A lightweight identification method for multi-source heterogeneous point review data and geographic grid alignment

Abstract

The application provides a lightweight identification method for multi-source heterogeneous point review data and geographic grid alignment, relates to the cross field of data geographic information, and comprises the following steps: performing feature extraction alignment mapping processing on multi-modal point review data to obtain a multi-modal general semantic feature vector; calculating semantic shares radiated by the multi-modal general semantic feature vector to each geographic grid, summing all the semantic shares of the same geographic grid to obtain grid semantic vectors of multiple geographic grids; inputting a demand tensor and a supply tensor to a bilinear attention model to obtain a demand and supply gap score of the geographic grid; and extracting a core business label to generate an interpretable quadruple. By using the above method, privacy can be ensured, bandwidth and calculation cost can be reduced, consumption demand distribution distortion caused by displacement can be reduced, edge side real-time rapid inference can be realized, and safe, real-time and transparent interpretable decisions can be provided.

Description

Technical Field

[0001] This invention relates to the field of data geographic information intersection, specifically a lightweight identification method for aligning multi-source heterogeneous review data with geographic grids. Background Technology

[0002] In the digital analysis of local life services, existing technologies mainly rely on structured data such as points of interest (POI) and customer flow to construct geographic grid features, making it difficult to utilize fine-grained semantic information such as "the taste is too salty" and "the queue is long" in user reviews. At the same time, the introduced general natural language processing models have insufficient understanding of the local business context and often misjudge specific expressions such as "extremely spicy" as negative, resulting in a distortion of the emotional polarity of demand.

[0003] In addition, the total volume of multimodal data accompanying the review is large, while the storage of field devices such as edge smart boxes is usually less than 8GB, making it difficult to deploy conventional large models; if the data is uploaded to the cloud for processing, it will face high bandwidth costs and data privacy compliance risks under regulations such as GDPR and the Cybersecurity Law.

[0004] Furthermore, existing methods often employ a direct "sentence-grid" assignment approach, binding the entire review to a single grid where the initial coordinates of the review were located. This ignores the actual range of consumer activity, resulting in a large number of semantically false grids and severely distorting the distribution of real demand.

[0005] Therefore, there is an urgent need for a lightweight identification method that aligns multi-source heterogeneous review data with geographic grids. Summary of the Invention

[0006] To overcome the existing technical problems, this invention provides a lightweight identification method for aligning multi-source heterogeneous review data with geographic grids.

[0007] The present invention adopts the following technical solution.

[0008] A lightweight identification method for aligning multi-source heterogeneous review data with geographic grids includes the following steps: S1. Receive multimodal review data containing the initial coordinates of the review, and perform feature extraction, alignment and mapping processing on the multimodal review data through a lightweight coding model to obtain a multimodal general semantic feature vector; S2. Based on the pre-trained spatial transition probability matrix, the initial coordinates of the evaluation, and multiple pre-divided geographic grids, calculate the semantic share of the multimodal general semantic feature vector radiating to each geographic grid. Summate all semantic shares of the same geographic grid to obtain the grid semantic vector of multiple geographic grids. S3. Obtain objective physical feature data of the geographic grid, construct a demand tensor by combining the grid semantic vector, extract commercial distribution feature data within the geographic grid to construct a supply tensor, input the demand tensor and supply tensor into the bilinear attention model to obtain the demand-supply gap score of the geographic grid. S4. Extract the links with high contribution weights from the output of the bilinear attention model, use an interpretable convolution algorithm to extract core business labels, and generate interpretable quadruples based on the core business labels, supply tensor, and demand-supply gap scores.

[0009] As a further improvement of the present invention, the specific steps for obtaining a multimodal general semantic feature vector by performing feature extraction, alignment, and mapping processing on multimodal review data using a lightweight coding model include: S11. Input the review text from the multimodal review data into the LocalBERT text encoding model to extract the text feature vector; The review images in the multimodal review data are scaled to a preset pixel size and then input into the MobileViT image model to extract visual theme vectors; The commentary voice data in the multimodal commentary data is sampled at a preset frequency and the Mel-frequency cepstral coefficient features are extracted. The samples are then input into the TinyConformer speech model for processing to extract the emotion feature vector. S12. Input the text feature vector, visual topic vector, and sentiment feature vector into the cross-modal alignment layer, perform contrastive learning using a contrastive loss function with a preset temperature coefficient, and perform quantization encoding using unified projection to obtain a multimodal general semantic feature vector.

