Deep learning model variable optimization selection method and device considering geographic causal relationship

Abstract

The invention discloses a deep learning model variable optimization selection method and device considering a geographic causal relationship. The method comprises the following steps: pre-training an open source large language model to obtain a first geographic large model; performing fine adjustment on the first geographic large model to obtain a second geographic large model; receiving user input comprising a to-be-studied variable set V and associated data, clearly processing a variable X and a target variable Y, and calling the first geographic large model to generate additional metadata for each to-be-studied variable in the V; determining a causal sequence among the to-be-researched variables in the V based on the second geographic big model; constructing a causal skeleton diagram for the V, and orienting edges in the causal skeleton diagram through the causal sequence to obtain a causal directed acyclic graph (DAG); screening nodes meeting an adjustment rule in the DAG to obtain an additional variable set Z; x and Z are used as input variables of the deep learning model, Y is used as an output variable, and optimization selection of variables of the deep learning model considering the geographic causal relationship is completed.

Description

Technical Field

[0001] The present invention relates to the fields of geographic modeling and causal science, and more specifically, to a method and device for optimizing the selection of variables for a deep learning model taking into account geographic causal relationships. Background Art

[0002] As a key tool for exploring complex geographical systems, geographic modeling plays a vital role in deepening our understanding of geographical phenomena. However, current deep learning models for geographical phenomena are often limited to mining statistical correlations between variables and fail to fully reveal the causal logic behind them, which, to a certain extent, limits the in-depth mining and application of geographical knowledge by the model.

[0003] In the context of machine learning, accurate modeling and inference of causal variables and their mechanisms can lead to more robust feature representations. This is crucial for explaining the generation principles of observational data. To achieve this goal, we usually need to have a deep understanding of geographical phenomena, construct a causal directed acyclic graph (DAG) between geographical variables, identify and exclude confounding variables, and select the remaining variables as additional inputs for modeling features. However, geographic causal modeling is highly dependent on expert knowledge.

[0004] The field of geography has a rich literature resource, and the knowledge contained in this literature provides valuable information for geographic causal modeling. Recently, large language models have shown great potential in simulating domain expert knowledge and performing causal reasoning. Large language models can act as a proxy for human domain knowledge, approximating the causal order provided by experts, thereby automating the causal reasoning process. In addition, when used in combination with existing causal discovery algorithms, large language models are expected to significantly improve the efficiency of automated causal inference, bringing revolutionary changes to the field of geographic modeling. In this way, we can more effectively utilize the knowledge in geographic literature and promote geographic research towards a deeper level of causal understanding. Summary of the invention

[0005] In order to solve the technical problems mentioned in the above background technology, the present invention proposes a method and device for optimizing the selection of variables of a deep learning model taking into account geographical causal relationships.

[0006] The present invention provides a method for optimizing and selecting variables of a deep learning model taking into account geographical causal relationships, comprising:

[0007] Step S1: Build a geographic professional knowledge base, pre-train the open source large language model, and obtain the first geographic large model;

[0008] Step S2: construct a causal relationship dataset containing geographic context information, and fine-tune the first geographic big model to obtain a second geographic big model;

[0009] Step S3: receiving user input including a variable set V to be studied and associated data, specifying a processing variable X and a target variable Y, and calling the first geographic macromodel to generate additional metadata for each variable to be studied in the variable set V to be studied;

[0010] Step S4: determining the causal order between the variables to be studied in the variable set V to be studied based on the second geographic macro model;

[0011] Step S5: construct a causal skeleton graph for the variable set V to be studied, and obtain a causal directed acyclic graph DAG by directing the edges in the causal skeleton graph through the causal order;

[0012] Step S6: Filter the nodes in the causal directed acyclic graph DAG that satisfy the adjustment rules to obtain an additional variable set Z;

[0013] Step S7: Using the processing variable X and the additional variable set Z as the input variables of the deep learning model and the target variable Y as the output variable of the deep learning model, the optimal selection of variables of the deep learning model taking into account the geographical causal relationship is completed.

[0014] Furthermore, in step S1, a geographic professional knowledge base is constructed, and pre-training is performed on an open source large language model to obtain a first geographic large model, including:

[0015] Geographic knowledge text data is widely extracted from various public resources. The public resources include but are not limited to professional books, academic papers, industry reports, encyclopedia entries, etc. The geographic knowledge text data includes but is not limited to geographic entity concepts, classical theories, geographic phenomena, etc.

