A sequence recommendation method based on big language model for bilateral semantic portrait construction and personalized intent perception

Abstract

This invention belongs to the fields of artificial intelligence, data mining, and e-commerce, and discloses a sequence recommendation method based on a large language model for constructing two-sided semantic profiles and personalized intent perception. Utilizing a large language model and user-generated experience reviews, it characterizes two-sided profiles of users and stores from multiple highly adapted dimensions. Through semantic embedding and dimensionality reduction, the semantic profiles are converted into embedded representations adapted to the sequence recommendation task. Using cross-attention or self-attention mechanisms, attention is calculated based on the embedded representations of the store profile and the user profile in the user's historical interaction sequence to learn the user's potential intent and preferences, forming a personalized intent-perceived interaction sequence. Finally, sequence modeling and recommendation prediction are performed based on this. This invention can improve the feature representation capabilities of users and stores, effectively alleviate the long-tail problem in the field of sequence recommendation, enhance the effect of personalized recommendations, and has good application value.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, data mining, and e-commerce, and in particular to a sequence recommendation method based on a large language model for constructing bilateral semantic profiles and personalized intent perception. Background Technology

[0002] With the widespread use of online consumption and service platforms, user behaviors such as consumption, browsing, and evaluation on these platforms are increasingly showing a trend towards large-scale and diversified development. On various platforms, while users leave feedback such as reviews after completing a consumption or service experience, the platforms also record multi-dimensional information related to their preferences, consumption tendencies, and experience descriptions. This type of semantic text content has become a common and important form of information on these platforms. Existing research shows that this user-generated text information can provide important references for users' consumption decision-making process, and to some extent reduce decision uncertainty and improve selection efficiency. Generative content not only reflects users' genuine evaluations of specific stores, but also contains users' subjective feelings and preferences in specific consumption scenarios, which is of great significance for understanding user behavior and assisting decision-making.

[0003] In recent years, sequence recommendation systems have become an important branch of the recommendation field. Their core goal is to analyze the dynamic evolution of user behavior and understand users' potential interests based on historical interaction feedback between users and stores (such as browsing, purchasing, and reviewing), thereby predicting stores that users may be interested in and providing them with accurate personalized recommendations.

[0004] In the existing field of sequence recommendation systems, research has attempted to incorporate textual information to assist in user behavior modeling. One approach encodes textual attributes such as store names and categories, and integrates them as side information with label-based interaction sequence representations. Another approach introduces textual content such as user reviews into sequence recommendation models to utilize the semantic descriptive information contained within them. However, these systems still have several limitations: (1) Online consumption and service platforms (such as Amazon and Yelp) commonly contain semantically rich textual information such as store names, store categories, and user reviews. Existing methods mostly utilize store and user identification information and their historical interaction logs for model construction, which is relatively lacking in the utilization of the semantic information contained in the text, making it difficult to fully characterize user interests and store features. At the same time, for some methods that introduce review information into sequence recommendation tasks, they fail to effectively extract key information from the reviews during feature representation, making the recommendation results susceptible to the influence of irrelevant information. Therefore, a new method is needed to effectively utilize the widely existing textual information on the platform and improve the ability to characterize user and store features.

[0005] (2) In the field of recommendation systems, the long-tail problem has always been a key challenge, and it is also a typical manifestation of the Pareto principle ("Pareto Law") in the recommendation field. Existing recommendation algorithms often over-focus on the top 20% of popular merchants or trending stores, resulting in the bottom 80% of niche stores failing to reach target users due to insufficient exposure. This bias not only exacerbates the limitations of user interest exploration, but also inhibits the potential value of long-tail stores and damages the diversity of the platform ecosystem. At the same time, the long-tail effect also exists on the user side, that is, only a small number of users have rich interaction information, while the interaction sequences of most users are very short, making it difficult to discover potential interest features.

