Ecological data graph representation method based on environment covariable correlation interaction

Abstract

The invention discloses an ecological data graph representation method based on environment covariable correlation interaction, and the method comprises the following steps: constructing a completely interconnected starting source graph structure based on an ecological correlation interaction theory; performing data normality test on the node variables; calculating data similarity among the node variables; generating a basic graph structure according to an analysis result; utilizing domain knowledge and an influence transfer theory to optimize a basic graph structure, and eliminating non-effective association; and generating a final adjacent matrix based on matrix operation to obtain an advanced graph structure. In conclusion, according to the ecological data graph representation method based on the environmental covariable association interaction, the data association expression ability can be improved, multi-dimensional data processing is supported, the interpretability and applicability of the model are improved, the comprehensiveness and accuracy of graph representation evaluation are enhanced, application of the deep learning technology in environmental science is promoted, and the method is suitable for popularization and application. The environment monitoring cost is reduced, and early warning and scientific decision making of pollution risks are supported.

Description

Technical Field

[0001] The present invention relates to the technical field of pollutant pollution risk prediction, and in particular to an ecological data graph representation method and method based on environmental covariate correlation interaction. Background Art

[0002] A graph is a collection of nodes and edges. It is a graphic composed of a number of given points and lines connecting two points, also known as a network. Graphs are a common mathematical method for modeling relationships between entities. The three types of information in a graph, namely nodes, edges, and overall topological structure, can jointly describe the intrinsic relationships between components in a complex system. Graphs can intuitively and efficiently present the relationships between super-multicomponents in complex systems, helping researchers to quickly understand data and transmit information, thereby discovering new patterns and potential connections hidden in complex data, making decisions, and expanding cognition. Graph technology has been widely used in fields such as species protein function research, polycrystalline material performance optimization, and molecular chemical feature extraction.

[0003] Ecological topological network refers to the representation of the interaction relationship of multi-level factors (environmental covariates) of complex ecosystems into a network diagram, which comprehensively presents the correlation relationship between factors. Environmental covariates refer to all environmental factors in the ecosystem that have an impact on biological organisms, which are comprehensive and interconnected. Ecological data graph representation technology is aimed at complex ecological data, and its purpose is to produce a reasonable, interpretable graph structure with certain properties. Currently reported network structures, such as bipartite networks, symbiotic networks, protein interaction networks, toxicological networks, etc., are the primary applications of graph representation technology in different scenarios. The above-mentioned graph representation networks provide new methods and perspectives for the description of specific problems, but they are generally based on graph representations of a single type of omics data and a single omics theory. Although the specific effects and relationships between components in a certain field of omics are well expressed and transmitted, they are not universal, have poor explanatory power, and cannot effectively describe multi-omics problems. In addition, the above methods focus on the shallow expression of numerical attributes between factors, and insufficient exploration of the ecological attributes between factors. Taking the symbiotic network as an example, it only considers the data similarity characteristics between environmental covariates, but does not consider the real transmission path and ecological attributes between them. In summary, the existing ecological domain graph representation methods cannot well characterize and learn the role of association relationships in pollutant risk prediction. Therefore, it is necessary to introduce new theories and develop new ecological data graph representation models. Based on effective association, this study starts from the data attributes and ecological attributes of environmental covariates, and constructs a graph representation DTGRep model for complex ecological data based on correlation analysis and association interaction theory.

[0004] Therefore, we urgently need to design an ecological data graph representation method based on the correlation and interaction of environmental covariates to solve the above problems. Summary of the invention

[0005] This paper proposes a new technical solution to solve the above problems through an ecological data graph representation method based on the correlation and interaction of environmental covariates. The specific technical solution includes steps such as data preprocessing, node generation, edge generation, graph generation, and graph representation evaluation. By constructing an ecological topological graph, the complex correlation between environmental covariates is converted into an operational graph structure, providing a solid foundation for the subsequent application of deep learning and graph learning technologies.

[0006] Specifically, it is a method for representing ecological data graphs based on the correlation and interaction of environmental covariates, the method comprising the following steps:

[0007] Based on the theory of ecological association and interaction, the multidimensional environmental covariates and pollutant risks in the ecosystem are represented as nodes in the graph, and the interactions are represented as edges, so as to construct a fully interconnected ecological topological origin graph structure.

[0008] Data normality tests were performed on environmental covariates and pollutant risk coefficients;

[0009] Based on data similarity theory, the data attributes between environmental covariates are calculated;

[0010] Generate the basic graph structure based on the data similarity analysis results;

[0011] By using domain knowledge and influence transfer theory, action layers are set to express the ecological attributes between environmental covariates, thereby optimizing the basic graph structure and eliminating ineffective associations between environmental covariates.

[0012] The final adjacency matrix is ​​generated based on matrix operations and visualized to obtain the advanced graph structure of ecological data.

[0013] As a technical solution of the present invention, the graph representation model is a DTGRep model based on ecological association interaction theory.

