, and filter the candidate negative samples whose cosine similarity is higher than a preset threshold to ensure that the negative samples and the positive samples are semantically similar.

[0031] By deeply analyzing user and item behavior patterns, negative examples similar to positive sample behavior patterns are further eliminated, ultimately forming a set of difficult negative examples. This set is updated after each training round to adapt to the model's latest representation capabilities. Taking e-commerce scenarios as an example, if a target positive sample item (such as a wireless headphone) is purchased by a user, candidate negative examples must not only be semantically similar to the positive sample (such as similar headphone products), but also meet the behavioral condition that the user did not purchase it. This dual filtering mechanism effectively avoids misclassifying potential positive samples as negative examples, ensuring that the samples in the difficult negative example set are both semantically similar and significantly different in user behavior feedback. This is achieved by comparing user historical behavior logs: first, items that the user has already interacted with positively (such as purchasing, favoriting, or prolonged browsing) are excluded; second, for samples that have been viewed but not clicked, further filtering is performed based on fine-grained behaviors such as dwell time (e.g., short-term exposure may reflect genuine disinterest). This refined behavior pattern analysis ensures that the final set of difficult negative examples represents items that the user may have encountered but did not actually select, while avoiding erroneous negative samples caused by data noise.

[0032] For the target item, extract the collaborative filtering embedding vector. and static image vector of items The collaborative filtering embedding vectors are obtained using collaborative filtering algorithms such as matrix factorization. A low-dimensional embedding representation is learned based on the user-item historical interaction matrix. A projection head (multilayer perceptron) maps the collaborative filtering embedding vectors and item static profile vectors to the same contrastive learning space, eliminating distributional differences. During contrastive loss calculation, the projected vectors are used for similarity matching, and the alignment is optimized using the InfoNCE loss function. Through hard-example contrastive learning, the model is forced to distinguish between semantically similar items but with different user feedback, enhancing the discriminative power of the representation.

[0033] Calculate positive sample pairs Similarity score Calculate the target item and the set of difficult examples. Similarity score for each negative sample The InfoNCE loss function is used for optimization, and contrastive learning is used to align the collaborative filtering embedding vectors with the item static image vectors. The contrastive loss... Represented as: ; in This indicates the degree to which the temperature parameter adjusts the sharpness of the similarity distribution. The temperature parameter decays with each training round, gradually focusing on difficult samples.

[0034] Comparative loss Compared with CTR prediction loss Regularization loss Joint optimization is represented as: ,in , The contribution strength is represented by the parameters of the fully connected layer of the collaborative filtering model, projection head, and CTR prediction module, which are updated synchronously through gradient descent. An adaptive optimizer is used to automatically adjust the learning rate of each parameter to ensure stable convergence of different modules.

[0035] It should be noted that the pre-computed user interest representation is read in real time from the vector database based on the user ID. This vector was generated through an offline interest aggregation module, incorporating collaborative filtering, item dynamics, and user dynamic attention features. Based on the candidate item ID list, item static profile vectors were batch-loaded from the cache. These item static profile vectors were generated offline and normalized by a large language model.

[0036] For each candidate item, calculate the item's static image vector. User interest representation inner product , is represented as: ; The inner product result is used as the recommendation score; a higher score indicates a stronger match with user interests. Candidate items are sorted in descending order of recommendation score, and filtered using business rules such as deduplication, inventory verification, and blacklist removal. Category diversity control is introduced into the sorting results to avoid excessive homogenization. A preset number of candidate items are selected for recommendation. Actual exposure and click data for this recommendation result are recorded and written to a real-time data warehouse for subsequent model iterations. If a user generates new interactions such as clicks or purchases, a large language model inference task is triggered, asynchronously updating the user's dynamic profile vector, which takes effect on the next request.

[0037] Figure 3 The architecture diagram of a recommender system based on user interest aggregation representations enhanced by a large model is shown.