[0016] The collected data is preprocessed to build a geographic knowledge base. The preprocessing methods include but are not limited to text cleaning, text deduplication, text segmentation, etc.

[0017] The open source language big model is further pre-trained based on the geographic knowledge base to obtain a first geographic big model. After pre-training, the first geographic big model has the general knowledge ability of geographic field knowledge.

[0018] Furthermore, in step S2, a causal relationship dataset containing geographic context information is constructed, and the first geographic big model is fine-tuned to obtain a second geographic big model, including:

[0019] Filter out text paragraphs containing clear causal relationship descriptions from the geographic text knowledge base, such as descriptions of geographic processes and geographic conclusions, extract the geographic context of events, manually identify causal relationship pairs, and construct a geographic causal relationship dataset D. Each data item d in the dataset D = {STA, GD, TAs, CRDs}, where:

[0020] STA = {Locaion, Time}, indicating spatial-temporal attributes, including location and time;

[0021] GD = domain, which indicates the identifier of the geographical discipline or research field (Geographical Domain), such as urban planning, environmental science, etc.

[0022] TAs={Attribute1,Attribute2,...,Attribute k}, representing thematic attributes, including a set of specific geographical features, such as climate, topography, etc.;

[0023] CRDs = {CRD1,CRD2,...,CRD n}, represents a causal relationship (Causal Relationship Description) set, where CRD = {Cause; Effect; Description}, Cause represents the cause variable, Effect represents the result variable, and Description represents the description of the causal relationship.

[0024] STA, GD and TAs in D are used to provide contextual information for geographic questions, CRDs are used to reason and analyze geographic questions, question-answer pairs are constructed, the first geographic model is fine-tuned, and the second geographic model is obtained.

[0025] Further, in step S3, user input including a variable set V to be studied and associated data is received, a processing variable X and a target variable Y are clarified, and the first geographic macromodel is called to generate additional metadata for each variable to be studied in the variable set V to be studied, including:

[0026] The user inputs the variable set to be studied and the associated data, and also needs to input the current modeling research field and the geographical context description of the research area. The variables to be studied are terms of various types in the geographical research field, such as temperature, precipitation, etc. The variable set should satisfy causal sufficiency, that is, the direct cause variables of any two variables in the variable set exist in the set;

[0027] The user needs to specify which pair of variables is the treatment variable X and the target variable Y. The XY variable pair is the variable that the user expects to model the causal effect.

[0028] The first geographic macromodel is called to generate additional metadata for each variable in the variable set and output in the form of <variable name, additional metadata>. The additional metadata at least includes a semantic description of the geographic phenomenon represented by the variable.

[0029] Furthermore, in step S4, based on the second geographic macro model, determining the causal order between the variables to be studied in the variable set V to be studied includes:

[0030] For a given set of geographic variables to be studied, all possible sets of triples are constructed, where each triple is composed of different elements in the set, represented by T = {<vi,vj,vk> |i,j,k∈{1,2,…,S} and i≠j≠k}, where vi,vj,vk are the i,j,kth variables to be studied respectively; S is the total number of elements in the set V;

[0031] Using the triple prompt word technique, based on the second geographic model, the causal relationship between the three variables to be studied in each triple in T is determined to form a directed acyclic graph subDAG;

[0032] Through the majority voting mechanism, the directed acyclic graph subDAG of all triples is merged to comprehensively obtain the overall causal structure; the majority voting mechanism is to count the number of times the causal relationship between each two variables to be studied is determined as a causal relationship in all triples, and select the direction with the most occurrences as the final causal direction;

[0033] The reasoning of the entire causal discovery process is recorded to ensure the transparency and traceability of the process and provide users with detailed documentation. The reasoning includes the response output of the geographic big model when performing causal orientation.

[0034] Furthermore, in step S5, a causal skeleton graph is constructed for the variable set V to be studied, and a causal directed acyclic graph DAG is obtained by directing the edges in the causal skeleton graph in the causal order, including:

[0035] A causal skeleton graph is created for the set of geographic variables according to a constraint-based PC (Peter-Clark) causal discovery algorithm. The edges in the causal skeleton graph are undirected or partially directed.

[0036] The causal skeleton graph is oriented using the causal order obtained in S4 to obtain a final causal directed acyclic graph DAG.