[0014] As a technical solution of the present invention, the graph representation model is a DTGRep model based on data similarity.

[0015] As a technical solution of the present invention, the DTGRep model based on data similarity includes the following steps:

[0016] The N environmental covariates included in the study were numbered and labeled, and were assigned consecutive index numbers from 1 to N;

[0017] Calculate the Pearson correlation coefficients between N environmental covariates and obtain the correlation coefficient matrix;

[0018] Convert the correlation coefficient matrix into a correlation matrix according to the set conditions;

[0019] Domain knowledge and relevant interactions are combined to generate an adjacency matrix and construct a DTGRep graph.

[0020] As a technical solution of the present invention, the Pearson correlation coefficient and the Spearman correlation coefficient are calculated by the following formulas respectively:

[0021]

[0022] As a technical solution of the present invention, the domain knowledge includes the impact transmission path between multidimensional environmental covariates in the ecosystem.

[0023] As a technical solution of the present invention, the transmission path is represented by layer block division, specifically including:

[0024] The output layer block can only transmit influence to other layer blocks;

[0025] Interactive layers can both transmit influences outwards and receive influences from other layers;

[0026] The receiving layer block can only receive influences from other layer blocks.

[0027] As a technical solution of the present invention, the evaluation indicators of the graph representation include the degree distribution and clustering coefficient of the nodes, and whether an effective pollution risk prediction model can be constructed based on the graph representation model.

[0028] As a technical solution of the present invention, the degree distribution of a node refers to the number of edges connecting each node, and the clustering coefficient is used to describe the degree of clustering of nodes in the graph, which is calculated by a formula, wherein the degree calculation formula is:

[0029] d v =∑ u∈V A[v,u];

[0030] The calculation formula of the clustering coefficient is:

[0031]

[0032] The method also includes constructing a deep learning model based on graph learning according to the graph representation model, which is used to predict and analyze the pollution risk (cumulative concentration or ecological risk) of heavy metals and other pollutants in the ecosystem, so as to verify the effectiveness of the graph representation method proposed in this study;

[0033] The deep learning model is implemented based on a graph neural network;

[0034] The graph neural network model includes the following steps:

[0035] Input graph representation adjacency matrix and node feature matrix;

[0036] Extract graph structure and node feature information through graph convolution layer;

[0037] The prediction results are output through the fully connected layer.

[0038] In the present invention, the origin graph structure is an undirected graph in which multidimensional environmental covariates are fully interconnected, and the edges between nodes only represent the interconnection relationship. The basic graph structure and the advanced graph structure are directed graphs, and the edges contain association weights. Among them, the basic graph structure introduces the data similarity characteristics between environmental covariates on the basis of the origin graph structure, eliminates a large number of shallow low-correlation edges, and significantly improves the effectiveness and focus of the graph structure; the advanced graph structure adds the ecological attribute characteristics between environmental covariates on the basis of the basic graph structure, further eliminates invalid correlation edges that have no practical significance, and improves the interpretability and comprehensibility of the graph structure.

[0039] Beneficial effects:

[0040] The present invention proposes an ecological data graph representation method based on the correlation and interaction of environmental covariates. By constructing a graph structure of ecological data, ecological data with complex correlation relationships are transformed from traditional non-Euclidean data format to Euclidean space form, providing a new technical means for pollutant risk prediction. Specific beneficial effects include:

[0041] Improve the ability to express data associations: By constructing an ecological topology map, the complex associations between various environmental covariates and effect attributes in the ecosystem are explicitly expressed. Compared with traditional linear data representation methods, the method of the present invention can better capture and describe the interactions and mutual influences between different environmental covariates in the ecosystem, improving the ability to express data associations.

[0042] Support for multi-dimensional data processing: The method of the present invention can effectively express multi-dimensional ecological data, including factors such as geography, climate, society, soil, and pollutants. Through comprehensive analysis of multi-dimensional data, we can fully and deeply understand the state and change patterns of the ecosystem, providing strong support for scientific decision-making.

[0043] Improve the interpretability and applicability of the model: By using data similarity theory, correlation interaction theory and layer block ideas to generate ecological maps, the method of the present invention can intuitively display the correlation strength and similarity between environmental covariates, thereby improving the interpretability of the model. At the same time, the method is applicable to different types of ecosystem data, including farmland ecosystems, water ecosystems and forest ecosystems, and has strong versatility and applicability.

[0044] Enhance the comprehensiveness and accuracy of the graphical representation evaluation: The present invention comprehensively evaluates the generated ecological map through multiple evaluation indicators to ensure the effectiveness and accuracy of the graphical representation. These evaluation indicators can help researchers better understand the structure and function of the ecosystem and guide the subsequent pollutant pollution risk prediction work.

