Multi-dimensional content recommendation method and system based on llm-rag-lbs fusion

Abstract

The application provides a multi-dimensional content recommendation method and system based on LLM-RAG-LBS fusion, which comprises the following steps: constructing a behavior feature vector according to a user's trajectory point sequence, inputting the behavior feature vector into a stage classification model to obtain a behavior stage label, and obtaining a structure template according to the behavior stage label and a label template mapping table. A language expression control vector is calculated according to the behavior feature vector, a regional semantic regulation matrix and a platform recommendation preference coefficient. A structured knowledge base comprises a plurality of candidate content segments, and each candidate content segment corresponds to a semantic vector and a structure label. A weighted adaptive score of the candidate content segment is obtained according to the language expression control vector, the semantic vector and the platform recommendation preference coefficient, and the candidate content segment with the highest score is selected as a selected content segment. A target content segment is selected from the selected content segment according to the structure template and the structure label, a recommended content segment is generated according to the target content segment and the language expression control vector, and the recommended content is recommended, thereby improving the accuracy of the recommended content.

Description

Technical Field

[0001] This invention belongs to the field of computer science, and in particular relates to a multi-dimensional content recommendation method and system based on LLM-RAG-LBS fusion. Background Technology

[0002] As generative AI technologies, represented by large language models, gradually enter the fields of search and information retrieval, users' information acquisition methods are shifting from traditional search centered on keyword matching to AI search models based on semantic understanding and generative responses. In this process, whether content can be correctly understood, fully utilized, and prioritized by AI systems is gradually becoming a key factor influencing information exposure and traffic allocation. Especially driven by mainstream platforms such as Baidu, Doubao, and DeepSeek, AI-generated results have become an important entry point for user decision-making, and the competition in content recommendation has shifted from simple retrieval ranking to a comprehensive adaptation capability centered around "understanding—generation—display." However, most existing content recommendation and generation systems still rely on keyword or static tag-based technologies, with their understanding of users' real needs remaining at a superficial semantic level. They struggle to conduct in-depth analysis based on users' actual behavioral patterns, leading to inaccurate recommendations. Therefore, improving the accuracy of recommended content has become an urgent technical problem to be solved. Summary of the Invention

[0003] The purpose of this invention is to design a multi-dimensional content recommendation method and system based on LLM-RAG-LBS fusion, which can improve the accuracy of recommended content.

[0004] To achieve the above objectives, a first aspect of the present invention provides a multi-dimensional content recommendation method based on LLM-RAG-LBS fusion, the method comprising: Obtain the user's trajectory point sequence, and construct a behavioral feature vector based on the trajectory point sequence; The behavioral feature vector is input into a preset stage classification model to obtain behavioral stage labels, and a structural template is matched according to the behavioral stage labels and a preset label template mapping table. The language expression control vector is calculated based on the behavioral feature vector, the preset regional semantic control matrix, and the preset platform recommendation preference coefficient; wherein, the preset structured knowledge base includes multiple candidate content fragments, and each candidate content fragment corresponds to a semantic vector and a structural label; Each candidate content segment is scored based on the language expression control vector, the semantic vector, and the platform recommendation preference coefficient to obtain a weighted adaptation score, and the top-performing segments are selected. The candidate content segment with the highest weighted adaptation score mentioned above is selected as the content segment. The selected content fragments are filtered according to the structural template and the structural tags to obtain target content fragments. Recommended content fragments are generated according to the target content fragments and the language expression control vector, and the recommended content fragments are recommended.

[0005] Further, the step of obtaining the user's trajectory point sequence and constructing a behavioral feature vector based on the trajectory point sequence includes: Obtain the user's trajectory point sequence, construct a service interest vector based on the trajectory point sequence, and extract the trajectory duration, total trajectory distance, and number of high-frequency dwell areas from the trajectory point sequence; The behavioral feature vector is obtained by concatenating the service interest vector, the trajectory duration, the total trajectory distance, and the number of high-frequency dwell areas.

[0006] Furthermore, the trajectory point sequence includes the geographic coordinates of each time point, and the step of constructing a service interest vector based on the trajectory point sequence includes: The city grid number is matched with each of the geographic coordinates and a preset coordinate grid mapping table; wherein each of the city grid numbers corresponds to a service category; The service interest vector is constructed based on all the service categories according to the city grid number.