[0037] Further, in step S6, the nodes satisfying the adjustment rules in the causal directed acyclic graph DAG are screened to obtain an additional variable set Z, including:

[0038] The nodes that meet the adjustment criteria in the causal directed acyclic graph are selected to form a variable set. The variable X and the variables in the set are used as input variables to train the deep learning model, and the causal effect is calculated through the observed data:

[0039]

[0040] Among them, E[Y|X=x] is the observation expectation, and E[Y|do(X=x)] is the intervention expectation, that is, the causal effect.

[0041] For the path from X to Y, 1) if the nodes in the set block all non-causal paths from X to Y, and 2) and the nodes in the set are not on the causal path from X to Y, then the adjustment criterion of (X, Y) is satisfied.

[0042] Furthermore, in step S7, the processing variable X and the additional variable set Z are used as input variables of the deep learning model, and the target variable Y is used as the output variable of the deep learning model to complete the optimal selection of variables of the deep learning model that takes into account geographical causal relationships.

[0043] The present invention also provides a device for optimizing and selecting variables of a deep learning model taking into account geographical causal relationships, the device comprising:

[0044] A large model pre-training module, used to construct a causal relationship dataset containing geographic context information, fine-tune the first geographic large model, and obtain a second geographic large model;

[0045] A large model fine-tuning module, used to construct a causal relationship data set containing geographic context information, and fine-tune the first geographic large model to obtain a second geographic large model;

[0046] A user interaction module is used to receive user input including a set of variables to be studied V and associated data, to specify a processing variable X and a target variable Y, and to call the first geographic macromodel to generate additional metadata for each variable to be studied in the set of variables to be studied V;

[0047] A causal order generation module is used to determine the causal order between the variables to be studied in the variable set V to be studied based on the second geographic macromodel and using a triplet prompt word technique;

[0048] A causal graph generation module is used to construct a causal skeleton graph for the variable set V to be studied, and obtain a causal directed acyclic graph DAG by directing the edges in the causal skeleton graph through the causal order;

[0049] The additional variable screening module is used to screen the nodes that meet the adjustment rules in the causal directed acyclic graph DAG to obtain an additional variable set Z.

[0050] The present invention also provides a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device, cause the computing device to perform the method as described above.

[0051] The present invention also provides an electronic device, comprising one or more processors, one or more memories and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing the method as described above.

[0052] After adopting the above scheme, the present invention has the following beneficial effects:

[0053] 1) The present invention simulates expert knowledge through a large language model to perform automated causal order inference, thereby reducing direct dependence on expert knowledge and improving the efficiency of causal reasoning.

[0054] 2) The present invention improves the interpretability and accuracy of the causal discovery process of geographic variables by integrating knowledge-driven causal discovery and statistical causal inference algorithms.

[0055] 3) The present invention can more accurately model the causal effects between geographical variables by considering the characteristic variables screened by causal relationships. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a schematic diagram of the technical process of the present invention.

[0057] Figure 2 The causal skeleton graph and the causal directed acyclic graph are illustrated in the embodiments of the present invention, where (a) is the causal skeleton graph and (b) is the causal directed acyclic graph.

[0058] Figure 3 A schematic diagram of a device for optimizing and selecting variables for a deep learning model taking into account geographical causal relationships provided by the present invention. DETAILED DESCRIPTION

[0059] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the protection scope of the present invention.

[0060] The purpose of the present invention is to provide a deep learning model variable optimization selection and device that takes into account geographical causal relationships. It simulates expert knowledge through a large language model to perform automated causal sequence inference, reduces direct dependence on expert knowledge, and improves the efficiency of causal reasoning. By integrating knowledge-driven causal discovery and statistical causal inference algorithms, it improves the interpretability and accuracy of the causal discovery process of geographical variables. By considering the characteristic variables screened by causal relationships, it can more accurately model the causal effects between geographical variables.

[0061] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0062] like Figure 1 As shown, the present invention provides a method for optimizing and selecting variables of a deep learning model taking into account geographical causal relationships, comprising:

[0063] Step S1: Collect geographic knowledge texts, preprocess the data, and build a geographic text knowledge base for further pre-training of the open source language model to obtain the first geographic model.

[0064] Specifically, geographical knowledge text data is widely extracted from various public resources. The public resources include but are not limited to professional books, academic papers, industry reports, encyclopedia entries, etc. The geographical knowledge text data includes but is not limited to geographical entity concepts, classical theories, geographical phenomena, etc.

