Geo generation engine optimization and intelligent recommendation search ranking method based on large model

Abstract

This invention discloses a method for optimizing a geo-generative engine and intelligent recommendation retrieval ranking based on a large model, relating to the fields of intelligent geographic information processing and artificial intelligence. This invention cleanses, normalizes, and extracts features from user geo-demands to form standardized demand information; it generates contextualized geographic content based on the GeoGeoLM model combined with multidimensional geographic data and cross-modal features; it recalls candidate points through multi-granularity geographic hierarchical indexing, and achieves comprehensive ranking by combining semantic matching, spatial matching, and dynamic weight allocation, followed by diversity-aware re-ranking; it outputs the generated content in association with the ranking results, and iteratively optimizes based on anomaly information and user behavior data. This method can improve the scenario adaptability of geo-content generation and retrieval ranking.

Description

Technical Field

[0001] This invention relates to the fields of intelligent geographic information processing and artificial intelligence technology, and in particular to a method for optimizing a geo-generative engine and intelligent recommendation, retrieval, and ranking based on a large model. Background Technology

[0002] With the rapid development of the digital economy, smart cities, location services, and local life services, content generation, intelligent recommendation, and retrieval ranking technologies for geographic information scenarios have been widely applied. Especially in scenarios such as map retrieval, point-of-interest recommendation, travel navigation, discovery of nearby services, cultural and tourism guidance, business site selection, and urban governance, users often expect the system to quickly obtain accurate, personalized, and interpretable geographic content and search results based on their location, query intent, time environment, and scenario requirements.

[0003] Existing geo-based retrieval and recommendation systems typically rely on keyword matching, static rule filtering, and shallow ranking models to retrieve and rank geographic locations. While these approaches can meet basic retrieval needs to some extent, they still have significant shortcomings in complex scenarios.

[0004] On the one hand, existing technologies have a weak ability to understand user geo-needs. Traditional methods typically treat user input as ordinary text and fail to adequately mine multi-dimensional features such as geographical terms, spatial constraints, temporal constraints, and scene preferences. In particular, for geo-needs with strong implicit intentions, such as "riverside parks suitable for nighttime walks," "nearby attractions for family trips during holidays," and "restaurants that can be quickly reached during peak hours," it is difficult to accurately identify and standardize their expression, resulting in incomplete and inaccurate input foundations for subsequent generation and sorting stages.

[0005] On the other hand, existing geo content generation and retrieval ranking are mostly independent of each other, lacking a collaborative mechanism. In existing systems, geographic content generation is usually based on template filling or shallow rule generation, making it difficult to combine user intent, scene characteristics, real-time geographic data, and multimodal information to generate contextualized and personalized content. Meanwhile, the retrieval ranking process is often based solely on distance, popularity, or general relevance for scoring, resulting in a lack of unified correlation between generated content and ranking results. This can easily lead to problems such as poorly described content for top-ranked locations and poorly ranked locations with rich content, affecting user experience and result consistency.

[0006] Furthermore, existing solutions also have shortcomings in terms of time awareness, scenario adaptation, and dynamic weight allocation. Most ranking models use fixed weights and lack the ability to dynamically adjust factors such as geospatial correlation, user intent matching, real-time rule data, and user preferences for different time and business scenarios. For example, in different scenarios such as holidays, weekday morning and evening rush hours, and nighttime consumption, users' focus on distance, travel time, real-time passenger flow, and weather factors is significantly different, and traditional methods are unable to adaptively change the ranking logic according to time and scenario characteristics, resulting in a lack of scenario sensitivity in search results.

[0007] Furthermore, many current geosystems make limited use of cross-modal geographic information. In addition to textual attributes, geographic locations are often accompanied by visual information such as satellite imagery, street view images, and topographic maps. This information is crucial for distinguishing the environment, landscape, spatial layout, and scene attributes of a location. However, existing technologies often rely solely on text tags or structured fields, failing to fully integrate cross-modal features such as images. This results in incomplete semantic representations of geographic locations, particularly in scenarios such as cultural tourism, commerce, and urban spatial experiences, which can easily reduce the accuracy of recommendations and generated results.

[0008] Furthermore, existing technologies also have significant limitations in integrating real-time data and geographic rules. In Geo application scenarios, information such as traffic, weather, passenger flow, business status, regional regulations, and ecological protection restrictions are characterized by rapid dynamic changes and strong rule constraints. Traditional methods often struggle to simultaneously ensure real-time performance and compliance, easily leading to situations where recommended results do not conform to geographic rules, content descriptions are inconsistent with real-time status, or outdated geographic data is referenced, affecting system reliability and availability.

[0009] Finally, existing geo-recommendation and ranking results typically overemphasize relevance and lack diversity control mechanisms. When multiple high-scoring locations are highly similar in geographical location, semantic attributes, or functional type, the system is prone to outputting homogeneous results, reducing user browsing efficiency and exploration experience. Furthermore, many systems do not fully utilize user behavior feedback and anomaly cases, resulting in lagging model and rule updates and making it difficult to form a continuous optimization loop for geo-based scenarios. Summary of the Invention

[0010] Therefore, there is an urgent need for a large-model-based Geo generative engine optimization and intelligent recommendation retrieval ranking method that can deeply perceive and standardize user needs for complex Geo business scenarios, and integrate large-model generation capabilities, geographic rule constraints, real-time data processing, dynamic recommendation ranking, and feedback iterative optimization mechanisms, in order to improve the quality of Geo content generation, retrieval ranking accuracy, scenario adaptability, and overall system intelligence.

[0011] To achieve the above objectives, this invention proposes a Geo-generative engine optimization and intelligent recommendation retrieval ranking method based on a large model, characterized by the following steps:

[0012] S1. Geo Requirement Awareness and Preprocessing: Receives user-inputted geo requirements and performs text cleaning and geographic terminology normalization. Based on named entity recognition and syntactic analysis, extracts geospatial features, user intent features, scene features, and temporal features from the standardized text, encoding the temporal features into a temporal feature vector. It is then concatenated with the spatial feature vector to form standardized Geo requirement information, which is sent to subsequent modules.

[0013] S2, Large Model-Driven Geo Content Generation: Calls the GeoGeoLM geo-specific large model, sequentially performs multi-dimensional geo data loading, cross-modal geo feature enhancement, deep integration of requirements and data, and scenario-based geo content generation. It also performs dual verification of Geo compliance and requirement matching on the generated content, and outputs scenario-based and personalized optimized geo content.

[0014] S3. Large-Scale Intelligent Recommendation, Retrieval, and Ranking: Based on a multi-granularity geographic hierarchical index, a set of candidate geographic points is quickly retrieved. The GeoGeoLM model performs dual screening on the candidate set using semantic similarity matching and geospatial topology matching. After filtering out invalid points, the GeoGeoLM model further refines the temporal feature vectors... With scene feature vector A common input dynamic weight allocation network is used to dynamically allocate weights to four dimensions: Geo-spatial association (W1), user intent matching (W2), geographic rules and real-time data (W3), and user behavior preferences (W4). The overall ranking score S is calculated and sorted, and then diversity-aware re-ranking is performed on the ranking results. The formula for calculating the overall ranking score is: ;

[0015] S4. Output of content generation and ranking results in a coordinated manner: Using the unique ID of the geographic location as the primary key, the geographic content output in step S2 is matched one-to-one with the ranking results in step S3. The content detail is adapted according to the ranking priority, and the linkage results are output through multi-terminal adaptation interfaces.

[0016] S5. Feedback Iterative Optimization: Collect system execution anomaly information and user behavior data, perform hierarchical diagnosis of anomalies, mine user preference features, use anomaly cases and behavior feedback cases as fine-tuning datasets, perform full-link iterative optimization on the GeoGeoLM model, generation engine and ranking mechanism, and update system parameters and rules.

[0017] The time dimension features mentioned in step S1 include the time period to which the query time belongs, the week type, the holiday identifier, and the season identifier. These time dimension features are encoded into a 128-dimensional time feature vector. The time feature vector is concatenated with the Geo space feature vector and then input into the GeoGeoLM model; the time feature vector The dynamic weight allocation network, which is synchronously transmitted to step S3, enables the network to adaptively adjust the weight allocation of each dimension in time scenarios such as holidays and specific periods, thereby achieving time-aware, scenario-based ranking.

