Model recommendation method and system based on geographic analysis model knowledge graph

Abstract

The application discloses a model recommendation method and system based on a geographic analysis model knowledge graph, relates to the technical field of geographic modeling and simulation, and comprises the following steps: keyword extraction is performed on a received geographic analysis model query request containing a user input model demand text, and attention weight is set; seed model nodes are obtained by performing node multidimensional matching on extracted geographic feature keywords; a plurality of geographic analysis model ripple sets are obtained by performing propagation diffusion on the seed model nodes as nodes, and a propagation score is calculated; a recommendation list is formed by sorting according to the propagation score, and a propagation path of the geographic analysis model is generated; and a large language model is used to generate a full-link interpretable personalized natural language recommendation reason. The model recommendation method based on the geographic analysis model knowledge graph can make full use of structured information in the knowledge graph, capture multi-dimensional and multi-path correlations between models and user demands, and significantly improve the accuracy and robustness of the recommendation result.

Description

Technical Field

[0001] This invention relates to the field of geographic modeling and simulation technology, and in particular to a model recommendation method and system based on geographic analysis model knowledge graphs. Background Technology

[0002] Geographic analysis models are crucial tools for simulating geographical phenomena and their evolution, and for solving geographical problems. In recent years, scholars both domestically and internationally have constructed numerous geographic analysis models for various geographical scenarios. With the collaborative coupling and integrated application of multiple models, the sharing of geographic analysis models has become a vital foundation for solving large-scale, complex geographical problems. However, due to the rapid increase in the number of shared geographic analysis models, users face the challenge of accurately matching their needs to authoritative and effective models.

[0003] Traditional search algorithms can only provide literal matching results based on keywords in a user's query, and cannot provide differentiated rankings based on the deeper intent of different users, resulting in all users receiving completely identical lists when faced with the same query. Furthermore, traditional recommendation system techniques such as collaborative filtering heavily rely on user interaction history, and the recommendation quality drops sharply when faced with sparse interaction data and the cold start problem. At the same time, these methods are mostly black-box models, with weak interpretability of recommendation results, making it difficult to meet the high requirements of users in the geographic analysis modeling field for the reliability and traceability of results.

[0004] A knowledge graph is a heterogeneous graph where nodes represent entities and edges represent relationships between entities. By mapping models, authors, keywords, scenarios, and their attributes to a knowledge graph, a rich semantic network can be constructed, laying the foundation for interpretable recommendations. However, traditional knowledge graph-based recommendation methods mostly employ knowledge graph embedding techniques, transforming entities and relationships into low-dimensional dense vectors and using vector similarity as the basis for recommendations. While this method can capture some semantic associations, it loses the explicit relationships and reasoning paths of the graph structure, degenerating the recommendation process into an uninterpretable black-box matching.

[0005] Therefore, there is an urgent need for a geographic analysis model recommendation method that can fully utilize the semantic and structural information of knowledge graphs, provide transparent and interpretable recommendation criteria, and dynamically adapt to user preferences. Summary of the Invention

[0006] This invention addresses the problems existing in the prior art by proposing a model recommendation method and system based on a geographical analysis model knowledge graph.

[0007] The technical means employed in this invention are as follows: A model recommendation method based on a geographical analysis model knowledge graph includes the following steps: Step 1: Using the geographic analysis model entity recognition model to receive the geographic analysis model query request containing the model requirement text input by the user, extract geographic feature keywords and set attention weights for each extracted geographic feature keyword; Step 2: Input each of the extracted geographic feature keywords into the geographic analysis model knowledge graph and perform multi-dimensional node matching to obtain the seed model node in the geographic analysis model knowledge graph corresponding to the geographic feature keyword; Step 3: In the knowledge graph of the geographic analysis model, propagation and diffusion are performed with each seed model node as a node to obtain multiple geographic analysis model ripple sets, and the propagation score of each geographic analysis model ripple set is calculated. Step 4: Sort the multiple ripple sets of the geographic analysis models according to their propagation scores, select a set number of geographic analysis models from the ripple sets of the geographic analysis models with high propagation scores to form a recommendation list, and generate the propagation path of the geographic analysis models. Step 5: Based on the propagation path and the recommendation list, use a large language model to generate fully interpretable personalized natural language recommendation reasons.

[0008] Furthermore, step 1 also includes classifying the extracted geographic feature keywords to generate geographic analysis model keywords, model terminology keywords, and model metadata vocabulary; When the geographic feature keyword belongs to the geographic analysis model keyword, the geographic feature keyword is input into the geographic analysis model knowledge graph and multi-dimensional matching of nodes is performed to obtain the seed model node, and the geographic analysis model obtained from the seed model node is the directly matched geographic analysis model. When the geographic feature keywords belong to model terminology keywords or model metadata vocabulary, the geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed to obtain seed model nodes. The geographic analysis model obtained from the seed model nodes is an indirect matching geographic analysis model.

[0009] Furthermore, the propagation score of the ripple set in the geographic analysis model is calculated as follows: (1) in, This represents the propagation score of the geographic analysis model; Hit model The number of different keywords; Indicates the geographic analysis model to which the diffusion occurs; This represents the initial attention weight of keywords in the geographic analysis model, with a value of [0, 1]. This represents the hop count decay factor, with a value range of (0, 1); This represents the propagation hop count from the seed model node corresponding to the geographic feature keyword to the current node, and is a non-negative integer; when h =0 indicates a direct match; when h ≥1 indicates that it has been passed h Secondary relationship propagation; This represents the type of relationship between two nodes in a knowledge graph and is an enumeration value. Representing relation type r The propagation weight is a non-negative real number that takes values ​​between [0, 1]. This represents the reward coefficient based on the number of keywords.