[0038] The second embodiment of the present invention provides a recommendation system based on large model-enhanced user interest aggregation representation. The system includes: a basic data module, a large language model service module, a semantic hard negative example contrast learning module, a user interest aggregation representation module, and an inference optimization module. The basic data module stores the name, description, and category attribute information of items, as well as user comment text and interaction behavior data; The large language model service module extracts explicit features based on the name, description, and category attributes of the item using the large language model, and infers implicit attributes based on world knowledge to generate static semantic vectors for the item. User comment text and interaction behavior data are input into the large language model to generate dynamic user profile vectors and dynamic item profile vectors. These vectors are then mapped to a unified representation space through an embedding layer and subjected to dimensional adjustment and normalization. The semantic hard negative example contrastive learning module introduces semantic hard negative example contrastive learning, selects the hard negative example set by using the InfoNCE loss function to align and collaboratively filter the embedded vector and the static image vector of the item, and updates the model parameters based on the contrastive loss through gradient descent. The user interest aggregation and representation module performs K-means clustering on the static profile vector of the item to generate several interest prototype centers. For the interest prototype centers of the target item, it calculates collaborative filtering specific attention, item dynamic attention, and user dynamic attention, and integrates the attention representation through a multilayer perceptron to generate user interest representation. The inference optimization module uses offline calculation and caching of profile vectors, interest prototype centers, and user interest representations. By quickly retrieving the pre-calculated vectors in the online stage, it obtains the inner product of the user interest representations and the static profiles of candidate items, generates a recommendation score, and recommends items based on the recommendation score.

[0039] A third embodiment of the present invention provides a computer-readable storage medium including a recommendation method program based on a large model-enhanced user interest aggregation representation. When the recommendation method program based on the large model-enhanced user interest aggregation representation is executed by a processor, it implements the steps of the recommendation method based on the large model-enhanced user interest aggregation representation.

[0040] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, and can be electrical, mechanical, or other forms. Furthermore, in the various embodiments of the present invention, all functional units can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0041] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0042] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A recommendation method based on large-model-enhanced user interest aggregation representation, characterized in that, Includes the following steps: The name, description, and category attributes of the item are input into a large language model to extract explicit features, and implicit attributes are inferred based on world knowledge to generate a static semantic vector of the item. Input user comment text and interaction behavior data into a large language model to generate user dynamic profile vectors and item dynamic profile vectors, and then map them to the vector space through an embedding layer; K-means clustering is performed on the static image vector of the item to generate several interest prototype centers. For the interest prototype centers of the target item, collaborative filtering specific attention, item dynamic attention and user dynamic attention are calculated. Attention representations are integrated through multilayer perceptron to generate user interest representations. In offline pre-training, semantic hard negative example contrastive learning is introduced. The set of hard negative examples is selected by InfoNCE loss function to align and collaboratively filter the embedding vector and the static portrait vector of the item. The model parameters are updated by gradient descent based on contrastive loss, and the portrait vector, interest prototype center and user interest representation are cached. During the online phase, a pre-computed vector is invoked to obtain the inner product of the user interest representation and the static profile of the candidate items, generating a recommendation score, and then recommending items based on the recommendation score.

2. The recommendation method based on large model-enhanced user interest aggregation representation according to claim 1, characterized in that, The name, description, and category attributes of an item are input into a large language model to extract explicit features. Based on world knowledge, implicit attributes are inferred to generate a static semantic vector for the item. Specifically: Collect structured metadata of items, including name, description and category attributes; perform data cleaning on the structured metadata; and design prompting strategies from the perspectives of explicit feature extraction and implicit attribute inference. In the explicit feature extraction perspective, a structured prompt template is constructed, and the large language model is required to extract explicit attributes based on the structured prompt template. The large language model returns key-value pairs of the explicit attributes, and the integrity is verified by the rule engine. In the implicit attribute reasoning perspective, the world knowledge of the large language model is stimulated through thought chain prompts, and the large language model outputs implicit attributes, which are then combined with a preset domain knowledge base for semantic calibration. The extracted attribute key-value pairs are concatenated into a text sequence, which is then input into a large language model to obtain a dense vector representation, which serves as the explicit attribute vector. The inferred implicit attribute text is then input separately into the same large language model to generate an implicit attribute vector. The explicit and implicit attribute vectors are weighted and summed, then mapped to a unified dimension through a linear layer. The high-dimensional vectors are then compressed into a low-dimensional space through a multilayer perceptron to generate static semantic vectors for items.