[0006] In conclusion, the long-tail problem is not only closely related to data sparsity, but also further exacerbates the insufficient characterization of user and store features. Therefore, balancing traffic distribution between the long tail and non-long tail segments and breaking the inherent constraints of the "Pareto Principle" (80 / 20 rule) is key to optimizing the fairness and user experience of recommendation systems. Summary of the Invention

[0007] To address the aforementioned problems, this invention proposes a sequence recommendation method based on a large language model for constructing bilateral semantic profiles and personalized intent perception, comprising the following steps: S1. In the profile construction stage enhanced by the large language model, based on the processing fluency theory and self-consistency theory, a prompt word engineering is designed to generate highly adapted dual-sided profiles for both the store and user sides. Specifically, based on the store's basic attribute information and overall store reviews, prompt word templates for the store-side profile are designed around three core dimensions: core store attributes, store feedback insights, and target customer groups. Based on users' historical review information, prompt word templates for the user-side profile are designed around three core dimensions: core user preference attributes, consumption scenario attributes, and consumption habit tendencies. The prompt words guide the large language model to generate store-side and user-side profiles from multiple core feature dimensions respectively. S2. In the semantic embedding stage of the user profile, the complex text information of the user profile and store profile is transformed into semantically rich embedded representations of the user profile and store profile respectively through the large language embedding model; the embedded representations are then linearly dimensionality reduced by principal component analysis. S3, the personalized intent perception stage, through cross-attention or self-attention mechanisms, performs attention calculation on the embedded representation of the user profile after dimensionality reduction in S2 and the embedded representation of the store profile contained in the user's historical interaction sequence, learns the user's potential intent and preferences, and forms a personalized intent perception interaction sequence. S4. In the sequence recommendation stage, the personalized intent-aware interaction sequence obtained in S3 is fed into the sequence recommender. Without changing the original structure of the sequence recommender, sequence modeling and recommendation prediction are performed. Specific patterns in the personalized intent-aware interaction sequence are learned and recommendation prediction is performed accordingly, thus achieving the adaptation of the personalized intent-aware module to different types of sequence recommenders.

[0008] The beneficial effects of this invention are as follows: This invention proposes a sequence recommendation method based on a large language model for constructing a two-sided semantic profile and perceiving personalized intent. First, through cue word engineering, the semantic understanding capability of the large language model is used to extract user-side profiles and store-side profiles with rich semantic information from user-generated content and store reviews, thereby improving the problem of insufficient utilization of textual semantic information in traditional sequence recommendation models. Second, this invention transforms the user-side profile and store-side profile into semantically rich embedded representations using a large language embedding model, and uses principal component analysis to reduce the feature dimension of the high-dimensional semantic embedded representations while retaining key information. Next, in the personalized intent perception stage, attention is calculated between the embedded representation of the dimensionality-reduced user-side profile and the embedded representation of the corresponding store-side profile in the user's historical interaction sequence through a cross-attention mechanism or a self-attention mechanism, to achieve effective fusion of user profile information and interaction sequence information, obtaining an overall intent preference perception representation. This overall intent preference perception at the sequence level is then weighted and fused with the user's original interaction sequence (the sequence of stores the user has visited historically) to obtain the user's personalized interaction sequence. Finally, the personalized intent perception result is introduced without changing the original sequence recommender structure, enhancing the adaptability of this invention to different types of sequence recommenders. During the model optimization phase, the main recommendation loss and auxiliary alignment loss are jointly introduced to ensure that the interest representations obtained from user behavior learning are consistent with the user semantic profiles, thereby improving the modeling effect for long-tail users and long-tail stores.

[0009] The method proposed in this invention can achieve personalized recommendations on real datasets and effectively alleviate the long-tail problem prevalent in the field of sequence recommendation. Its overall performance outperforms existing benchmark methods on commonly used sequence recommendation evaluation metrics such as NDCG@N, HR@N, and MRR. Furthermore, regarding recommendation performance for the long-tail problem, compared with other benchmark models, the framework proposed in this invention exhibits superior long-tail recommendation performance, and also significantly improves performance for non-long-tail users or stores without compromising recommendation performance on the non-long-tail side.

[0010] This invention also possesses significant practical application value. By fully mining the semantic information contained in user-generated content and attribute information, constructing semantic profiles on both the user and store sides, and introducing a personalized intent perception mechanism, it can provide online consumption and service platforms with a more accurate understanding of user preferences, thereby supporting the generation of more personalized recommendation results. Simultaneously, this invention can effectively alleviate the problem of insufficient characterization of long-tail users and stores in recommendation methods, helping platforms to more comprehensively identify potential needs and value objects from massive user and store data. This provides data-driven technical support for improving the overall effectiveness of recommendation methods, optimizing platform content distribution strategies, and promoting the overall balance and sustainable operation of the platform's recommendation system. Attached Figure Description

[0011] Figure 1 This is an architecture diagram of a sequence recommendation method based on a large language model for constructing two-sided semantic profiles and personalized intent perception. Detailed Implementation

[0012] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0013] A method for constructing bilateral semantic profiles and personalized intent perception based on a large language model is proposed. It realizes the construction of user and store profiles and personalized sequence recommendations through four core modules, namely: a profile construction module enhanced by a large language model, a profile semantic embedding module, a personalized intent perception module, and a sequence recommendation module.