[0010] Furthermore, the geographical analysis model knowledge graph includes nodes for the full Chinese name of the model, the abbreviation of the model in Chinese, the full English name of the model, the abbreviation of the model in English, the purpose node, model classification information, keyword nodes, model applicable scope information nodes, developer information nodes, distributor information nodes, release date nodes, model reference system information nodes, model performance information nodes, model coupling information nodes, model operating hardware condition nodes, model operating software condition nodes, and the relationships between the nodes; When the geographic feature keyword belongs to the geographic analysis model keyword, the extracted geographic feature keyword is input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed, including: calculating the similarity between each geographic feature keyword and the full Chinese name node, abbreviated Chinese name node, full English name node, and abbreviated English name node of the model in the geographic analysis model knowledge graph, and obtaining the graph direct model node with the highest matching degree with each standardized model name; When the geographic feature keywords belong to model terminology keywords, the extracted geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed, including: matching each geographic feature keyword with the target node, model classification information, keyword node, and model application scope information node in the geographic analysis model knowledge graph, and using them as seed nodes to participate in diffusion; When the geographic feature keywords belong to the model metadata vocabulary, the extracted geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed, including: each geographic feature keyword is matched with the developer information node, distributor information node, release date node, model reference system information node, model performance information node, model coupling information node, model running hardware condition node, and model running software condition node in the geographic analysis model knowledge graph, and is used as a seed node to participate in diffusion.

[0011] Furthermore, it also includes obtaining users' historical demand preferences and setting attention weights for the geographic feature keywords based on the users' historical demand preferences.

[0012] A system for a model recommendation method based on a geographic analysis model knowledge graph as described in this application includes: The query parsing module is used to receive a geographic analysis model query request containing the model requirement text input by the user, and to extract geographic feature keywords from the model requirement text using a geographic analysis model entity recognition model, and to set an attention weight for each extracted geographic feature keyword. The knowledge graph processing module is used to input each of the extracted geographic feature keywords into the geographic analysis model knowledge graph and perform multi-dimensional matching of nodes to obtain the seed model nodes in the geographic analysis model knowledge graph corresponding to the geographic feature keywords. The recommendation engine module is used to obtain multiple geographic analysis model ripple sets by propagating and diffusing with each seed model node as a node in the geographic analysis model knowledge graph, and to calculate the propagation score of each geographic analysis model ripple set by performing propagation score calculation on each geographic analysis model ripple set. The sorting output module is used to sort multiple ripple sets of the geographic analysis models according to their propagation scores, select a set number of geographic analysis models from the ripple sets of the geographic analysis models with high propagation scores to form a recommendation list, and generate the propagation path of the geographic analysis models. An interpretability generation module is used to generate fully interpretable personalized natural language recommendation reasons based on the propagation path and the recommendation list using a large language model.

[0013] Furthermore, the query parsing module is also used to classify the extracted geographic feature keywords to generate geographic analysis model keywords, model terminology keywords, and model metadata vocabulary. When the geographic feature keyword belongs to the geographic analysis model keyword, the geographic feature keyword is input into the geographic analysis model knowledge graph and multi-dimensional matching of nodes is performed to obtain the seed model node, and the geographic analysis model obtained from the seed model node is the directly matched geographic analysis model. When the geographic feature keywords belong to model terminology keywords or model metadata vocabulary, the geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed to obtain model metadata nodes, which are further diffused to obtain seed model nodes. The geographic analysis model obtained from the seed model nodes is an indirectly matched geographic analysis model.

[0014] Furthermore, the recommendation engine module calculates the propagation score for each of the geographic analysis model ripple sets, as follows: (2) in, This represents the propagation score of the geographic analysis model; Hit model The number of different keywords; Indicates the geographic analysis model to which the diffusion occurs; This represents the initial attention weight of keywords in the geographic analysis model, with a value of [0, 1]. This represents the hop count decay factor, with a value range of (0, 1); This represents the propagation hop count from the seed model node corresponding to the geographic feature keyword to the current node, and is a non-negative integer; when h =0 indicates a direct match; when h ≥1 indicates that it has been passed h Secondary relationship propagation; This represents the type of relationship between two nodes in a knowledge graph and is an enumeration value. Representing relation type r The propagation weight is a non-negative real number that takes values ​​between [0, 1]. This represents the reward coefficient based on the number of keywords.