[0045] Supporting modeling of multiple relationships: The method of the present invention supports the generation of ecological maps based on different relationships (such as time series relationship and spatial distribution relationship), which can flexibly adapt to different research needs. By comprehensively considering the time and space dimensions, the dynamic changes and spatial distribution characteristics of the ecosystem can be more comprehensively revealed.

[0046] Promote the application of deep learning technology in environmental science: By converting ecological data into a graph structure in the form of Euclidean space, the method of the present invention provides technical support for deep learning technology based on graph representation. Researchers can use advanced graph neural network technology to deeply mine and analyze ecological data, further improving the accuracy and efficiency of pollutant pollution risk prediction.

[0047] Reduce environmental monitoring costs and secondary pollution: Traditional pollutant cumulative concentration assessment methods, such as quantitative analysis, often require a large amount of chemical reagents and human resources, and have high costs and secondary pollution problems. The method of the present invention can effectively reduce environmental monitoring costs and reduce secondary pollution during the experiment through computer-assisted assessment technology, thus contributing to environmental protection.

[0048] Support early warning of pollution risks and scientific decision-making: Through comprehensive analysis of various environmental covariates and effect attributes in the ecosystem, the method of the present invention can achieve early warning of pollution risks, help decision makers take timely response measures, and prevent further deterioration of the ecological environment. At the same time, based on scientific data analysis results, it can provide a strong basis for environmental management and policy formulation, and promote the sustainable development of the ecological environment.

[0049] In summary, the ecological data graph representation method based on the correlation interaction of environmental covariates proposed in the present invention has significant beneficial effects in improving the ability to express data correlation, supporting multidimensional data processing, improving model interpretability and applicability, enhancing the comprehensiveness and accuracy of graph representation evaluation, supporting multiple correlation relationship modeling, promoting the application of deep learning technology, reducing environmental monitoring costs and secondary pollution, supporting early warning of pollution risks and scientific decision-making, etc., and provides an innovative and efficient technical means for the monitoring, evaluation and management of the ecological environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Provide a framework for ecological data graph representation;

[0051] Figure 2This is the technical framework diagram of DTGRep;

[0052] Figure 3 It is the basic graph structure built based on DTGRep;

[0053] Figure 4 It is an advanced graph structure built based on DTGRep. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention, taking the pollution of heavy metals in the ecosystem of the production area polluted by heavy metals as an example to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] Graph representation can represent rich and complex associations. It extracts relationships from raw data through relational modeling and represents these relationships with graphs. In the graph, nodes represent feature labels that contain specific attributes, and edges represent relationships such as dependency, subordination, and correlation between nodes. Ecological data graph representation is based on graph theory, integrating multi-omics ecological data such as geographic information, geochemistry, and environmental science, and using generalized ecological networks such as symbiotic networks, bipartite networks, and interaction networks to construct an ecological network topology structure for graph representation learning tasks ( Figure 1 Specifically, the environmental covariates are taken as source nodes, the pollutant concentrations are taken as target nodes, and certain attributes between environmental covariates and between environmental covariates and pollutant concentrations are taken as edges to construct an ecological network topology graph, which is used to represent the complex relationship between the environmental covariates of the ecosystem.

[0056] Studies have shown that the ecosystem is an organic whole composed of a number of environmental factors that interact and depend on each other. After heavy metals enter the ecosystem, they will show trade-offs, synergies, and bundling with environmental factors, forming a complex dynamic network. Different environmental factors can change the impact path and risk intensity of the enrichment of pollutants such as heavy metals, greatly increasing the difficulty of targeted prevention and control of heavy metals. The scientific assessment of heavy metal pollution risk is based on the comprehensive collection of information on its enrichment influencing factors and the reasonable simulation of its role. When facing the problem of pollution risk assessment, existing studies are generally based on structured Euclidean spatial data, focusing on the mathematical analysis of shallow characteristics of environmental covariates, while the potential ecological attributes and intrinsic associations between environmental covariates have not been fully explored and demonstrated. This leads to the direct neglect of some valuable information in the ecosystem, resulting in information loss, which means that there are shortcomings in the existing feature representation methods and modeling technologies. In this context, we explored the ecological data graph representation method for complex correlation expression of multidimensional environmental factors, driving effect disclosure, and graph learning technology modeling. Based on data similarity, we proposed the concept of factor action block layering and constructed a DTGRep model based on correlation analysis and correlation interaction theory.

[0057] Before exploring the graph representation techniques for complex association-based ecological data, we first briefly introduce the main theories involved in the research process, including environmental covariates, graphs, graph generation, correlation analysis, and association interactions.