[0014] In the first stage, the profile construction stage enhanced by the large language model, based on the processing fluency theory and self-consistency theory, through a carefully designed prompt word project, the large language model is guided to use its semantic understanding capabilities to extract semantic information from user evaluation information to form a user profile, and to extract semantic information from the basic attribute information and overall evaluation information of the store to form a store profile.

[0015] In the second stage, during the semantic embedding stage of the profile, the user-side profile and store-side profile formed in the first stage are converted into embedded representations suitable for sequence recommendation using a large language embedding model. Given that such embedded representations usually have high dimensionality, while the embedding vectors required by the sequence recommendation model have low dimensionality, in order to adapt to subsequent steps, this invention introduces principal component analysis to perform linear dimensionality reduction on the embedded representations, thereby obtaining low-dimensional embedded representations while retaining key semantic information.

[0016] The third stage, the personalized intent perception stage, achieves deep integration of user profiles and interaction sequences. This stage constructs a personalized intent perception mechanism, combining cross-attention and self-attention attention structures to perform attention calculations on the embedded representations of the user profile and the corresponding embedded representations of the store profiles in the user's historical interaction sequences. Specifically, the attention mechanism learns the user's potential interests and preferences, enabling the semantic information of the user profile to effectively enhance the interaction sequence representation, thereby forming a personalized intent perception interaction sequence.

[0017] The fourth stage involves personalized recommendation prediction during the sequence recommendation phase. This stage inputs the personalized intent-aware interaction sequence obtained in the third stage into the sequence recommender. By introducing personalized intent-aware results without altering the original sequence recommender structure, this enhances the method's adaptability to different types of sequence recommenders. The sequence recommender processes the interaction sequence sequentially through an embedding layer, a sequence encoder, and an optimization and prediction module, outputting recommendation results to predict which stores the user might be interested in later.

[0018] This invention relates to a sequence recommendation method based on a large language model for constructing bilateral semantic profiles and personalized intent perception, such as... Figure 1 As shown, this method mainly includes a large language model-enhanced profile construction module, a profile semantic embedding module, a personalized intent perception module, and a sequence recommendation module. The large language model-enhanced profile construction module, based on the basic attribute information of users and stores and user-generated content, uses pre-designed prompt word templates to guide the large language model to generate user profiles and store profiles from multiple semantic dimensions, thereby obtaining user-side profiles representing user preference features and store-side profiles representing store attribute features. Based on this, the profile semantic embedding module converts the text descriptions of the user-side profiles and store-side profiles into continuous embedded representations through the large language embedding model, and reduces the complexity of high-dimensional semantic embedding through dimensionality reduction processing, making it more suitable for sequence recommenders. Next, the personalized intent perception module constructs two attention mechanisms, self-attention or cross-attention, to weightedly fuse the embedded representations of the user-side profiles with the embedded representations of the corresponding store-side profiles in the user's historical interaction sequences, thereby learning an interaction sequence representation that reflects the user's personalized intent features, thus overcoming the problem of insufficient characterization of individual user differences in traditional sequence modeling. Finally, the sequence recommender module takes the personalized intent-aware interaction sequence as input and, without changing the original sequence recommender structure, introduces the personalized intent-aware representation to achieve integration with the existing sequence recommendation model. Through the embedding layer, sequence encoder, and optimization and prediction modules, it models user interaction behavior and outputs the final recommendation results.

[0019] (1) A profile building module enhanced by a large language model; Process fluency theory suggests that information with clear structure and explicit semantics is easier to understand and process; self-consistency theory indicates that users tend to choose stores that align with their own characteristics. Therefore, to construct clearly defined user and store profiles and achieve a high degree of adaptation between the two profiles, this invention designs store-side prompts from three aspects: core store attributes, store feedback insights, and target customer groups, to characterize the store-side profile; and designs user-side prompts from three aspects: core user preference attributes, consumption scenario attributes, and consumption habit tendencies, to characterize the customer-side profile.

