GeoShapley engineering acceleration method and system for large-scale high-dimensional space interpretation task

Abstract

The invention discloses a GeoShapley engineering acceleration method and system for large-scale high-dimensional space interpretation tasks, and the method comprises the steps: obtaining and processing an input sample, and obtaining a candidate interpretation point set of the input sample; performing representative screening on the candidate interpretation point set to obtain representative interpretation points; feature subsets in the representative explanation points are screened through an on-demand generation and complementary pairwise strategy; performing hierarchical reconstruction and weight endowing on the screened feature subsets according to a scale layer and a feature subset inner layer, and calculating the weights of the reconstructed feature subsets; and carrying out weighted regression solution according to the reconstructed weight of the feature subset to obtain a feature contribution value of each feature in the feature subset. Through representative explanation point screening, feature subset screening and hierarchical reconstruction of the kernel weight of the feature subset, the construction time and storage pressure of the feature subset are reduced, the peak memory occupation is remarkably reduced, the stability is improved, and the calculation efficiency of an explanation calculation project is improved.

Description

Technical Field

[0001] This invention relates to the fields of computer software and geographic information systems, and in particular to a GeoShapley engineering acceleration method and system for large-scale high-dimensional spatial interpretation tasks. Background Technology

[0002] Among existing interpretable artificial intelligence technologies, the SHAP method, based on the Shapley value idea, decomposes the model output into the additive contribution of each input feature, thereby achieving interpretable attribution for single-sample predictions. Its typical implementation estimates the contribution of each feature by sampling different feature subsets and fitting a weighted linear model.

[0003] In geospatial data scenarios, spatial autocorrelation and spatial heterogeneity make it impossible to fully express the background effects corresponding to geographical location using the traditional feature independence assumption. Therefore, geographical location needs to be introduced as an explicit feature into the Shapley game framework to achieve a decomposable interpretation of environmental feature contributions, location contributions, and their interaction contributions.

[0004] On the other hand, several general feature attribution calculation and interpretation generation schemes have been proposed in published patents regarding the computational efficiency and interpretation generation of Shapley values ​​or SHAP. However, their improvement focuses on general tabular data or general model interpretation processes. For the memory and runtime bottlenecks caused by the explosion of feature subset combinations, construction of kernel weight objects, generation of design matrices, and batch calculation of spatial interpretation points at the national scale in GeoShapley scenarios that use geographical location as a feature, there is still a lack of targeted engineering solutions.

[0005] In engineering implementation, existing GeoShapley interpretation computations typically use kernel-weighted regression to estimate Shapley class contributions. This approach requires generating several feature subsets for each interpreted sample and constructing corresponding binary design matrices and kernel weight objects. Then, the contribution coefficients are obtained by weighted least squares.

[0006] When the number of explanatory points is large and the feature variables have high dimensionality, i.e., the explanatory task has a large computational scale, directly constructing the index set, design matrix, and long weight vector related to the feature subset combination at the beginning of the operation will generate a huge array in memory, leading to a significant increase in memory consumption, slowing down the overall computation process, and even causing the program to fail. This type of problem is more prominent in spatial planning tasks because explanatory points often come from high-resolution rasters at the national scale, and feature variables often reach tens of dimensions and include geographical location features.

[0007] Furthermore, common feature subset generation methods in existing implementations include explicit enumeration or approximate enumeration of power sets, which causes the feature subset index object to grow exponentially with the number of features, making it difficult to operate stably under high-dimensional feature variable conditions. Summary of the Invention

[0008] To address the problems of large-scale computational engineering tasks, high memory consumption, and low computational efficiency in existing technologies, this invention provides a GeoShapley engineering acceleration method and system for large-scale high-dimensional space interpretation tasks, which reduces computational data memory and improves computational efficiency.