[0065] The collected data is preprocessed to build a geographic knowledge base. The preprocessing methods include but are not limited to text cleaning, text deduplication, text segmentation, etc.

[0066] Based on the geographic text knowledge base, the open source language model is further pre-trained to obtain the first geographic model. After pre-training, the first geographic model has the general knowledge of geographic field knowledge.

[0067] Step S2: construct a high-quality causal relationship dataset containing geographic context information to fine-tune the first geographic big model to obtain a second geographic big model.

[0068] Specifically, text paragraphs containing clear descriptions of causal relationships, such as descriptions of geographic processes and conclusions, are selected from the geographic text knowledge base, the geographic context of events is extracted, causal relationship pairs are manually identified, and a geographic causal relationship dataset D is constructed. Each data item d in the dataset D = {STA, GD, TAs, CRDs}, where:

[0069] STA = {Locaion, Time}, indicating spatial-temporal attributes, including location and time;

[0070] GD = domain, which indicates the identifier of the geographical discipline or research field (Geographical Domain), such as urban planning, environmental science, etc.

[0071] TAs={Attribute1,Attribute2,...,Attribute k}, representing thematic attributes, including a set of specific geographical features, such as climate, topography, etc.;

[0072] CRDs = {CRD1,CRD2,...,CRD n}, represents a causal relationship (Causal Relationship Description) set, where CRD = {Cause; Effect; Description}, Cause represents the cause variable, Effect represents the result variable, and Description represents the description of the causal relationship.

[0073] STA, GD and TAs in D are used to provide contextual information for geographic questions, CRDs are used to reason and analyze geographic questions, question-answer pairs are constructed, the first geographic model is fine-tuned, and the second geographic model is obtained.

[0074] In this example, assuming that we are studying the impact of urbanization on air quality, we can construct the following dataset D entries:

[0075] STA: {Location:"Beijing",Time:"January 2023"}

[0076] GD: "Environmental Science"

[0077] TAs:

[0078] "PM2.5 concentration": 40 micrograms per cubic meter

[0079] "Temperature": -5℃

[0080] "Relative humidity": 40%

[0081] "Wind speed": 2 m / s

[0082] CRDs:

[0083] CRD1: {Cause: "Increased industrial emissions"; Effect: "Increased PM2.5 concentration";

[0084] Description:"Increased industrial activity has led to higher concentrations of fine particles in the air."}

[0085] CRD2: {Cause: "PM2.5 concentration increased"; Effect: "Air quality decreased";

[0086] Description:"Increases in fine particulate matter concentrations will reduce the Air Quality Index."}

[0087] Step S3: receiving user input including a variable set V to be studied and associated data, clarifying a processing variable X and a target variable Y, and calling the first geographic macromodel to generate additional metadata for each variable to be studied in the variable set V to be studied;

[0088] Specifically, the user inputs the variable set to be studied and the associated data, and also needs to input the current modeling research field and the geographical context description of the research area. The variables to be studied are terms of various types in the geographical research field, such as temperature, precipitation, etc. The variable set should satisfy causal sufficiency, that is, the direct cause variables of any two variables in the variable set exist in the set;

[0089] The user needs to specify which pair of variables is the treatment variable X and the target variable Y. The XY variable pair is the variable that the user expects to model the causal effect.

[0090] The first geographic macromodel is called to generate additional metadata for each variable in the variable set and output in the form of <variable name, additional metadata>. The additional metadata at least includes a semantic description of the geographic phenomenon represented by the variable.

[0091] In this example, if the user studies the urban heat island effect in Shanghai, a possible input is:

[0092] Variable set V: {temperature (T), green coverage (GC), building density (BD), industrial emissions (IE), population density (PD)}

[0093] Related data: Contains the observed values ​​of the above variables in different cities and time periods.

[0094] Research area: Environmental science, especially urban climatology.

[0095] Description of the geographical context of the study area: e.g., Shanghai, 2023.