[0018] The GeoGeoLM geographic-specific large model described in step S2 adopts an improved Transformer architecture, which includes a three-layer core structure with sequential collaboration:

[0019] The first layer is the Geo semantic understanding layer: introducing a multi-granularity geosemantic mask matrix. According to city level (granularity coefficient) ), regional level ( ), street level ( The three-level hierarchical structure is constructed and weighted for fusion. The formula for calculating the attention weight is as follows:

[0020] ;

[0021] in Higher attention weights are given to geographical features such as latitude and longitude, terrain, and scene; at the same time, a 768-dimensional user intent vector is constructed by integrating historical user behavior data, along with a time feature vector. After being pieced together, they jointly participate in the analysis of personalized needs;

[0022] The second layer is the geographic rule fusion layer: it employs a knowledge embedding attention mechanism to vectorize and embed the geographic professional rule base in the form of a knowledge graph. The embedding formula is as follows:

[0023] ;

[0024] in For geographic basic feature vectors, For the demand context vector, The rule weight matrix is ​​used; and expired geographic data with an update time exceeding 24 hours is filtered out through a geographic data masking mechanism.

[0025] The third layer is the scenario-based generation layer: based on the improved Transformer decoder, it configures a dedicated decoding template matrix for each Geo business scenario. The decoding formula is:

[0026] ;

[0027] in As a fusion vector of demand and geographic data, The scene bias term; the GeoGeoLM model uses contrastive learning as its pre-training objective, and the loss function is:

[0028] ;

[0029] in This is the semantic representation vector of the anchor point's geographic location. To and The geographic location semantic representation vector that constitutes a positive sample pair (the representation of the same geographic location under different data augmented views, or the geographic semantic representation of locations that are highly similar in the same scenario). Within the same training batch, except Semantic representation vectors of other geographic locations besides themselves (including positive samples) (and negative samples). The temperature coefficient enhances the model's semantic discrimination ability for emerging geo business scenarios.

[0030] The cross-modal geographic feature enhancement mentioned in step S2 specifically involves: extracting 512-dimensional visual feature vectors from satellite imagery or street view images of geographic locations stored in the Geo data warehouse using a pre-trained convolutional neural network. This is compared with the textual geographic feature vector of the GeoGeoLM model. After fusion and layer normalization, the enhanced geographic feature vector is obtained:

[0031] ;

[0032] in The cross-modal fusion coefficients are automatically learned through end-to-end training; when image data is missing... Automatically set to 0 to degenerate into plain text mode; the enhanced feature vector It is used for both content generation in step S2 and semantic matching in step S3, thereby improving the semantic distinguishability of geographic locations in terms of visual attributes.

[0033] In step S2, the formula for calculating the demand matching degree during the dual verification of demand matching degree is:

[0034] ;

[0035] in For explicit demand quantity, For implicit demand quantity, For the first Explicit requirement coverage (1 for a successful match, 0 for otherwise). For the first The coverage of implicit requirements (if the cosine similarity between the user intention vector and the geographical location feature vector is ≥ 0.8, take 1, otherwise take 0); matching degree It is regarded as qualified when [condition], otherwise the model readjusts the generation logic and performs secondary generation; after secondary generation When it is still unqualified, a low-confidence mark is triggered and the case is stored in the abnormal fine-tuning data set.

[0036] The semantic and spatial dual matching described in step S3 is specifically as follows: Semantic matching calculates the cosine similarity between the user demand semantic vector and the geographical location scenario adaptation vector , and filters out geographical locations with a similarity lower than 0.8; Spatial matching is based on the analysis of geographical spatial topological relationships, and sequentially performs spatial range matching, geographical constraint condition matching, and geographical rule matching (filtering out points within ecological protection areas and restricted access areas); Dual matching comprehensive judgment:

[0037]

[0038] ;

[0039] The geographical locations that meet [condition] enter the ranking layer.

[0039] In the dynamic weight allocation described in step S3, the GeoGeoLM model calculates the weights of the four major dimensions with the formula based on the concatenated vector of the scenario feature vector and the time feature vector , and the weights satisfy to , and follow the following rules: for distance-sensitive demands ; for scenario-based demands ; for real-time data-sensitive demands ; for personalized correction values:

[0040] ;

[0041] Among them is the user intention vector composed of the user intention features extracted in step S1, is the geographical feature vector of the candidate geographical location ( after cross-modal enhancement in step S2 or its subset); is the user intention similarity, when the user intention similarity exceeds 0.8 , otherwise ; Among them, the Geo spatial association dimension score , is the normalized value of the geographical straight-line distance, is the traffic accessibility, and the calculation formula is:

[0042] ;

[0043] To integrate real-time traffic congestion data to calculate actual travel time, The theoretical shortest travel time is calculated based on road network topology; geographical locations with the same score are ranked as follows: Perform a second sort.

[0044] The diversity-perceived reordering in step S3 specifically involves: after the comprehensive score S is sorted, the Top-M candidate points are sorted using the maximum marginal relevance algorithm. Perform a reordering process, selecting the highest-scoring point in each iteration and adding it to the selected set. The scoring rules are as follows:

[0045] ;

[0046] in , Calculation based on geographic straight-line distance normalization; The diversity-relevance balance coefficient is dynamically assigned by the GeoGeoLM model based on the scene type, and is used for browsing exploratory scenes. Navigation destination-type scenarios The final output set size is Top-N ( This ensures that the result set has a reasonable geographical distribution and diverse semantic coverage.

[0047] The feedback iterative optimization in step S5 includes:

[0048] Anomaly classification: based on formula Calculate processing priority. Trigger emergency optimization; otherwise, include it in the regular iteration.

[0049] User preference modeling: using formulas ( (Based on data from the past 90 days) accumulated user behavior characteristics, a personalized preference model was constructed and integrated into the generation logic. and ranking weight adjustments;

[0050] Model fine-tuning: Periodically perform mini-batch fine-tuning training on the GeoGeoLM model using anomalous cases and user behavior feedback cases as the fine-tuning dataset. , ), and simultaneously optimize cross-modal fusion coefficients. and diversity balance coefficient .

[0051] It also includes a Geo data warehouse as the underlying data support, which comprises a geographic foundation database, a geographic professional rule base, a real-time spatiotemporal database, a user behavior database, and a multi-granularity geographic hierarchical index library; the multi-granularity geographic hierarchical index library stores a three-level R-tree spatial index, and the candidate point fast retrieval query complexity is [missing value]. It supports filtering by geographic category hash tags, and can handle queries exceeding a threshold in the past 7 days. The system establishes a pre-computed cache for hotspot areas and links it with real-time data updates; the real-time spatiotemporal database establishes a second-level linkage with the Geo generation engine, and when real-time data such as traffic, passenger flow or weather in the target area changes, the engine automatically triggers a content update mechanism to dynamically adjust the generated geographic content; the geographic professional rule base is stored in the form of a knowledge graph and is used to generate geographic rule matching for compliance verification and retrieval ranking of the engine.

[0052] Compared with existing technologies, the Geo-generative engine optimization and intelligent recommendation retrieval ranking method based on large models provided by this invention has at least the following beneficial effects:

[0053] This invention extracts geospatial features, user intent features, scene features, and time dimension features by performing text cleaning, geographic terminology normalization, named entity recognition, and syntactic analysis on user geo requirements. The time dimension features are then encoded and used in subsequent processing, enabling geo requirements to be input into subsequent modules in a more unified and complete form. This provides a foundation for geographic content generation and intelligent recommendation, retrieval, and ranking.

[0054] In terms of geographic content generation, this invention is based on the GeoGeoLM geographic-specific large model, combining multi-dimensional geographic data loading, cross-modal geographic feature enhancement, deep integration of requirements and data, and scenario-based generation processing to generate geographic content that is adapted to user needs and geographic scenarios. At the same time, by constraining the generated results through Geo compliance verification and requirement matching verification, it helps to improve the consistency between the generated content and Geo business rules and user needs.

[0055] In terms of retrieval and ranking, this invention utilizes a multi-granularity geographic hierarchical index to quickly recall a set of candidate geographic locations, and combines semantic similarity matching and geospatial topology matching for dual screening, ensuring that the candidate locations entering the ranking stage have good relevance. Based on this, time feature vectors and scene feature vectors are jointly input into a dynamic weight allocation network to dynamically assign weights to dimensions such as geospatial association, user intent matching, geographic rules and real-time data, and user behavior preferences, and calculate a comprehensive ranking score. This allows the ranking results to be adjusted according to changes in time and business scenarios, improving the scenario adaptability of geo retrieval and ranking.

[0056] Furthermore, this invention introduces a diversity-aware re-ranking mechanism after comprehensive ranking to further optimize candidate points, so that the results maintain relevance while taking into account the rationality of geographical distribution and semantic coverage, thereby improving the problem of overly similar points in the result set.