[0020] Next, using a large language model ( LLM ) and carefully designed prompt templates , Store information is formed by combining basic store attribute information with overall evaluation information. Convert into store profile This will record all the stores visited by each customer and their corresponding reviews. Transform into user profile The above processes satisfy the following relationships: .

[0021] (2) Image semantic embedding module; The description texts of the store-side profile and user-side profile, enhanced by the large language model, are input into the large language embedding model and transformed into embedded representations containing deep semantic information from the profile text; thus forming the embedding matrix of the store profile respectively. Embedding matrix of user profiles , These represent the number of stores and the number of users, respectively. b The dimension of embedded representation; the use of a large language embedding model ( Construct an embedding matrix that satisfies: .

[0022] Next, principal component analysis is used to perform linear dimensionality reduction on the embedded representation to obtain the store profile embedding matrix. and user profile embedding matrix ,in, d It is the dimension of the embedded representation after dimensionality reduction; the dimensionality reduction mapping relationship satisfies: .

[0023] (3) Personalized intent perception module; Embedded representation of user profiles ) through linear projection matrix Mapped to query Q in cross-attention; embedding the store profile in the user interaction sequence ( , (Through two transformation matrices) The mapping is represented by key K and value V in cross-attention; an attention score is calculated based on the query Q and key K, and the score is applied to value V to achieve effective fusion of user profile information and interaction sequence information, resulting in an overall intent preference perception representation. ; to perceive the overall intent preference Broadcast to sequence level and the original user interaction sequence The user's personalized interaction sequence is obtained by weighting and fusing the sequence of stores visited in the past. : in, For fusion weighting coefficients.

[0024] Next, the original interaction sequence Embedded representation of user profiles The fusion representation is used as the query Q and key K for self-attention, while simultaneously incorporating the original interaction sequence. Here, as the value V, The self-attention mechanism introduces user profiles as auxiliary information into the interaction sequence, forming a personalized user perception sequence. When calculating the self-attention score, both Q and K incorporate user profiles to calculate the relationships between stores in the original user interaction sequence. Specifically, this can be represented as: in, Represents the original interaction sequence The attention score between the i-th store and the j-th store. To measure the semantic relevance between two stores, and Measure the relevance between each store and the user profile separately. It is a contextual bias term of user preferences; personalized interaction sequences are calculated through a self-attention mechanism. The process is as follows: Both the aforementioned cross-attention mechanism and self-attention mechanism are implemented using a multi-head attention structure.

[0025] (4) Sequence recommendation module; In step (3), the user profile is embedded to represent and the original user interaction sequence The personalized interactive sequence representation formed by fusion As input to the sequence recommender, this is a decoupled fusion mechanism. Specifically, the decoupled fusion enhances the semantic expression of user interaction sequences by introducing user profile semantic information on the input side, while maintaining the original structure of the sequence recommender. This achieves decoupling between the personalized intent perception module and the sequence recommendation model, and adapts to different types of sequence recommender models.

[0026] The sequence recommender mainly consists of three key parts: the embedding layer, the sequence encoder, and optimization and prediction.

[0027] ①Embedded layer; In the embedding layer, the position of the store in the user interaction sequence reflects the order in which users visit the store, which is extremely important for obtaining user behavior patterns. Therefore, the embedding of the store's position in the interaction sequence is crucial. With personalized interaction sequences Adding them together enhances the ability to express the order of each store at different time steps in the interaction sequence. The process is represented as follows: in, This indicates the first step in step S3 after semantic enhancement based on the user profile. k An embedded representation of each store.

[0028] ② Sequence encoder; In the sequence encoder, a Transformer-based coding structure is used to process the sequence representation to obtain the sequence representation. and sequence interest intention representation It satisfies the following relationship: in, This represents the sequence representation after passing through L layers of Transformers, specifically the representation of its last time step. This typically represents the sequence of interactions perceived by the recommender from the current personalized intent. The learned intent representation.