[0007] Furthermore, the process of recommending content fragments includes: The recommended content fragments are input into a preset content language style analysis model to obtain semantic dimension values; The platform adaptation score is obtained by scoring based on the semantic dimension value, the language expression control vector, and the preset platform adaptation parameters. If the platform adaptation score is lower than a preset platform threshold, the recommended content segment is adjusted according to a preset language style template library to obtain a first recommended content adjusted segment, and the first recommended content adjusted segment is recommended; otherwise, the recommended content segment is recommended directly.

[0008] Furthermore, the process of recommending content fragments includes: The recommended content fragments are matched against a preset regional tag library to obtain the high-frequency word hit frequency; The regional adaptation score is calculated based on the hit frequency of the high-frequency words, the number of embedded fields in the regional tag library, and the preset platform weights. If the regional adaptation score is less than a preset regional threshold, the recommended content segment is adjusted according to a preset regional tag library to obtain a second recommended content adjusted segment, and the second recommended content adjusted segment is recommended; otherwise, the recommended content segment is recommended directly.

[0009] Further, the step of calculating the language expression control vector based on the behavioral feature vector, the preset regional semantic control matrix, and the preset platform recommendation preference coefficient includes: A language fusion term is constructed based on the behavioral feature vector and the regional semantic regulation matrix; An imbalance penalty term is constructed based on the platform recommendation preference coefficient and the behavioral feature vector; The language fusion term is subtracted from the imbalance penalty term to obtain the language expression control coefficients; wherein, all the language expression control coefficients constitute the language expression control vector.

[0010] Further, the step of scoring each candidate content segment based on the language expression control vector, the semantic vector, and the platform recommendation preference coefficient to obtain a weighted adaptation score includes: Construct a weighted term for expression style based on the language expression control vector and the semantic vector; Construct a platform recommendation sensitivity constraint term based on the platform recommendation preference coefficient and the language expression control vector; The local weighted score is obtained by subtracting the expression style weighting term from the platform recommendation sensitivity constraint term. The local weighted scores of all semantic dimensions are then added together to obtain the weighted adaptation score.

[0011] A second aspect of the present invention provides a multi-dimensional content recommendation system based on LLM-RAG-LBS fusion, the system comprising: The acquisition unit is used to acquire the user's trajectory point sequence and construct a behavioral feature vector based on the trajectory point sequence; The matching unit is used to input the behavior feature vector into a preset stage classification model to obtain behavior stage labels, and match the structure template according to the behavior stage labels and a preset label template mapping table. The calculation unit is used to calculate the language expression control vector based on the behavioral feature vector, the preset regional semantic control matrix and the preset platform recommendation preference coefficient; wherein, the preset structured knowledge base includes multiple candidate content fragments, and each candidate content fragment corresponds to a semantic vector and a structural label; The scoring unit is used to score each candidate content segment based on the language expression control vector, the semantic vector, and the platform recommendation preference coefficient, to obtain a weighted adaptation score, and select the top [segment]. The candidate content segment with the highest weighted adaptation score mentioned above is selected as the content segment. The recommendation unit is used to filter the selected content fragments according to the structure template and the structure tag to obtain target content fragments, generate recommended content fragments according to the target content fragments and the language expression control vector, and recommend the recommended content fragments.

[0012] In a third aspect of the invention, an electronic device is provided, the electronic device including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the method described in the first aspect above.

[0013] In a fourth aspect of the invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect.

[0014] The beneficial technical effects of the present invention are at least as follows: To address the aforementioned issues, this invention provides a multi-dimensional content recommendation method and system based on LLM-RAG-LBS fusion. Its core lies in introducing a collaborative modeling mechanism integrating user behavior, regional semantics, and platform preferences throughout the entire content generation and delivery process. This allows content to be precisely guided before generation, gain authoritative support during generation, and achieve effective structural and scenario adaptation after generation. Based on user trajectory point sequences, this invention structurally characterizes their service interests and behavioral stages, providing semantic control for content generation that aligns with real-world scenarios. Furthermore, by combining user intent with regional linguistic features, a generation control structure reflecting semantic differences across cities is constructed, ensuring content expression naturally conforms to local cognitive habits. Further, by integrating structured retrieval information with the generation control mechanism, generated content maintains consistent expression while possessing traceable factual support and a stable structural form. Finally, by combining recommendation preferences and regional tags from different platforms, the generated content undergoes structural reorganization and expression enhancement, making it more consistent with the platform's recommendation system's evaluation logic for content quality and presentation without altering its core semantics. Through the above overall technical design, this invention achieves continuous optimization from understanding user behavior to content generation and then to platform recommendation and display, effectively improving the adaptability and priority display capability of content in the recommendation system. It is particularly suitable for application scenarios with high requirements for geographic accuracy, credibility of expression and recommendation stability. Attached Figure Description

[0015] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0016] Figure 1This is a flowchart of a multi-dimensional content recommendation method based on LLM-RAG-LBS fusion provided in an embodiment of this application.