[0015] Furthermore, the geographical analysis model knowledge graph includes nodes for the full Chinese name of the model, the abbreviation of the model in Chinese, the full English name of the model, the abbreviation of the model in English, the purpose node, model classification information, keyword nodes, model applicable scope information nodes, developer information nodes, distributor information nodes, release date nodes, model reference system information nodes, model performance information nodes, model coupling information nodes, model operating hardware condition nodes, model operating software condition nodes, and the relationships between the nodes; The knowledge graph processing module is used to input each extracted geographic feature keyword into the geographic analysis model knowledge graph and perform multi-dimensional node matching as follows: When the geographic feature keyword belongs to the geographic analysis model keyword, the extracted geographic feature keyword is input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed, including: calculating the similarity between each geographic feature keyword and the full Chinese name node, abbreviated Chinese name node, full English name node, and abbreviated English name node of the model in the geographic analysis model knowledge graph, and obtaining the graph direct model node with the highest matching degree with each standardized model name; When the geographic feature keywords belong to model terminology keywords, the extracted geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed, including: matching each geographic feature keyword with the target node, model classification information, keyword node, and model application scope information node in the geographic analysis model knowledge graph, and using them as seed nodes to participate in diffusion; When the geographic feature keywords belong to the model metadata vocabulary, the extracted geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed, including: each geographic feature keyword is matched with the developer information node, distributor information node, release date node, model reference system information node, model performance information node, model coupling information node, model running hardware condition node, and model running software condition node in the geographic analysis model knowledge graph, and is used as a seed node to participate in diffusion.

[0016] Compared with existing technologies, the model recommendation method and system based on geographic analysis model knowledge graph disclosed in this invention have the following beneficial effects: 1. The model recommendation method based on a geographic analysis model knowledge graph disclosed in this invention parses user queries into multiple geographic feature keywords, and then performs a wave diffusion operation on the knowledge graph starting from each geographic feature keyword. Finally, the diffusion results of multiple keywords are combined to score the model. This multi-source cross-diffusion mechanism can fully utilize the structured information in the knowledge graph to capture the multi-dimensional and multi-path association between the model and user needs. Compared with traditional single-vector retrieval or graph-based random walk methods, it significantly improves the accuracy and robustness of the recommendation results.

[0017] 2. During the diffusion process, this invention dynamically adjusts the propagation strength of different relationship paths. This mechanism enables nodes that are semantically closer to the user query to obtain higher propagation gains, while semantically irrelevant nodes are effectively suppressed. Thus, without increasing the training data, adaptive optimization of relationship weights is achieved, enhancing the model's ability to understand natural language queries.

[0018] 3. This invention not only outputs a list of recommended models but also provides a breakdown of the total score for each model (displaying the contribution score for each keyword and the keyword quantity reward), and can automatically generate natural language recommendation descriptions for top-ranked models (by calling a large language model). Furthermore, it generates the shortest path from keywords to recommended models, allowing users to visually verify the recommendation basis. These interpretable designs greatly enhance user trust in the recommendation results and facilitate developers in diagnosing and optimizing models.

[0019] 4. This invention is based on a general knowledge graph representation and relation diffusion algorithm, and does not depend on data formats or rules specific to any particular domain. By simply replacing the node types and relation definitions in the knowledge graph, it can be applied to various scenarios such as product recommendation, document recommendation, and expert discovery, demonstrating excellent cross-domain transferability. Attached Figure Description

[0020] Figure 1 This is a flowchart of the model recommendation method based on a geographical analysis model knowledge graph disclosed in this invention; Figure 2 This is a schematic diagram of the module structure of the model recommendation system based on geographic analysis model knowledge graph disclosed in this invention; Figure 3 A detailed flowchart of an embodiment of the method recommended by the present invention. Figure 4 This is a schematic diagram of the knowledge graph diffusion algorithm in this invention; Figure 5 This is a partial structural diagram of a knowledge graph according to a specific embodiment of the present invention; Figure 6 This is a schematic diagram of a recommended process according to a specific embodiment of the present invention. Detailed Implementation

[0021] like Figure 1 As shown, the model recommendation method based on geographic analysis model knowledge graph disclosed in this invention includes the following steps: Step 1: Using a geographic analysis model entity recognition model to extract geographic feature keywords from the received geographic analysis model query request containing the user-input model requirement text, this embodiment assigns attention weights to each keyword according to the following attention rules: The keyword categories for geographic analysis models are prioritized from highest to lowest as follows: model metadata entry information > model classification information > model application scope information > model modeling principle information > spatial data / reference system information > model developer / distributor information > model operating conditions information, with different weight coefficients for different categories. Keywords located in the subject or core object position of the query text have a higher weight than keywords that appear as modifiers. Keywords containing spatial information such as river basin names and topographic region names are given additional weight increments.

[0022] The above rules are used to address the technical problem of inaccurate identification of the importance of keywords in geographic analysis models in spatial analysis scenarios by general models, and to improve the recognition accuracy of keywords in geographic analysis models.

[0023] In this embodiment, the entity recognition model of the geographic analysis model can be an existing model or it can be trained based on the input geographic analysis model literature corpus. Step 2: Input each of the extracted geographic feature keywords into the geographic analysis model knowledge graph and perform multi-dimensional node matching to obtain the seed model node in the geographic analysis model knowledge graph corresponding to the geographic feature keyword; Step 3: In the knowledge graph of the geographic analysis model, propagation and diffusion are performed with each seed model node as a node to obtain multiple geographic analysis model ripple sets, and the propagation score of each geographic analysis model ripple set is calculated. Step 4: Sort the multiple ripple sets of the geographic analysis models according to their propagation scores, select a set number of geographic analysis models from the ripple sets of the geographic analysis models with high propagation scores to form a recommendation list, and generate the propagation path of the geographic analysis models. Step 5: Based on the propagation path and the recommendation list, a large language model is used to generate fully interpretable personalized natural language recommendation reasons. That is, the models are sorted in descending order according to the final score, and a preset number of models are selected as the recommendation results. Explainable information of the recommendation results is generated and output. The explainable information includes: displaying the contribution score of the model under each keyword and the score composition information of the keyword quantity reward, generating the shortest graph traversal path from the keyword node to the recommendation model node and presenting it as a knowledge graph diffusion path, and calling the large language model to generate natural language recommendation reasons of a preset length based on the query text, model name and description, and the recommendation logic formed by the shortest path.

