Method and device for analyzing innovative elements of medical and invasive gathering area based on graph attention network

Abstract

The invention discloses a graph attention network-based innovation element analysis method for a medical innovation gathering area, and belongs to the technical field of urban innovation space analysis. The method comprises the following steps: dividing a research area into 250m * 250m grids and coding; the method comprises the following steps of: acquiring three types of heterogeneous data of patents, innovation carriers and space support elements, and calculating innovation main body density, carrier quantity, population density, accessibility, POI richness and dominant functions of each grid after coordinate matching and standardization to form 12-dimensional node features; marking the grids as a high number-high influence, a high number-low influence and a low number-low influence by taking the median of the innovation number and the cooperation influence as a threshold value; constructing an undirected graph according to a Queen adjacency rule, and generating a binary adjacency matrix; and building a double-layer graph attention network, processing class imbalance by using a focus loss function, and outputting a grid classification result and an element importance value through training.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of urban planning and construction, and particularly relates to a scientific and technological innovation agglomeration area innovation element identification, correlation analysis and spatial distribution research method based on a graph attention network (GAT), which is suitable for innovation ecological evaluation and planning decision support of various scientific and technological innovation agglomeration areas. BACKGROUND

[0002] In recent years, graph neural networks (GNN) have been gradually introduced into innovation geography research due to their advantages of "fusing multi-source heterogeneous data + explicitly modeling topological relationships". In 2023, the Chinese Academy of Sciences constructed an "enterprise-university-technology" heterogeneous graph in Beijing-Tianjin-Hebei, used RGCN to identify cross-border technology fusion paths, and the prediction accuracy was improved by 19% compared with traditional SVM. Fudan University used Temporal GNN to predict the enterprise derivative trend in Shenzhen High-tech Zone, and the link prediction F1 was improved by 0.21. However, in the face of the "scientific and technological innovation agglomeration area" scenario, which is multi-agent, high-dynamic and strongly policy-intervened, there are still the following research gaps: (1) Coarse spatial granularity - existing GNN research units are ≥1 km, which is difficult to capture street-level innovation heterogeneity; (2) Single relationship - only using "cooperative patents" or "road adjacency" single relationship, without considering multiple edges of "geographical adjacency + functional complementarity + knowledge cooperation"; (3) Lack of sample imbalance processing - a small number of high-activity units are often submerged, making it difficult for planning departments to accurately lock "true hotspots".

[0003] Therefore, there is an urgent need for a 250 m-level refined, nonlinear, and class-robust innovation element identification method for scientific and technological innovation agglomeration areas to fill the above gaps.

[0004] With the intensification of global technological competition and the deepening of China's innovation-driven development strategy, scientific and technological innovation agglomeration areas have become the "new engine" of regional economic high-quality development. However, the rapid expansion has exposed three major pain points: (1) "Spatial continuity loss" - traditional research often takes innovation agents as nodes, ignoring the comprehensive effects of 250 m-level micro grids, leading to "one-size-fits-all" innovation spatial policies; (2) "Linearization of element relationships" - grid assignment and OLS regression models assume that element flow is linear, which is insufficient to explain nonlinear mechanisms such as knowledge spillover and industry-university-research collaboration; (3) "Class imbalance" - the area proportion of high-activity units is often less than 10%, but they account for more than 40% of innovation output. Traditional cross-entropy loss easily leads the model to fall into the "majority class trap", resulting in low accuracy in identifying key hotspots. SUMMARY

[0005] In view of this, the purpose of the present application is to provide a scientific and technological innovation agglomeration area innovation element analysis method based on a graph attention network, to solve the problems of spatial continuity loss, insufficient description of nonlinear correlation, and low key innovation unit recognition accuracy caused by class imbalance in the prior art.