[0058] Environmental covariates:

[0059] Environmental covariates refer to environmental factors within the ecosystem. They are the low-level basic units of the ecological environment and are also the basic characteristic variables of our research. In the origin ecosystem, environmental factors emphasize other environmental components besides organisms, which have direct and indirect effects on the behavior and distribution of heavy metals in organisms. Through data collection, literature research and other methods, we believe that multiple factors such as geography, climate, society, and soil can be included. Among them, geographical factors can include longitude, latitude, elevation, water system density, etc.; climate factors include temperature, light intensity, ultraviolet intensity, annual average precipitation, sunshine hours, etc.; social factors include information such as road density, gross domestic product, population density, village density, as well as mining area attributes such as mine type, mining scale, mining method, and number of mining areas; soil factors include different heavy metal contents in the soil, soil nutrients such as nitrogen, phosphorus, and potassium, as well as soil physical and chemical properties such as soil organic carbon, exchangeable sodium salt, conductivity, basic saturation, exchange base, soil pH, organic matter content, cation exchange capacity, and soil mechanical composition. The heavy metal data in the soil, rice grains and other media in the study area can be obtained through laboratory chemical analysis, and the latitude and longitude data of the sampling points can be obtained using GPS equipment during sampling. The basic data of other environmental covariates can be obtained from public websites such as the website of the Resource and Environmental Science and Data Center of the Chinese Academy of Sciences, the website of the Soil Science Database, and the National Mineral Exploration and Mining Information Disclosure System. The feature definition is then performed using the characteristics of the factor attributes to obtain the corresponding point values. This process is implemented using interpolation technology on the ArcGIS platform. The experimental data set obtained can be defined as: D = {(X1, Y1), (X2, Y2), ···, (X M ,Y M )}, M represents the number of sampling points. i represents the environmental covariate vector of the i-th sampling point, X ij represents the jth environmental covariate at the ith sampling point. Similarly, Y il Represents the content of the lth pollutant at the ith sampling point.

[0060] This experiment obtained a multidimensional environmental covariate dataset D in the ecosystem of a certain study area through data collection, chemical analysis, spatial interpolation, etc. The dataset D contains a total of 1890 sets of data, and the environmental covariates include: longitude (Long), latitude (Lat), elevation (DEM), village density (VD), annual average precipitation (AnnAP), sunshine hours (SunH), water system length (WLen), water system area (WArea), soil acidity (pH), sand mass fraction (Sand), loam mass fraction (Silt), clay mass fraction (Clay), gravel mass fraction (Gravel), soil bulk density (Bu lk), soil organic carbon content (TOC), exchangeable sodium salt (ESP), electrical conductivity (ECE), basic saturation (BSa), exchange base (TEB), cation exchange capacity (CEC), calcium carbonate (CaCO3), gross domestic product (GDP), population density (POP), geochemical background values ​​of five heavy metal pollutants in the soil of the study area, lead (Pb), cadmium (Cd), chromium (Cr), arsenic (As), and mercury (Hg) (respectively expressed as bHM), and soil cadmium concentration (Cd). Therefore, the environmental covariates included in this study include: Long, Lat, AnnAP, SunH, DEM, WLen, WArea, pH, VD, Sand, Silt, Clay, Gravel, Bulk, TOC, ESP, ECE, BSa, TEB, CEC, CaCO3, GDP, POP, bCd, bPb, bCr, bAs, bHg, and Cd.

[0061] Graph: Graph is one of the most powerful frameworks in data structure and algorithm, which can be used to represent all types of structures or systems. Graph not only provides a theoretical framework, but also provides a mathematical foundation on which to analyze, understand and learn complex systems in the real world. Formally, a graph is a two-tuple, denoted by G = (V, ξ), where V and ξ represent the set of nodes and the set of edges respectively. In research, the adjacency matrix A∈R is usually used. n×n To represent a certain graph structure. The adjacency matrix A represents the matrix of connection relationships between adjacent vertices. For node v i and v j The edge (v i ,v j ), if (v i ,v j )∈ξ then A[v i ,v j ]=1, otherwise A[v i ,v j ] = 0. In practical application scenarios, each node is usually mapped into a potential low-dimensional representation, and the structural information of the graph is encoded into the graph using edges.

[0062] Similarity measurement: Similarity measurement is an analysis of the correlation between multiple variables, which can reflect the closeness of the correlation between the data attributes of the variables. The correlation coefficient measurement method is the most widely used correlation evaluation method. Correlation analysis provides a practical basis for exploring the complex relationship between multidimensional environmental covariates in ecosystems.

[0063] Correlation coefficient: The Pearson correlation coefficient r(x,y) between two variables x and y is calculated as follows:

[0064]

[0065] The Spearman correlation coefficient r(x,y) between two variables x and y is calculated as follows:

[0066]

[0067] For the correlation coefficient r, it is generally believed that: when |r|<0.2, it indicates a very low correlation between the two variables; when 0.2≤|r|<0.4, it indicates a low correlation; when 0.4≤|r|<0.7, it indicates a moderate correlation; when 0.7≤|r|<0.9, it indicates a high correlation; and when 0.9≤|r|≤1, it indicates a very high correlation.