[0024] This invention discloses a model recommendation method based on a geographic analysis model knowledge graph. This method parses user queries into multiple geographic feature keywords and then performs a wave-like diffusion across the knowledge graph, starting with each keyword. Finally, it integrates the diffusion results of multiple keywords to score the model. This multi-source cross-diffusion mechanism fully utilizes the structured information in the knowledge graph, capturing multi-dimensional and multi-path connections between the model and user needs. Compared to traditional single-vector retrieval or graph-based random walk methods, it significantly improves the accuracy and robustness of recommendation results. During the diffusion process, this invention dynamically adjusts the propagation intensity of different relationship paths. This mechanism allows nodes semantically closer to the user query to receive higher propagation gains, while semantically irrelevant nodes are effectively suppressed. This achieves adaptive optimization of relationship weights without increasing training data, enhancing the model's understanding of natural language queries. This invention not only outputs a list of recommended models but also provides a total score decomposition for each model (displaying the contribution score of each keyword and the keyword quantity reward), and can automatically generate natural language recommendation descriptions for top-ranked models (by calling a large language model). Furthermore, it generates the shortest path from keywords to the recommended models, allowing users to visually verify the recommendation basis. These interpretable designs greatly enhance user trust in the recommendation results and facilitate developers in diagnosing and optimizing the model. This invention is based on a general knowledge graph representation and relation diffusion algorithm, independent of domain-specific data formats or rules. Simply by replacing the node types and relation definitions in the knowledge graph, it can be applied to various scenarios such as product recommendation, literature recommendation, and expert discovery, demonstrating excellent cross-domain transferability.

[0025] Furthermore, step 1 also includes classifying the extracted geographic feature keywords to generate geographic analysis model keywords, model terminology keywords, and model metadata vocabulary; When the geographic feature keyword belongs to the geographic analysis model keyword, the geographic feature keyword is input into the geographic analysis model knowledge graph and multi-dimensional matching of nodes is performed to obtain the seed model node, and the geographic analysis model obtained from the seed model node is the directly matched geographic analysis model. When the geographic feature keywords belong to model terminology keywords or model metadata vocabulary, the geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed to obtain seed model nodes. The geographic analysis model obtained from the seed model nodes is an indirect matching geographic analysis model.

[0026] This step considers that model terminology keywords or model metadata terms in a single user input will propagate through an additional layer in the graph structure compared to geographic analysis model keywords, resulting in a reduction in the final model score reached by model terminology keywords or model metadata terms. By classifying the model terminology keywords and the seed nodes they point to, placing them at the same initial layer, the system's usability and intelligence are greatly enhanced.

[0027] Furthermore, the propagation score of the ripple set in the geographic analysis model is calculated as follows: (3) in, This represents the propagation score of the geographic analysis model; Hit model The number of different keywords; Indicates the geographic analysis model to which the diffusion occurs; This represents the initial attention weight of keywords in the geographic analysis model, with a value of [0, 1]. This represents the hop count decay factor, with a value range of (0, 1); This represents the propagation hop count from the seed model node corresponding to the geographic feature keyword to the current node, and is a non-negative integer; when h =0 indicates a direct match; when h ≥1 indicates that it has been passed h Secondary relationship propagation; This represents the type of relationship between two nodes in a knowledge graph and is an enumeration value. Representing relation type r The propagation weight is a non-negative real number that takes values ​​between [0, 1]. This represents the reward coefficient based on the number of keywords.

[0028] like Figure 4 As shown, the initial attention weights of the seed node The weights of the relation after r The next step j is calculated, where j is an intermediate node and can be any node. If j is a model node, the propagation score needs to be calculated. After multiple propagation steps, the final node m is obtained. i The scores of a certain model are accumulated and a bonus score is added before being sent to the sorting algorithm.

[0029] The aforementioned propagation score calculation formula addresses the semantic drift problem in multi-hop reasoning of knowledge graphs by using a hop count decay factor γ, ensuring that models farther from the seed node score lower than those closer. It distinguishes between different semantic relationships using relation type weights θ and guarantees fairness in multi-keyword comprehensive matching through keyword quantity rewards. This algorithm ensures the semantic relevance of recommendation results while accommodating both precise and generalized queries. Compared to existing recommendation methods based on graph traversal or pure vector similarity, it significantly improves retrieval accuracy, personalized adaptation capabilities, and interpretability.