[0017] Figure 2 This is a schematic diagram of the structure of a multi-dimensional content recommendation system based on LLM-RAG-LBS fusion provided in an embodiment of this application. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0019] Please refer to Figure 1 , Figure 1 This is a flowchart of a multi-dimensional content recommendation method based on LLM-RAG-LBS fusion provided in this application embodiment. The Large Language Model (LLM) is used to understand and generate content based on user needs and text semantics. Retrieval-Augmented Generation (RAG) is used to improve content accuracy by combining external knowledge bases or authoritative information during content generation. Location-Based Service (LBS) is used to acquire and utilize users' geographic location information to achieve region-related content recommendations. Figure 1 The method may include, but is not limited to, steps S101 to S105.

[0020] Step S101: Obtain the user's trajectory point sequence and construct a behavioral feature vector based on the trajectory point sequence; Step S102: Input the behavior feature vector into the preset stage classification model to obtain the behavior stage label, and match the structure template according to the behavior stage label and the preset label template mapping table. Step S103: Calculate the language expression control vector based on the behavioral feature vector, the preset regional semantic control matrix, and the preset platform recommendation preference coefficient; wherein, the preset structured knowledge base includes multiple candidate content fragments, each of which corresponds to a semantic vector and a structural label. Step S104: Each candidate content segment is scored based on its language expression control vector, semantic vector, and platform recommendation preference coefficient to obtain a weighted adaptation score. The top-performing segments are then selected. The candidate content segment with the highest weighted adaptation score is selected as the content segment. Step S105: Filter the selected content fragments according to the structure template and structure tags to obtain the target content fragments, generate recommended content fragments according to the target content fragments and language expression control vectors, and recommend the recommended content fragments.

[0021] In step S101 of some embodiments, the user's trajectory data may come from the positioning module in the mobile application. After authorization by the user, coordinate points are periodically collected through the mobile device to form a time-ordered sequence of trajectory points. As shown below: ; in, , Indicates the user's position in the first month. Time points Geographic coordinates This represents the total number of trajectory points. Each trajectory point is searched for in a pre-defined coordinate grid mapping table based on its geographic coordinates, and matched with a city grid number. The city was divided using a regular grid method. All grids have been processed offline and labeled as different service category areas, denoted as... Indicates service category The corresponding grid sets, such as education and training, legal services, and healthcare.

[0022] Based on the city grid number, a service interest vector is constructed for all service categories. This involves statistically analyzing the frequency of user behavior on each service grid category to build the service interest vector. Each dimension Indicates the user's service category The activity level on the platform is calculated using the following formula: ; in, Indicates the user's service category Activity levels reflect the extent to which user behavior is covered within that service area; For indicator functions, when Belongs to the service category Grid set hour, The value is 1 if it is not 0 otherwise; Service Interest Vector This indicates the user's interest in various service types within the current region. Number of service categories.

[0023] To determine the current stage of a user's behavior, three behavioral features are extracted from the trajectory point sequence: trajectory duration. Total distance traveled on the trajectory Number of high-frequency dwell areas The trajectory duration is calculated by the difference between the first and last timestamps in the trajectory point sequence, i.e. The total trajectory distance is calculated by summing the Euclidean distances between adjacent trajectory points, as follows: ; in, The total distance traveled on the trajectory indicates the degree of movement of the area covered by the trajectory. By traversing the sequence of trajectory grid numbers, the number of grids whose repetition exceeds a set threshold is counted to measure whether a user repeatedly enters a certain type of service area.

[0024] Ultimately, the service interest vector will be... Concatenate the three behavioral features to form a behavioral feature vector. As shown below: ; In step S102 of some embodiments, the behavior feature vector is... Input a predefined stage classification model and output behavior stage labels. Specifically, the model structure is a two-layer fully connected network with 32-dimensional hidden layers. The output is a probability distribution of 6 behavior labels, and the behavior label corresponding to the highest probability is taken as the final output behavior stage label. This indicates the current stage of the user's behavior. Possible values ​​include: browsing, comparison, consultation, appointment, on-site visit, follow-up, etc. The model is trained using existing user behavior data on the platform and is used for static inference after deployment.