[0009] Therefore, the technical solution adopted by the present invention is as follows: A GeoShapley engineering acceleration method for large-scale high-dimensional space interpretation tasks is provided, the method comprising: The input samples of the GeoShapley interpretation computation project are obtained, the coordinate system and resolution of the input samples are unified, and pixel-by-pixel alignment is performed to obtain a set of candidate interpretation points for the input samples; the input samples are specifically geographic raster samples containing environmental features and geographic location features. A similarity metric is calculated for each explanatory point in the candidate explanatory point set, and representative explanatory points are obtained by representative screening based on the calculation results; The feature subsets in each representative explanatory point are screened using an on-demand generation and complementary pairing strategy. The on-demand generation and complementary pairing strategy specifically sets a threshold for the number of feature subsets and a specific type, and screens feature subsets belonging to a specific type. When the number of screened feature subsets does not reach the threshold, supplementary feature subsets are extracted from the remaining feature subsets, and complementary subsets of the supplementary feature subsets are generated. Each selected feature subset is reconstructed and weighted according to the scale layer and the inner layer of the feature subset, and the weight of the reconstructed feature subset is calculated. We perform weighted regression based on the weights of the reconstructed feature subsets to obtain the feature contribution value of each feature in the feature subset.

[0010] According to the above scheme, representative explanatory points are obtained in the following ways: The input samples are divided into spatial units according to a preset spatial resolution. Within each spatial cell, all interpretation points that satisfy the mask and data validity conditions are selected, and their standardized feature vector set is constructed. Calculate the similarity measure of the explanatory points within each spatial unit and the average vector of the standardized feature vector set. Based on the calculation results, select representative explanatory points and integrate the representative explanatory points of all spatial units to form representative explanatory points.

[0011] According to the above scheme, the similarity measure is specifically a distance measure; the selection of representative explanatory points based on the calculation results is specifically as follows: the pixel with the closest Euclidean distance from the average value vector is selected as the representative explanatory point, or the representative explanatory point is determined according to the median sample or center sample criterion; among them, the center sample criterion is specifically as follows: when the spatial unit is a square grid, the center point of the square grid is taken as the representative explanatory point.

[0012] The spatial resolution preset according to the above scheme is specifically latitude and longitude grid resolution or projected coordinate grid resolution.

[0013] According to the above scheme, specific types of feature subsets include the empty set, the universal set, all feature subsets whose size does not exceed a preset size threshold, and their complementary sets; the specific process for selecting feature subsets from representative explanatory points is as follows: Set a threshold for the number of feature subsets and a specific type to cover feature subsets of a specific type; Determine whether the number of covered feature subsets has reached the threshold. If not, randomly select the remaining feature subsets from the feature subsets whose size exceeds the preset threshold, and generate their complementary sets to obtain the filtered feature subsets.

[0014] According to the above scheme, the weighted regression solution is specifically calculated based on the binary matrix of the feature subset, the weight matrix of the feature subset, and the model prediction vector of the feature subset.

[0015] According to the above scheme, the model prediction vector of the feature subset is obtained by predicting each feature subset through the model prediction function and arranging the predicted values ​​in the order of the feature subsets; the weight matrix of the feature subset is specifically a kernel weight diagonal matrix, and its diagonal elements are the scalar weights corresponding to the weights of each reconstructed feature subset.

[0016] According to the above scheme, the reconstructed feature subset weights are specifically calculated from the scale layer weights and the inner weights of the feature subsets; among them, the scale layer weights are calculated from the total number of features and the size of the feature subsets; the inner weights of the feature subsets are calculated from the number of combinations of feature subset sizes selected from the total number of features.