[0006] A scientific and technological innovation agglomeration area innovation element analysis method based on a graph attention network has the following technical solutions: A scientific and technological innovation agglomeration area innovation element analysis method based on a graph attention network, characterized by comprising the following steps: (1) Research area definition and grid division: lock the target scientific and technological innovation agglomeration area closed boundary with latitude and longitude coordinates, divide with 250 m x 250 m equidistant grid, remove invalid grid in non-built-up area, give unique ID to the remaining valid grid and record the latitude and longitude coordinates of the lower left corner vertex; (2) Multi-source data collection and preprocessing: collect patent data, innovation carrier data and spatial support element data, after data cleaning, coordinate matching to WGS-84 coordinate system and standardization processing, form a standard data set which can be superimposed on the corresponding grid; the patent data is preferentially derived from the State Intellectual Property Office patent database, the innovation carrier data is preferentially derived from the Qichacha platform, and the spatial support element data is preferentially derived from the mobile phone signaling, OSM road network and POI database; (3) Node feature matrix and classification label construction: a. Calculate three types of 12 indicators including enterprise subject density, university subject density, enterprise subject density, strategy source carrier number, R&D carrier number, space carrier number, residential population density, working population density, spatial accessibility, POI richness and dominant function one-hot encoding for each grid, to form a 12-dimensional node feature vector; b. Statistic the innovation number and cooperation influence of all grids, and divide the grids into three categories of double-high type, single-high type and double-low type and give labels according to the median of the two as dynamic threshold; (4) Graph network construction: determine the adjacency relationship between grids based on the Queen adjacency rule and generate a binary adjacency matrix, integrate the node feature matrix and the adjacency matrix to form an integrated undirected graph of "element-relation-space"; (5) GAT model training: build a graph attention network containing two layers of graph convolution layer and one layer of fully connected classification layer, use focal loss function and Adam optimizer, divide the training set, validation set and test set by stratified sampling in the proportion of 3:1:1, complete the model training; the focal loss function parameters are α=[5.0, 4.0, 0.5], γ=2, the initial learning rate of Adam is 0.02, and the dynamic learning rate adjustment and early stopping mechanism are set; (6) Result output and analysis: output the grid cell innovation active type classification results, global and type-specific innovation factor importance values, and visualize the spatial distribution characteristics of innovation factors through GIS for innovation hotspot identification and spatial flow effect analysis.

[0007] Specifically, the standardization processing of step (2) includes: for continuous indicators, using Min-Max standardization to map to [0, 1], and for classification indicators, using one-hot encoding.

[0008] Specifically, the Queen adjacency rule of step (4) means that two grids are adjacent in the horizontal, vertical or diagonal direction, that is, it is determined that there is an edge connection, which satisfies Δx=250 m and Δy=0 m, Δx=0 m and Δy=250 m, or Δx=250 m and Δy=250 m.

[0009] Specifically, the first layer of the two-layer graph convolution layer of step (5) is configured with 4 attention mechanisms, which aggregate 12-dimensional input into a 128-dimensional feature vector, and the activation function uses ELU; the second layer continues to aggregate into a 128-dimensional feature vector; the full connection classification layer maps the 128-dimensional feature to a 3-dimensional output vector and converts it into a probability distribution through Softmax.

[0010] Specifically, step (5) further comprises: outputting the accuracy and weighted average F1-score of the training set and the validation set every 10 rounds, if the F1-score of the "double high type" sample is lower than 0.5, then increase the number of attention heads or adjust the focus parameter γ of the focal loss function; if the overall accuracy is lower than 0.85, then supplement the node features or introduce distance attenuation weight to adjust the adjacency matrix.

[0011] Specifically, step (6) generates an innovation factor distribution heat map through GIS tools, and labels the high-activity grid aggregation area, innovation network extension direction and transverse innovation corridor position, to realize planning decision support.