[0025] Furthermore, the behavioral stage labels Convert to structural template number This is used to select a structural template for recommended content, controlling the logical order, information density, and layout of the output content. The structural template number can be matched to a tag template mapping table to obtain the corresponding structural template, or it can be transformed using an adjustable function to achieve stage sensitivity control. Let... ,in A set of platform-tunable structure mapping functions, for example: like =Browse, then =1 (Only generate a brief description); like =In comparison, then =3 (Generate multi-brand comparison paragraphs); like =Consultation, then =4 (Generation scheme description + data support); like =On-site visit, then =5 (Generate address map + operation process); The mapping function is set by the platform and can be optimized in stages based on the results of A / B experiments, so that the content structure not only meets the semantic expression requirements, but also matches the user behavior conversion path.

[0026] In step S103 of some embodiments, the aim is to determine the service interest vector. Behavioral stage labels By combining the user's current location and expression style preferences, a language expression control vector with regional semantic adaptability and behavioral context consistency is constructed. It is used to drive large language models to generate content.

[0027] Specifically, the input for this step includes service interest vectors. and behavioral stage labels . This indicates the degree of attention a user pays to different service types during their current activity trajectory, such as... This may indicate that users are highly active in using "intellectual property" services; These variables represent the user's current behavioral stage; for example, when the user is in the "consultation" stage, recommended content needs to emphasize comparison and specificity. These variables determine what content should be generated and how it should be expressed. Furthermore, a new regional semantic control matrix is ​​introduced. Its structure is It is obtained through regional pre-training, among which Indicates the first The service category in the current city is numbered as follows. The control coefficient for each language control dimension. Dimension This corresponds to key semantic features in the content, such as tone of voice, paragraph length preference, keyword richness, and whether local entity information is embedded.

[0028] Because recommendation platforms do not have consistent preferences for "localized expression," for example, Baidu's recommendation system prefers content with a complete structure and clear authoritative expression, while Doubao platform prefers content that is more lifelike, colloquial, and has a vivid scene. Therefore, this step introduces a platform recommendation preference coefficient. This is used to dynamically adjust the importance of different dimensions to different platforms in the language expression control vector. The platform recommendation preference coefficient is modeled according to the platform recommendation weight curve and does not change during model operation, thus being regarded as a static platform parameter. A language fusion term is constructed based on the service interest vector and the regional semantic regulation matrix in the behavioral feature vector; an imbalance penalty term is constructed based on the platform recommendation preference coefficient and the behavioral feature vector; the language expression control coefficient is obtained by subtracting the language fusion term from the imbalance penalty term. As shown below: ; in, Let be the language expression control coefficient, representing the first digit. The fusion output value of the semantic dimension is used to control the first semantic dimension of the recommended content. One semantic direction; Indicates the user's opinion on the first Activity levels for each service category; For the first Under the service category, the first The control coefficient for each semantic dimension; This represents the average activity level across all service categories. The platform recommendation preference coefficient represents the platform's tolerance for fluctuations in the expression of this semantic dimension. If a certain dimension has a significant impact on the platform's recommendations (e.g., structural integrity is highly important on Baidu), then its... The values ​​are relatively small to suppress fluctuations in this direction. All language expression control coefficients constitute the language expression control vector. , of which each This is a language expression control coefficient used to control the generated recommended content in the first... Expression style in several semantic dimensions, such as whether to use exclamatory sentences, whether to emphasize numbers, and whether to embed local examples.

[0029] The first term of the formula is a language fusion term, which is a weighted combination of user interest weights and regional language parameters. The second term is an imbalance penalty term; when the differences in user interests across categories are too large, the penalty value increases to suppress the negative impact of expression imbalance on platform recommendations. The introduction of this platform penalty term reflects the combined control of platform preferences and semantic expression stability in this step, and also embodies the integrated design concept of the "content-as-a-service recommendation strategy" in this invention.

[0030] The above steps involve, for the first time, integrating and modeling users' multi-service attention with regional language expression habits to form a semantic expression control vector; secondly, considering the sensitivity of the platform's recommendation system to language style and information structure, introducing a platform penalty term to regulate the generated input; and thirdly, the selection of the structural template depends not only on the user's stage tags but also on the mapping function configured by the platform. Achieve dynamic matching of behavior and structure. This step is not simply about constructing a Prompt, but rather about structuring linguistic expressions into adjustable control vectors. This is the structural embedding point of LLM generation in the regional recommendation system, serving as the core bridge in the "understanding-generation-adaptation" closed loop.