[0017] A GeoShapley engineering acceleration system for large-scale high-dimensional space interpretation tasks is also provided, the system comprising: The input sample processing module is used to acquire input samples from the GeoShapley interpretation computation project, unify the coordinate system and resolution of the input samples, and perform pixel-by-pixel alignment to obtain a set of candidate interpretation points for the input samples; the input samples are specifically geographic raster samples containing environmental features and geographic location features. The representative explanatory point filtering module is used to calculate the similarity measure of each explanatory point in the candidate explanatory point set, and to perform representative filtering based on the calculation results to obtain representative explanatory points; The feature subset filtering module is used to filter feature subsets in each representative explanatory point using an on-demand generation and complementary pairing strategy. The on-demand generation and complementary pairing strategy specifically sets a feature subset quantity threshold and a specific type, filters feature subsets belonging to a specific type, and when the number of filtered feature subsets does not reach the quantity threshold, it extracts supplementary feature subsets from the remaining feature subsets and generates complementary subsets of the supplementary feature subsets. The hierarchical reconstruction module is used to perform hierarchical reconstruction and assign weights to each selected feature subset according to the scale layer and the inner layer of the feature subset, and to calculate the weights of the reconstructed feature subsets. The solution module performs weighted regression based on the weights of the reconstructed feature subsets to obtain the feature contribution value of each feature in the feature subset.

[0018] A computer storage medium is also provided, which stores a computer program that can be executed by a processor, the computer program performing the GeoShapley engineering acceleration method described above for large-scale high-dimensional space interpretation tasks.

[0019] The beneficial effects of this invention are as follows: By representatively selecting representative interpretation points, selecting feature subsets on demand, and hierarchically reconstructing the kernel weights of feature subsets, this invention reduces the construction time and storage pressure of feature subsets. It transforms the long vector storage of all feature subset kernel weights into a short vector storage for each feature subset size, and calculates the size layer weight coefficient and the corresponding inner layer weight coefficient after selecting specific feature subsets. This transforms the weight storage from an exponential object enumerated by feature subsets to a linear object enumerated by feature subset size, thereby significantly reducing peak memory usage and improving stability, and improving the computational efficiency of the GeoShapley interpretive computation project.

[0020] Furthermore, this invention uses an on-demand generation and complementary pairing strategy to screen feature subsets. It uses a small-scale feature subset plus a complementary set as the basic coverage, and then supplements feature subsets on demand in a scale-layered manner, while generating their complementary sets. This ensures that the representative interpretation points of the feature subsets maintain symmetrical coverage at the scale level, avoids the construction of large-scale feature subset index objects, significantly reduces the scale of computation, and ensures the stability of interpretation. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the method flow of the GeoShapley engineering acceleration method for large-scale high-dimensional space interpretation tasks according to an embodiment of the present invention; Figure 2This is a schematic diagram of the system architecture of the GeoShapley engineering acceleration system for large-scale high-dimensional space interpretation tasks, according to an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0023] To address the problems of low computational efficiency and unstable operation of GeoShapley interpretation computation engineering in existing technologies when the number of interpretation points is large and the dimensionality of feature variables is high, this invention provides a GeoShapley engineering acceleration method for large-scale high-dimensional space interpretation tasks, such as... Figure 1 As shown, the method includes: S1. Obtain the input samples from the GeoShapley interpretation computation project, unify the coordinate system and resolution of the input samples, and perform pixel-by-pixel alignment to obtain the set of candidate interpretation points for the input samples; S2. Calculate the similarity measure for each explanatory point in the candidate explanatory point set, and perform representative screening based on the calculation results to obtain representative explanatory points; S3. Use on-demand generation and complementary pairing strategies to filter feature subsets from each representative explanatory point; S4. Reconstruct and assign weights to each selected feature subset according to the scale layer and the inner layer of the feature subset, and calculate the weights of the reconstructed feature subsets. S5. Perform weighted regression based on the weights of the reconstructed feature subsets to obtain the feature contribution value of each feature in the feature subset.

[0024] Specifically, the input sample is a geographic raster sample containing environmental and geographic location features; the interpretation point can be the center point of the raster cell, the observation station, the sampling point, or a representative point of any spatial object, and the interpretation point set can be the rasterized result of point features or area features, or a representative cell set extracted from a high-resolution raster according to rules.