[0068] Correlation interaction: The correlation interaction theory highlights the interconnected, mutually reinforcing and restrictive relationship between environmental covariates in the ecosystem, that is, the effect of any factor on heavy metals is not a single one, but multiple interconnected factors influence, interact and act in combination. The combined effect of each factor is not a simple accumulation of the effects of a single factor, but a complex process of mutual superposition and mutual offset between effects. The correlation interaction theory based on ecological impact transmission is an important theoretical basis for ecological data graph representation technology.

[0069] Ecological data graph representation: Ecological data graph representation technology is aimed at complex ecological data, and its purpose is to generate a reasonable graph structure with certain properties. Therefore, unlike general graph generation scenarios, domain knowledge is of great help in the design of graph generation models and evaluation strategies in this study. We define the graph structure G = (V, ξ) of the origin ecosystem, where the node V represents the environmental covariate that has a potential impact on cadmium adsorption, transfer, enrichment and other behaviors, and the edge ξ represents the correlation between environmental covariates. The technical framework of DTGRep is as follows: Figure 2 shown.

[0070] Graph representation: Graph representation requires defining a generation step to give the process of constructing the edges in the graph. This process can be regarded as a certain edge existing at the node v. i and v jOne challenge faced by graph representation is how to accurately and efficiently represent effective association interactions. To address this challenge, we define whether there is an edge between any two environmental covariate nodes based on data similarity and action partitioning as follows:

[0071] (1) Defined by data attributes between environmental covariate nodes

[0072] In this study, the initial structure of the graph representation model was created based on the data similarity between environmental covariates. Specifically, the basic graph representation of DTGRep was created based on the Pearson correlation coefficient or Spearman correlation coefficient between the nodes of the environmental covariates. The reason why the Pearson correlation coefficient and the Spearman correlation coefficient were used as data correlation measures is mainly based on the following considerations: the essence of the Pearson correlation coefficient is the cosine of the vector angle between two sets of data after standardization, and it is the most widely used data similarity measurement method in the ecological field; while the Spearman correlation coefficient is used to measure the rank correlation between two vectors, and is widely used in machine learning technology to measure the correlation between sample vectors.

[0073] Definition: Only when the data correlation between node elements meets the characteristic conditions is it considered that (vi, vj)∈ξ, that is, A[vi, vj]=1, otherwise A[vi, vj]=0. Data correlation analysis gives the intrinsic correlation relationship between nodes from the perspective of data attributes.

[0074] (2) Defined by ecological attributes between environmental covariate nodes

[0075] The above graph structure created based on data similarity contains many worthless or even erroneous associations in the field of ecological environment. For example, the internal associations between geographical environmental covariates such as longitude, latitude, and elevation displayed in the basic graph structure are meaningless. In this context, this study incorporated the ecological attributes of the nodes themselves and proposed the concept of effective associations. It is worth noting that in the field of ecology, the influence relationship between environmental covariates has not been fully revealed. To solve this problem, this study summarized and block-by-block the influence relationships between multidimensional environmental covariates in the origin ecosystem, and divided them into three types of blocks: output block B out , Interaction layer block B inter and receiving layer block B in The specific definitions are as follows:

[0076] B out For the output layer block, the environmental covariates in this layer block can only transmit influences to the environmental covariates in other layer blocks, and cannot receive influences from itself and the environmental covariate nodes in other layer blocks.

[0077] B inter It is an interactive layer block, which can affect Bin The environmental covariate nodes within a layer block can also receive influences from environmental covariate nodes in other layer blocks.

[0078] B in To receive layer blocks, only layer blocks B can be received out and B inter The role of the internal environment covariate node cannot transmit its influence to the outside.

[0079] Therefore, this study believes that for an environmental covariate node, it is necessary to first determine the layer block to which the environmental covariate belongs, and then determine the edge relationship between it and the environmental covariate nodes in its own layer block and other layers. The layer block concept not only simplifies and speeds up the association matching problem between environmental covariates without affecting the knowledge contained in the association relationship, but also clarifies and reflects the inherent similarities between environmental covariates.

[0080] Graph representation evaluation: Evaluating graph generation models is a difficult task, because the generation model has no concept of accuracy and error. Therefore, in addition to the difficulty of characterizing correlation interactions, another challenge facing graph representation is how to reasonably and quantitatively evaluate the reliability of graph representation models. The evaluation method used in this study is to calculate the degree distribution, clustering coefficient distribution and other statistics of the generated graph, and compare them with the statistics of the basic graph structure, as well as whether an effective pollution risk prediction model can be constructed based on the graph representation model. It should be noted that on the basis of conventional statistics, this study incorporates the concept of effective association for a more scientific evaluation of graph representation models.