[0012] Another aspect of the present application discloses a scientific and technological innovation agglomeration area innovation factor analysis device based on a graph attention network, and the technical scheme is as follows: A scientific and technological innovation agglomeration area innovation factor analysis device based on a graph attention network, characterized by comprising: a regional grid division module, used for defining the closed boundary of a target scientific and technological innovation agglomeration area, dividing the region with 250m×250m equidistant grids, removing invalid grids in non-built-up areas, giving a unique ID to the valid grids and recording the longitude and latitude coordinates of the lower left corner vertex; A data processing module is configured to collect patent data, innovation carrier data and spatial support element data, perform data cleaning, WGS-84 coordinate system coordinate matching and standardization processing (continuous type index Min-Max standardization mapping to [0, 1], and classification index one-hot encoding), and output a standardized data set that can be superimposed on a grid; The patent data is preferentially derived from the patent database of the State Intellectual Property Office, the innovation carrier data is preferentially derived from the Qichacha platform, and the spatial support element data is preferentially derived from mobile phone signaling, OSM road network and POI database; a feature and label construction module is configured to calculate 12 indexes of innovation subject features, innovation carrier features and spatial support element features of each grid to generate a 12-dimensional node feature matrix, and to divide the grid into double-high type, single-high type and double-low type and assign labels by taking the median of innovation quantity and cooperation influence as a dynamic threshold; A graph network construction module is configured to determine the adjacency relationship of the grid based on the Queen adjacency rule (satisfying Δx=250m and Δy=0m, Δx=0m and Δy=250m, or Δx=250m and Δy=250m) and generate a binary adjacency matrix, integrate the node feature matrix and the adjacency matrix, and form an integrated undirected graph of "element - relationship - space"; A model training module is configured to build a graph attention network containing two layers of graph convolution layers and one layer of fully connected classification layers, adopt a focal loss function (α=[5.0, 4.0, 0.5], γ=2) and an Adam optimizer (initial learning rate 0.02), divide the data set by stratified sampling in a ratio of 3:1:1, perform model training and set dynamic learning rate adjustment and early stopping mechanism; the first layer of graph convolution layer is configured with 4 attention mechanisms, which aggregate the 12-dimensional input into a 128-dimensional feature vector (the activation function is ELU), the second layer aggregates into a 128-dimensional, and the fully connected classification layer maps the 128-dimensional feature to a 3-dimensional output and converts it into a probability distribution through Softmax; A result output and analysis module is configured to output grid innovation activity type classification results, global and type-specific innovation element importance values, generate an innovation element distribution heat map through GIS tools, and label high-activity areas, innovation network extension directions and horizontal innovation corridors.

[0013] Specifically, the model training module further includes an optimization unit configured to output the accuracy and weighted average F1-score of the training set and the validation set every 10 rounds, increase the number of attention heads or adjust the γ value if the "double-high type" sample F1-score is lower than 0.5, and supplement node features or introduce distance decay weight to adjust the adjacency matrix if the overall accuracy is lower than 0.85.

[0014] Specifically, the data processing module can supplement and collect various types of data from legitimate channels such as government open data platforms, industry databases, and statistical yearbooks.

[0015] Specifically, the results output and visualization results of the analysis module are used for innovation hotspot identification, spatial flow effect analysis, and planning decision support.

[0016] The beneficial effects are: (1) A 250 m × 250 m fine grid is used as the nodes of the graph network to ensure spatial continuity and accurately aggregate multi-dimensional innovation information, effectively capturing the spatial clustering characteristics of innovation activities; (2) Based on graph attention network, the nonlinear relationship between nodes is dynamically captured, which breaks through the limitations of traditional linear model and accurately describes the knowledge flow and spatial spillover effect; (3) By using the focus loss function and the hierarchical sampling strategy, the weight of the majority class samples is significantly reduced, and the feature learning of a few highly active units is strengthened. The overall accuracy of the test set reaches 0.9025. (4) It integrates three types of heterogeneous data: patents, innovation carriers, and spatial support elements, covering the entire ecological dimension of "innovation subject-carrier-environment", and has strong scene adaptability; (5) Based on the Queen adjacency rule, spatial adjacency relationships are constructed, and the geographical flow effect is embedded into the graph network to fit the geographical attributes of the innovation space. Attached Figure Description

[0017] Figure 1. Schematic diagram of the method flow of the present invention.

[0018] Figure 2. Schematic diagram of the graph attention network model of the present invention.

[0019] Figure 3. Spatial distribution heat map of innovative elements in the embodiment. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0021] The example uses the Guangzhou Zhongda Science and Technology Innovation Transformation Cluster Area (36.8 km²) as the experimental area, with data spanning from January 2020 to December 2020. All raster operations were performed using GDAL 3.4.0 + PostGIS 3.2, and the GAT model was implemented using PyTorch 1.13 and PyTorch-Geometric 2.2. Experimental hardware included two Intel Xeon Gold 6226R processors, four NVIDIA RTXA6000 GPUs, and 256 GB of RAM. The following provides a detailed 11-step process and parameter table for the repeatable experiment. For guidance on the specific implementation process, please refer to... Figure 1The diagram shown is a schematic representation of the method flow of the present invention. Figure 2 This is a diagram of the attention network model structure of the present invention.