[0057] Furthermore, this invention associates the generated geographic content with the ranking results based on the unique ID of the geographic location, and outputs the linked results after adapting the content detail according to the ranking priority, so that there is a good correspondence between the generated content and the search ranking results, which is conducive to improving the consistency of the output results.

[0058] Meanwhile, this invention collects system execution anomaly information and user behavior data to perform hierarchical diagnosis of anomalies, and uses anomaly cases and behavior feedback cases for iterative optimization of the GeoGeoLM model, generation engine and ranking mechanism, thereby forming a technical link that runs through demand perception, content generation, retrieval ranking and feedback optimization, which is conducive to improving the system's optimization capability and scenario adaptability during continuous operation.

[0059] In summary, this invention enables collaborative coordination among Geo demand understanding, geographic content generation, candidate location selection, intelligent recommendation retrieval and ranking, result linkage output, and feedback iterative optimization, thereby improving the adaptability of content generation and recommendation retrieval in Geo scenarios and the overall processing effect. Attached Figure Description

[0060] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort. In the drawings:

[0061] Figure 1 This is an overall flowchart of a large-model-based Geo-generative engine optimization and intelligent recommendation retrieval ranking method.

[0062] Figure 2 GeoGeoLM is a core computational graph for Geo generative engine optimization and intelligent recommendation retrieval ranking method based on a large model. Detailed Implementation

[0063] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0064] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0065] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0066] Example 1

[0067] This embodiment provides a method for optimizing a Geo generative engine and intelligently recommending and ranking based on a large model, applied to the end-to-end intelligent processing scenario of a comprehensive geographic information service platform. The platform supports various Geo business scenarios, including map navigation, tourism recommendations, and local life services, with diverse user needs, complex spatial constraints, and high real-time requirements. This embodiment uses a user's search query "nearby mountain scenic areas suitable for family camping" (initiated at 10:00 AM on Saturday) via a map app as a case study to detail the complete execution process of the five core steps and each functional module of this invention.

[0068] To achieve intelligent processing of the aforementioned Geo requirements throughout the entire process, the method described in this embodiment includes five steps: Geo requirement perception and preprocessing, large-model-driven Geo content generation, large-model intelligent recommendation, retrieval, and ranking, collaborative output of generated content and ranking results, and feedback iterative optimization. A Geo data warehouse serves as the underlying data support. Each step relies on the GeoGeoLM geo-specific large-scale model to operate collaboratively, forming a closed-loop technical chain of "input—processing—execution—feedback—optimization."

[0069] The module receives geo-requests in multiple formats and converts speech to text. It interfaces with user clients such as map apps, geographic information service platforms, and cultural tourism recommendation systems to receive geo-requests in both text and speech formats. For speech input, it integrates an ASR (Automatic Speech Recognition) model that is customized and optimized for geographical terminology, place names, and scene vocabulary to transcribe speech into initial text requests. In this embodiment, the user inputs "nearby mountain scenic areas suitable for family camping" in text format, and the system then proceeds to the text cleaning process.

[0070] Text cleaning and standardization of geographical terms. The following processing is performed on the initial text in sequence: First, redundant character removal, deleting meaningless symbols and repetitive expressions; Second, geographical term normalization, converting non-standard expressions into standard terms based on the geographical professional term normalization dictionary. For example, "Magic City" is unified as "Shanghai", and "camping" is unified as "camping"; Third, sentence structure optimization, adjusting the structure of texts with ambiguous expressions through syntactic analysis. In this embodiment, after normalization processing, "suburbs near the city" is combined with the user's current location and standardized as "the suburban area within a radius of 50 km from the current location", "mountain scenic area" is standardized as the geographical type label "mountainous scenic area", and "parent-child camping" is retained as a composite demand label.

[0071] [[ID=B]]Multi-dimensional feature extraction. Based on a pre-trained geographical semantic feature extraction model, combined with natural language processing technologies such as named entity recognition, syntactic analysis, and requirement decomposition, four core features are extracted from the standardized text:

[0072] First, Geo spatial features: The geographical area is "the suburban area within a radius of 50 km from the current location", the geographical type is "mountainous scenic area", the spatial range constraint is "within 50 km", and the geographical constraint conditions are "the terrain is suitable for camping (slope requirement), there is a water source, and there is a parking condition".

[0073] Second, user intention features: The explicit demands are "camping" and "parent-child travel"; The implicit demands mined through semantic reasoning of the GeoGeoLM model include "perfect children's supporting facilities", "gentle terrain (slope < 15°)", "the camp has potable water sources", "tent pitching and open fire use are allowed", "the scenic area has no dangerous terrain", and "there is no ecological reserve restriction", a total of 6 implicit demands, denoted as to The number of implicit demands ; The number of explicit demands .

[0074] Third, scenario features: The business scenario type is "cultural and tourism camping parent-child scenario", and the scenario constraint condition is "travel on weekends".

[0075] Fourth, time dimension features: The query time is in the morning period, the week type is weekend (Saturday), the holiday identifier is no (non-statutory holiday), and the season identifier is summer. The above time dimension features are encoded into a 128-dimensional time feature vector , which is concatenated with the Geo spatial feature vector and then jointly input into the GeoGeoLM model.

[0076] The extraction results are encapsulated in JSON format as standardized Geo demand information, synchronized and sent to the Geo generation engine optimization module and the intelligent recommendation retrieval ranking module, and stored in the Geo data warehouse.

[0077] The operation of GeoGeoLM's dedicated geo-model: GeoGeoLM adopts an improved Transformer architecture, which includes a three-layer core structure that processes standardized geo-requirement information sequentially.

[0078] In the Geo semantic understanding layer, the model introduces a multi-granularity geographic semantic mask matrix. According to city level (granularity coefficient) ), regional level ( ), street level ( The attention weights are constructed in a three-tiered structure and then weighted and fused. The formula for calculating the attention weights is:

[0079] ;

[0080] In this embodiment, the user requirement "mountainous scenic area within 50km of the suburbs" corresponds to the regional granularity ( At this granularity level, the model prioritizes activating attention weights for geographical features within the target area, such as mountainous terrain, slope, and vegetation cover. Simultaneously, the model incorporates user historical behavior data (the user has 3 family trips and 2 outdoor camping trips in the past 90 days) to construct a 768-dimensional user intent vector, along with a time feature vector. (The code "Weekend / Morning / Non-Holiday / Summer") is concatenated to form a personalized demand analysis result, which identifies that the user has a strong preference for "niche mountain campsites with moderate foot traffic".

[0081] In the geographic rule fusion layer, the model embeds the geographic professional rule base into a vectorized form of a knowledge graph. The embedding formula is as follows:

[0082] ;

[0083] In this embodiment, This represents the geographic basic feature vector of the target area's mountainous scenic area. For the context vector of the "family camping" requirement, This is the rule weight matrix. The model uses knowledge graph retrieval to identify rules within the target area related to ecological protection (3 scenic areas are located in the core area of ​​an ecological protection zone, prohibiting camping, and are marked as filtered objects), scenic area campsite management rules (some scenic areas restrict the use of open flames), and geographical safety rules (areas with slopes exceeding 15° are marked as unsuitable for family camping). A total of 15 rule constraints are embedded into the matrix. The geographic data masking mechanism synchronously filters and updates outdated traffic data that has been updated for more than 24 hours.

[0084] At the scenario generation layer, the model calls the exclusive decoding template matrix corresponding to the "cultural tourism camping parent-child scenario". The decoding formula is:

[0085] ;

[0086] in As a fusion vector of demand and geographic data, This is a biased item for family camping scenarios, which strengthens the weight of dimensions such as children's facilities, safety tips, and campsite environment in the output content.

[0087] Cross-modal geographic feature enhancement: 512-dimensional visual feature vectors are extracted from satellite imagery of candidate scenic spots stored in the Geo data warehouse using a pre-trained convolutional neural network.

[0088] ;

[0089] Taking candidate scenic spot A as an example, after its satellite imagery was extracted using CNN, the activation values ​​of three dimensions—vegetation coverage, terrain flatness, and water proximity—in its visual feature vector were significantly higher than those of other scenic spots, indicating a high degree of visual semantic matching with the demand for "family camping." The visual feature vector was then compared with the textual geographic feature vector. After fusion and layer normalization, the enhanced geographic feature vector is obtained:

[0090] ;

[0091] In this embodiment, the cross-modal fusion coefficient of mountain scenic area sites is... After end-to-end training, the value is 0.58, indicating that visual features significantly enhance the semantic meaning of this type of geographic location. Enhanced geographic feature vector of scenic area A. The semantic expression of the three visual attribute dimensions of "flat terrain", "dense vegetation" and "near water source" is significantly enhanced, which helps to achieve more accurate alignment with the user's implicit needs of "gentle terrain" and "near water source" in subsequent semantic matching.