[0030] Furthermore, the geographical analysis model knowledge graph includes nodes for the full Chinese name of the model, the abbreviation of the model in Chinese, the full English name of the model, the abbreviation of the model in English, the purpose node, model classification information, keyword nodes, model applicable scope information nodes, developer information nodes, distributor information nodes, release date nodes, model reference system information nodes, model performance information nodes, model coupling information nodes, model operating hardware condition nodes, model operating software condition nodes, and the relationships between the nodes; When the geographic feature keyword belongs to the geographic analysis model keyword, the extracted geographic feature keyword is input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed, including: calculating the similarity between each geographic feature keyword and the full Chinese name node, abbreviated Chinese name node, full English name node, and abbreviated English name node of the model in the geographic analysis model knowledge graph, and obtaining the graph direct model node with the highest matching degree with each standardized model name; When the geographic feature keywords belong to model terminology keywords, the extracted geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed, including: matching each geographic feature keyword with the target node, model classification information, keyword node, and model application scope information node in the geographic analysis model knowledge graph, and using them as seed nodes to participate in diffusion; When the geographic feature keywords belong to the model metadata vocabulary, the extracted geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed. This includes: mapping each geographic feature keyword to developer information nodes, distributor information nodes, release date nodes, model reference system information nodes, model performance information nodes, model coupling information nodes, model operating hardware condition nodes, and model operating software condition nodes in the geographic analysis model knowledge graph, serving as seed nodes for diffusion. This step defines diffusion directions for different types of keywords, ensuring accuracy, and decomposes user needs into multiple directions for propagation, ensuring the system can handle diverse queries.

[0031] Furthermore, it also includes obtaining users' historical demand preferences and setting attention weights for the geographic feature keywords based on the users' historical demand preferences.

[0032] This step involves acquiring and utilizing users' historical needs and preferences to personalize the keyword attention weights. This mechanism allows the system's keyword extraction results to be continuously optimized as the number of times users use the system increases. Furthermore, the system's relationship weight factor mapping table can be updated synchronously based on user preferences, thereby improving the recommendation effect.

[0033] like Figure 2 As shown, the system disclosed in this application for a model recommendation method based on a geographic analysis model knowledge graph includes: The query parsing module, whose core is the geographic analysis model keyword recognizer, is used to receive geographic analysis model query requests containing user-input model requirement text. It uses a geographic analysis model entity recognition model to extract geographic feature keywords from the model requirement text and sets attention weights for each extracted geographic feature keyword. Furthermore, the geographic analysis model keyword recognizer is trained using a corpus of literature related to the geographic analysis model field and the geoscience field. The input is a natural language query, and the output is a geographic analysis model keyword and its corresponding attention weight. This weight is used to express the importance of a keyword in the query.

[0034] The knowledge graph processing module is used to input each of the extracted geographic feature keywords into the geographic analysis model knowledge graph and perform multi-dimensional matching of nodes to obtain the seed model nodes in the geographic analysis model knowledge graph corresponding to the geographic feature keywords. The recommendation engine module is used to obtain multiple geographic analysis model ripple sets by propagating and diffusing with each seed model node as a node in the geographic analysis model knowledge graph, and to calculate the propagation score of each geographic analysis model ripple set by performing propagation score calculation on each geographic analysis model ripple set. The sorting output module is used to sort multiple ripple sets of the geographic analysis models according to their propagation scores, select a set number of geographic analysis models from the ripple sets of the geographic analysis models with high propagation scores to form a recommendation list, and generate the propagation path of the geographic analysis models. An interpretability generation module is used to generate fully interpretable personalized natural language recommendation reasons based on the propagation path and the recommendation list using a large language model.

[0035] Furthermore, the query parsing module is also used to classify the extracted geographic feature keywords to generate geographic analysis model keywords, model terminology keywords, and model metadata vocabulary. When the geographic feature keyword belongs to the geographic analysis model keyword, the geographic feature keyword is input into the geographic analysis model knowledge graph and multi-dimensional matching of nodes is performed to obtain the seed model node, and the geographic analysis model obtained from the seed model node is the directly matched geographic analysis model. When the geographic feature keywords belong to model terminology keywords or model metadata vocabulary, the geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed to obtain model metadata nodes, which are further diffused to obtain seed model nodes. The geographic analysis model obtained from the seed model nodes is an indirectly matched geographic analysis model.

[0036] Furthermore, the recommendation engine module calculates the propagation score for each of the geographic analysis model ripple sets, as follows: (4) in, This represents the propagation score of the geographic analysis model; Hit model The number of different keywords; Indicates the geographic analysis model to which the diffusion occurs; This represents the initial attention weight of keywords in the geographic analysis model, with a value of [0, 1]. This represents the hop count decay factor, with a value range of (0, 1); This represents the propagation hop count from the seed model node corresponding to the geographic feature keyword to the current node, and is a non-negative integer; when h =0 indicates a direct match; when h ≥1 indicates that it has been passed h Secondary relationship propagation; This represents the type of relationship between two nodes in a knowledge graph and is an enumeration value. Representing relation type r The propagation weight is a non-negative real number that takes values ​​between [0, 1]. This represents the keyword quantity reward coefficient; Furthermore, the geographical analysis model knowledge graph includes nodes for the full Chinese name of the model, the abbreviation of the model in Chinese, the full English name of the model, the abbreviation of the model in English, the purpose node, model classification information, keyword nodes, model applicable scope information nodes, developer information nodes, distributor information nodes, release date nodes, model reference system information nodes, model performance information nodes, model coupling information nodes, model operating hardware condition nodes, model operating software condition nodes, and the relationships between the nodes; The knowledge graph processing module is used to input each extracted geographic feature keyword into the geographic analysis model knowledge graph and perform multi-dimensional node matching as follows: When the geographic feature keyword belongs to the geographic analysis model keyword, the extracted geographic feature keyword is input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed, including: calculating the similarity between each geographic feature keyword and the full Chinese name node, abbreviated Chinese name node, full English name node, and abbreviated English name node of the model in the geographic analysis model knowledge graph, and obtaining the graph direct model node with the highest matching degree with each standardized model name; When the geographic feature keywords belong to model terminology keywords, the extracted geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed, including: matching each geographic feature keyword with the target node, model classification information, keyword node, and model application scope information node in the geographic analysis model knowledge graph, and using them as seed nodes to participate in diffusion; When the geographic feature keywords belong to the model metadata vocabulary, the extracted geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed, including: each geographic feature keyword is matched with the target node, model classification node, model application scope information node, keyword node, concept classification node, developer information node, distributor information node, model reference system information node, model performance information node, model coupling information node, and model operation condition node in the geographic analysis model knowledge graph, and is used as a seed node to participate in diffusion.