[0025] Specifically, the explanatory point set is representatively screened, and representative explanatory points are retained in a spatially stratified manner, which significantly reduces the number of representative explanatory points on which kernel regression depends while maintaining spatial coverage balance; this includes: The input sample is divided into spatial units according to a preset spatial resolution; in this embodiment, the spatial resolution can be latitude and longitude grid resolution or projected coordinate grid resolution. Within each spatial cell, all interpretation points that satisfy the mask and data validity conditions are selected, and their standardized feature vector set is constructed. Calculate the similarity measure of the explanatory points within each spatial unit and the average vector of the standardized feature vector set. Select representative explanatory points based on the calculation results and integrate the representative explanatory points of all spatial units.

[0026] Specifically, the similarity measure is a distance measure; the selection of representative explanatory points based on the calculation results is as follows: the pixel with the closest Euclidean distance to the mean vector is selected as the representative explanatory point, or the representative explanatory point is determined according to the median sample or center sample criterion; wherein, the center sample criterion is as follows: when the spatial unit is a square grid, the center point of the square grid is selected as the representative explanatory point.

[0027] Specifically, in this embodiment, considering the high spatial autocorrelation of environmental features of all 0.0083° pixels within a 0.1° resolution grid, the GeoShapley value is calculated using a 0.1° resolution grid. Specifically, within each 0.1° grid, all 0.0083° pixels that meet the mask and data validity conditions are first selected and their standardized feature vector set is constructed. Then, the average vector of the features within the grid is calculated. Finally, the medoid pixel with the closest Euclidean distance to the average vector in the high-dimensional space is selected as the representative interpretation point, which significantly reduces the scale of computation.

[0028] The method in this embodiment uses representative interpretation points of each spatial unit to form a compressed set of interpretation points, which is used to reduce the number of interpretation points and reduce the computation time.

[0029] Preferably, a set of feature variable vectors can be constructed and the mean vector can be calculated within a preset coarse grid cell. Then, the cell with the smallest Euclidean distance to the mean vector can be selected as the representative explanatory point of the cell to reduce the scale of computation. The selection criteria for the representative explanatory point can also be the median vector nearest neighbor criterion, the density peak criterion, or the cluster center nearest neighbor criterion.

[0030] Preferably, the spatial resolution of the coarse grid can be a fixed-angle grid, a fixed-distance grid, or a grid adaptively divided according to administrative regions and topographical zones.

[0031] Specifically, the method filters the feature subsets from the representative explanatory points to avoid generating a full power set index object. Instead, it uses a small, mandatory feature subset plus a complementary set as the basic coverage, and then supplements the feature subsets as needed in a hierarchical manner based on the size of the feature subsets, thus avoiding the problem at its root. The construction of the level feature subset index object; specifically including: Set a threshold for the number of feature subsets. This threshold represents the upper limit of the number of feature subsets used for regression fitting for each interpreted sample. Set the specific type of feature subsets that must be included, and cover feature subsets of that specific type, including the empty set, the universal set, all feature subsets whose size does not exceed a preset size threshold, and their complementary sets; If the number of feature subsets included in the screening scope does not reach the threshold, then according to the size layer weight of the feature subsets, additional shares are generated in each size layer. The remaining number of feature subsets are randomly selected from the feature subsets whose size exceeds the preset size threshold. In addition, for each additional feature subset, its complementary set is generated and added to the regression sample set in pairs to ensure that the regression samples maintain symmetrical coverage at the size level, thereby ensuring a low estimation variance caused by random sampling and improving interpretability. Thus, the screened feature subsets are obtained.

[0032] Specifically, the two-level weighting form of the feature subset size layer weight and the feature subset inner layer weight can be expressed as:

[0033]

[0034]

[0035] In the formula, Representative feature subset The kernel weights, The total number of representative features, Represents the size of the feature subset; Representative from Select from the features The number of combinations of features. When the size of the feature subset is 0, and when the size of the feature subset is equal to the total number of feature variables, the denominator of the kernel weight expression is 0. In implementation, separate kernel weight values ​​should be used for the empty set and the whole set to avoid division by zero. Representative size The weighting coefficients assigned to the combination layer (i.e., the scaling layer). This represents a fixed feature subset of size. Under the condition of specific feature subset The weighting coefficients within the layer.