[0029] ③ Optimization and prediction; In obtaining the user sequence interest representation output by the sequence encoder and sequence interest intention representation Then, the sequence recommender is trained and recommendation results are generated through the optimization and prediction module; The binary cross-entropy loss function is used as the main loss of the sequence recommender. The calculation is as follows: in, Indicates user c Personalized interactive sequences Capture to time step t The resulting sequence representation, Positive samples are those from stores that actually interact with the customer. Negative samples; Let be the scoring function that matches the sequence representation with the candidate store embedding; its output is the predicted score. For the Sigmoid function; In the main loss In addition, auxiliary alignment loss is introduced. , minimize and The formula for minimizing the Euclidean distance in user semantic profile embedding is as follows: in, This represents the embedded representation of the user profile obtained after dimensionality reduction by principal component analysis based on the embedded representation obtained in step S2 from the large language semantic embedding model. Based on the aforementioned main loss and auxiliary alignment loss, the overall optimization objective of this research method is defined as: During the recommendation phase, the user sequence interest representation output by the sequence encoder is used. Dot-matrix with candidate stores Calculate the score ranking and sort based on the predicted score to generate a Top-N recommendation result set.

[0030] Model validation and analysis phase; ①Dataset description; Six real-world datasets were selected from the Yelp platform: four state datasets (AZ: Arizona, USA; IN: Indiana, USA; MO: Missouri, USA; ON: Ontario, Canada) and two city datasets (QC-M: Montreal, Quebec, Canada; NC-C: Charlotte, North Carolina, USA). To evaluate the prediction results of the recommendation method, user sequences with fewer than three interactions were removed. Basic descriptive statistical analysis was performed on the datasets, and long-tail data statistics are reported. A data overview is shown in Table 1.

[0031] Table 1 Overview and Descriptive Statistical Analysis of the Yelp Dataset

[0032] ②Performance analysis; SASRec was used as the sequence recommendation model in this invention. Considering that the personalized preference perception module contains two personalized attention mechanisms, we supplemented all experiments with a comparative analysis of cross-attention and self-attention. We evaluated the recommendation performance of LLM4Rec using three evaluation metrics (NDCG@N, HR@N, and MRR) on six datasets (Table 1). The best results are indicated in bold, the second-best results are indicated by underline, and the percentage improvement (values ​​in parentheses) is the result calculated compared to the second-best model.

[0033] The results are shown in Table 2. On all datasets, LLM4Rec significantly outperforms general recommendation models (such as NCF and LightGCN), benchmark sequence recommendation models (such as SASRec, BERT4Rec, and GRU4Rec), long-tail sequence recommendation models (such as MELT), personalized sequence recommendation models (such as SSE-PT), and sequence recommendation models based on large language models (such as LLMInit and LLMEmb). Compared with suboptimal models, the overall performance improvement of this invention (LLM4Rec) ranges from 1.6% to 15.4%. Furthermore, the percentage improvement in NDCG@N and MRR metrics is greater than that in HR@N, and the HR@1 metric shows an improvement of over 9% compared to suboptimal models. This indicates that the model described in this invention can recommend real stores to higher positions.

[0034] Table 2. Overall performance of the present invention and the benchmark model on 6 datasets.

[0035] To further examine the impact of large language models with different performance levels on the recommendation performance of the framework of this invention, we selected... As a baseline model, Arizona (AZ) and Charlotte, North Carolina (NC-C) were selected as experimental data to compare and analyze six large language models used to form profiles and six large language models used to form embedded representations.

[0036] For constructing large language models for text profiling, six large language models were added to the original glm-4-air: qwen3:30b, gemini2.5-flash, llama3.1:8b, Mistral-v0.3:7b, and Gemma3-4b, for a total of six. The performance changes of different large language models on the recommended method of this invention were compared and analyzed. Among them, glm-4-air and gemini2.5-flash are closed-source models, with gemini2.5-flash being a high-performance model; while qwen3:30b, llama3.1:8b, Mistral-v0.3:7b, and Gemma3-4b are open-source models with relatively fewer parameters. In this comparison, the semantic profiling was transformed into an embedded large language model (OpenAI:text-embedding-3-large). The results are shown in Table 3. Regardless of whether it is an open-source or closed-source model, the framework of this invention can bring similar performance, with a maximum fluctuation of 0.64%. Therefore, the profiling process is less affected by the performance of the large language models.