[0022] Step 1: Precise definition of the study area boundary Within the area bounded by Jiangnan Avenue (113.24°E) to the west, Jianghai Avenue (113.33°E) to the east, and Xinji Road (23.07°N) to the south, a 36.8 km² closed polygon was constructed using the 2020 Guangzhou 1:2000 topographic map as the base map and the Albers equal-area projection, with an error of <0.3 m.

[0023] Step 2: Mesh Generation and Unique Encoding A 250 m × 250 m grid, totaling 1,652 grids, was generated using the Fishnet tool. This grid was then intersected with the 2020 Guangzhou City Land and Sea Use Status Layer (Guangdong Natural Resources Letter

[2021] No. 782). 66 non-construction land areas, including rivers, lakes, and mountains, were removed, leaving 1,586 valid grids. These grids were coded as “G-0001~G-1586”, and the coordinates of the lower left corner were written into the MongoDB “mesh_id” field to create a spatial index (2dsphere).

[0024] Step 3: Multi-source data alignment and cleaning (see Table 1) Table 1. Overview of Multi-Source Data Cleaning Rules .

[0025] Step 4: Calculate the 12-dimensional features of the nodes (see Table 2) Table 2. Node Feature Quantization Formulas and Value Ranges .

[0026] Step 5: Dynamic Threshold Division of Labels The median number of innovations was calculated to be 8, and the median number of collaborative impacts was 12 for 1,586 grids. Labels were generated according to the following rules: 33 (2.1%) were high-innovation (both high-value and low-value), 102 (6.4%) were high-value (single high-value), and 1,451 (91.5%) were low-innovation (both low-value and low-value). To analyze the activity level of the innovation space in greater detail, we used the median of two key indicators—the number of innovations and the impact of collaboration—as thresholds to divide all grid units into three categories. The specific classification rules are as follows: High quantity and high impact (double high type) refers to grid units where the number of innovations exceeds the median number of innovations, and the collaborative impact also exceeds the median collaborative impact. These grid units excel in both innovation output and collaborative capability, and are the core areas of innovation activity.

[0027] High quantity – low impact (single-high type) refers to grid units where the number of innovations exceeds the median number of innovations, but the collaborative impact is lower than the median collaborative impact. While these grid units have high innovation output, their collaborative impact is relatively weak, potentially indicating dispersed innovation resources or imperfect collaborative mechanisms.

[0028] Low quantity and low impact (double low type) means that the number of innovations is lower than the median number of innovations, and the impact of cooperation is also lower than the median impact of cooperation. Innovation activities in these grid units are relatively inactive and may require more external support and resource investment to stimulate innovation.

[0029] Table 3 Dataset Partitioning Table .

[0030] Step 6: Constructing the multi-edge graph In addition to the Queen space adjacency (1586×1586), two heterogeneous edges are added: (1) Functional complementary edge: If the dominant functions of the two grids are different and the cosine similarity of POI is >0.6, then add a "functional complementary" edge; (2) Knowledge cooperation edge: If there are ≥1 jointly applied patents between two grids, then add a "knowledge cooperation" edge with weight = number of cooperative patents / total number of patents in grid i.

[0031] The final result is a heterogeneous graph with "1,586 nodes + 9,840 spatial edges + 3,126 functional edges + 692 knowledge edges", with an edge type dimension of 3.

[0032] Step 7: Refine the model architecture (see Table 4) Table 4 GAT Hyperparameters and Hardware Consumption .

[0033] Step 8: Class Imbalance Training Strategy Improved focus loss is adopted: L FL=−Σ c=1 3 α c (1−pc ) γ log pc , where α=[5.0,4.0,0.5], γ=2; Training set: validation set: test set = 3:1:1 (stratified sampling), 1000 epochs in total, early stop patience 200 rounds, learning rate decay coefficient = 0.5, minimum lr = 1×10⁻5.

[0034] Step 9: Accuracy Verification and Error Analysis Test set results: Overall accuracy = 0.9025, weighted F1 = 0.916; among them, F1 = 0.533 for the double-high type, F1 = 0.557 for the single-high type, and F1 = 0.950 for the double-low type.