[0081] (1) Degree

[0082] Degree is a measure of the adjacency between nodes and domain nodes, and plays an important role in determining graph connectivity. The number of edges connected to a node in the graph structure is the degree of the node, and the distribution of the degrees of each node in the graph is the degree distribution. For a node v, the degree d v It is the most intuitive statistic, reflecting the number of edges connected to node v. The calculation formula is as follows:

[0083]

[0084] For an undirected graph, the number of nonzero elements in the i-th row / column of the adjacency matrix A is exactly the degree of the i-th vertex (d vi ); for a directed graph, the degree is divided into out-degree (d out ) and in-degree (d in ).

[0085] (2) Clustering coefficient

[0086] The clustering coefficient can describe the clustering of a graph, and is divided into local clustering coefficient and global clustering coefficient. The clustering coefficient is based on node triples and summarizes the clusters in the entire graph. It can be applied to describe undirected graphs and wired graphs. The calculation formula is as follows:

[0087]

[0088] Where N closed and N open Respectively represent the number of closed triples and open triples in the graph, N u is the domain node combination of node u. During the calculation process, N closed Add N open Equivalent to the number of combinations of any two nodes in the domain node, N closed It is the number of edges between two nodes in the domain node.

[0089] Specific steps: DTGRep is created based on the Pearson correlation coefficient and Spearman correlation coefficient between data. The specific steps are now introduced.

[0090] (1) Assign an index number to each of the N environmental covariates included in the study to obtain a sorting index.

[0091] (2) Perform data normality test and log-normal distribution test on environmental covariates and pollutant concentrations.

[0092] (3) Based on the data set D, the data similarity (Pearson correlation coefficient or Spearman correlation coefficient) between N environmental covariates is calculated to obtain the correlation coefficient matrix R = (r ij ) n*n , the element r in the matrix ij Represents the correlation coefficient between the i-th and j-th environmental covariates. Calculate the data similarity between environmental covariates. For covariates that obey normal distribution (including lognormal distribution), the Pearson correlation coefficient is used to measure the data similarity. For covariates that do not obey normal distribution (including lognormal distribution), the Spearman correlation coefficient is used to measure the data similarity.

[0093] (4) The correlation coefficient matrix R is transformed according to the condition L to obtain the correlation matrix R'. The condition L is defined as: the element |r in the matrix R ij |(i≠j) is greater than the corresponding conversion coefficient K={k1,k2,…,k n}(k i The conversion coefficient corresponding to the i-th feature) is set to 1 if it is significantly correlated at the 0.01 level, otherwise it is set to 0. The initial value of the conversion coefficient K can be set according to the following classification standards.

[0094] Standard: When |r|<0.2, it indicates a very low correlation between the two variables; when 0.2≤|r|<0.4, it indicates a low correlation; when 0.4≤|r|<0.7, it indicates a moderate correlation; when 0.7≤|r|<0.9, it indicates a high correlation; when 0.9≤|r|≤1, it indicates a very high correlation.

[0095] (5) According to the following definition, N environmental covariates are divided into B out , B inter and B in Three layer blocks, and use this to create a path matrix P = (p ij ) n*n . The element p in the matrix P ij =1 indicates that there is a transmission path from the i-th environmental covariate to the j-th environmental covariate, p ij =0 means there is no interaction between environmental covariates i and j.

[0096] Definition: B out It is the output layer block, which has the ability to send data to other layer blocks (B inter and B in ) ability to transmit influences, but not to receive influences from itself and B inter and B in The influence of nodes within the layer block; B inter It is an interactive layer block, which can affect B in Nodes in a layer block can also receive data from other layer blocks (B out and B inter ) The influence of internal nodes; B in To receive layer blocks, only layer blocks B can be received out and B inter The role of internal nodes cannot transmit influence to the outside.

[0097] (6) Performing element-wise multiplication operation on the correlation matrix R' and the path matrix P yields the adjacency matrix A. R', P and A are all 0-1 matrices.

[0098] (7) Construct a DTGRep graph based on the adjacency matrix A, that is, a graph representation G = (V, ξ) based on the Pearson correlation coefficient and the influence path. Specifically, G is a directed graph, and a node (e.g., v j ∈V) represents an environmental covariate (environmental covariate or pollutant concentration), and the edge (for example (v i ,v j )∈ξ) indicates that there is an association between two environmental covariates.

[0099] Experimental analysis: During the experiment, in order to evaluate the performance of the proposed graph representation model, the performance of the basic graph structure and the advanced graph structure created based on DTGRep were compared.

[0100] DTGRep: The DTGRep graph representation model is based on the theory of data similarity and effective association interaction. During the experiment, the Pearson correlation coefficient or Spearman correlation coefficient between all environmental covariates Long, Lat, AnnAP, SunH, DEM, WLen, WArea, pH, VD, Sand, Silt, Clay, Gravel, Bulk, TOC, ESP, ECE, BSa, TEB, CEC, CaCO3, GDP, POP, bCd, bPb, bCr, bAs, bHg, and Cd in the data set D was calculated to obtain the correlation coefficient matrix R. Then, the matrix R was converted into the correlation matrix R' according to the condition L defined above. Combined with the representational significance of the correlation coefficient r and the above-mentioned grading criteria, during the experiment, we converted the conversion coefficient K={k1,k2,…,k n} Medium k i The initial value of is uniformly set to 0.2.