[0036] The equivalent decomposition formed using this two-level weighting form shows that, since the product of the hierarchical weights and the intra-layer weights is equivalent to the original kernel weights at the feature subset level... To maintain consistency, only the computational path was changed without altering the definition of the explanatory variables. Building upon this, we restructured the weight objects, which previously required explicit storage for each subset, into a mechanism of "first hierarchical arrangement based on feature subset size, then generation as needed." During the regression solution process, only the generated feature subsets need to be used to construct the design matrix. The corresponding rows are transformed from explicit long vector representations of kernel function weights into scalar weights computed on the spot by subset size, thereby significantly reducing peak memory usage and improving runtime efficiency.

[0037] Preferably, the mechanism of first stratifying by feature subset size and then generating on demand can be replaced by a block generation method. The block division is based on the feature subset size range, the batch number of representative explanatory points, or the background sample block number.

[0038] Specifically, the solution for weighted regression using weighted least squares can be expressed as:

[0039] In the formula, Representative sample The regression solution vector and the corresponding set of contribution coefficients to be estimated. This represents a binary design matrix composed of many feature subset masks. The kernel weight matrix representing the feature subset. The vector of model predictions for sample i under the binary mask Z input defined by each feature subset. Each component is constructed from the input defined by the feature subset mask. Unselected features are replaced by background samples, and the background prediction results are aggregated to form the corresponding component. (The symbol is indicated by superscript.) This represents the transpose operation.

[0040] In this context, each row of Z corresponds to a binary indicator vector of a feature subset S, indicating whether each feature is included in that subset. Specifically, the row contains 0 / 1: 1 indicates that the feature is "preserved" in the subset, and 0 indicates that it is "masked and replaced with the background".

[0041] Among them, the model predicted value vector Specifically, the model prediction function predicts each feature subset, and the predicted values ​​are arranged in the order of the feature subsets. Specifically, for a given explanatory sample, we construct many input versions that "retain only some features, replacing the rest with background sample values." For each feature subset, a model prediction value is obtained. Arranging the predicted values ​​of all subsets into a column vector in a fixed order yields the model prediction. That is, the set of prediction results given by the model under many mask inputs.

[0042] Specifically, this weighted regression solution transforms the intermediate object of the regression solution from a large matrix object that grows linearly with the number of feature subsets to one whose size is mainly determined by the number of features. The decision is made to use low-dimensional cumulative objects, thereby further reducing peak memory and improving solution speed.

[0043] Preferably, the weighted least squares solver can be a matrix decomposition solver, an iterative solver, or a solver with regularization terms to improve numerical stability.

[0044] Preferably, the weighted regression solution for each sample can be processed in parallel. Multi-process parallel isomorphic operations are performed on batches of explanatory samples to achieve scalable explanatory computation throughput under national scale and high-dimensional feature variables. The parallel partitioning granularity can be partitioned by explanatory point, partitioned by feature subset generation task, or partitioned by input sample block. At the same time, fixed feature subset generation seeds and cache files can be set to reproduce experimental results under the same input and the same generation sequence. The output results are written to disk in batches or in a streaming manner to avoid aggregating all explanatory results in memory.

[0045] Preferably, the high-dimensional feature variables in this embodiment can be continuous raster factors, discrete raster factors after one-hot encoding, and composite factor vectors obtained by fusing multi-source data.

[0046] Preferably, the geographical location features in this embodiment can be composed of plane coordinates after unified projection, or they can be composed of standardized longitude and latitude.

[0047] Preferably, the background sample set in this embodiment can be the full set of valid pixels, or it can be a subset extracted from the valid pixels in a fixed number.

[0048] Specifically, the set of contribution coefficients output in this embodiment includes the baseline contribution, geographical location contribution, intrinsic contribution of environmental features, and spatial interaction effect of environmental effects; among which, geographical location contribution participates in contribution decomposition as an independent feature and is stored separately from environmental feature contribution in the output to support the distinction between spatial effects and inherent environmental effects.