[0031] In step S104 of some embodiments, the aim is to construct a content generation module that integrates structured retrieval and generation control mechanisms. The recommendation scenario addressed by this invention emphasizes the goal of "platform-priority recommendation," requiring that the generated recommendation content not only meets users' personalized expression needs but also adapts to the structural preferences and information density requirements of the platform's recommendation system. Therefore, the generation process in this step must address three key challenges: first, ensuring that the semantic style of the generated content is consistent with the platform's recommendation logic; second, ensuring that the generated structure meets the preset template and does not experience structural drift; and third, ensuring that the generated content has credible data support and meets the platform's scoring requirements for content authority (such as the EEAT standard). To this end, a RAG-style generation scheme that integrates control over the generation structure, semantic offset penalty, and authoritative information injection mechanisms is proposed.

[0032] The scheme includes three sub-processes: one is based on Structured semantic matching retrieval; secondly, combining The structure template controls the generation of content paragraphs; thirdly, semantic and structural regular expressions are introduced for each paragraph to ensure that the generation process is controllable and friendly to the final recommendation system.

[0033] Specifically, a structured knowledge base is pre-set, which includes multiple candidate content fragments. Each candidate content fragment corresponds to a semantic vector. and structural tags Among them, candidate content fragments are pre-selected. Through keyword extraction, sentence structure analysis, and tag alignment processing, 3D semantic vector Structural tags This indicates the template paragraph position where the corresponding content segment can be inserted, such as "Problem Introduction Paragraph", "Solution Paragraph", "Case Study Paragraph", etc.

[0034] Each candidate content segment is scored based on its language expression control vector, semantic vector, and platform recommendation preference coefficient to obtain a weighted fit score. Specifically, an expression style weighting term is constructed based on the language expression control vector and semantic vector; a platform recommendation sensitivity constraint term is constructed based on the platform recommendation preference coefficient and language expression control vector; the expression style weighting term and the platform recommendation sensitivity constraint term are subtracted to obtain a local weighted score; and the local weighted scores of all semantic dimensions are summed to obtain the weighted fit score. As shown in the following formula: ; The first term of the formula The second item is a weighted item for the style of expression. Recommend sensitivity constraints for the platform. For candidate content fragments Weighted adaptation score, language expression control vector , of which each This is a control factor for language expression. for The mean value reflects the overall level of semantic balance. The platform recommends a preference coefficient. This is a semantic vector. When the expression in a certain semantic dimension fluctuates too much (i.e., (If the content deviates too much from the overall style), the platform's recommendation system may consider it style instability and lower the recommendation score. Therefore, this item was added to adjust the expressive balance when selecting content segments. Based on the weighted adaptation score... Sort from highest to lowest, and select the top... The candidate content fragments are selected as the selected content fragments to form the information injection pool. .

[0035] In step S105 of some embodiments, according to the structural template number Corresponding structural template The generated content is divided into Each paragraph is defined by a template structure, specifically including a problem background paragraph, a solution paragraph, a data reference paragraph, and a case study or evaluation paragraph. For example... The corresponding structure might be: "Problem Background Section + Three-Section Solution Section + Data Reference Section + Customer Evaluation Section", that is... It is divided into 6 paragraphs, with the 5th paragraph being the "data reference paragraph".

[0036] Indicates the first The structural tags corresponding to the content of the paragraph are used to generate the first paragraph. When writing paragraph content, from Select structural label matching as The selected content segment is used as the target content segment, and recommended content segments are generated based on the target content segment and the language expression control vector. Specifically, the following paragraph generation function is used to generate recommended content segments: ; in, Indicates the first A recommended excerpt. A function to generate paragraphs. The target content segment, that is, the information segment used in the current paragraph. For semantic expression control vectors, For semantic offset tolerance set, These are structural integrity constraint parameters. Recommended content fragment set. Each content segment includes semantic style control tags, information fragment source annotations, and platform recommendation structure hint fields. The generated results not only reflect user interests and regional expression preferences in style but also have real data support at the content level and meet the platform's optimization rules in structure. Finally, all recommended content segments are recommended. It should be noted that in generating the first... When processing segment content, first inject information from the information pool based on the corresponding structural tags of the segment. Selected content segments with consistent structural tags are filtered. For example, when the first... When the structure tag for a paragraph is "Problem Background Paragraph", it should be selected first from... Select content segments also labeled "Problem Background" as target content segments. When multiple content segments meet the criteria, they are sorted according to their weighted adaptation scores, and the content segment with the highest score is selected as the target content segment. If no content fragment completely matches the structure tag, then a semantically similar and structurally compatible selected content fragment is chosen as the target content fragment. For example, a fragment that can describe the current state of the industry or user pain points can be selected to supplement the background of the problem. This ensures the continuity of the content generation process and the integrity of the overall structure.