[0101] According to the transmission mechanism and correlation interaction between environmental covariates, the environmental covariates are divided into B out , B inter and B in Three layers. Among them, layer B out Includes Long, Lat, DEM, VD, AnnAP, SunH, WLen and WArea; Layer Block B inter Includes pH, Sand, Silt, Clay, Gravel, Bulk, TOC, ESP, ECE, BSa, TEB, CEC, CaCO3, GDP, POP, bCd, bPb, bCr, bAs and bHg; Layer B in Including Cd. From the above definition of the three layers, we can get the path matrix P = (p ij ) n*n .

[0102] The adjacency matrix A is obtained by performing element-wise multiplication of the correlation matrix R' and the path matrix P. Then, DTGRep is visualized based on A. The basic graph structure obtained during the experiment is shown in Figure 3 , the final advanced graph structure can be found in Figure 4 .

[0103] Since the data similarity between environmental covariates is taken into account, DTGRep removes the shallow low correlations in the origin graph structure, highlights the core correlations, and improves the effectiveness and expression efficiency of the graph structure. Since the effective correlation interactions of environmental covariates are taken into account, DTGRep removes the ineffective correlations in the basic graph structure, further improving the interpretability and effectiveness of the graph structure. Figure 3 and Figure 4, we can intuitively find that: when the correlation coefficient is the same, DTGRep has a more concise, efficient and focused graph structure than the basic graph structure.

[0104] Table 1 shows the degree statistics of each node (environmental covariate) in the DTGRep advanced graph structure and the basic graph structure. It can be seen that in the basic graph structure, there are 171 association paths. After layer blockization, 55 ineffective associations are eliminated, accounting for 32.2%. It fully shows that only considering the data attributes between environmental covariates to build a map has obvious defects, because there are many ineffective associations that have no ecological significance in the associations created only based on data similarity.

[0105] For example Figure 3 In the study, the association between Long, Lat and DEM is of no practical significance for exploring the pollution risk of heavy metals. There are a large number of association paths for Cd pointing to other environmental covariates (such as AnnAP, Bulk, VD), which are obviously wrong. Because environmental covariates can affect the heavy metal content in the soil, but the heavy metal content obviously does not affect environmental covariates such as AnnAP, Bulk, and VD. The existence of invalid paths in the basic graph structure will not only reduce the effectiveness and interpretability of the graph structure, but also affect the effective feature extraction in the graph learning process based on the graph representation technology. The DTGRep graph representation technology proposed in this study, based on the similarity between data, further incorporates the ecological significance of the data into the study based on the association interaction, which solves the above problems well.

[0106] In order to better evaluate the proposed graph representation technology, we calculated the clustering coefficient of DTGRep and the corresponding basic graph structure. According to graph theory, the clustering coefficient is a coefficient that represents the degree of node clustering in a graph. The results show that the average clustering coefficient of DTGRep is 57.20%, which is lower than the average clustering coefficient of the corresponding basic graph structure (71.1%). Specifically for different layers, it can be seen that the average clustering coefficient is higher in B in The decrease in the layer block is the highest. The above results show that the graph created based on the combination strategy of "data correlation + correlation interaction" reduces the degree of clustering of nodes in the graph, and this effect is more significant for the target variable. Therefore, the average clustering coefficient proves the efficiency of DTGRep from another perspective. Further observation of the graph structure of DTGRep shows that it shows rich attribute relationships in the ecological dataset, which indirectly indicates the limitations of structured data.

[0107] By comparing DTGRep with the basic graph structure created only based on data similarity, it can be concluded that DTGRep, which is based on domain knowledge and comprehensively considers the ecological impact transmission path, is efficient and can effectively eliminate ineffective associations in traditional graph representation technology and improve the efficiency of graph representation.

[0108] In order to further verify the effectiveness of the ecological data graph representation method DTGRep based on the interaction of environmental covariates proposed in this study, we used the DTGRep method to construct the topological structure G of the dataset D. We took G as input and heavy metal concentration as label, and built a heavy metal concentration prediction model GNN-HMC based on the graph neural network (GNN) technology. The network structure of the GNN-HMC model is input component→Graph embedding→Linear transformation layers→Output layer. Before using the model, the Min-Max normalization method was used to scale the environmental covariate data. Feature scaling is based on the following considerations: first, unify the dimensions and evaluation criteria at the application level; second, improve the convergence speed of the model; in addition, it can also avoid small values ​​in the data from being swallowed, and avoid large values ​​from causing numerical problems. In addition, to prevent data snooping errors, the sample data set was randomly divided into a training set and a test set in advance at a ratio of 4:1. Among them, the training set is used to adjust the hyperparameters of the model, and the test set is used to evaluate the performance of the model. The experimental results show that the heavy metal concentration prediction model GNN-HMC constructed based on the DTGRep method is effective in predicting Cd in soil. 2 As high as 0.806. It can be seen that the prediction model has excellent performance and the results are very satisfactory, which proves that the ecological data graph representation method DTGRep based on the interaction of environmental covariates is effective.