[0092] Multi-dimensional geographic data loading: The generation engine loads the following data from the Geo data warehouse for the target area (within 50km of the current location): basic geographic data (latitude and longitude, terrain slope, geomorphological features, and details of supporting facilities for 27 mountain scenic areas); geographic professional rules (ecological protection rules for the target area, campsite management rules for scenic areas, and geographic safety rules, totaling 15 rules); real-time spatiotemporal data (real-time traffic congestion coefficient from 08:00 on the same day, real-time visitor capacity of each scenic area, weather data for the day, and remaining parking resources for each scenic area); historical generation cases of the same scenario (12 generation cases of the "parent-child camping" scenario in the past 30 days).

[0093] Deep Integration of Demand and Data: The GeoGeoLM model deeply integrates standardized Geo demand information with the aforementioned multi-dimensional geographic data through a three-layer core structure, combined with a user personalized preference model, to generate a demand-data feature fusion vector.

[0094] ;

[0095] in This is the feature vector representing the requirements of the current query. The personalized preference vector is constructed based on the user's behavioral data over the past 90 days. The fusion coefficient of 0.2 ensures that the preference model does not excessively interfere with the current requirement analysis in scenarios where the requirement is clear.

[0096] Contextualized geographic content generation: Based on the "cultural tourism camping family scenario," the model calls the corresponding contextualized generation strategy to generate the following multi-format content for the filtered candidate scenic spots (3 scenic spots located in ecological protection areas have been filtered, leaving 24):

[0097] For each candidate scenic area, a personalized introduction is generated, including scene suitability assessment (campsite terrain slope value, parent-child suitability rating), details of children's facilities (distribution of children's play area, medical emergency point, family toilet), water source and parking conditions (distance of drinking water source, parking lot capacity), real-time visitor capacity of the scenic area (ratio of current visitor flow to maximum capacity), and visual environment description (campsite environment description generated based on enhanced visual features of satellite imagery); the optimal driving route plan is generated by combining real-time traffic data (integrating the current congestion coefficient); personalized suggestions are generated for parent-child camping scenarios (sun protection suggestions for children, campsite equipment list, summer insect prevention tips).

[0098] Geo compliance and demand matching are verified simultaneously: The compliance verification is based on geo-specific rules, checking whether the generated content of the 24 candidate scenic spots involves ecological protection zones (pre-filtered) and whether it complies with the scenic spot campsite management rules (filtering out 2 scenic spots that explicitly prohibit tenting). After the compliance verification, the remaining 22 candidate scenic spots proceed to the demand matching calculation.

[0099] The formula for calculating the demand matching degree is:

[0100] ;

[0101] Taking scenic area A as an example, the explicit demand coverage is calculated as follows: Explicit demand 1 "camping", scenic area A has compliant campsites, ; Explicit demand 2: "Family travel"; Scenic spot A has a children's play area and family-only camping sites. ; , .

[0102] The implicit demand coverage is calculated as follows: (The children's facilities are well-equipped), and the cosine similarity between the user's intent vector and the feature vector of the children's facilities in scenic area A is 0.91 ≥ 0.8. ; (Gentle terrain with a slope <15°), the average slope of the main camp area A in the scenic area is 8°, and the cosine similarity is 0.87 ≥ 0.8. ; (With available drinking water sources), Scenic Area A has 3 water intake points, with a cosine similarity of 0.93 ≥ 0.8. ; (Tents and open flames are permitted), the rules for Camp A in the scenic area allow this, and the cosine similarity is 0.89 ≥ 0.8. ; (No dangerous terrain), the main campsite A in the scenic area has been fenced off, and the cosine similarity is 0.84 ≥ 0.8. ; (Without ecological protection zone restrictions), scenic area A is located outside the buffer zone of the protection zone, and its cosine similarity is 0.92 ≥ 0.8. ; , .

[0103] Substitute into the formula: ;

[0104] The model determined that the content generated by scenic spot A did not match the requirements. The reason was that the description of children's facilities in scenic spot A was insufficient. The generation logic was readjusted, and the distribution of children's medical emergency points, details of parent-child amusement projects, and nighttime safety lighting conditions were added before a second generation was performed.

[0105] Recalculate after secondary generation: Since the implicit requirement description coverage is fully met, the cosine similarity of all terms is higher than 0.8, and the explicit requirement coverage remains unchanged, substituting into the formula:

[0106] ;

[0107] The above results indicate that the numerator in the matching degree calculation formula is limited by the upper limit that all six implicit requirements have been met. Substituting the formula into the actual explicit requirement weights of this embodiment (because the secondary generation supplements the parent-child suitability description dimension, the model decomposes the scene matching dimension of the explicit requirement of "parent-child travel" into two sub-requirements, namely...) ), recalculate:

[0108] ;

[0109] The model determined that the secondary generation still fell below the threshold, triggering a low-confidence marker. Scenic spot A was downgraded and its core information was output. This case was then stored in the anomaly fine-tuning dataset for subsequent fine-tuning and optimization of the GeoGeoLM model's threshold judgment mechanism. Double matching verification was simultaneously performed on the remaining 21 scenic spots, with 18 of them showing a good matching score. The system outputs complete and optimized geographic content; 3 scenic spots meet the standards after secondary generation, and a total of 21 scenic spots pass the double verification and enter the subsequent ranking module.

[0110] Rapid Retrieval from Multi-Granularity Geographic Hierarchical Index: The system uses an R-tree structure to perform regional (1-10km) level searches within the Geo data warehouse's multi-granularity geographic hierarchical index, combining the user's current location with the query spatial range (50km), and then calls the geographic category hash label. Filter out non-mountain scenic area locations and quickly retrieve a set of candidate locations:

[0111] ;

[0112] in The query space is defined by a circle centered at the user's current location with a radius of 50km. This is a set of hash labels for mountain scenic area categories. In this embodiment, multi-granularity indexing is used... The recall was completed within the time complexity, and a total of candidate locations were obtained. This includes 27 mountain scenic areas within the target area. The system synchronously checks the pre-calculated cache; the region-scene combination (family camping / weekend within 50km) queried this time has been queried 327 times in the past 7 days, which does not exceed the hotspot threshold. If the pre-computed cache is not hit, the process will proceed to online processing.

[0113] Precise matching based on both semantics and spatiality: Semantic matching calculates the semantic vector of user needs. (Fusion of cross-modal enhancement features) Adaptation vectors for each candidate scenic area Cosine similarity:

[0114] ;

[0115] Semantic similarity was calculated for each of the 27 candidate scenic spots, and then filtered. In this embodiment, 6 scenic spots were filtered out of the candidate set because they lacked facilities for children or did not allow camping, and their semantic similarity was below 0.8 (the lowest value was 0.61). The remaining 21 scenic spots entered the spatial matching stage.

[0116] Spatial matching is based on geospatial topology analysis and performs three filtering steps in sequence: First, spatial range matching filters scenic spots beyond a 50km range. In this example, two scenic spots were filtered because their actual distances exceeded the range, leaving 19. Second, geographical constraint matching filters mountainous scenic spots with steep terrain (slope ≥ 15°). In this example, three scenic spots with average slopes exceeding 15° were filtered, leaving 16. Third, geographical rule matching filters scenic spots located in ecological protection zones (core areas or buffer zones) and restricted areas. In this example, one scenic spot was filtered because it was located in the buffer zone of a protection zone, leaving 15 that enter the dual matching comprehensive judgment stage.

[0117] Dual-match comprehensive judgment: ;

[0118] Calculations were performed on the 15 candidate scenic spots separately. ,filter In this embodiment, all 15 scenic spots have a comprehensive similarity score higher than 0.8 and are included in the ranking.

[0119] Multi-dimensional dynamic weight allocation: The GeoGeoLM model assigns weights based on the scene feature vector of the current query. (Encoding "Cultural Tourism Camping Family Scene") and Temporal Feature Vector Calculate the weights of the four dimensions from the concatenated vector (encoded as "weekend / morning / non-holiday / summer"):

[0120] ;

[0121] In this embodiment, due to the scenario-based requirement (family camping), the model is dynamically assigned... (User intent matching dimension has the highest weight); Since the query time is a weekend morning (peak travel time), real-time traffic and passenger flow data have a significant impact on the ranking, and the model assigns... Geographical distance imposes certain constraints on family travel, giving... The user's behavioral preference data is relatively abundant, giving them... ; Verify weight constraints: It satisfies the normalization condition.