[0037] In one example, to ensure consistency in semantic style and structural information among the generated recommended content fragments, the following paragraph penalty regular expression is defined. : ; in, For the generation of the first The semantic dimension value of the recommended content fragment is calculated through the semantic vector extraction layer built into the model; This is a control coefficient for language expression; This is the semantic tolerance threshold; This indicates whether the segment has filled in the information fields required by the structure template. If the structure is complete, it is 1; otherwise, it is 0. This indicates the number of semantic dimensions contained in the semantic expression control vector. This represents the structural integrity penalty coefficient.

[0038] This regular expression is not used for backpropagation, but rather for runtime evaluation and as the trigger for the paragraph replacement mechanism. When When the threshold is exceeded, it will be replaced. The used content fragments are used to regenerate recommended content fragments. This mechanism ensures that even during the unsupervised inference stage of the model, content can be corrected through structure-semantic coupling penalties, improving the adaptability of content in platform structure review and recommendation models.

[0039] Steps S101 to S105, as illustrated in this embodiment, involve acquiring the user's trajectory point sequence and constructing a behavioral feature vector based on the sequence. The behavioral feature vector is input into a preset stage classification model to obtain behavioral stage labels. A structural template is then matched to the behavioral stage labels using a preset label template mapping table. A language expression control vector is calculated based on the behavioral feature vector, a preset regional semantic control matrix, and a preset platform recommendation preference coefficient. The structured knowledge base includes multiple candidate content fragments, each corresponding to a semantic vector and a structural label. Each candidate content fragment is scored based on the language expression control vector, the semantic vector, and the platform recommendation preference coefficient to obtain a weighted adaptation score. The top-performing fragments are then selected. The candidate content fragment with the highest weighted adaptation score is selected as the chosen content fragment. The selected content fragment is then filtered based on the structural template and structural tags to obtain the target content fragment. Recommended content fragments are then generated based on the target content fragment and the language expression control vector, and these recommended content fragments are then submitted, thus improving the accuracy of the recommended content.

[0040] In some embodiments, after step S105, the deployment and structural adaptation of recommended content for multiple platforms aims to optimize the set of recommended content fragments without altering the core semantics and factual content. Reconstruct the code based on platform preferences and regional tags to better align with the language style, structural requirements, and regional expression strategies of the recommendation platform.

[0041] Specifically, the sources of platform recommendation preference data include two categories: first, platform-side content specification documents, such as Baidu's content review documents explaining structured fields, content length, and keyword density; and second, historical recommendation sample data obtained through the platform's content delivery interface. The platform's API typically returns metrics such as the number of impressions, click-through rate, and indexing status for each piece of content. After clustering and semantic vectorization of this data, the platform's recommendation weight configuration under different service categories and regions can be summarized, constructing a set of platform adaptation parameters. Each of them Indicates the first Segment content in semantic dimension The degree to which content is favored by the platform. For example, Doubao platform gives higher weight to "local cases" and "life scenario descriptions", while Baidu platform prefers content with "qualification certification" and "clear structure".

[0042] Taking "Guangzhou Intellectual Property Service Recommendation" as an example, the generated target content fragment The phrase "Bangzhuan Intellectual Property Services has served over 3,000 companies in Guangzhou" leans towards a more conversational tone on the Doubao platform, while on the Baidu platform, its structural tags, such as "qualification type" and "customer type," need to be strengthened. Therefore, this step requires rewriting and structurally expanding the content according to platform preferences without changing the meaning of the sentence. This process involves not only identifying the corresponding semantic dimension values ​​within the paragraph, but also... It also needs to be based on the language expression control coefficient defined by the platform. Perform semantic difference alignment.

[0043] The platform adaptation score is calculated based on semantic dimension values, language expression control vectors, and preset platform adaptation parameters. As shown below: ; in, For the first The platform adaptation score of the recommended content on the target platform. Let be the language expression control coefficient, which is the th in the language expression control vector. The fusion output value of each semantic dimension For content segment The actual value of the current expressive style in this dimension is calculated using a content-based language style analysis model. This model is a three-layer feedforward network: the first layer is the word vector input, the second layer is the syntactic structure encoding, and the third layer outputs the same dimension as the input. The same semantic distribution vector.

[0044] When the platform compatibility score Below the preset platform threshold When the platform's adaptation score is reached, a paragraph adjustment process will be triggered. This process is implemented by the content style replacement module, which has a built-in platform-specific language style template library. The recommended content segment is adjusted according to the language style template library to obtain the first recommended content segment with adjustments, which is then recommended. Otherwise, the original recommended content segment is recommended directly. For example, if the structure is "too simplistic," the content segment will be enhanced into a composite structure including background description, indicator data, and customer reviews. The replacement method combines rule-based replacement with template filling to ensure controllability and consistency. Greater than or equal to the preset platform threshold At that time, recommended content snippets can be recommended directly.

