A method and system for endangered animal crime risk assessment based on big data

Abstract

The application discloses a kind of endangered animal crime risk assessment method and system based on big data, it is related to risk assessment technical field, including: constructing dynamic semantic association graph;The total amount of resource input and resource output of each node is obtained, according to the niche differentiation strategy, the effective scale, connection efficiency, constraint degree and control ability of each node are calculated based on the neighbor set formed, and the multi-dimensional function atlas of criminal network is formed in combination with dynamic semantic association graph;Extract the cause-effect chain of cause-effect association network, carry out real-time risk analysis to the node reserved by cause-effect chain, and divide risk level.The total amount of resource input of node is quantified by constructing resource flow matrix, etc., the node is mapped to three-dimensional function space to form three-dimensional point cloud, and the function role in criminal network is naturally emerged from data by using density clustering algorithm combined with niche differentiation, and the emerging function role and hidden division mode that cannot be found by traditional method are identified.

Description

Technical Field

[0001] This invention relates to the field of risk assessment technology, specifically to a method and system for risk assessment of endangered animal crimes based on big data. Background Technology

[0002] The illegal trade in endangered wildlife has a long industrial chain, is highly concealed, and has significant cross-regional characteristics. Moreover, with the development of the Internet and logistics industry, the criminal patterns are constantly iterating and upgrading, showing new characteristics of networking, fragmentation, and dynamism, which brings great challenges to the supervision of operations.

[0003] With the development of big data and artificial intelligence technologies, crime risk assessment methods based on data analysis have been gradually applied. However, in practical applications, existing technologies lack in-depth analysis of functional division and resource flow, as well as causal reasoning and multi-dimensional assessment capabilities. They cannot distinguish true causal relationships, resulting in high false alarm and false negative rates. At the same time, existing methods are mostly static assessments that do not consider the time delay effect of criminal activity transmission and cannot predict the development trend and transmission path of criminal activities. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for assessing the risk of endangered animal crimes based on big data, comprising the following specific steps: S1. Based on the multi-source data obtained from big data, perform word segmentation and feature extraction, and construct a feature co-occurrence matrix with all extracted candidate features. Based on the feature co-occurrence matrix, perform feature clustering to generate entities, extract all entities containing at least two different entities, and perform relationship discovery. Combine the obtained confidence scores with entities and automatically discovered effective relationships to construct a dynamic semantic association graph. S2. Based on the dynamic semantic association graph, extract the semantics of all edges and the corresponding relational verbs. After classifying the edge types, construct a node resource flow matrix and calculate the total resource input and total resource output of each node. Calculate the resource conversion coefficient of each node based on the total resource output and total resource input. According to the niche differentiation strategy, establish a three-dimensional point cloud with the total resource input, total resource output, and resource conversion coefficient of each node. Based on the standardization processing and clustering calculation of the three-dimensional point cloud, quantify the functional coverage breadth and functional similarity of each node based on the obtained clusters. Then find the neighbor nodes of each node. Calculate the effective scale, connection efficiency, constraint degree, and control capability of each node based on the formed neighbor set to complete the stability assessment of each node. Combine all the quantified values ​​with the dynamic semantic association graph to form a multi-dimensional functional graph of the crime network. S3. Construct node activity time series based on the multidimensional functional graph of the crime network. After smoothing the node activity time series, generate node pair activity sequence combinations and perform causal relationship testing. The first and second prediction models combine the prediction results with the lag order obtained from the causal relationship test to calculate statistics. Construct a causal association network with the obtained statistics, extract the causal chains of the causal association network, perform real-time risk analysis on the nodes retained in the causal chains, and classify the risk levels.

[0005] Preferably, S1 performs sentence-by-sentence word segmentation on all text data in the multi-source data, extracts all nouns, verbs and quantifiers as candidate features, converts all field values ​​in all non-text data in the multi-source data into candidate features in string form, assigns a unique identifier to each candidate feature, and records the total number of times each candidate feature appears in all data sources, the number of data sources in which each candidate feature appears, and the content of the context window for each appearance of each candidate feature. Based on the multi-source data obtained from big data, including customs seizure records, publicly available data from social media, cross-border logistics trajectory data, financial transaction flow data, animal habitat monitoring data, etc.; the above operations will uniformly convert the raw data of different formats and sources into a computable feature representation, retaining all elements of information that the data may contain. A two-dimensional co-occurrence matrix is ​​constructed using all candidate features as rows and columns. The total number of times any two candidate features co-occur within the same context window is used as the element value corresponding to any two candidate features in the two-dimensional co-occurrence matrix. Specifically, the size of the context window is set to a single sentence or a single data record. That is, if two features appear in the same action record, the same social media post, the same logistics trajectory, or the same financial transaction, it is considered as one co-occurrence. Based on the feature co-occurrence matrix, the point mutual information similarity between any two candidate features is calculated as follows: the ratio of the total number of occurrences of each of the two candidate features to the total number of occurrences of all candidate features is calculated as the probability of each of the two candidate features occurring; the ratio of the total number of co-occurrences of the two candidate features to the total number of occurrences of all candidate features is calculated as the probability of the two candidate features co-occurring; the natural logarithm of the result of dividing the probability of co-occurrence of the two features by the product of the probabilities of occurrence of the two features individually is taken to obtain the point mutual information similarity value. Each candidate feature is treated as an independent cluster, and a hierarchical clustering algorithm is used to cluster all candidate features based on the point mutual information similarity value. The two clusters with the highest similarity are merged into a new cluster. After repeated merging, the feature clustering is completed, and each final cluster is treated as an entity.