[0037] Table 3. Impact of profiles generated by different large language models on recommendation performance

[0038] For the process of transforming text profiles into embedded representations, six LLM embedding models were added to the original OpenAI:text-embedding-3-large: Google:gemini-embedding-exp-03-07, Qwen3-embedding-4B, snowflake-arctic-embed2, bge-m3, and jina-embeddings-v3, for a total of six models. The impact of different LLM embedding models on the performance of the push-line method was compared and analyzed. During this process, the large language model used to construct the profile (glm-4-air) remained unchanged. The results are shown in Table 4. After using the above six LLM embedding models within the framework of this invention, their performance varied slightly, with a maximum performance fluctuation of 1.66%. It can be seen that the two closed-source state-of-the-art (SOTA) embedding models showed relatively good recommendation performance. However, the performance of the other four open-source models differed significantly. Therefore, in the process of transforming complex text profiles into embedded representations, it is necessary to select high-performance LLM embedding models to form high-quality semantic embedded representations to adapt to the subsequent sequence recommender.

[0039] Table 4. Impact of different LLM embedding models on recommendation performance

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A sequence recommendation method based on a large language model for constructing bilateral semantic profiles and personalized intent perception, characterized in that, Includes the following steps: S1. In the profile building stage of large language model enhancement, based on the processing fluency theory and self-consistency theory, the prompt word project is designed to generate highly adapted dual profiles on the store side and the user side. Specifically, based on the store's basic attribute information and overall store reviews, design store profile prompt word templates around three core dimensions: core store attributes, store feedback insights, and target customer groups; Based on users' historical comments, user profile prompt templates were designed around three core dimensions: core user preference attributes, consumption scenario attributes, and consumption habit tendencies. The prompt words guide the large language model to generate store-side profiles and user-side profiles from multiple core feature dimensions; S2. In the semantic embedding stage of the user profile, the complex text information of the user profile and store profile is transformed into semantically rich embedded representations of the user profile and store profile respectively through the large language embedding model; the embedded representations are then linearly dimensionality reduced by principal component analysis. S3, the personalized intent perception stage, through cross-attention or self-attention mechanisms, performs attention calculation on the embedded representation of the user profile after dimensionality reduction in S2 and the embedded representation of the store profile contained in the user's historical interaction sequence, learns the user's potential intent and preferences, and forms a personalized intent perception interaction sequence. S4, Sequence Recommendation Stage: The personalized intent-aware interaction sequence obtained in S3 is fed into the sequence recommender. Sequence modeling and recommendation prediction are performed without changing the original structure of the sequence recommender. Learn specific patterns in personalized intent-aware interaction sequences and make recommendation predictions accordingly, thereby adapting the personalized intent-aware module to different types of sequence recommenders.

2. The sequence recommendation method based on a large language model for constructing bilateral semantic profiles and personalized intent perception according to claim 1, characterized in that, In step S1, through prompt word engineering, based on the processing fluency theory and self-consistency theory, prompt words are designed from three aspects: core store attributes, store feedback insights, and target customer groups to depict the store profile. User-side prompts are designed from three aspects: core user preference attributes, consumption scenario attributes, and consumption habit tendencies, in order to create a customer profile.

3. The sequence recommendation method based on a large language model for constructing bilateral semantic profiles and personalized intent perception according to claim 2, characterized in that, In step S1, a large language model and prompt word templates are used. , Store information is formed by combining basic store attribute information with overall evaluation information. Convert into store profile This will record all the stores visited by each customer and their corresponding reviews. Transform into user profile The above processes satisfy the following relationships: ， 。 4. The sequence recommendation method based on a large language model for constructing bilateral semantic profiles and personalized intent perception according to claim 3, characterized in that, In step S2, the description text of the store-side profile and user-side profile enhanced by the large language model is input into the large language embedding model and converted into an embedded representation containing deep semantic information in the profile text, forming the embedding matrix of the store profile respectively. Embedding matrix of user profiles , These represent the number of stores and the number of users, respectively. b The dimension of the embedded representation; the large language embedding model Constructing an embedding matrix satisfies: 。 5. The sequence recommendation method based on a large language model for constructing bilateral semantic profiles and personalized intent perception according to claim 4, characterized in that, In step S2, principal component analysis is used to perform linear dimensionality reduction on the embedded representation to obtain the store profile embedding matrix. and user profile embedding matrix ,in, d It is the dimension of the embedded representation after dimensionality reduction; the dimensionality reduction mapping relationship satisfies: 。 6. The sequence recommendation method based on a large language model for constructing bilateral semantic profiles and personalized intent perception according to claim 5, characterized in that, In step S3, the user profile is embedded in the representation. Through linear projection matrix Mapped to query Q in cross-attention; Embedded representation of store profiles in user history interaction sequences , ; through two transformation matrices The mapping is the key K and value V in cross-attention; An attention score is calculated based on the query Q and key K, and this attention score is applied to the value V to effectively integrate user profile information and interaction sequence information, thereby obtaining an overall intent preference perception representation. The overall intent preference perception representation Broadcast to sequence level And embedded representation of store profiles in the user's historical interaction sequence. By performing weighted fusion, the personalized interaction sequence of the current user is obtained. : in, For fusion weighting coefficients.