[0122] Taking scenic spot B, which ranks highly among the 15 candidate scenic spots that entered the ranking layer, as an example, the features of each dimension are quantified as follows:

[0123] Geo spatial correlation dimensions The straight-line distance from scenic spot B to the user's current location is 38km, after normalization. Actual travel time Minutes (calculated using the real-time traffic congestion coefficient of 1.35 for the morning of the day), theoretically the shortest travel time. Traffic accessibility (minutes, calculated based on road network topology):

[0124] ;

[0125] ;

[0126] User intent matching dimension Scenic Area B achieved a perfect score in explicit demand coverage. Among implicit demands, all six aspects—children's facilities, terrain conditions, water source, tent erection permits, safety conditions, and no protected area restrictions—met the standards. The scene suitability score was 0.92 (based on cross-modal enhanced feature vector matching). The overall quantitative assessment... .

[0127] Geographic rules and real-time data dimensions Scenic Area B passed all compliance checks. At 10:00 AM that day, the real-time passenger flow was 43% of the camp's carrying capacity, indicating a good passenger flow situation. The real-time traffic condition is mild congestion. The weather was sunny and the temperature was 28°C, suitable for camping. ); Quantization after weighting each sub-feature .

[0128] User behavior preference dimension Based on the user's behavioral data over the past 90 days, a personalized preference model was constructed. The user showed a high preference for "less popular mountain campsites with moderate crowds," and scenic spot B falls into this preference category. .

[0129] Perform comprehensive score calculation; calculate personalized adjustment value: user intent vector. Geographic feature vector of scenic area B similarity ,therefore:

[0130] ;

[0131] Substitute into the comprehensive score calculation formula:

[0132] ;

[0133] ;

[0134] ;

[0135] The comprehensive scores of the 15 candidate scenic spots were calculated one by one, and the initial ranking results were obtained by sorting them from high to low scores. Scenic spot B ranked first with a comprehensive score of 0.914, scenic spot C (Songling Family Camp, score 0.901) ranked second, and scenic spot D (Cuigu Mountain Camp, score 0.897) ranked third.

[0136] Diversity-perceived re-ranking: Perform diversity-perceived re-ranking on the Top-15 candidate scenic spots, parameters The GeoGeoLM model dynamically assigns values ​​based on the fact that "family camping" is a browsing and exploration-type scene. Starting with an empty set as the initial selection set, the first round selects the scenic spot B with the highest overall score from the 15 scenic spots and adds it to the selected set. .

[0137] The second round of calculations covers the remaining 14 scenic areas. Taking scenic area C (Songling Family Camp) as an example:

[0138] ;

[0139] Semantic similarity between scenic area C and scenic area B (Both belong to the family-friendly mountain campground category), geographical proximity (The two scenic areas are 12km apart in a straight line, and their proximity is relatively high after normalization):

[0140] ;

[0141] ;

[0142] Calculate the remaining 14 scenic spots one by one. Scenic spot E (Fengling Outdoor Camp, overall score 0.881, semantic similarity with scenic spot B 0.71, geographical proximity 0.42):

[0143] ;

[0144] ;

[0145] Scenic Area E If the score is higher than that of scenic spot C, then scenic spot E will be selected in the second round and added to the selected set. And so on, to complete the final output set of the Top-10. The diversity reordering ensures that the geographical distribution of scenic spots in the recommendation results is reasonable and that the mountain ranges to which they belong are diverse, without the phenomenon of multiple scenic spots being concentrated in the same mountain range.

[0146] Deep integration and matching of generated content with ranking results: Using the unique ID of the geographic location as the primary key, the geographic content output by the Geo generation engine is precisely matched one-to-one with the final ranking results after diversity reordering. Content priority is adapted based on the comprehensive score S of the ranked location: The top-ranked scenic spot B generates comprehensive content including details of children's facilities (including the location of children's medical emergency points and a map of family-only campsites), a camping area topographic map (generated based on satellite imagery visual features), real-time visitor flow information (current occupancy rate 43%, advance booking recommended), driving route planning (integrating real-time traffic data, estimated 68 minutes), summer family camping equipment suggestions, and insect and sun protection tips; scenic spots ranked 2nd to 5th generate standard-dimensional content including core facilities, accessibility information, and visitor flow status; scenic spots ranked 6th to 10th generate core information including basic geographic information and suitability descriptions.

[0147] For scenic spots whose ranking changes during the diversity re-ranking (e.g., scenic spot C is adjusted from the initial 2nd place to the 3rd place), the original comprehensive score of 0.901 and the adjusted 3rd place are retained in the output structure for front-end differentiated display and user feedback collection.

[0148] The system features standardized output across multiple platforms. The visual map displays the Top-10 scenic spots in ranking order, using different colors to differentiate between comprehensive score ranges. Users can click on a scenic spot's marker to view either full-dimensional or standard-dimensional geographical content. The text and image display shows the ranking results in a list format, including the comprehensive score (Scenic Spot B: 0.914), ranking position, suitability description, and concise geographical content. The voice display converts the core information of the top-ranked scenic spot B into a voice broadcast: "We recommend Qingyun Mountain Family Campsite, 38 kilometers from you, approximately a 68-minute drive. The campsite has a gentle slope, complete children's facilities, and currently has moderate visitor numbers, making it suitable for family weekend camping." Users can also ask follow-up questions via voice.

[0149] Real-time status monitoring and anomaly detection: During this query execution, all system monitoring metrics were within the normal range: semantic matching calculation took 32ms, spatial matching took 18ms, comprehensive score calculation took 25ms, the linkage output interface response time was 97ms, and the total processing time was 172ms. The system did not detect any anomalies and completed the query processing normally. Execution logs and user behavior data were synchronously stored in the Geo data warehouse.

[0150] User behavior data collection and preference modeling: After receiving the recommendation results, the user browsed the details of the 1st (Scenic Spot B), 3rd (Scenic Spot C), and 5th ranked scenic spots on the visual map, and finally clicked to book the 3rd ranked scenic spot C (Songling Family Camp). The system collected the above behavioral data and identified that the user chose the 3rd ranked scenic spot C over the 1st ranked scenic spot B, judging that the actual attractiveness of scenic spot C to the user in a certain dimension was higher than the level reflected by the overall score ranking.

[0151] User preference modeling uses a formula to accumulate the user's behavioral characteristics over the past 90 days:

[0152] ;

[0153] in The time decay coefficient, This data represents the number of behaviors over the past 90 days. The behavioral data from this instance (selecting the 3rd ranked scenic spot instead of the 1st) was added to the preference model as a new behavioral sample. The model identified the differences between scenic spot C and scenic spot B: scenic spot C is geographically more remote (further from the city center, with lower visitor traffic), consistent with the user's historical preference for "niche, low-traffic campsites." The updated preference model was stored in the Geo data warehouse and integrated into subsequent generation logic and ranking weight adjustments.

[0154] Ranking mechanism and weight adjustment; based on the above behavioral feedback, the system identifies the direction for adjusting the ranking mechanism for this user: in the dimension of user behavior preferences. In the calculation, the weighting of the "low-traffic niche scenic spot" attribute was increased by 0.05; at the same time, the updated results from the user preference model were incorporated into the personalized adjustment value. The calculation benchmark will be used to adjust the overall score of scenic spot C upwards in the user's next similar query.

[0155] Meanwhile, the system recorded the operation of adjusting scenic spot C from the 2nd to the 3rd position in the diversity perception re-ranking, and the user ultimately selected scenic spot C, indicating that the result of the diversity adjustment matches the user's actual preference, thus verifying the... (Exploratory) Adaptability to this user. The system's diversity balance coefficient for this user. Keep the current assignment and continue using it in the next query. Used as an initial assignment reference.

[0156] Anomaly Classification Diagnosis: During this query execution, the system captured an anomaly where the demand matching degree of scenic spot A after secondary generation was still lower than the threshold. The anomaly classification is as follows: The scope of impact is the generation of content for a single scenic spot (partial functions), and the level is Level 2 (…). The affected users are all users who searched for this scenic spot (a portion of users). Processing priority:

[0157] ;

[0158] The case was determined to be included in the regular iteration and recorded in the abnormal fine-tuning dataset for processing during the next batch fine-tuning. The GeoGeoLM model diagnosed the anomaly in this case as "insufficient coverage of the generation engine's description dimensions of parent-child facilities" and formulated the adjustment strategy as "decoding the template matrix in the parent-child scene". "Increase the weight of the children's facility description dimension in the text."