[0096] Treatment variable X: Industrial emissions (IE)

[0097] Target variable Y: Temperature (T)

[0098] The first geographic model is called, and additional metadata is generated for each variable in the variable set V, for example:

[0099] <Temperature (T), "the average temperature of an urban area, affected by many factors, including green space coverage, building density, etc.">

[0100] <Green Coverage (GC), "the percentage of green area in an urban area, which plays a positive role in regulating temperature.">

[0101] <Building density (BD), “the degree to which buildings are concentrated per unit area, which may increase the absorption and storage of heat.”>

[0102] <Industrial Emissions (IE), “Greenhouse gas emissions from industrial activities that may increase heat loads in urban areas.”>

[0103] <Population density (PD), "the number of inhabitants per unit area, is related to energy consumption and heat generation.">

[0104] Step S4: determining the causal order between the variables to be studied in the variable set V to be studied based on the second geographic macro model;

[0105] Specifically, for a given set of geographic variables to be studied, all possible sets of triples are constructed, where each triple is composed of different elements in the set, represented by T = {<vi,vj,vk> |i,j,k∈{1,2,…,S} and i≠j≠k}, where vi,vj,vk are the i,j,kth variables to be studied respectively; S is the total number of elements in the set V;

[0106] Using the triple prompt word technique, the nodes of each triple in the set are oriented based on the geographic model to form a directed acyclic graph subDAG;

[0107] In this example, given the variable set = {temperature (T), green coverage (GC), building density (BD), industrial emissions (IE), population density (PD)}, construct all possible triple sets T. Each triple t consists of different elements in, for example, T = {<t1={T,GC,BD}> ,t2={T,BD,IE}>,...}, using triplet prompts, the nodes in each triplet in the set T are oriented through the second geographic model.

[0108] The triple prompt word technology provided by the present invention uses the in-context learning (ICL) method in the prompt construction process, injects examples into the prompt, and allows the large model to focus on the reasoning process. This method realizes the instant learning capability of the model without updating the model parameters. The following table 1 is an example of a prompt template:

[0109] Table 1

[0110]

[0111]

[0112] Specifically, through the majority voting mechanism, all triples of subDAGs are merged to obtain the overall causal structure. For each variable pair, the number of times it is determined as a causal relationship in all subDAGs is counted, and the direction with the most occurrences is selected as the final causal direction. If a conflict occurs (i.e., the number of votes for two or three possible edge orientations is the same), the second geographic model is used again to resolve the conflict through the chain of thought prompt (CoT) and make a final decision.

[0113] In the example, suppose that in the triple<T1={T,GC,BD}> In the model, the causal relationship between BD→T and GC→T is determined.<T2={T,IE,PD}> In the model, it is possible to determine the causal relationship of IE→T and PD→T. Through majority voting, if the majority of subDAGs show that T is affected by BD and GC, the final causal direction will be determined as BD→T and GC→T.

[0114] Specifically, the reasoning basis of the entire causal discovery process is recorded to ensure the transparency and traceability of the process and provide users with detailed documentation. The reasoning basis includes the response output of the geographic large model when performing causal orientation. It includes the causal analysis results of all triples, the model response output, and the final causal structure diagram.

[0115] Step S5, constructing a causal skeleton graph for the variable set V to be studied, and obtaining a causal directed acyclic graph DAG by directing the edges in the causal skeleton graph through the causal order.

[0116] Specifically, according to the constraint-based PC (Peter-Clark) causal discovery algorithm, a causal skeleton diagram is created for the set of geographic variables, such as Figure 2 As shown in (a) in the figure, the edges in the causal skeleton graph are undirected or partially directed. The causal skeleton graph is directed using the causal order obtained in S4 to obtain the final causal directed acyclic graph DAG, as shown in Figure 2 As shown in (b).

[0117] In the example, an undirected edge is added to each variable pair in the variable set V. The PC algorithm conditional independence test is used to test each pair of variables in the variable set V to determine whether there is a direct causal relationship between them, and a causal skeleton graph is obtained. The causal order information obtained by the S4 algorithm is used to further direct the edges in the causal skeleton graph. The introduction of the PC algorithm increases the robustness of causal discovery.

[0118] Step S6, screening the nodes satisfying the adjustment rules in the causal directed acyclic graph DAG to obtain an additional variable set Z.

[0119] Specifically, the nodes that meet the adjustment criteria in the causal directed acyclic graph are selected to form a variable set Z. The variables X and the variables in the set Z are used as input variables to train the deep learning model to calculate the causal effect through observation data:

[0120]

[0121] Among them, E[Y|X=x] is the observation expectation, and E[Y|do(X=x)] is the intervention expectation, that is, the causal effect.

[0122] For the path from X to Y, 1) if the nodes in set Z block all non-causal paths from X to Y, and 2) and the nodes in set Z are not on the causal path from X to Y, then Z satisfies the adjustment criteria of (X, Y).