[0109] The research results provided by the embodiments of the present invention are of great significance. The solution of the present invention introduces the concept of graph representation into the ecological field for the first time, proposes an effective graph representation model, and verifies it in the field of heavy metal pollution risk assessment. It not only transforms ecological data with complex correlation relationships from traditional non-Euclidean data format to Euclidean space form, but also provides a technical basis for further exploring the application of deep learning technology based on graph representation.

[0110] Table 1 Node degree statistics of DTGRep and basic graph structure in graph structure

[0111]

[0112] Note: d out is the out-degree; d in is the in-degree; invalid is the number of invalid associations; R validIt is the ratio of valid edges to the total number of edges of a single node.

[0113] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0114] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for representing ecological data graphs based on the interaction of environmental covariates, characterized in that: The method comprises the following steps: Based on the ecological association interaction theory, a fully interconnected origin graph structure is constructed, in which the multidimensional environmental covariates and pollutant risk coefficients in the ecosystem are represented as nodes in the graph, and the interactions are represented as edges. Data normality test and log-normal distribution test were performed on environmental covariates and pollutant risks; The data similarity between environmental covariates was calculated. The Pearson correlation coefficient was used to measure the data similarity for covariates that followed a normal distribution, and the Spearman correlation coefficient was used to measure the data similarity for covariates that did not follow a normal distribution. Generate the basic graph structure based on the data similarity analysis results; By using domain knowledge and influence transfer theory, action layers are set to express the ecological attributes between environmental covariates, thereby optimizing the basic graph structure and eliminating ineffective associations between environmental covariates. The final adjacency matrix is ​​generated based on matrix operations and visualized to obtain the advanced graph structure of ecological data.

2. The method for representing ecological data graphs based on the interaction of environmental covariates according to claim 1, characterized in that: The graphical representation model is a DTGRep model based on ecological association interaction theory.

3. The method for representing ecological data graphs based on the interaction of environmental covariates according to claim 2, characterized in that: The graph representation model is a DTGRep model based on data similarity.

4. The method for representing ecological data graphs based on the interaction of environmental covariates according to claim 3, characterized in that: The DTGRep model based on data similarity includes the following steps: The N environmental covariates included in the study were numbered and labeled, and were assigned consecutive index numbers from 1 to N; Data normality test and log-normal distribution test were performed on environmental covariates and pollutant risk coefficients; Calculate the data similarity between environmental covariates. For covariates that obey normal distribution, the data similarity is measured by Pearson correlation coefficient, and for covariates that do not obey normal distribution, the data similarity is measured by Spearman correlation coefficient to obtain the correlation coefficient matrix. Convert the correlation coefficient matrix into a correlation matrix according to the set conditions; Combining domain knowledge and influence transfer theory, the adjacency matrix is ​​generated and the DTGRep graph is constructed.

5. The method for representing ecological data graphs based on the interaction of environmental covariates according to claim 4, characterized in that: The Pearson correlation coefficient and the Spearman correlation coefficient are calculated by the following formula (1) and formula (2), respectively:

6. The method for representing ecological data graphs based on the interaction of environmental covariates according to claim 1, characterized in that: The domain knowledge includes the impact transmission paths between multidimensional environmental covariates in the ecosystem.

7. The method for representing ecological data graphs based on the interaction of environmental covariates according to claim 6, characterized in that: The impact transfer path is represented by layer blocks, specifically including: The output layer block can only transmit influence to other layer blocks; Interactive layers can both transmit influences outwards and receive influences from other layers; The receiving layer block can only receive influences from other layer blocks.

8. The method for representing ecological data graphs based on the interaction of environmental covariates according to claim 1, characterized in that: The evaluation indicators of the graph representation include the degree distribution and clustering coefficient of the nodes, and whether an effective pollution risk prediction model can be constructed based on the graph representation model.

9. The method for representing ecological data graphs based on the interaction of environmental covariates according to claim 8, characterized in that: The degree distribution of a node refers to the number of edges connecting each node, while the clustering coefficient is used to describe the degree of clustering of nodes in the graph. It is calculated using the following formula, where the degree calculation formula is: d v =∑ u∈V A[v,u]; The calculation formula of the clustering coefficient is: The method also includes constructing a deep learning model based on the graph representation model DTGRep for predicting and analyzing the pollution risk of pollutants such as heavy metals in the ecosystem; The deep learning model is implemented based on a graph neural network; The graph neural network model includes the following steps: Input graph representation adjacency matrix and node feature matrix; Extract graph structure and node feature information through graph convolution layer; The prediction results are output through the fully connected layer.

Images