[0010] As a further improvement of the present invention, the specific steps of S2 include: selecting the transition probability from the geographic grid corresponding to the initial coordinates of the evaluation to other geographic grids based on the spatial transition probability matrix constructed by training the walking network and the cycling network; multiplying the multimodal general semantic feature vector by the corresponding transition probability to obtain the semantic share radiating to the corresponding geographic grid; and summing all semantic shares of the same geographic grid to obtain the grid semantic vector of multiple geographic grids.

[0011] As a further improvement of the present invention, the specific steps for training and constructing the spatial transition probability matrix include: constructing a spatial dual-graph network model of a 15-minute walking network and a cycling network to determine the maximum physical boundary of the transition probability of each geographic grid. Extract nighttime light remote sensing brightness data that characterizes the commercial attractiveness of each geographic grid, combine it with historical desensitized trajectory big data, statistically analyze the personnel movement flow of each geographic grid, perform joint probability calculation and normalization processing on personnel movement flow and nighttime light remote sensing brightness data, and generate a spatial transition probability matrix stored in sparse matrix form.

[0012] As a further improvement to the present invention, the specific steps for constructing the demand tensor and the supply tensor include: S31. Obtain objective physical feature data of the geographic grid, including passenger flow data, nighttime light remote sensing brightness data and average house price data. Concatenate the grid semantic vector, passenger flow data, nighttime light remote sensing brightness data and average house price data to obtain the demand tensor. S32. Extract commercial distribution feature data within the geographic grid. The commercial distribution feature data includes the density of points of interest (POIs) of the same category, the average order value, and the historical rating variance. Concatenate the density of POIs of the same category, the average order value, and the historical rating variance to obtain the supply tensor.

[0013] As a further improvement to the present invention, the specific steps for inputting the demand tensor and supply tensor into a bilinear attention model to obtain the demand-supply gap score of the geographic grid include: configuring a learnable cross-mapping weight matrix, and using the matrix multiplication formula... Perform bilinear cross-reaction calculations and input the hyperbolic tangent function. The demand-supply gap score is obtained. in It is the transpose of the demand tensor D. It is the cross-mapping weight matrix. It is the supply tensor.

[0014] As a further improvement of the present invention, the specific steps for extracting core business tags include: extracting the link with the highest weight contribution ratio in the cross-mapping weight matrix, performing reverse backpropagation analysis along the link to the input layer through an interpretable convolution algorithm, projecting the corresponding high contribution vector features inversely to the locally stored high-frequency business keyword library for matching, and filtering and extracting the matched keywords as core business tags.

[0015] As a further improvement to the present invention, the specific steps for generating interpretable quadruples based on core business tags, supply tensors, and demand-supply gap scores include: The core business tags are filled into the preset text semantic template to generate demand gap text. The density of POIs of the same category in the supply tensor is used as the supply density feature. The demand-supply gap score is used as the opportunity score. The historical score variance in the supply tensor is combined with the score probability distribution to calculate the confidence interval. The demand gap text, supply density features, opportunity scores, and confidence intervals are concatenated and packaged to generate an interpretable quadruple.

[0016] As a further improvement of the present invention, S5 is included after S4: calling a local string concatenation script to embed the feature names and values ​​in the interpretable quadruples into a preset radar chart code text specification template, generating a plain text visualization code string, and pushing it to the business terminal for offline front-end rendering.

[0017] The beneficial effects of this invention are as follows: 1. This invention employs a lightweight coding model to extract features and perform cross-modal alignment mapping on multimodal comment data, including initial comment coordinates, outputting a compact multimodal general semantic feature vector. This model has a simplified structure and low computational overhead, allowing it to run locally on edge smart boxes with less than 8GB of storage. It eliminates the need to upload the entire original text, image, and voice packet to the cloud, fundamentally reducing bandwidth costs and data privacy compliance risks. Therefore, it enables rapid parsing and semantic feature extraction of multimodal data packets, significantly compressing data volume and meeting the real-time processing requirements of the edge.

[0018] 2. The multimodal general semantic feature vector is distributed radially to multiple pre-divided geographic grids, and the semantic share of each geographic grid is summed to generate a grid semantic vector. This mechanism simulates the probability of consumer movement within a real-world activity range, such as a 15-minute living circle, breaking through the limitations of traditional direct "sentence-grid" binding and effectively compensating for the displacement bias of "a customer making a purchase in geographic grid A but leaving a review in geographic grid B." This reduces the degree of demand distribution distortion caused by false grid placement.