[0159] GeoGeoLM model fine-tuning: The system periodically (every 7 days) aggregates abnormal cases and user behavior feedback cases into a fine-tuning dataset, and performs mini-batch fine-tuning training on the GeoGeoLM model, with a learning rate of... , Simultaneous optimization of cross-modal fusion coefficients (Fine-tuning was performed after evaluating the visual feature integration effect of family-friendly mountain scenic areas) and diversity balance coefficient The scene mapping rules are then implemented. The optimized model parameters, engine rules, and ranking mechanism parameters are synchronously updated to each core module, forming a complete closed-loop optimization chain.

[0160] Example 2

[0161] Based on the method framework described in Embodiment 1, this embodiment takes the specific operation of the spatiotemporal demand prediction and content pre-generation mechanism in a high-frequency query scenario as an example to explain in detail the pre-generation trigger conditions, cache hits, and response latency reduction process.

[0162] The scenario in this example is as follows: The day after the execution of Example 1 (Sunday), the spatiotemporal demand prediction submodule built into the GeoGeoLM model analyzes the historical query sequence of the target area and finds that the query count for the area-scenario-time combination "family camping mountain scenic area within 50km / weekend / morning" has been 312, 389, 401, and 427 times respectively on each weekend in the past four weeks, showing a continuous upward trend. The submodule models the historical query sequence based on LSTM and predicts the demand intensity distribution of this combination on Sunday morning:

[0163] ;

[0164] in The time hidden state of the LSTM (encodes the query sequence features of this time period in the past 4 weeks). The target region's geographic feature vector. , For learnable weight matrix, For bias, This is the Sigmoid activation function. Substituting the current parameters, the predicted value is... Exceeding the trigger threshold The system will complete the triggering and execution of the pre-generated tasks before the first query arrives on Sunday (estimated at 08:30).

[0165] Pre-generated task execution: The system performs a complete content generation and ranking process in advance for the scenario combination of "family camping mountain scenic area within 50km / weekend / morning / summer" in a low-priority asynchronous manner. This includes multi-dimensional geographic data loading, cross-modal feature enhancement, deep integration of demand data, scenario-based content generation, dual verification, multi-dimensional dynamic weight ranking, and diversity perception re-sorting. The pre-generated results (including the full-dimensional / standard-dimensional / core information content and comprehensive score ranking of the Top-10 scenic spots) are cached in the Geo data warehouse. The cache timestamp is recorded as Sunday 08:24:17. The cache validity period is linked to real-time data updates (default 30 minutes, which will expire immediately if the real-time passenger flow or traffic data changes exceed the threshold).

[0166] Cache hit processing: At 08:51 on Sunday, User A entered "nearby mountainous areas suitable for camping with children" through the map APP. After processing by the Geo demand perception and preprocessing module, the spatial characteristics (mountainous scenic spots within 50km), intent characteristics (family camping), scene characteristics (cultural tourism camping family scene), and time characteristics (weekend / morning / summer) of the standardized Geo demand information highly matched the scene combination of the pre-generated cache. The system retrieved the pre-calculated cache and hit the valid cache (27 minutes from the cache generation time, within the validity period).

[0167] Real-time compliance verification: The system performs lightweight real-time compliance verification on the cached content to confirm that there have been no new changes in the control of ecological protection zones since the pre-generation time, and that the real-time visitor flow data has not triggered the update threshold (the visitor flow change of each scenic spot is less than 10%). The cached content is still valid and the cached result is returned directly.

[0168] Response latency comparison: The total response time for this query was 48ms (including 21ms for request preprocessing and 27ms for cache retrieval and verification), which is approximately 72% lower than the 172ms online computation in Example 1. For high-frequency scenarios with an average daily query volume exceeding 5000 times, the pre-generation mechanism can reduce online inference resource consumption by more than 60%, while reducing the average response latency on the user side from the 500ms level to less than 50ms, significantly improving the user experience.

[0169] Cache update mechanism: At 09:58 on Sunday, the real-time spatiotemporal database detected a sudden increase in the real-time visitor flow of the second-ranked scenic spot (Fengling Outdoor Camp) in the target area to 87% of the camp's carrying capacity (exceeding the pre-generated 43%). The change exceeded the trigger threshold, and the system automatically invalidated the current cache, triggering an incremental re-ranking task, recalculating only the scenic spots whose visitor flow data changed. (Geographical rules and real-time data dimensions) and comprehensive score S, update the ranking results and corresponding generated content (add a visitor flow warning prompt "The current scenic spot is crowded, it is recommended to choose an alternative campsite" to the content of scenic spot B), complete the cache update and record the new cache timestamp 09:59:34, the entire incremental re-ranking took 41ms, which did not affect the online query service.

[0170] This embodiment verifies the actual effect of the spatiotemporal demand prediction and content pre-generation module in high-frequency query scenarios. The prediction hit rate (weekend morning parent-child camping scenario) reaches 83%. With the cooperation of the closed-loop feedback optimization mechanism in Embodiment 1, the prediction accuracy continues to improve as query data accumulates.

[0171] Example 3

[0172] This embodiment, based on the method framework described in Embodiment 1, introduces a multi-dimensional uncertainty perception score correction mechanism and a bidirectional coupling correction mechanism for generation-ranking semantic consistency to address the following two issues: First, the quantitative scores of the four major feature dimensions themselves are uncertain (due to real-time data fluctuations, sparse user behavior data, missing geographic images, etc.). Directly using point estimates in the comprehensive score calculation will lead to an imbalance in score confidence, causing the ranking results to produce false high scores for points with high uncertainty. Second, there is a potential semantic deviation between the geographic content semantic vector generated in step S2 and the geographic feature vector on which the ranking in step S3 depends. If the semantic expressions of the two at the same geographic point are inconsistent, the linked output with the unique ID as the primary key, although aligned at the data level, will have a problem of content and ranking logic being disconnected at the semantic level, affecting the content perception quality on the user side.

[0173] This embodiment uses a user's input of "highly rated boutique homestays near the city, preferably with mountain views" as a query example (initiated at 7 PM on Friday evening) to provide a detailed explanation of the complete execution process of the two correction mechanisms mentioned above.

[0174] After processing by the Geo Demand Perception and Preprocessing module, the standardized Geo Demand information extraction results are as follows: Geo spatial characteristics are "within 15km of the city center, with abundant mountain scenery"; explicit demands are "characteristic homestays" and "positive reviews". Implicit requirements include "rooms with mountain views", "quiet and noise-free", "parking available", "breakfast service available", and "convenient transportation". Time feature vector The coding is "weekday / evening / non-holiday / autumn". The evening time period characteristics have an upward impact on the weight of real-time traffic data in the subsequent weight allocation (evening is the travel peak).

[0175] After the conventional four-dimensional feature quantification is completed in step S3, this embodiment introduces Monte Carlo Dropout uncertainty estimation for the scores of each dimension of each candidate geographic location, quantifies the confidence of each dimension score, and incorporates the uncertainty as a correction term into the calculation of the comprehensive score.

[0176] For the scoring subnetwork in the GeoGeoLM model responsible for quantizing features in each dimension, the Dropout layer activation state is maintained during the inference phase (Dropout rate). Repeat the process for each candidate geographic location. This forward inference process yields the score sampling sequence for each dimension. (Based on dimension...) ( For candidate geographic locations The The score of the second sampling is denoted as Calculate the mean and variance of the scores for each dimension:

[0177] ;

[0178] ;

[0179] With mean Alternate origin estimate Incorporate into subsequent comprehensive score calculations, using variance This characterizes the degree of uncertainty in the score for that dimension. The overall uncertainty level is defined as the dynamic weighted sum of the variances of each dimension:

[0180] ;

[0181] in The dimension weights of the network output are dynamically assigned to the weights in step S3, which are consistent with those in the comprehensive score calculation, so that the uncertainty correction is aligned with the weight system.

[0182] In the original comprehensive score Based on this, an uncertainty penalty correction term is introduced to obtain the comprehensive score after uncertainty perception correction. :

[0183] ;

[0184] in This is the uncertainty penalty coefficient, dynamically assigned by the GeoGeoLM model based on scene characteristics, with a default value of [value missing]. For scenarios sensitive to real-time data ( Since real-time data fluctuations are the main source of uncertainty, [the following is taken]: For scenarios with sparse user behavior data (fewer than 100 historical behavior counts) (times), due to low confidence in preference modeling, take Uncertainty penalty correction term Geographic locations with low confidence scores across all dimensions will be penalized to suppress false high scores and make the ranking results more robust.