[0037] The overall recommendation process for this method is as follows: Figure 3As shown: First, user input text is received and fed into the geographic analysis model keyword recognizer to extract a list of geographic analysis model keywords. Then, the keywords in the list are evaluated: if the current keyword is a model name, a layer of diffusion is performed, and the resulting new words are written back to the geographic analysis model keyword list, continuing the iteration; if the current keyword is not a model name, it is used as a seed node, and graph diffusion is performed on the seed node based on the geographic analysis model knowledge graph to obtain diffused candidate entities.

[0038] The candidate entities obtained from the diffusion are again checked to see if they are model names: if they are not model names, the diffusion operation is returned and the map diffusion is iterated; if the candidate entity is a model name, the scoring stage is entered, the model score corresponding to the geographic analysis model is calculated, and the ripple set corresponding to this diffusion is obtained.

[0039] Next, determine whether there is a common geographic analysis model in the ripple set: if there is a common model, add bonus points after adding all scores on the basis of the original model score, and retain the diffusion path of the highest score; if there is no common model, directly retain the current model score, and retain the diffusion path of the top-5 highest scores.

[0040] After scoring and filtering all candidate models, the top 5 diffusion paths with the highest scores are selected (top-5). A top-5 recommendation list is generated based on these top-5 paths. Simultaneously, a recommendation reason is generated for each recommended model based on the selected diffusion paths. Finally, the ranked recommendation list and recommendation reasons are integrated, and the geographic analysis model recommendation results are output.

[0041] Example Step 1: Users input natural language queries through the web interface, such as: "Our team is studying the impact of stormwater runoff and flooding on surface water quality in City A. We have already built a pipe network model and now need a model that can couple sewage overflow with river water quality simulation. Can you recommend a model that can simulate the convection and diffusion of pollutants?" The user's input is then fed into the geographic analysis model and entity recognition model to extract structured keywords and attention weight factors. Example of analysis results: {("Rainwater runoff", 0.90), ("Water quality", 0.80), ("Pipeline model", 0.70), ("Sewage overflow", 0.80),}.

[0042] Step 2: Input each of the extracted geographic feature keywords into the geographic analysis model knowledge graph and perform multi-dimensional node matching to obtain the seed model node in the geographic analysis model knowledge graph corresponding to the geographic feature keyword; The three geographic feature keywords, “rainwater runoff”, “water quality”, “pipeline network model” and “sewage overflow”, were respectively input into the geographic analysis model knowledge graph for multi-dimensional node matching to obtain the seed model nodes corresponding to the geographic feature keywords in the geographic analysis model knowledge graph. The knowledge graph structure in this embodiment is as follows: Figure 5 As shown, the knowledge graph of the complete geographic analysis model includes nodes for the model's full Chinese name, abbreviated Chinese name, full English name, abbreviated English name, keywords, model classification, applicable scope, concept classification, developer information, distributor information, release date, model reference system information, model performance information, model coupling information, hardware conditions for model operation, and software conditions for model operation, as well as relation types such as AUTHOR_SHIP, HAS_KEYWORD, SIMILAR_TO, DEVELOPED_BY, DISTRIBUTED_BY, DERIVED_FROM, CATEGORIZED_AS, APPLICABLE_TO, INSTANCE_OF, ACHIEVES, INTERFACES_WITH, REQUIRES, and SERVES_AS. The graph contains approximately 24,000 nodes and 100,000 relations, stored in the Neo4j graph database. In this embodiment, the propagation process is illustrated using the nodes reached in this propagation.

[0043] Step 3: In the knowledge graph of the geographic analysis model, the geographic analysis model ripple set is obtained by propagating and diffusing the keyword nodes "rainwater runoff", "water quality", "pipeline model" and "sewage overflow" as seed nodes respectively. In this embodiment, the keyword "rainwater runoff" (weight 0.90) is used as an example. It is used as the keyword type seed node and propagates along the "CATEGORIZED_AS" relationship to the concept classification node "hydrological geography". From the matched concept classification node, it continues to spread outward along all relationships to obtain the keyword node "urban runoff" and continues to spread along the "HAS_KEYWORD" relationship to the model node "SWMM".

[0044] If a non-model node (such as the concept classification node "Hydrogeography") is encountered in the propagation path, the propagation continues. Within the maximum number of diffusion layers (set to 3), the propagation reaches a model node each time, and the model node score is calculated.

[0045] In this embodiment, the relational weighting factor The model defaults to the same settings, so this is not reflected. Supervised learning is performed using data collected after extensive user interaction. A relationship weight mapping table can be set based on user preferences to further improve recommendation performance.