[0019] 3. The bilinear attention model significantly reduces the number of parameters through features such as low-rank decomposition, allowing the scoring model parameters for demand-supply gaps to be controlled below 35M, and the computational complexity of forward inference for a single grid is extremely low. Therefore, sub-millisecond-level rapid inference of demand-supply gaps for single geographic grids can be achieved at the edge, supporting real-time analysis of large-scale urban grids.

[0020] 4. The interpretable convolutional algorithm extracts core business tags and integrates the supply tensor and demand gap scores to generate interpretable quadruples. These quadruples intuitively reflect key decision-making information, and the entire generation process is completed within edge devices without transmitting any intermediate data back to the cloud. Therefore, merchants can directly obtain interpretable business analysis results through terminals such as the DianMaster App, balancing data security, real-time response, and decision transparency. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the present invention; Figure 2 This is a flowchart illustrating the process of calculating the multimodal general semantic feature vector of the present invention; Figure 3 This is an architectural diagram of the present invention. Detailed Implementation

[0023] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. To better illustrate this embodiment, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product.

[0024] It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings. The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] Reference Figure 1 and Figure 2 As can be seen, a lightweight identification method for aligning multi-source heterogeneous review data with geographic grids includes the following steps: S1. Receive multimodal review data containing the initial coordinates of the review, and perform feature extraction, alignment and mapping processing on the multimodal review data through a lightweight coding model to obtain a multimodal general semantic feature vector; S2. Based on the pre-trained spatial transition probability matrix, the initial coordinates of the evaluation, and multiple pre-divided geographic grids, calculate the semantic share of the multimodal general semantic feature vector radiating to each geographic grid. Summate all semantic shares of the same geographic grid to obtain the grid semantic vector of multiple geographic grids. S3. Obtain objective physical feature data of the geographic grid, construct a demand tensor by combining the grid semantic vector, extract commercial distribution feature data within the geographic grid to construct a supply tensor, input the demand tensor and supply tensor into the bilinear attention model to obtain the demand-supply gap score of the geographic grid. S4. Extract the links with high contribution weights from the output of the bilinear attention model, use an interpretable convolution algorithm to extract core business labels, and generate interpretable quadruples based on the core business labels, supply tensor, and demand-supply gap scores.

[0026] 1. This invention employs a lightweight coding model to extract features and perform cross-modal alignment mapping on multimodal comment data, including initial comment coordinates, outputting a compact multimodal general semantic feature vector. This model has a simplified structure and low computational overhead, allowing it to run locally on edge smart boxes with less than 8GB of storage. It eliminates the need to upload the entire original text, image, and voice packet to the cloud, fundamentally reducing bandwidth costs and data privacy compliance risks. Therefore, it enables rapid parsing and semantic feature extraction of multimodal data packets, significantly compressing data volume and meeting the real-time processing requirements of the edge.

[0027] 2. The multimodal general semantic feature vector is distributed radially to multiple pre-divided geographic grids, and the semantic share of each geographic grid is summed to generate a grid semantic vector. This mechanism simulates the probability of consumer movement within a real-world activity range, such as a 15-minute living circle, breaking through the limitations of traditional direct "sentence-grid" binding and effectively compensating for the displacement bias of "a customer making a purchase in geographic grid A but leaving a review in geographic grid B." This reduces the degree of demand distribution distortion caused by false grid placement.

[0028] 3. The bilinear attention model significantly reduces the number of parameters through features such as low-rank decomposition, allowing the scoring model parameters for demand-supply gaps to be controlled below 35M, and the computational complexity of forward inference for a single grid is extremely low. Therefore, sub-millisecond-level rapid inference of demand-supply gaps for single geographic grids can be achieved at the edge, supporting real-time analysis of large-scale urban grids.

[0029] 4. The interpretable convolutional algorithm extracts core business tags and integrates the supply tensor and demand gap scores to generate interpretable quadruples. These quadruples intuitively reflect key decision-making information, and the entire generation process is completed within edge devices without transmitting any intermediate data back to the cloud. Therefore, merchants can directly obtain interpretable business analysis results through terminals such as the DianMaster App, balancing data security, real-time response, and decision transparency.