[0049] Furthermore, the interaction contribution terms between location features and various environmental features can be output.

[0050] Specifically, the storage format of the contribution output can be stored line by line according to interpretation points, aggregated by raster window, or aggregated by spatial unit.

[0051] During the computation process, the model prediction adopts a dual block and filtering strategy of background samples and feature subsets, strictly limiting the number of input rows in a single prediction, avoiding the construction of a giant direct product matrix with a scale of "number of subsets × number of background samples", and ensuring that peak memory is controllable.

[0052] Furthermore, preferably, the generation of the feature subset depends only on the feature set size and sampling strategy, random seed, and sample size budget, and does not depend on the feature values ​​of specific samples. Therefore, the number of features, the size of the feature subset, and the weights are fixed under the above conditions, and the weighted least squares solution can be pre-cached. and Reuse its numerical decomposition results; only updates are needed for each explained sample. This allows for the solution to be completed quickly, significantly reducing redundant computations. Finally, by performing isomorphic operations on batches of interpreted samples in parallel through multiple processes, repeatable and scalable GeoShapley interpretation computations were achieved under national scale and high-dimensional feature variables, providing stable input for subsequent spatial statistical summarization of location and interaction contributions.

[0053] Specifically, a comparative test experiment was conducted using the method of this embodiment, and the specific test environment is shown in Table 1: Table 1 Experimental Environment Configuration

[0054] The test results are shown below: Table 2 Comparison of Feature Subset Kernel Weight Matrix Construction

[0055] Table 3 Explanation and Comparison of Total Project Operation Time

[0056] Table 4 Accuracy Verification Comparison

[0057] in, Representing the The intrinsic contribution of each environmental factor after removing spatial background effects The marginal contribution of geographical location features as joint features. Representing geographical location characteristics and the first The interaction contributions between environmental features and Together, they depicted the path by which spatial processes influence prediction. Represents the quantity of environmental characteristics.

[0058] As shown in the table, the contribution calculation results of the improved version are used as "predicted values," and the contribution calculation results of the baseline version are used as "true values" for accuracy verification. The error stems from differences in floating-point rounding and accumulation order: when the two implementations perform weighted averaging and linear equation solving on the same batch of input samples, the order of operations changes due to block / parallel processing (and the potential use of float32 to reduce memory usage), resulting in errors of approximately [number missing]. The acceptable numerical difference. Specifically, in this embodiment, the range of the dependent variable y being explained is the logistic probability value [0, 1], so an error of this order of magnitude is acceptable.

[0059] Furthermore, this embodiment of the invention also provides a GeoShapley engineering acceleration system for large-scale high-dimensional space interpretation tasks, used to implement the GeoShapley engineering acceleration method for large-scale high-dimensional space interpretation tasks described in this embodiment, such as... Figure 2 As shown, the system includes: The input sample processing module is used to acquire input samples from the GeoShapley interpretation computation project, unify the coordinate system and resolution of the input samples, and perform pixel-by-pixel alignment to obtain a set of candidate interpretation points for the input samples; the input samples are specifically geographic raster samples containing environmental features and geographic location features. The representative explanatory point filtering module is used to calculate the similarity measure of each explanatory point in the candidate explanatory point set, and to perform representative filtering based on the calculation results to obtain representative explanatory points; The feature subset filtering module is used to filter feature subsets in each representative explanatory point using an on-demand generation and complementary pairing strategy. The on-demand generation and complementary pairing strategy specifically sets a feature subset quantity threshold and a specific type, filters feature subsets belonging to a specific type, and when the number of filtered feature subsets does not reach the quantity threshold, it extracts supplementary feature subsets from the remaining feature subsets and generates complementary subsets of the supplementary feature subsets. The hierarchical reconstruction module is used to perform hierarchical reconstruction and assign weights to each selected feature subset according to the scale layer and the inner layer of the feature subset, and to calculate the weights of the reconstructed feature subsets. The solution module performs weighted regression based on the weights of the reconstructed feature subsets to obtain the feature contribution value of each feature in the feature subset.