[0046] Figure 6 The figure shows the paths and corresponding scores obtained after each keyword diffusion. After the four keywords have diffused, the model "SWMM" is simultaneously matched by "rainwater runoff", "water quality", and "pipeline model". An additional bonus of +0.20 is awarded for matching all three keywords simultaneously. The model has a default hop count decay factor. The score is 0.5. For model "SWMM", the scores for each diffusion to this node can be calculated using the formula as 0.11, 0.10, and 0.35, respectively, and the total score after all are added together is 0.76. Similarly, the scores for model "EFDC" are 0.40, model "Pipedream-WQ" is 0.20, model "swmm_mcp" is 0.18, and model "WEB-SWMM" is 0.18.

[0047] Step 4: Sort the ripple set of geographic analysis models by propagation score, and select the top 5 models to form a recommendation list. Generate a score decomposition, and normalize the correlation between the scores obtained by the SWMM model recommendation process and other models to be recommended, showing contributions of "rainwater runoff: 0.11," "water quality: 0.10," "pipeline model: 0.35," and "keyword reward: +0.20." Simultaneously, visualize the recorded knowledge graph paths for easy user verification.

[0048] Step 5, as follows Figure 6 As shown, based on the propagation path and the recommendation list, a large language model is used to generate a fully interpretable personalized natural language recommendation reason. In this embodiment, the model is invoked, and combined with the user query, model name and description, and the above path logic, a 30 to 60-word natural language recommendation reason is generated using the large language model, such as: "SWMM is an authoritative hydrological model specifically used for urban stormwater and flood simulation. It has been cited in many urban flooding studies and is highly matched with the scenario of the impact of stormwater runoff and urban flooding on surface water quality that you are concerned about." The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A model recommendation method based on a geographical analysis model knowledge graph, characterized in that, Includes the following steps: Step 1: Using the geographic analysis model entity recognition model to receive the geographic analysis model query request containing the model requirement text input by the user, extract geographic feature keywords and set attention weights for each extracted geographic feature keyword; Step 2: Input each of the extracted geographic feature keywords into the geographic analysis model knowledge graph and perform multi-dimensional node matching to obtain the seed model node in the geographic analysis model knowledge graph corresponding to the geographic feature keyword; Step 3: In the knowledge graph of the geographic analysis model, propagation and diffusion are performed with each seed model node as a node to obtain multiple geographic analysis model ripple sets, and the propagation score of each geographic analysis model ripple set is calculated. Step 4: Sort the multiple ripple sets of the geographic analysis models according to their propagation scores, select a set number of geographic analysis models from the ripple sets of the geographic analysis models with high propagation scores to form a recommendation list, and generate the propagation path of the geographic analysis models. Step 5: Based on the propagation path and the recommendation list, use a large language model to generate fully interpretable personalized natural language recommendation reasons.

2. The model recommendation method based on geographical analysis model knowledge graph according to claim 1, characterized in that: Step 1 further includes classifying the extracted geographic feature keywords to generate geographic analysis model keywords, model terminology keywords, and model metadata vocabulary. When the geographic feature keyword belongs to the geographic analysis model keyword, the geographic feature keyword is input into the geographic analysis model knowledge graph and multi-dimensional matching of nodes is performed to obtain the seed model node, and the geographic analysis model obtained from the seed model node is the directly matched geographic analysis model. When the geographic feature keywords belong to model terminology keywords or model metadata vocabulary, the geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed to obtain seed model nodes. The geographic analysis model obtained from the seed model nodes is an indirect matching geographic analysis model.

3. The model recommendation method based on geographical analysis model knowledge graph according to claim 1, characterized in that: The propagation score of the ripple set in the geographic analysis model is calculated as follows: When the geographic feature keywords belong to the geographic analysis model keywords, the propagation score of the direct matching geographic analysis model is calculated as follows: （1） in, This represents the propagation score of the geographic analysis model; Hit model The number of different keywords; Indicates the geographic analysis model to which the diffusion occurs; This represents the initial attention weight of keywords in the geographic analysis model, with a value of [0, 1]. This represents the hop count decay factor, with a value range of (0, 1); This represents the propagation hop count from the seed model node corresponding to the geographic feature keyword to the current node, and is a non-negative integer; when h =0 indicates a direct match; when h ≥1 indicates that it has been passed h Secondary relationship propagation; This represents the type of relationship between two nodes in a knowledge graph and is an enumeration value. Representing relation type r The propagation weight is a non-negative real number that takes values ​​between [0, 1]. This represents the reward coefficient based on the number of keywords.

4. The model recommendation method based on geographical analysis model knowledge graph according to claim 1, characterized in that: The geographical analysis model knowledge graph includes nodes for the full Chinese name of the model, the abbreviation of the model in Chinese, the full English name of the model, the abbreviation of the model in English, the purpose node, model classification information, keyword nodes, model applicable scope information nodes, developer information nodes, distributor information nodes, release date nodes, model reference system information nodes, model performance information nodes, model coupling information nodes, model operating hardware condition nodes, model operating software condition nodes, and the relationships between the nodes. When the geographic feature keyword belongs to the geographic analysis model keyword, the extracted geographic feature keyword is input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed, including: calculating the similarity between each geographic feature keyword and the full Chinese name node, abbreviated Chinese name node, full English name node, and abbreviated English name node of the model in the geographic analysis model knowledge graph, and obtaining the graph direct model node with the highest matching degree with each standardized model name; When the geographic feature keywords belong to model terminology keywords, the extracted geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed, including: matching each geographic feature keyword with the target node, model classification information, keyword node, and model application scope information node in the geographic analysis model knowledge graph, and using them as seed nodes to participate in diffusion; When the geographic feature keywords belong to the model metadata vocabulary, the extracted geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed, including: each geographic feature keyword is matched with the developer information node, distributor information node, release date node, model reference system information node, model performance information node, model coupling information node, model running hardware condition node, and model running software condition node in the geographic analysis model knowledge graph, and is used as a seed node to participate in diffusion.