[0185] Taking candidate homestay X in this embodiment as an example, the estimated scores of the four dimensions calculated by the conventional process are as follows: (9km from the city center, with good accessibility). (High coverage of implicit needs such as mountain view rooms and breakfast service). (Sufficient evaluation data and complete real-time room status data). (The user has two homestay booking records in their history). Dynamic weight is... , , , Personalized correction value Overall score:

[0186] ;

[0187] ;

[0188] Monte Carlo Dropout Uncertainty Estimation ( (times), the variance results for each dimension are as follows: (Traffic accessibility data is highly real-time and has low variance.) (The information on mountain view rooms is from the platform and may contain some subjective description bias, with a slightly high variance.) (The real-time room status data update of the homestay X was delayed by more than 1 hour, which reduced the confidence of the model's score for this dimension and resulted in the highest variance.) (The sample size of user behavior is small, and the variance is moderate.)

[0189] Overall level of uncertainty:

[0190] ;

[0191] ;

[0192] This embodiment belongs to a real-time data sensitive scenario. (not exceeding the 0.4 threshold), take the default value. Overall score after uncertainty perception correction:

[0193] ;

[0194] Uncertainty estimation was performed simultaneously on homestay Y, which had the highest overall score in the candidate set. The variance of homestay Y was low in all dimensions. (Due to its complete real-time data and sufficient user review samples), the corrected score is:

[0195] ;

[0196] Before the correction, homestay X had a composite score of 0.875, higher than homestay Y's 0.871. After the correction, homestay Y's score of 0.868 was higher than homestay X's 0.866, resulting in a ranking reversal. This reversal reasonably reflects the delay in real-time room status data updates for homestay X. The objective risk of low confidence in dimensional scores is addressed by correcting the ranking results to make them more robust and reliable.

[0197] The system Exceeding the variance threshold The system automatically adds data quality labels to geographic locations and retains an uncertainty level field in the linked output structure. This field is used by the front end to inform users of the confidence level when displaying the data. In the feedback iteration optimization module of step S5, the system prioritizes triggering data integrity supplementation tasks for high uncertainty locations.

[0198] In this embodiment, after the content is generated and output in step S2 and before the linkage matching in step S4, a semantic consistency calculation for generation and ranking is introduced. The semantic expression of the generated content and the geographic feature vector on which the ranking depends are explicitly aligned and checked. The consistency measurement result is fed back to the comprehensive score in the form of a correction term, thereby realizing true semantic bidirectional coupling between the generation module and the ranking module.

[0199] Step S2 is for each candidate geographic location Output contextualized geographic content text Then, the Geo semantic understanding layer (encoder part) of the GeoGeoLM model is called to encode the generated content text and extract the semantic vector of the generated content:

[0200] ;

[0201] The Geographic location scene adaptation vector used in semantic matching in step S3 (Ranking-side feature vectors) share the same encoder, but have different input sources: It is generated by co-coding structured geographic data and historical generation cases in the Geo data warehouse, and The content text generated in this instance is encoded in real time. Ideally, the two should be highly consistent, but due to the scene bias during the generation process... Due to the influence of user preferences, there is a possibility of semantic shift.

[0202] The cosine similarity between the generated content semantic vector and the ranking-side feature vector is calculated and defined as the generation-ranking semantic consistency score.

[0203] ;

[0204] A higher value indicates that this is a geographical location. The more consistent the generated content is with the geographic semantic features on which the ranking depends, the more guaranteed the quality of the linked output content.

[0205] The consistency score is added as a correction term to the overall score after uncertainty perception correction. To obtain the final overall score :

[0206] ;

[0207] in As the semantic consistency benchmark threshold, Let the consistency correction coefficient be taken as... ;when When the correction term is positive, a score bonus is awarded to points with a high degree of semantic consistency between generation and ranking; when... When the correction term is negative, a penalty is imposed on points with semantic deviations, and a consistency anomaly flag is triggered.

[0208] Taking homestay Y as an example in this embodiment, step S2 generates content for homestay Y that includes highly contextualized descriptions such as "mountain view facing double room, private terrace, tranquil mountain environment, breakfast incorporating local ingredients." The semantic vector of the generated content is then extracted. , and the ranking-side feature vector Calculate cosine similarity: Above the threshold The correction item is positive:

[0209] ;

[0210] Perform the same calculation for guesthouse X. When generating content for guesthouse X in step S2, the real-time room status data for guesthouse X is incomplete ( (High dimensionality variance) The generation engine uses a rather vague description of the real-time room information of homestay X in the contextual generation layer, resulting in a deviation between the generated content and the ranking-side feature vector in the semantic dimension of "room availability". The correction term is negative:

[0211] ;

[0212] Final overall score ranking: Homestay Y ( > Homestay X ( The ranking results are consistent with the ranking after uncertainty correction, and the gap has further widened, making the semantic interpretability of the ranking results stronger: Homestay Y, due to the high consistency between the generated content and the semantics of geographical features, can provide users with a consistent experience in the final output where the content description and ranking logic corroborate each other; Homestay X, due to the semantic deviation between the generated content and the ranking features, has its score reasonably constrained without compromising the accuracy of data linkage.

[0213] right For geographic locations, the system triggers the generation-ranking semantic consistency anomaly flag, encapsulating the following information as a consistency anomaly feedback case and storing it in the anomaly fine-tuning dataset: the query demand vector for that location. Rank-side feature vector Generate content semantic vectors Consistency score And deviation direction analysis (the GeoGeoLM model automatically diagnoses which semantic dimension the generated content deviates from the ranking characteristics).

[0214] In the GeoGeoLM model fine-tuning phase of step S5, the aforementioned consistency anomaly feedback cases are added to the fine-tuning dataset, with semantic consistency alignment loss used as an additional fine-tuning objective.

[0215] ;

[0216] in This is the set of all geographic locations in the current batch that have triggered consistency anomaly flags. Learning loss compared to the original Combined as a fine-tuning of the total loss:

[0217] ;

[0218] To ensure consistency in aligning loss weights, this fine-tuning objective drives the GeoGeoLM model to proactively align with the semantic space of the ranking feature vector during content generation, fundamentally reducing the frequency of generation-ranking semantic bias, rather than relying solely on post-scoring corrections for remedies.

[0219] In this embodiment, the consistency anomaly cases of homestay X are stored in the fine-tuning dataset. The model diagnosis bias direction is "insufficient semantic dimension description of real-time room status information". In the next batch fine-tuning, a special optimization will be carried out on the "content generation strategy in the scenario of incomplete real-time data" to reduce the semantic offset problem of generated content caused by the lack of real-time data.

[0220] After uncertainty perception correction and generation-ranking semantic consistency correction, the final comprehensive score for all candidate homestays is calculated. Sort by high to low, then perform diversity-aware reordering ( (Browse exploratory scenarios), and finally output the Top-10 recommendation results. In this embodiment, the top-ranked homestay Y carries the following additional fields in the output structure: uncertainty level. (Low uncertainty, high confidence) Generation-ranking semantic consistency score (High consistency); Additional fields for the second-ranked homestay X: (Medium uncertainty due to real-time data delay) (Inconsistency anomaly markers have been stored in the fine-tuning dataset).

[0221] The aforementioned additional fields serve as content quality confidence indicators in the linked output structure. They are displayed as detailed annotations on the visualization map and presented as data credibility explanations at the bottom of the content details page on the text and image display side. This allows users to have a basis for perceiving the confidence of the recommendation results while obtaining recommended content, thereby improving system transparency and user trust.