[0060] The various modules or mechanisms of the system are mainly used to implement the various steps of the above method embodiments, and will not be described in detail here.

[0061] In addition, this embodiment also provides a computer storage medium storing a computer program that can be executed by a processor. This computer program executes the GeoShapley engineering acceleration method described above for large-scale high-dimensional space interpretation tasks.

[0062] The engineering acceleration method provided in this invention focuses on kernel-weighted regression as the solution framework, but systematically reconstructs the feature subset generation strategy, kernel weight object organization, design matrix construction, and parallel execution methods. This achieves low memory consumption and high throughput computation in scenarios with high-dimensional feature variables and a large number of explanatory points. Without changing the theoretical form of GeoShapley kernel weights and the definition of explanatory quantities, this method reduces the storage size of kernel weight objects and feature subset index objects, avoiding the exponential memory pressure caused by feature subset enumeration. Without compromising explanatory repeatability and consistency, it reconstructs the feature subset generation path, making it independent of full power set enumeration and maintaining a weight allocation mechanism consistent with kernel weights. It improves throughput efficiency under large-scale batch processing conditions, enabling stable operation under national-scale spatial raster and high-dimensional feature variable conditions, and supports parallel expansion.

[0063] Specifically, according to the method described in this embodiment and the experimental comparison results shown in the figure, the technical effects of the present invention include: by hierarchically reconstructing the feature subset size of the kernel weight object, the weight storage is transformed from an exponential object enumerated by feature subsets to a linear object enumerated by feature subset size, thereby significantly reducing peak memory usage and improving stability; by using a feature subset on-demand generation and complementary pairing strategy, the generation of a full power set index object is avoided, reducing the feature subset index construction time and storage pressure, and improving the symmetric coverage and repeatability of the interpretation results; by designing an incremental construction method for matrices and weights, the regression solution process is transformed from explicit large matrix construction to low-dimensional cumulative solution, thereby significantly reducing memory usage and improving solution speed; by using candidate interpretation point space compression and parallel batch processing scheduling, the method can support large-scale GeoShapley interpretation calculations under national-scale spatial raster and high-dimensional feature variable conditions, improving engineering usability and applicability.

[0064] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0065] The order of the steps in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0066] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A GeoShapley engineering acceleration method for large-scale high-dimensional space interpretation tasks, characterized in that, The method includes: The input samples of the GeoShapley interpretation computation project are obtained, the coordinate system and resolution of the input samples are unified, and pixel-by-pixel alignment is performed to obtain a set of candidate interpretation points for the input samples; the input samples are specifically geographic raster samples containing environmental features and geographic location features. A similarity metric is calculated for each explanatory point in the candidate explanatory point set, and representative explanatory points are obtained by representative screening based on the calculation results; The feature subsets in each representative explanatory point are screened using an on-demand generation and complementary pairing strategy. The on-demand generation and complementary pairing strategy specifically sets a threshold for the number of feature subsets and a specific type, and screens feature subsets belonging to a specific type. When the number of screened feature subsets does not reach the threshold, supplementary feature subsets are extracted from the remaining feature subsets, and complementary subsets of the supplementary feature subsets are generated. Each selected feature subset is reconstructed and weighted according to the scale layer and the inner layer of the feature subset, and the weight of the reconstructed feature subset is calculated. We perform weighted regression based on the weights of the reconstructed feature subsets to obtain the feature contribution value of each feature in the feature subset.

2. The GeoShapley engineering acceleration method for large-scale high-dimensional space interpretation tasks according to claim 1, characterized in that, Representative explanatory points are obtained in the following ways: The input samples are divided into spatial units according to a preset spatial resolution. Within each spatial cell, all candidate interpretation points that satisfy the mask and data validity conditions are selected, and their standardized feature vector set is constructed. Calculate the similarity metric of candidate explanatory points within each spatial unit and the average vector of the standardized feature vector set. Select representative explanatory points based on the calculation results and integrate the representative explanatory points of all spatial units.