5. The model recommendation method based on geographical analysis model knowledge graph according to claim 1, characterized in that: It also includes obtaining users' historical needs and preferences, and setting attention weights for the geographic feature keywords based on the users' historical needs and preferences.

6. A system for the model recommendation method based on a geographic analysis model knowledge graph as described in any one of claims 1-5, characterized in that, include: The query parsing module is used to receive a geographic analysis model query request containing the model requirement text input by the user, and to extract geographic feature keywords from the model requirement text using a geographic analysis model entity recognition model, and to set an attention weight for each extracted geographic feature keyword. The knowledge graph processing module is used to input each of the extracted geographic feature keywords into the geographic analysis model knowledge graph and perform multi-dimensional matching of nodes to obtain the seed model nodes in the geographic analysis model knowledge graph corresponding to the geographic feature keywords. The recommendation engine module is used to obtain multiple geographic analysis model ripple sets by propagating and diffusing with each seed model node as a node in the geographic analysis model knowledge graph, and to calculate the propagation score of each geographic analysis model ripple set by performing propagation score calculation on each geographic analysis model ripple set. The sorting output module is used to sort multiple ripple sets of the geographic analysis models according to their propagation scores, select a set number of geographic analysis models from the ripple sets of the geographic analysis models with high propagation scores to form a recommendation list, and generate the propagation path of the geographic analysis models. An interpretability generation module is used to generate fully interpretable personalized natural language recommendation reasons based on the propagation path and the recommendation list using a large language model.

7. The system for model recommendation based on geographic analysis model knowledge graph according to claim 6, characterized in that: The query parsing module is also used to classify the extracted geographic feature keywords to generate geographic analysis model keywords, model terminology keywords, and model metadata vocabulary. When the geographic feature keyword belongs to the geographic analysis model keyword, the geographic feature keyword is input into the geographic analysis model knowledge graph and multi-dimensional matching of nodes is performed to obtain the seed model node, and the geographic analysis model obtained from the seed model node is the directly matched geographic analysis model. When the geographic feature keywords belong to model terminology keywords or model metadata vocabulary, the geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed to obtain model terminology keyword nodes or model metadata vocabulary nodes. Seed model nodes are then obtained through further diffusion, and the geographic analysis model obtained from the seed model nodes is the indirectly matched geographic analysis model.

8. The system for model recommendation based on geographic analysis model knowledge graph according to claim 7, characterized in that: The recommendation engine module calculates the propagation score for each ripple set of the geographic analysis model, as follows: Furthermore, the recommendation engine module calculates the propagation score for each of the geographic analysis model ripple sets, as follows: （2） in, This represents the propagation score of the geographic analysis model; Hit model The number of different keywords; Indicates the geographic analysis model to which the diffusion occurs; This represents the initial attention weight of keywords in the geographic analysis model, with a value of [0, 1]. This represents the hop count decay factor, with a value range of (0, 1); This represents the propagation hop count from the seed model node corresponding to the geographic feature keyword to the current node, and is a non-negative integer; when h =0 indicates a direct match; when h ≥1 indicates that it has been passed h Secondary relationship propagation; This represents the type of relationship between two nodes in a knowledge graph and is an enumeration value. Representing relation type r The propagation weight is a non-negative real number that takes values ​​between [0, 1]. This represents the reward coefficient based on the number of keywords.

9. The system for model recommendation based on geographic analysis model knowledge graph according to claim 8, characterized in that: The geographical analysis model knowledge graph includes nodes for the full Chinese name of the model, the abbreviation of the model in Chinese, the full English name of the model, the abbreviation of the model in English, the purpose node, model classification information, keyword nodes, model applicable scope information nodes, developer information nodes, distributor information nodes, release date nodes, model reference system information nodes, model performance information nodes, model coupling information nodes, model operating hardware condition nodes, model operating software condition nodes, and the relationships between the nodes. The knowledge graph processing module is used to input each extracted geographic feature keyword into the geographic analysis model knowledge graph and perform multi-dimensional node matching as follows: When the geographic feature keyword belongs to the geographic analysis model keyword, the extracted geographic feature keyword is input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed, including: calculating the similarity between each geographic feature keyword and the full Chinese name node, abbreviated Chinese name node, full English name node, and abbreviated English name node of the model in the geographic analysis model knowledge graph, and obtaining the graph direct model node with the highest matching degree with each standardized model name; When the geographic feature keywords belong to model terminology keywords, the extracted geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed, including: matching each geographic feature keyword with the target node, model classification information, keyword node, and model application scope information node in the geographic analysis model knowledge graph, and using them as seed nodes to participate in diffusion; When the geographic feature keywords belong to the model metadata vocabulary, the extracted geographic feature keywords are input into the geographic analysis model knowledge graph and multi-dimensional node matching is performed, including: each geographic feature keyword is matched with the developer information node, distributor information node, release date node, model reference system information node, model performance information node, model coupling information node, model running hardware condition node, and model running software condition node in the geographic analysis model knowledge graph, and is used as a seed node to participate in diffusion.

Images