3. The recommendation method based on large model-enhanced user interest aggregation representation according to claim 1, characterized in that, User comment text and interaction behavior data are input into a large language model to generate dynamic user profile vectors and dynamic item profile vectors. These are then mapped to the vector space through an embedding layer, specifically: Collect user comments on items, standardize the comments, segment the standardized comments into sentences, extract structured semantic units, integrate user behavior logs, divide behavior sessions by time windows, normalize discrete behaviors, and perform one-hot encoding on categorical behaviors. We construct time-series correlated text templates, combine behavioral data with comment text, and use a large language model to analyze the consistency between behavior and text to identify potential preferences. All user interaction data are input into the large language model. Based on the potential preference prompting large language model, interest tags are generated and a probabilistic interest distribution vector is output. The vector is then mapped to the vector space through the embedding layer to generate a user dynamic semantic vector. In addition, the global data of the item being interacted with is used to prompt the large language model to extract group perception features, which are then mapped to the vector space through the embedding layer to generate an item dynamic semantic vector. The user dynamic semantic vector and the item dynamic semantic vector are transformed using a multilayer perceptron to obtain the user dynamic profile vector and the item dynamic profile vector.

4. The recommendation method based on large model-enhanced user interest aggregation representation according to claim 1, characterized in that, K-means clustering is performed on the static image vectors of items to generate several interest prototype centers, specifically: Collect static image vectors of all items and perform data verification. Introduce particle swarm optimization algorithm to optimize the number of interest prototypes and obtain the optimal number of interest prototypes. Use cosine similarity as distance metric. Based on the optimal number of interest prototypes, the K-means++ algorithm is used to initialize the interest prototype centers. The cosine distance from the static portrait vector of each item to all prototype centers is calculated and assigned to the cluster to which the nearest center belongs. The mean vector of each cluster is recalculated as the new interest prototype center. Clustering stops when the change in the center of interest is less than the threshold or the maximum number of iterations is reached. Clusters with fewer than the preset threshold of samples are removed and reassigned to the nearest valid cluster to obtain the final center of interest. Interpretable labels are then generated through a large language model.

5. The recommendation method based on large model-enhanced user interest aggregation representation according to claim 1, characterized in that, For the interest prototype center of the target item, collaborative filtering specific attention, item dynamic attention, and user dynamic attention are calculated. Attention weights are then fused using a multilayer perceptron to generate a user interest representation, specifically: Based on the static profile vector of the target item, the nearest interest prototype center is retrieved from several pre-calculated interest prototype centers and used as the target interest prototype center. The user's historical interaction item set is extracted, which includes three types of features: collaborative filtering embedding vector, item dynamic profile vector, and user dynamic profile vector. For each type of feature, the attention weights with the target interest prototype center are calculated separately. In the specific attention calculation of collaborative filtering, the target interest prototype center is used as the query vector and the collaborative filtering embedding vector is used as the key value. The weights are calculated by scaling the dot product attention and then summed to obtain the collaborative filtering interest representation. In the dynamic attention calculation of items, the target interest prototype center is used as the query vector and the dynamic profile vector of the item is used as the key value. When calculating similarity, a time decay factor is introduced to generate attention weights and output the dynamic interest representation of the item in a weighted manner. In the calculation of user dynamic attention, the target interest prototype center is used as the query vector, the user dynamic profile vector is used as the key value, and the attention weight is calculated by adding the behavior type weight, and the user dynamic interest representation is output in a weighted manner. By integrating collaborative filtering interest representations, dynamic interest representations of items, and dynamic interest representations of users through a multilayer perceptron, a user interest representation is output.