3. The GeoShapley engineering acceleration method for large-scale high-dimensional space interpretation tasks according to claim 2, characterized in that, The similarity measure is specifically a distance measure; the selection of representative explanatory points based on the calculation results is specifically as follows: the pixel with the closest Euclidean distance from the mean vector is selected as the representative explanatory point, or the representative explanatory point is determined according to the median sample or center sample criterion; among them, the center sample criterion is specifically as follows: when the spatial unit is a square grid, the center point of the square grid is selected as the representative explanatory point.

4. The GeoShapley engineering acceleration method for large-scale high-dimensional space interpretation tasks according to claim 2, characterized in that, The preset spatial resolution is specifically the latitude and longitude grid resolution or the projected coordinate grid resolution.

5. The GeoShapley engineering acceleration method for large-scale high-dimensional space interpretation tasks according to claim 1, characterized in that, Specific types of feature subsets include the empty set, the universal set, all feature subsets whose size does not exceed a preset size threshold, and their complementary sets; the selection of feature subsets from representative explanatory points specifically involves: Set a threshold for the number of feature subsets and a specific type to cover feature subsets of a specific type; Determine whether the number of covered feature subsets has reached the threshold. If not, randomly select the remaining feature subsets from the feature subsets whose size exceeds the preset threshold, and generate their complementary sets to obtain the filtered feature subsets.

6. The GeoShapley engineering acceleration method for large-scale high-dimensional space interpretation tasks according to claim 1, characterized in that, The weighted regression solution is calculated based on the binary matrix of the feature subset, the weight matrix of the feature subset, and the model prediction vector of the feature subset.

7. The GeoShapley engineering acceleration method for large-scale high-dimensional space interpretation tasks according to claim 6, characterized in that, The model prediction vector of the feature subset is obtained by predicting each feature subset using a prediction function and arranging the predicted values ​​in the order of the feature subsets; the weight matrix of the feature subset is a kernel weight diagonal matrix, and its diagonal elements are the scalar weights corresponding to the weights of each reconstructed feature subset.

8. The GeoShapley engineering acceleration method for large-scale high-dimensional space interpretation tasks according to claim 1, characterized in that, The reconstructed feature subset weights are specifically calculated from the scale layer weights and the inner weights of the feature subsets; the scale layer weights are calculated from the total number of features and the feature subset size; the inner weights of the feature subsets are calculated from the number of combinations of feature subset sizes selected from the total number of features.

9. A GeoShapley engineering acceleration system for large-scale high-dimensional space interpretation tasks, characterized in that: The system includes: The input sample processing module is used to acquire input samples from the GeoShapley interpretation computation project, unify the coordinate system and resolution of the input samples, and perform pixel-by-pixel alignment to obtain a set of candidate interpretation points for the input samples; the input samples are specifically geographic raster samples containing environmental features and geographic location features. The representative explanatory point filtering module is used to calculate the similarity measure of each explanatory point in the candidate explanatory point set, and to perform representative filtering based on the calculation results to obtain representative explanatory points; The feature subset filtering module is used to filter feature subsets in each representative explanatory point using an on-demand generation and complementary pairing strategy. The on-demand generation and complementary pairing strategy specifically sets a feature subset quantity threshold and a specific type, filters feature subsets belonging to a specific type, and when the number of filtered feature subsets does not reach the quantity threshold, it extracts supplementary feature subsets from the remaining feature subsets and generates complementary subsets of the supplementary feature subsets. The hierarchical reconstruction module is used to perform hierarchical reconstruction and assign weights to each selected feature subset according to the scale layer and the inner layer of the feature subset, and to calculate the weights of the reconstructed feature subsets. The solution module performs weighted regression based on the weights of the reconstructed feature subsets to obtain the feature contribution value of each feature in the feature subset.

10. A computer storage medium, characterized in that, It contains a computer program that can be executed by a processor, which performs the GeoShapley engineering acceleration method for large-scale high-dimensional space interpretation tasks as described in any one of claims 1-8.

Images