[0030] Reference Figure 2 As can be seen, the specific steps for obtaining a multimodal general semantic feature vector by performing feature extraction, alignment, and mapping on multimodal review data using a lightweight coding model include: S11. Input the review text from the multimodal review data into the LocalBERT text encoding model to extract the text feature vector; The review images in the multimodal review data are scaled to a preset pixel size and then input into the MobileViT image model to extract visual theme vectors; The commentary voice data in the multimodal commentary data is sampled at a preset frequency and the Mel-frequency cepstral coefficient features are extracted. The samples are then input into the TinyConformer speech model for processing to extract the emotion feature vector. S12. Input the text feature vector, visual topic vector, and sentiment feature vector into the cross-modal alignment layer, perform contrastive learning using a contrastive loss function with a preset temperature coefficient, and perform quantization encoding using unified projection to obtain a multimodal general semantic feature vector.

[0031] Specifically, the LocalBERT text encoding model consists of three layers and is trained using an efficient fine-tuning technique for Adapter parameters. The vocabulary retains only 18k high-frequency local business words, and the LocalBERT text encoding model is kept at 12M, outputting a 128-dimensional text feature vector.

[0032] The MobileViT image model is structured as MobileViT-Edge 0.5×, used to output 128-dimensional visual subject vectors for "dishes / environment / prices". The MobileViT image model is kept at 14M.

[0033] The emotion feature vector is sampled at 16 kHz and the Mel-Cepstral Coefficient features are extracted. The TinyConformer speech model is maintained at 4 M and can recognize dialect accents. The emotion feature vector is specifically 64-dimensional.

[0034] The cross-modal alignment layer shares a 256-dimensional quantized semantic space, which is mapped to a unified 256-dimensional quantized semantic space after contrastive learning. The contrastive learning temperature is 4. In actual tests, the forward time of the multimodal general semantic feature vector of RTX3050 hardware at the edge is less than 80ms.

[0035] It should be noted that the source of multimodal review data can be the multimedia packets of "group purchase coupon redemption + voice evaluation" uploaded by users collected by the DianMaster App, which are received by the edge side via HTTPS and written to a temporary NVMe partition.

[0036] As a further improvement of the present invention, the specific steps of S2 include: selecting the transition probability from the geographic grid corresponding to the initial coordinates of the evaluation to other geographic grids based on the spatial transition probability matrix constructed by training the walking network and the cycling network; multiplying the multimodal general semantic feature vector by the corresponding transition probability to obtain the semantic share radiating to the corresponding geographic grid; and summing all semantic shares of the same geographic grid to obtain the grid semantic vector of multiple geographic grids.

[0037] As a further improvement of the present invention, the specific steps for training and constructing the spatial transition probability matrix include: constructing a spatial dual-graph network model of a 15-minute walking network and a cycling network to determine the maximum physical boundary of the transition probability of each geographic grid. Extract nighttime light remote sensing brightness data that characterizes the commercial attractiveness of each geographic grid, combine it with historical desensitized trajectory big data, statistically analyze the personnel movement flow of each geographic grid, perform joint probability calculation and normalization processing on personnel movement flow and nighttime light remote sensing brightness data, and generate a spatial transition probability matrix stored in sparse matrix form.

[0038] The size of the spatial transition probability matrix is ​​equal to the number of urban grid cells. The number of grid cells is 1600 in this embodiment of the invention. 1600, where one of the geographic grids is 250m in size. The size is 250m, stored in a sparse matrix format, which effectively reduces the size; the actual spatial transition probability matrix is ​​less than 6MB. The generated spatial transition probability matrix has the structure P(grid_i→grid_j).

[0039] As a further improvement to the present invention, the specific steps for constructing the demand tensor and the supply tensor include: S31. Obtain objective physical feature data of the geographic grid, including passenger flow data, nighttime light remote sensing brightness data and average house price data. Concatenate the grid semantic vector, passenger flow data, nighttime light remote sensing brightness data and average house price data to obtain the demand tensor. S32. Extract commercial distribution feature data within the geographic grid. The commercial distribution feature data includes the density of points of interest (POIs) of the same category, the average order value, and the historical rating variance. Concatenate the density of POIs of the same category, the average order value, and the historical rating variance to obtain the supply tensor.

[0040] Aggregation within a 250m grid: Demand tensor D_{k×1} = [grid semantic vector, passenger flow data, nighttime light remote sensing brightness data, average house price data], supply tensor S_{k×1} = [density of points of interest (POIs) of the same category, average transaction value, historical rating variance]. The demand-supply gap score is subsequently output through a Bilinear Attention mechanism with ≤1k parameters. The overall model is less than or equal to 30M, and the inference time per geographic grid is <10 ms.