[0123] Step S7: Using the processing variable X and the additional variable set Z as the input variables of the deep learning model and the target variable Y as the output variable of the deep learning model, the optimal selection of variables of the deep learning model taking into account the geographical causal relationship is completed.

[0124] The present invention simulates expert knowledge through a large language model to perform automated causal sequence inference, thereby reducing direct reliance on expert knowledge and improving the efficiency of causal reasoning. By integrating knowledge-driven causal discovery and statistical causal inference algorithms, the interpretability and accuracy of the causal discovery process of geographic variables are improved. By considering characteristic variables screened for causal relationships, the causal effects between geographic variables can be modeled more accurately.

[0125] The present invention also provides a device for optimizing and selecting variables of a deep learning model taking into account geographical causal relationships, specifically, Figure 3 As shown, the device comprises:

[0126] A large model pre-training module, used to construct a causal relationship dataset containing geographic context information, fine-tune the first geographic large model, and obtain a second geographic large model;

[0127] A large model fine-tuning module, used to construct a causal relationship data set containing geographic context information, and fine-tune the first geographic large model to obtain a second geographic large model;

[0128] A user interaction module is used to receive user input including a set of variables to be studied V and associated data, to specify a processing variable X and a target variable Y, and to call the first geographic macromodel to generate additional metadata for each variable to be studied in the set of variables to be studied V;

[0129] A causal order generation module is used to determine the causal order between the variables to be studied in the variable set V to be studied based on the second geographic macromodel and using a triplet prompt word technique;

[0130] A causal graph generation module is used to construct a causal skeleton graph for the variable set V to be studied, and obtain a causal directed acyclic graph DAG by directing the edges in the causal skeleton graph through the causal order;

[0131] The additional variable screening module is used to screen the nodes that meet the adjustment rules in the causal directed acyclic graph DAG to obtain an additional variable set Z.

[0132] The technical solution of the above-mentioned deep learning model variable optimization selection device taking into account geographical causal relationships is similar to the aforementioned method and will not be repeated here.

[0133] Based on the same technical solution, the present invention also discloses a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device, enable the computing device to execute the above-mentioned deep learning model variable optimization selection method taking into account geographic causal relationships.

[0134] Based on the same technical solution, the present invention also discloses a computing device, including one or more processors, one or more memories and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing the above-mentioned deep learning model variable optimization selection method taking into account geographic causal relationships.

[0135] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0136] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0137] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0138] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0139] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for optimizing the selection of variables in a deep learning model taking into account geographical causal relationships, characterized in that: include: Step S1: Build a geographic professional knowledge base, pre-train the open source large language model, and obtain the first geographic large model; Step S2: construct a causal relationship dataset containing geographic context information, and fine-tune the first geographic big model to obtain a second geographic big model; Step S3: receiving user input including a variable set V to be studied and associated data, specifying a processing variable X and a target variable Y, and calling the first geographic macromodel to generate additional metadata for each variable to be studied in the variable set V to be studied; Step S4: determining the causal order between the variables to be studied in the variable set V to be studied based on the second geographic macro model; Step S5: construct a causal skeleton graph for the variable set V to be studied, and obtain a causal directed acyclic graph DAG by directing the edges in the causal skeleton graph through the causal order; Step S6: Filter the nodes in the causal directed acyclic graph DAG that satisfy the adjustment rules to obtain an additional variable set Z; Step S7: Using the processing variable X and the additional variable set Z as the input variables of the deep learning model and the target variable Y as the output variable of the deep learning model, the optimal selection of variables of the deep learning model taking into account the geographical causal relationship is completed.

2. According to claim 1, a method for optimizing and selecting variables of a deep learning model taking into account geographical causal relationships is characterized in that: The step S1 specifically includes: Geographic knowledge text data obtained from various public resources, including professional books, academic papers, industry reports and encyclopedia entries, and the geographic knowledge text data includes geographic entity concepts, classical theories and geographical phenomena; Preprocessing the acquired geographic knowledge text data to build a geographic knowledge base; wherein the preprocessing method includes text cleaning, text deduplication and text segmentation; The open source language big model is pre-trained based on the geographic professional knowledge base to obtain a first geographic big model.