6. The recommendation method based on large model-enhanced user interest aggregation representation according to claim 1, characterized in that, In offline pre-training, semantically difficult negative example contrastive learning is introduced. The set of difficult negative examples is selected by aligning and collaboratively filtering the embedding vectors and the static image vectors of the items using the InfoNCE loss function. Specifically: For each target item, candidate negative samples are screened from the set of uninterrupted items, items that the user is clearly not interested in are excluded, and the cosine similarity between the target item and the candidate negative samples is calculated based on the static profile vector of the item. Candidate negative samples with cosine similarity higher than a preset threshold are screened, and negative examples with similar behavior patterns to positive samples are excluded, thus forming a set of difficult negative examples. For the target item, extract the collaborative filtering embedding vector. and static image vector of items By using a projection head, the collaborative filtering embedding vector and the static image vector of the item are mapped to the same contrast space, thus eliminating distribution differences; Calculate positive sample pairs Similarity score Calculate the target item and the set of difficult examples. Each negative sample Similarity score The InfoNCE loss function is used for optimization, and contrastive learning is used to align the collaborative filtering embedding vectors with the item static image vectors. The contrastive loss... Represented as: ； in This represents the temperature parameter.

7. The recommendation method based on large model-enhanced user interest aggregation representation according to claim 1, characterized in that, In the online phase, a pre-computed vector is invoked to obtain the inner product of the user interest representation and the static profile of the candidate items, generating a recommendation score. Item recommendations are then made based on the recommendation score. Specifically: Based on the user ID, pre-computed user interest representations are read from the vector database in real time, and static profile vectors of items are loaded in batches from the cache based on the candidate item ID list. For each candidate item, the inner product of the item's static profile vector and the user's interest representation is calculated. The result of the inner product is used as the recommendation score. The candidate item list is sorted in descending order of the recommendation score. Category diversity control is introduced into the sorting result to avoid excessive homogenization. A preset number of candidate items are selected for recommendation.

8. A recommendation system based on large model-enhanced user interest aggregation representation, characterized in that, The recommendation method based on large model-enhanced user interest aggregation representation as described in any one of claims 1-7 includes: a basic data module, a large language model service module, a semantic hard negative example contrastive learning module, a user interest aggregation representation module, and an inference optimization module. The basic data module stores the name, description, and category attribute information of items, as well as user comment text and interaction behavior data; The large language model service module extracts explicit features based on the name, description, and category attributes of the item using the large language model, and infers implicit attributes based on world knowledge to generate static semantic vectors for the item. User comment text and interaction behavior data are input into the large language model to generate dynamic user profile vectors and dynamic item profile vectors. These vectors are then mapped to a unified representation space through an embedding layer and subjected to dimensional adjustment and normalization. The semantic hard negative example contrastive learning module introduces semantic hard negative example contrastive learning, selects the hard negative example set by using the InfoNCE loss function to align and collaboratively filter the embedded vector and the static image vector of the item, and updates the model parameters based on the contrastive loss through gradient descent. The user interest aggregation and representation module performs K-means clustering on the static profile vector of the item to generate several interest prototype centers. For the interest prototype centers of the target item, it calculates collaborative filtering specific attention, item dynamic attention, and user dynamic attention, and integrates the attention representation through a multilayer perceptron to generate user interest representation. The inference optimization module caches the profile vector, interest prototype center, and user interest representation offline. During the online phase, it calls the pre-computed vector to obtain the inner product of the user interest representation and the static profile of the candidate item, generates a recommendation score, and recommends items based on the recommendation score.

Images