[0041] As a further improvement to the present invention, the specific steps for inputting the demand tensor and supply tensor into a bilinear attention model to obtain the demand-supply gap score of the geographic grid include: configuring a learnable cross-mapping weight matrix, and using the matrix multiplication formula... Perform bilinear cross-reaction calculations and input the hyperbolic tangent function. The demand-supply gap score is obtained. in It is the transpose of the demand tensor D. It is the cross-mapping weight matrix. It is the supply tensor.

[0042] The bilinear attention model, or Bilinear Attention, operates on the same principle, so the calculation process will not be elaborated further.

[0043] As a further improvement of the present invention, the specific steps for extracting core business tags include: extracting the link with the highest weight contribution ratio in the cross-mapping weight matrix, performing reverse backpropagation analysis along the link to the input layer through an interpretable convolution algorithm, projecting the corresponding high contribution vector features inversely to the locally stored high-frequency business keyword library for matching, and filtering and extracting the matched keywords as core business tags.

[0044] As a further improvement to the present invention, the specific steps for generating interpretable quadruples based on core business tags, supply tensors, and demand-supply gap scores include: The core business tags are filled into the preset text semantic template to generate demand gap text. The density of POIs of the same category in the supply tensor is used as the supply density feature. The demand-supply gap score is used as the opportunity score. The historical score variance in the supply tensor is combined with the score probability distribution to calculate the confidence interval. The demand gap text, supply density features, opportunity scores, and confidence intervals are concatenated and packaged to generate an interpretable quadruple.

[0045] As an example, the interpretable convolutional algorithm maps the keywords with the highest weights to "grape milk tea," "sweet," and "queue," automatically generating: Demand gap = "Insufficient supply of low-sugar grape milk tea" Supply density = 12 units / km 2 Chance score = 0.74 The confidence interval is [0.68–0.80]. Mermaid radar chart strings are generated synchronously and rendered directly by the DianMaster App, without requiring any cloud access.

[0046] As a further improvement of the present invention, S5 is included after S4: calling a local string concatenation script to embed the feature names and values ​​in the interpretable quadruples into a preset radar chart code text specification template, generating a plain text visualization code string, and pushing it to the business terminal for offline front-end rendering.

[0047] In addition, as seen in the example above, Mermaid radar chart strings can also be generated for direct rendering by the front end, without the need for the cloud.

[0048] The technical solution of this invention was tested on 32,000 text, image, and voice messages in 6 districts of a special economic zone: 1. Edge parsing of 10 MB data packets averaged 23 seconds, with a peak memory usage of 1.4 GB; 2. Single geographic grid inference took 6.8 ms, and the Top-100 chance grid hit rate was 29% higher than the traditional "sentence hard-grid" method; 3. The supervised adoption rate of the explained quadruple was 81%, which was 37% higher than the black-box vector solution.

[0049] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A lightweight identification method for aligning multi-source heterogeneous review data with a geographic grid, characterized in that, Includes the following steps: S1. Receive multimodal review data containing the initial coordinates of the review, and perform feature extraction, alignment and mapping processing on the multimodal review data through a lightweight coding model to obtain a multimodal general semantic feature vector; S2. Based on the pre-trained spatial transition probability matrix, initial coordinates, and pre-divided geographic grids, calculate the semantic share of the multimodal general semantic feature vector radiating to each geographic grid. Sum all semantic shares of the same geographic grid to obtain the grid semantic vector of multiple geographic grids. S3. Obtain the objective physical feature data of the geographic grid, construct a demand tensor by combining the grid semantic vector, extract the commercial distribution feature data within the geographic grid to construct a supply tensor, and input the demand tensor and the supply tensor into a bilinear attention model to obtain the demand-supply gap score of the geographic grid. S4. Extract the links with high contribution weights from the output of the bilinear attention model, extract the core business labels using an interpretable convolution algorithm, and generate interpretable quadruples based on the core business labels, the supply tensor, and the demand-supply gap score.