[0045] In one example, the regional information supplementation module is implemented through regional tag matching and linkage with hot keyword templates. In the generated recommended content snippets, if the regional embedding is insufficient (e.g., the content does not contain "Guangzhou" or its landmark terms "Zhujiang New Town" or "Tianhe District"), the regional tag enhancement mechanism will be triggered. Regional Tag Library Derived from the local service database and the platform's recommended thesaurus, such as "Guangzhou Intellectual Property Service Center" and "Guangzhou Pure Enjoy Fitness Tianhe Flagship Store," the system will insert these phrases, where semantic location allows, to enhance the density of local expression. The regional adaptation score is calculated based on the frequency of high-frequency word hits, the number of embedded fields in the regional tag library, and preset platform weights. As shown below: ; in, Indicates the first The regional adaptation score of the content segment. This refers to the frequency of high-frequency word hits, that is, the frequency of local high-frequency words hitting within this segment. This represents the number of embedded fields from the geographic tag library; and The weights assigned to the platform are used to control the balance between language naturalness and regional density. If If the recommended content segment is adjusted based on the regional tag library, a second adjusted recommended content segment is obtained and recommended. Otherwise, the original recommended content segment is recommended directly. The adjustment method can be to add localized scenario sentences or replace some general expressions based on the regional tag library. It should be noted that when multiple local high-frequency words exist, the hit frequency is obtained by counting and summing the occurrences of all local high-frequency words in the content segment; specifically, it can be the sum of the hit counts of each local high-frequency word.

[0046] Please see Figure 2 This application also provides a multi-dimensional content recommendation system based on LLM-RAG-LBS fusion, which can implement the above-mentioned multi-dimensional content recommendation method based on LLM-RAG-LBS fusion. The system includes: The acquisition unit 201 is used to acquire the user's trajectory point sequence and construct a behavior feature vector based on the trajectory point sequence; The matching unit 202 is used to input the behavior feature vector into the preset stage classification model to obtain the behavior stage label, and match the structure template according to the behavior stage label and the preset label template mapping table. The calculation unit 203 is used to calculate the language expression control vector based on the behavioral feature vector, the preset regional semantic control matrix and the preset platform recommendation preference coefficient; wherein, the preset structured knowledge base includes multiple candidate content fragments, and each candidate content fragment corresponds to a semantic vector and a structural label; Scoring unit 204 is used to score each candidate content segment based on the language expression control vector, semantic vector, and platform recommendation preference coefficient, obtaining a weighted adaptation score, and selecting the top... The candidate content segment with the highest weighted adaptation score is selected as the content segment. Recommendation unit 205 is used to filter selected content fragments based on structural templates and structural tags to obtain target content fragments, generate recommended content fragments based on target content fragments and language expression control vectors, and recommend the recommended content fragments.

[0047] The specific implementation of this multi-dimensional content recommendation system based on LLM-RAG-LBS fusion is basically the same as the specific implementation of the multi-dimensional content recommendation method based on LLM-RAG-LBS fusion described above, and will not be repeated here.

[0048] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A multi-dimensional content recommendation method based on LLM-RAG-LBS fusion, characterized in that, The method includes: Obtain the user's trajectory point sequence, and construct a behavioral feature vector based on the trajectory point sequence; The behavioral feature vector is input into a preset stage classification model to obtain behavioral stage labels, and a structural template is matched according to the behavioral stage labels and a preset label template mapping table. The language expression control vector is calculated based on the behavioral feature vector, the preset regional semantic control matrix, and the preset platform recommendation preference coefficient; wherein, the preset structured knowledge base includes multiple candidate content fragments, and each candidate content fragment corresponds to a semantic vector and a structural label; Each candidate content segment is scored based on the language expression control vector, the semantic vector, and the platform recommendation preference coefficient to obtain a weighted adaptation score, and the top-performing segments are selected. The candidate content segment with the highest weighted adaptation score mentioned above is selected as the content segment. The selected content fragments are filtered according to the structural template and the structural tags to obtain target content fragments. Recommended content fragments are generated according to the target content fragments and the language expression control vector, and the recommended content fragments are recommended.