Claims

1. A method for optimizing a geo-generative engine and intelligently recommending and ranking search results based on a large model, characterized in that: Includes the following steps: S1, Geo Requirement Awareness and Preprocessing: Receives Geo requirements input by the user, performs text cleaning and geographic terminology normalization; based on named entity recognition and syntactic analysis, extracts Geo spatial features, user intent features, scene features and time dimension features from the standardized text, encodes the time dimension features into a time feature vector and concatenates it with the spatial feature vector to form standardized Geo requirement information, which is then sent to subsequent modules. S2, Large Model-Driven Geo Content Generation: Calls a geo-specific large model, sequentially performs multi-dimensional geo data loading, cross-modal geo feature enhancement, deep integration of requirements and data, and scenario-based geo content generation. It also performs dual verification of geo compliance and requirement matching on the generated content, outputting scenario-based and personalized optimized geo content. S3, Large-Scale Intelligent Recommendation, Retrieval, and Ranking: Based on multi-granularity geographic hierarchical indexing, a set of candidate geographic points is quickly retrieved. The geographic-specific large-scale model performs dual screening of the candidate set by semantic similarity matching and geospatial topology matching. After filtering out invalid points, the geographic-specific large-scale model inputs time feature vectors and scene feature vectors into a dynamic weight allocation network to dynamically allocate weights to four dimensions: Geospatial Association (W1), User Intent Matching (W2), Geographic Rules and Real-Time Data (W3), and User Behavior Preferences (W4). The comprehensive ranking score S is calculated and sorted, and then diversity-aware re-sorting is performed on the ranking results. The formula for calculating the overall ranking score is as follows: ; S4. Output of content generation and ranking results in a coordinated manner: Using the unique ID of the geographic location as the primary key, the geographic content output in step S2 is matched one-to-one with the ranking results in step S3. The content detail is adapted according to the ranking priority, and the linkage results are output through a multi-terminal adaptation interface. S5. Feedback Iterative Optimization: Collect system execution anomaly information and user behavior data, perform hierarchical diagnosis of anomalies, mine user preference features, use anomaly cases and behavior feedback cases as fine-tuning datasets, perform full-link iterative optimization of the geographic-specific large model, generation engine and ranking mechanism, and update system parameters and rules.

2. The method for optimizing a Geo-generative engine and intelligent recommendation retrieval ranking based on a large model as described in claim 1, characterized in that, The time dimension features mentioned in step S1 include the time period to which the query time belongs, the week type, the holiday identifier, and the season identifier. The time feature vector and the Geo spatial feature vector are concatenated and input into the geographic-specific large model. The time feature vector is synchronously transmitted to the dynamic weight allocation network in step S3, so that the network can adaptively adjust the weight allocation of each dimension in time scenarios such as holidays and specific time periods, so as to realize time-aware scenario-based ranking.

3. The method for optimizing a Geo-generative engine and intelligent recommendation retrieval ranking based on a large model as described in claim 1, characterized in that, The geographic-specific large model described in step S2 adopts an improved Transformer architecture, which includes a three-layer core structure with sequential collaboration: The first layer is the Geo semantic understanding layer: a multi-granularity geographic semantic mask matrix is ​​introduced, constructed and weighted in three levels: city, region, and street. The attention weight calculation formula is as follows: ； in It assigns higher attention weight to geographical features such as latitude and longitude, terrain and scene; at the same time, it integrates user historical behavior data to construct user intent vectors, which are then concatenated with time feature vectors to participate in personalized demand analysis. The second layer is the geographic rule fusion layer: it employs a knowledge embedding attention mechanism to vectorize and embed the geographic professional rule base in the form of a knowledge graph. The embedding formula is as follows: ； in For geographic basic feature vectors, For the demand context vector, The rule weight matrix is ​​used; and expired geographic data with an update time exceeding 24 hours is filtered out through a geographic data masking mechanism. The third layer is the scenario-based generation layer: based on the improved Transformer decoder, it configures a dedicated decoding template matrix for each Geo business scenario. The decoding formula is: ； in As a fusion vector of demand and geographic data, The scene bias term; the geo-specific large model uses contrastive learning as the pre-training objective, and the loss function is: ； in This is the semantic representation vector of the anchor point's geographic location. To and The geographic location semantic representation vector that constitutes a positive sample pair Within the same training batch, except Semantic representation vectors of other geographic locations besides itself; This is the temperature coefficient.

4. The method for optimizing a Geo-generative engine and intelligent recommendation retrieval ranking based on a large model as described in claim 1, characterized in that, The cross-modal geographic feature enhancement described in step S2 specifically involves: extracting visual feature vectors from satellite imagery or street view images of geographic locations stored in the Geo data warehouse using a pre-trained convolutional neural network, fusing them with textual geographic feature vectors from a large geographic-specific model, and obtaining the enhanced geographic feature vectors after layer normalization. ； in The cross-modal fusion coefficients are automatically learned through end-to-end training; when image data is missing... Automatically set to 0 to degenerate into plain text mode; the enhanced feature vector It is used for both content generation in step S2 and semantic matching in step S3, thereby improving the semantic distinguishability of geographic locations in terms of visual attributes.

5. The method for optimizing a Geo-generative engine and intelligent recommendation retrieval ranking based on a large model according to claim 1, characterized in that, In step S2, the formula for calculating the demand matching degree during the dual verification of demand matching degree is: ； Among them is the quantity of explicit demand, is the quantity of implicit demand, is the coverage of the is the coverage of the th implicit demand; the matching degree is regarded as qualified when otherwise, the model adjusts the generation logic again and performs secondary generation; after secondary generation if it is still unqualified, a low-confidence mark is triggered and this case is stored in the abnormal fine-tuning dataset.

6. The method for optimizing a Geo-generative engine and intelligent recommendation retrieval ranking based on a large model according to claim 1, characterized in that, The semantic and spatial dual matching mentioned in step S3 specifically involves: semantic matching to calculate the semantic vector of user requirements. Adaptation vectors for geographic locations cosine similarity Filtering geographic locations with a similarity score below 0.8; spatial matching is based on geospatial topology analysis, sequentially performing spatial range matching, geographic constraint matching, and geographic rule matching (filtering locations within ecological protection zones and restricted areas); comprehensive judgment based on dual matching: ； Geographic locations are ranked.

7. The method for optimizing a Geo-generative engine and intelligent recommendation retrieval ranking based on a large model according to claim 1, characterized in that, In the dynamic weight allocation described in step S3, the geographic-specific large model is based on the scene feature vector. The concatenated vector with the time feature vector, using the formula Calculate the weights of the four dimensions to The weights satisfy And follow these rules: for distance-sensitive needs ; for scenario-based needs ; Real-time data sensitive requirements The personalized correction value is: ； in The user intent vector is composed of the user intent features extracted in step S1. This represents the geographic feature vector of the candidate geographic location; User intent similarity; when the user intent similarity exceeds 0.8... ,otherwise Among them, the Geo spatial correlation dimension score , This is the normalized value of the geographic straight-line distance. For traffic accessibility, the calculation formula is as follows: ； To integrate real-time traffic congestion data to calculate actual travel time, The theoretical shortest travel time is calculated based on road network topology; geographical locations with the same score are ranked as follows: Perform a second sort.

8. The method for optimizing a Geo-generative engine and intelligent recommendation retrieval ranking based on a large model according to claim 1, characterized in that, The diversity-perceived reordering in step S3 specifically involves: after the comprehensive score S is sorted, the Top-M candidate points are sorted using the maximum marginal relevance algorithm. Perform a reordering process, selecting the highest-scoring point in each iteration and adding it to the selected set. The scoring rules are as follows: ； in , Calculation based on geographic straight-line distance normalization; The diversity-relevance balance coefficient is dynamically assigned by the geographic-specific large model based on the scene type; the final output set size is Top-N (…). This is used to ensure that the result set has a reasonable geographical distribution and diverse semantic coverage.

9. The method for optimizing a Geo-generative engine and intelligent recommendation retrieval ranking based on a large model according to claim 1, characterized in that, The feedback iterative optimization in step S5 includes: Anomaly classification: based on formula Calculate processing priority. Trigger emergency optimization; otherwise, include it in the regular iteration. User preference modeling: using formulas Accumulate user behavior characteristics, construct a personalized preference model, and integrate it into the generation logic. and ranking weight adjustments; Model fine-tuning: Regularly use anomalous cases and user behavior feedback cases as fine-tuning datasets to perform small-batch fine-tuning training on the geo-specific large model, and simultaneously optimize the cross-modal fusion coefficients. and diversity balance coefficient .

10. The method for optimizing a Geo-generative engine and intelligent recommendation retrieval ranking based on a large model according to claim 1, characterized in that, It also includes a Geo data warehouse as the underlying data support, which comprises a geographic foundation database, a geographic professional rule base, a real-time spatiotemporal database, a user behavior database, and a multi-granularity geographic hierarchical index library; the multi-granularity geographic hierarchical index library stores a three-level R-tree spatial index, and the candidate point fast retrieval query complexity is [missing value]. It supports filtering by geographic category hash labels, and can handle queries with a past query count exceeding a threshold. The system establishes a pre-computed cache for hotspot areas and links it with real-time data updates; the real-time spatiotemporal database establishes a second-level linkage with the Geo generation engine, and when real-time data such as traffic, passenger flow or weather in the target area changes, the engine automatically triggers a content update mechanism to dynamically adjust the generated geographic content. The geographic rule base is stored in the form of a knowledge graph and is used to generate geographic rule matching for engine compliance verification and retrieval ranking.

Images