7. The sequence recommendation method based on a large language model for constructing bilateral semantic profiles and personalized intent perception as described in claim 6, characterized in that, In step S3, the embedded representation of the store profile in the user's historical interaction sequence is... Embedded representation of user profiles The fusion representation is used as the query Q and key K for self-attention, while simultaneously embedding the store profile representation from the user's historical interaction sequence. As the value V, ; The self-attention mechanism incorporates user profiles as auxiliary information into the user's historical interaction sequence, forming a personalized user perception sequence. When calculating the self-attention score, both Q and K incorporate user profiles to calculate the relationships between stores in the original user interaction sequence, specifically, as follows: in, Embedded representation of store profiles in the user's historical interaction sequence The attention score between the i-th store and the j-th store. To measure the semantic relevance between two stores, and Measure the relevance between each store and the user profile separately. It is a contextual bias term of user preferences; personalized interaction sequences are calculated through a self-attention mechanism. The process is as follows: Furthermore, both the aforementioned cross-attention mechanism and self-attention mechanism are implemented using a multi-head attention structure.

8. The sequence recommendation method based on a large language model for constructing bilateral semantic profiles and personalized intent perception according to claim 7, characterized in that, In step S4, the user profile embedded representation from step S3 is used... and the original user interaction sequence The personalized interactive sequence representation formed by fusion As input to the sequence recommender, it serves as a decoupled fusion mechanism. Specifically, this decoupled fusion mechanism enhances the semantic expression of user interaction sequences by introducing user profile semantic information on the input side, while maintaining the original structure of the sequence recommender. This achieves decoupling between the personalized intent perception module and the sequence recommendation model, and adapts to different types of sequence recommender models.

9. The sequence recommendation method based on a large language model for constructing bilateral semantic profiles and personalized intent perception according to claim 8, characterized in that, In step S4, the sequence recommender includes an embedding layer, a sequence encoder, and optimization and prediction. In the embedding layer, the position of a store in the user interaction sequence reflects the order in which users visit stores, which is crucial for obtaining user behavior patterns. The embedding of the store's position in the user interaction sequence... With personalized interaction sequences The addition enhances the sequential representation of each store at different time steps in the user interaction sequence, and the process is expressed as follows: in, This indicates the first step in step S3 after semantic enhancement based on the user profile. k Embedded representation of each store; In the sequence encoder, a Transformer-based coding structure is used to process the sequence representation to obtain the sequence representation. and sequence interest intention representation It satisfies the following relationship: in, This represents the sequence representation after passing through L layers of Transformers, specifically the representation of its last time step. This typically represents the sequence of interactions perceived by the recommender from the current personalized intent. The learned intent representation.

10. The sequence recommendation method based on a large language model for constructing bilateral semantic profiles and personalized intent perception according to claim 9, characterized in that, In step S4, the user sequence interest representation output by the sequence encoder is obtained. and sequence interest intention representation Then, the sequence recommender is trained and recommendation results are generated through the optimization and prediction module; The binary cross-entropy loss function is used as the main loss of the sequence recommender. The calculation is as follows: in, Indicates user c Personalized interactive sequences Capture to time step t The resulting sequence representation, Positive samples are those from stores that actually interact with the customer. Negative samples; Let be the scoring function that matches the sequence representation with the candidate store embedding; its output is the predicted score. For the Sigmoid function; In the main loss In addition, auxiliary alignment loss is introduced. , minimize and The formula for minimizing the Euclidean distance in the user semantic profile embedding is as follows: in, This represents the embedded representation of the user profile obtained after dimensionality reduction by principal component analysis based on the embedded representation obtained in step S2 from the large language semantic embedding model. Based on the aforementioned main loss and auxiliary alignment loss, the overall optimization objective of this research method is defined as: During the recommendation phase, the user sequence interest representation output by the sequence encoder is used. Dot-matrix with candidate stores Calculate the predicted score, sort the results based on the predicted score, and generate a Top-N recommendation result set.

Images