2. The multi-dimensional content recommendation method based on LLM-RAG-LBS fusion according to claim 1, characterized in that, The step of obtaining the user's trajectory point sequence and constructing a behavioral feature vector based on the trajectory point sequence includes: Obtain the user's trajectory point sequence, construct a service interest vector based on the trajectory point sequence, and extract the trajectory duration, total trajectory distance, and number of high-frequency dwell areas from the trajectory point sequence; The behavioral feature vector is obtained by concatenating the service interest vector, the trajectory duration, the total trajectory distance, and the number of high-frequency dwell areas.

3. The multi-dimensional content recommendation method based on LLM-RAG-LBS fusion according to claim 2, characterized in that, The trajectory point sequence includes the geographic coordinates of each time point, and the step of constructing a service interest vector based on the trajectory point sequence includes: The city grid number is matched with each of the geographic coordinates and a preset coordinate grid mapping table; wherein each of the city grid numbers corresponds to a service category; The service interest vector is constructed based on all the service categories according to the city grid number.

4. The multi-dimensional content recommendation method based on LLM-RAG-LBS fusion according to claim 1, characterized in that, The process of recommending content fragments includes: The recommended content fragments are input into a preset content language style analysis model to obtain semantic dimension values; The platform adaptation score is obtained by scoring based on the semantic dimension value, the language expression control vector, and the preset platform adaptation parameters. If the platform adaptation score is lower than a preset platform threshold, the recommended content segment is adjusted according to a preset language style template library to obtain a first recommended content adjusted segment, and the first recommended content adjusted segment is recommended; otherwise, the recommended content segment is recommended directly.

5. The multi-dimensional content recommendation method based on LLM-RAG-LBS fusion according to claim 1, characterized in that, The process of recommending content fragments includes: The recommended content fragments are matched against a preset regional tag library to obtain the high-frequency word hit frequency; The regional adaptation score is calculated based on the hit frequency of the high-frequency words, the number of embedded fields in the regional tag library, and the preset platform weights. If the regional adaptation score is less than a preset regional threshold, the recommended content segment is adjusted according to a preset regional tag library to obtain a second recommended content adjusted segment, and the second recommended content adjusted segment is recommended; otherwise, the recommended content segment is recommended directly.

6. The multi-dimensional content recommendation method based on LLM-RAG-LBS fusion according to claim 1, characterized in that, The process of calculating the language expression control vector based on the behavioral feature vector, the preset regional semantic control matrix, and the preset platform recommendation preference coefficient includes: A language fusion term is constructed based on the behavioral feature vector and the regional semantic regulation matrix; An imbalance penalty term is constructed based on the platform recommendation preference coefficient and the behavioral feature vector; The language fusion term is subtracted from the imbalance penalty term to obtain the language expression control coefficients; wherein, all the language expression control coefficients constitute the language expression control vector.

7. The multi-dimensional content recommendation method based on LLM-RAG-LBS fusion according to claim 1, characterized in that, The step of scoring each candidate content segment based on the language expression control vector, the semantic vector, and the platform recommendation preference coefficient to obtain a weighted adaptation score includes: Construct a weighted term for expression style based on the language expression control vector and the semantic vector; Construct a platform recommendation sensitivity constraint term based on the platform recommendation preference coefficient and the language expression control vector; The local weighted score is obtained by subtracting the expression style weighting term from the platform recommendation sensitivity constraint term. The local weighted scores of all semantic dimensions are then added together to obtain the weighted adaptation score.

8. A multi-dimensional content recommendation system based on LLM-RAG-LBS fusion, characterized in that, The system includes: The acquisition unit is used to acquire the user's trajectory point sequence and construct a behavioral feature vector based on the trajectory point sequence; The matching unit is used to input the behavior feature vector into a preset stage classification model to obtain behavior stage labels, and match the structure template according to the behavior stage labels and a preset label template mapping table. The calculation unit is used to calculate the language expression control vector based on the behavioral feature vector, the preset regional semantic control matrix and the preset platform recommendation preference coefficient; wherein, the preset structured knowledge base includes multiple candidate content fragments, and each candidate content fragment corresponds to a semantic vector and a structural label; The scoring unit is used to score each candidate content segment based on the language expression control vector, the semantic vector, and the platform recommendation preference coefficient, to obtain a weighted adaptation score, and select the top [segment]. The candidate content segment with the highest weighted adaptation score mentioned above is selected as the content segment. The recommendation unit is used to filter the selected content fragments according to the structure template and the structure tag to obtain target content fragments, generate recommended content fragments according to the target content fragments and the language expression control vector, and recommend the recommended content fragments.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the multi-dimensional content recommendation method based on LLM-RAG-LBS fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the multi-dimensional content recommendation method based on LLM-RAG-LBS fusion as described in any one of claims 1 to 7.

Images