[0035] The ablation experiment shows that if the "functional complementarity" edge is removed, the F1 score of the dual-high type decreases by 0.08; if the "knowledge cooperation" edge is removed, the Recall score of the single-high type decreases by 0.11, proving that heterogeneous edges are effective.

[0036] Step 10: GIS Visualization and Planning Recommendations In ArcGIS Pro 3.0, Natural Breaks (Jenks) is used to divide the predicted probability into 5 levels, resulting in... Figure 3 "Spatial Distribution Heat Map of Innovation Elements": The high-value area of ​​the dual high type forms a "π"-shaped corridor along Xingang West Road-Sun Yat-sen University-Guangzhou International Bio Island; Single-high type plaques mainly appear along Guangzhou Avenue South. It is recommended to add "industry-university-research joint laboratories" to enhance the influence of cooperation. The low-density grid is concentrated in the old industrial area of ​​Lijiao-Daganwei, and it is recommended to activate it through "industrial buildings moving upstairs + incubator implantation".

[0037] Step 11 Timeliness Correction and Rolling Forecast To address the 4-5 year time lag of patents, this embodiment further extrapolates the time series of "2020 patents + new carriers added from 2021 to 2023": using the 25 new incubators and 2 new unicorns added from 2021 to 2023 as dynamic features, the innovation hotspots in 2025 are re-predicted. After on-site verification in 2023, the hotspot grid hit rate is 84%, meeting the forward-looking needs of land and space planning.

Claims

1. A method for analyzing innovation elements in science and technology innovation clusters based on graph attention networks, characterized in that, Includes the following steps: (1) Study area definition and grid division: The closed boundary of the target science and technology innovation cluster area is locked by latitude and longitude coordinates. The grid is divided into 250m×250m equally spaced grids. Invalid grids in non-built areas are eliminated. The remaining valid grids are assigned a unique ID and the latitude and longitude coordinates of the lower left corner vertex are recorded. (2) Multi-source data acquisition and preprocessing: Collect patent data, innovation carrier data and spatial support element data, clean the data, match the coordinates to the WGS-84 coordinate system and standardize the data to form a standardized dataset that can be superimposed on the corresponding grid; the patent data is primarily sourced from the State Intellectual Property Office patent database, the innovation carrier data is primarily sourced from the Qichacha platform, and the spatial support element data is primarily sourced from mobile signaling, OSM road network and POI database; (3) Construction of node feature matrix and classification labels: a. For each grid computing innovation entity characteristics, innovation carrier characteristics, and spatial support element characteristics, a total of 12 indicators in three categories are used, including enterprise entity density, university entity density, public institution entity density, number of policy source carriers, number of R&D carriers, number of spatial carriers, residential population density, working population density, spatial accessibility, POI richness, and dominant function one-hot encoding, forming a 12-dimensional node feature vector; b. Statistically analyze the number of innovations and the collaborative impact of all grids, and use the median of the two as a dynamic threshold to classify the grids into three categories: high-innovation, high-cooperative, and low-innovation, and assign them labels; (4) Graph network construction: Based on the Queen adjacency rule, the adjacency relationship between grids is determined and a binary adjacency matrix is ​​generated. The node feature matrix and the adjacency matrix are integrated to form an integrated undirected graph of "elements-relationships-space". (5) GAT model training: A graph attention network consisting of two graph convolutional layers and one fully connected classification layer is constructed. The focus loss function and Adam optimizer are used. The training set, validation set and test set are divided into training set, validation set and test set in a 3:1:1 ratio to complete the model training. The parameters of the focus loss function are α=[5.0,4.0,0.5] and γ=2. The initial learning rate of Adam is 0.

02. Dynamic learning rate adjustment and early stopping mechanism are set. (6) Results output and analysis: Output the classification results of innovation activity types of grid units, the importance values ​​of global and sub-type innovation elements, and visualize the spatial distribution characteristics of innovation elements through GIS for innovation hotspot identification and spatial flow effect analysis.

2. The method according to claim 1, wherein the standardization process in step (2) comprises: For continuous indicators, Min-Max normalization is used to map them to [0,1], and one-hot encoding is used for categorical indicators.