2. The lightweight identification method for aligning multi-source heterogeneous review data with geographic grids according to claim 1, characterized in that, The specific steps for obtaining a multimodal general semantic feature vector by performing feature extraction, alignment, and mapping processing on the multimodal review data using a lightweight coding model include: S11. Input the review text in the multimodal review data into the LocalBERT text encoding model to extract the text feature vector; The review images in the multimodal review data are scaled to a preset pixel size and then input into the MobileViT image model to extract visual theme vectors; The commentary speech in the multimodal commentary data is sampled at a preset frequency and the Mel-frequency cepstral coefficient features are extracted. The samples are then input into the TinyConformer speech model for processing to extract the emotional feature vector. S12. Input the text feature vector, visual topic vector and emotion feature vector into the cross-modal alignment layer, perform contrastive learning using a contrastive loss function with a preset temperature coefficient, and perform quantization encoding using unified projection to obtain the multimodal general semantic feature vector.

3. The lightweight identification method for aligning multi-source heterogeneous review data with geographic grids according to claim 1, characterized in that, The specific steps of S2 include: selecting the transition probability from the geographic grid corresponding to the initial coordinates of the evaluation to other geographic grids based on the spatial transition probability matrix constructed by training the walking network and cycling network; multiplying the multimodal general semantic feature vector by the corresponding transition probability to obtain the semantic share radiating to the corresponding geographic grid; summing all semantic shares of the same geographic grid to obtain the grid semantic vector of multiple geographic grids.

4. The lightweight identification method for aligning multi-source heterogeneous review data with geographic grids according to claim 3, characterized in that, The specific steps for training and constructing the spatial transition probability matrix include: constructing a spatial bi-graph network model of a 15-minute walking network and a cycling network to determine the maximum physical boundary of the transition probability for each geographic grid. Extract nighttime light remote sensing brightness data that characterizes the commercial attractiveness of each geographic grid, combine it with historical desensitized trajectory big data, statistically analyze the personnel movement flow of each geographic grid, and perform joint probability calculation and normalization processing on the personnel movement flow and nighttime light remote sensing brightness data to generate a spatial transition probability matrix stored in sparse matrix form.

5. The lightweight identification method for aligning multi-source heterogeneous review data with geographic grids according to claim 1, characterized in that, The specific steps for constructing the demand and supply tensors include: S31. Obtain the objective physical feature data of the geographic grid, which includes passenger flow data, nighttime light remote sensing brightness data and average house price data. Concatenate the grid semantic vector, passenger flow data, nighttime light remote sensing brightness data and average house price data to obtain the demand tensor. S32. Extract the commercial distribution feature data within the geographic grid. The commercial distribution feature data includes the density of points of interest (POIs) of the same category, the average order value, and the historical rating variance. Concatenate the density of POIs of the same category, the average order value, and the historical rating variance to obtain the supply tensor.

6. The lightweight identification method for aligning multi-source heterogeneous review data with geographic grids according to claim 1, characterized in that, The specific steps for inputting the demand tensor and the supply tensor into a bilinear attention model to obtain the demand-supply gap score for a geographic grid include: configuring a learnable cross-mapping weight matrix, and using matrix multiplication formulas... Perform bilinear cross-reaction calculations and input the hyperbolic tangent function. The demand-supply gap score is obtained. in It is the transpose of the demand tensor D. It is the cross-mapping weight matrix. It is the supply tensor.

7. A lightweight identification method for aligning multi-source heterogeneous review data with a geographic grid, as described in claim 6, is characterized in that... The specific steps for extracting core business tags include: extracting the link with the highest weight contribution ratio in the cross-mapping weight matrix, performing reverse backpropagation analysis along the link to the input layer using an interpretable convolution algorithm, projecting the corresponding high contribution vector features inversely to a locally stored high-frequency business keyword library for matching, and filtering and extracting the matched keywords as the core business tags.

8. A lightweight identification method for aligning multi-source heterogeneous review data with a geographic grid, as described in claim 7, is characterized in that... The specific steps for generating interpretable quadtuples based on the core business tags, supply tensor, and demand-supply gap score include: The core business tags are filled into a preset text semantic template to generate demand gap text. The density of points of interest (POIs) of the same category in the supply tensor is used as the supply density feature. The demand-supply gap score is used as the opportunity score. The historical score variance in the supply tensor is combined with the score probability distribution to calculate the confidence interval. The demand gap text, supply density features, chance scores, and confidence intervals are concatenated and packaged to generate the interpretable quadruple.

9. A lightweight identification method for aligning multi-source heterogeneous review data with a geographic grid according to claim 1, characterized in that, S4 is followed by S5: calling a local string concatenation script to embed the feature names and values ​​in the interpretable quadruples into a preset radar chart code text specification template, generating a plain text visualization code string, and pushing it to the business terminal for offline front-end rendering.

Images