3. The method for optimizing and selecting variables of a deep learning model taking into account geographic causal relationships according to claim 1, characterized in that: The step S2 specifically includes: The text paragraphs containing clear causal relationship descriptions are selected from the geographic text knowledge base, the geographic contexts are extracted, the causal relationship pairs are manually identified, and a causal relationship dataset D is constructed; each data d in the causal relationship dataset D = {STA, GD, TAs, CRDs}, where: Spatiotemporal attribute STA = {Locaion, Time}, Location is the geographical location, Time is the time; Geographical discipline or research field GD = domain, where domain is an identifier; Thematic attributes TAs = {Attribute1, Attribute2, ..., Attribute t }, Attribute k is the tth geographical feature; Causal relationship set CRDs = {CRD1,CRD2,...,CRD N }, where CRD n ={Cause n ;Effect n ;Description n }, Cause n The cause variable representing the nth causal relationship, Effect n Indicates the result variable of the nth causal relationship, Description n Represents the description of the nth causal relationship. STA, GD and TAs in D are used to provide contextual information for geographic questions, CRDs are used to reason and analyze geographic questions, question-answer pairs are constructed, the first geographic model is fine-tuned, and the second geographic model is obtained.

4. The method for optimizing and selecting variables of a deep learning model taking into account geographic causal relationships according to claim 1, characterized in that: The step S3 specifically includes: Receive user input, the user input includes a set of variables V to be studied, associated data, and a description of the current modeling research field and the geographical context of the research area, wherein V should satisfy causal sufficiency, that is, the direct cause variables of any two variables to be studied in V exist in V; The user specifies the processing variable X and the target variable Y to form an XY variable pair, which serves as the variable for modeling the causal effect that the user expects; The first geographic macromodel is called to generate additional metadata for each variable to be studied, and output in the form of <variable name, additional metadata>; wherein the additional metadata at least includes a semantic description of the geographic phenomenon represented by the variable.

5. The method for optimizing and selecting variables of a deep learning model taking into account geographic causal relationships according to claim 4 is characterized in that: The step S4 comprises: For V, construct all possible triple sets T, T = {<vi,vj,vk> |i,j,k∈{1,2,…,S} and i≠j≠k}, where vi,vj,vk are the i,j,kth variables to be studied respectively; S is the total number of elements in the set V; Using the triple prompt word technique, based on the second geographic model, the causal relationship between the three variables to be studied in each triple in T is determined to form a directed acyclic graph subDAG; Through the majority voting mechanism, the directed acyclic graph subDAG of all triples is merged to comprehensively obtain the overall causal structure; the majority voting mechanism is to count the number of times the causal relationship between each two variables to be studied is determined as a causal relationship in all triples, and select the direction with the most occurrences as the final causal direction.

6. The method for optimizing and selecting variables of a deep learning model taking into account geographic causal relationships according to claim 5, characterized in that: The step S5 comprises: creating a causal skeleton graph for the said according to a constraint-based PC causal discovery algorithm; The causal skeleton graph is oriented using the causal order obtained in S4 to obtain a final causal directed acyclic graph DAG.

7. The method for optimizing and selecting variables of a deep learning model taking into account geographic causal relationships according to claim 6, characterized in that: The adjustment rule in step S6 is: whether the node blocks all non-causal paths from X to Y and is not on the causal path from X to Y.

8. A device for optimizing and selecting variables of a deep learning model taking into account geographical causal relationships, characterized in that: The device comprises: The large model pre-training module is used to build a geographic professional knowledge base, pre-train the open source large language model, and obtain the first geographic large model; A large model fine-tuning module, used to construct a causal relationship data set containing geographic context information, and fine-tune the first geographic large model to obtain a second geographic large model; A user interaction module is used to receive user input including a set of variables to be studied V and associated data, specify a processing variable X and a target variable Y, and call the first geographic macromodel to generate additional metadata for each variable to be studied in the set of variables to be studied V; A causal order generation module is used to determine the causal order between the variables to be studied in the variable set V to be studied based on the second geographic macromodel and using a triplet prompt word technique; A causal graph generation module is used to construct a causal skeleton graph for the variable set V to be studied, and obtain a causal directed acyclic graph DAG by directing the edges in the causal skeleton graph through the causal order; The additional variable screening module is used to screen the nodes that meet the adjustment rules in the causal directed acyclic graph DAG to obtain an additional variable set Z.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that When the processor executes the computer program, the steps of the method described in any one of claims 1 to 7 are implemented.

10. A computer storage medium storing a computer program, characterized in that When the computer program is executed by a processor, the steps of the method described in any one of claims 1 to 7 are implemented.

Images