3. The method according to claim 1, wherein the Queen adjacency rule in step (4) refers to the following: two grids are determined to have an edge connection if they are adjacent in the horizontal, vertical or diagonal direction, satisfying Δx=250 m and Δy=0 m, Δx=0 m and Δy=250 m, or Δx=250 m and Δy=250 m.

4. The method according to claim 1, wherein the first layer of the two graph convolutional layers in step (5) is configured with a 4-head attention mechanism to aggregate the 12-dimensional input into a 128-dimensional feature vector, and the activation function is ELU; the second layer continues to aggregate into 128 dimensions; the fully connected classification layer maps the 128-dimensional features to a 3-dimensional output vector and converts it into a probability distribution through Softmax.

5. The method according to claim 1 or 4, wherein step (5) further comprises: Every 10 rounds, output the accuracy and weighted average F1-score of the training and validation sets. If the F1-score of the "high-performing" samples is below 0.5, increase the number of attention heads or increase the focus parameter γ of the focus loss function. If the overall accuracy is below 0.85, supplement node features or introduce distance decay weights to adjust the adjacency matrix.

6. The method according to claim 1, wherein step (6) generates a heat map of the distribution of innovation elements using GIS tools, and marks the high-activity grid clustering areas, the direction of innovation network extension and the location of horizontal innovation corridors to achieve planning decision support.

7. A device for analyzing innovation elements in science and technology innovation clusters based on graph attention networks, characterized in that, include: The regional grid division module is used to define the closed boundary of the target science and technology innovation cluster area. It uses a 250m×250m equally spaced grid to divide the area, removes invalid grids in non-built-up areas, assigns a unique ID to the valid grids, and records the latitude and longitude coordinates of the lower left corner vertex. The data processing module is used to collect patent data, innovation carrier data and spatial support element data, perform data cleaning, WGS-84 coordinate system coordinate matching and standardization processing, and output a standardized dataset that can be superimposed on the grid. The patent data is primarily sourced from the State Intellectual Property Office's patent database, the innovation carrier data is primarily sourced from the Qichacha platform, and the spatial support element data is primarily sourced from mobile signaling, OSM road network, and POI databases. The feature and tag construction module is used to calculate 12 indicators for each grid, including innovation subject characteristics, innovation carrier characteristics, and spatial support element characteristics, to generate a 12-dimensional node feature matrix. The median of the number of innovations and the cooperative influence are used as dynamic thresholds to divide the grids into high-innovation, high-cooperation, and low-innovation types and assign them tags. The graph network construction module is used to determine grid adjacency relationships based on the Queen adjacency rule and generate a binary adjacency matrix. It integrates the node feature matrix and the adjacency matrix to form an integrated undirected graph of "elements-relationships-space". The model training module is used to build a graph attention network containing two graph convolutional layers and one fully connected classification layer. It adopts the focus loss function and Adam optimizer, divides the dataset into layers in a 3:1:1 ratio, performs model training, and sets up dynamic learning rate adjustment and early stopping mechanisms. The first graph convolutional layer is configured with a 4-head attention mechanism to aggregate the 12-dimensional input into a 128-dimensional feature vector. The second layer aggregates the features into 128 dimensions. The fully connected classification layer maps the 128-dimensional features to a 3-dimensional output and converts them into a probability distribution through Softmax. The results output and analysis module is used to output the classification results of grid innovation activity types, the importance values ​​of global and categorized innovation elements, and to generate a heat map of innovation element distribution using GIS tools, marking high-activity areas, innovation network extension directions, and horizontal innovation corridors.

8. The apparatus according to claim 7, characterized in that, The model training module also includes an optimization unit, which outputs the accuracy and weighted average F1-score of the training and validation sets every 10 rounds. If the F1-score of the "high-performing" samples is lower than 0.5, the number of attention heads is increased or the γ value is increased. If the overall accuracy is lower than 0.85, node features are added or distance decay weights are introduced to adjust the adjacency matrix.

9. The apparatus according to claim 7, characterized in that, The data processing module can supplement and collect various types of data from legitimate channels such as government open data platforms, industry databases, and statistical yearbooks.

10. The apparatus according to claim 7, characterized in that, The visualization results from the output and analysis module are used for identifying innovation hotspots, analyzing spatial flow effects, and supporting planning decisions.

Images