[0006] Preferably, in step S1, all context windows containing at least two different entities are extracted, the frequency of occurrence of verb features connecting the two entities in each context window is counted, the top three verb features with the highest frequency in each pair of entities are taken as candidate relationships between each pair of entities, and the point mutual information confidence of each candidate relationship is calculated. The point mutual information confidence of each candidate relationship is taken as the effective relationship between each pair of entities. The point mutual information confidence score is calculated as follows: First, the probability of the verb feature and the pair of entities appearing together is calculated. Then, the product of the probability of the verb feature appearing together and the probability of the pair of entities appearing together is calculated. Next, the probability of the verb feature and the pair of entities appearing together is divided by the product of the probability of the verb feature appearing together and the probability of the pair of entities appearing together. Finally, the natural logarithm of the result of the product of the probability of the verb feature and the pair of entities appearing together and the probability of the pair of entities appearing together is taken to obtain the point mutual information confidence score. Then, with entities as nodes and effective relationships as edges, the hyperbolic tangent transform value of the point mutual information confidence of each effective relationship is used as the weight of the edge. A timestamp attribute is added to each node and each edge to record the time of the first appearance of each node and / or each edge, the time of the last appearance of each node and / or each edge, and the time point of each appearance of each node and / or each edge, forming a dynamic semantic association graph with a complete time dimension.

[0007] Preferably, in step S2, all edges are extracted based on the dynamic semantic association graph. According to the semantics of the verbs corresponding to all extracted edges, all extracted edges are classified into types including input edges, output edges, and bidirectional edges. A two-dimensional resource flow matrix is ​​constructed with all nodes as rows and columns. The element value at any position in the two-dimensional resource flow matrix is ​​used as the sum of the weights of all output edges flowing from the corresponding row node to the corresponding column node. In some other specific embodiments, for bidirectional edges, the weights of the bidirectional edges are evenly distributed to the corresponding positions in the two directions. Input edges represent relationships where resources flow from other nodes to the corresponding node, such as purchasing, receiving, and importing; output edges represent relationships where resources flow from the corresponding node to other nodes, such as selling, sending, and exporting; bidirectional edges represent relationships where resources flow in both directions, such as exchanging and cooperating. Based on the two-dimensional resource flow matrix, the sum of all element values ​​in the corresponding column of each node is used to obtain the total amount of resources flowing from all other nodes to each node, which is used as the total resource input; the sum of all element values ​​in the corresponding row of each node is used to obtain the total amount of resources flowing from each node to all other nodes, which is used as the total resource output. Based on the ratio of total resource output to total resource input, a hyperbolic tangent transform is performed on the ratio. The hyperbolic tangent transform value is then normalized to obtain the resource conversion coefficient. There exists a situation where the total resource input is zero and the resource conversion coefficient is one when the total resource output is zero. S2 constructs a three-dimensional functional space using the total resource input, total resource output, and resource conversion coefficient of each node as three coordinate axes. Each node is mapped to the three-dimensional functional space constructed by the total resource input, total resource output, and resource conversion coefficient of each node, forming a three-dimensional point cloud. The three-dimensional functional space refers to the ability to integrate the three core resource features of a node into a unified spatial representation, so that nodes with different functional roles can naturally form different clusters in the space. After standardizing the 3D point cloud, clustering calculations are performed on the standardized 3D point cloud to obtain clusters.

[0008] Preferably, the clustering calculation of S2 specifically includes: Based on each point in the three-dimensional functional space, calculate the number of other points contained within a preset radius centered on each point, and use this as the density of each point; Mark all points with a density greater than the average density as core points; Randomly select unvisited core points, mark the randomly selected unvisited core points as visited, and create new clusters; Add all points within a preset radius of randomly selected unvisited core points to the queue to be processed; If a current processing point in the queue has not been visited, mark it as visited. If a current processing point in the queue is a core point, add all unvisited points within a preset radius of the current processing point to the queue. Add the current processing point to the current cluster. Process each point in the queue sequentially until the queue is empty, thus completing the construction of a cluster. Repeat the steps until all core points have been visited. Mark all points that are not assigned to any cluster as noise points; S2 assigns each final cluster to a functional role type based on the central position of each cluster in the three-dimensional functional space and the niche differentiation strategy. Then, based on the Euclidean distance from each node to the center of all clusters, the functional coverage breadth of each node is obtained after normalization. Based on the Chebyshev distance between corresponding points of each pair of nodes in the three-dimensional functional space, the functional similarity between each pair of nodes is calculated.

[0009] Preferably, step S2 forms a corresponding neighbor set based on all directly connected neighbor nodes of each node. The neighbor set is the basis for analyzing the structural position of a node in the network and directly reflects the local connectivity of the node. Based on any two neighbor nodes in the neighbor set corresponding to each node, a directly connected edge is found. When a directly connected edge exists, any two neighbor nodes with a directly connected edge are regarded as redundant connections, thereby reflecting the tightness of the connection between the node's neighbors. The more redundant connections, the less obvious the structural hole advantage of the node is. Based on the total number of neighbors of each node, the number of all redundant connections in the neighbor set corresponding to each node is calculated. The result of subtracting the natural logarithm of the number of all redundant connections in the neighbor set corresponding to each node from the total number of neighbors of each node is subjected to hyperbolic tangent transformation to obtain the effective size of each node. The larger the effective size, the more different functional groups connected to each node. The hyperbolic tangent transformation follows the same calculation logic as the above scheme, which will not be repeated here. The connection efficiency is obtained by taking the natural logarithm of the ratio of the effective size of each node to the total number of neighbors of each node. Based on each neighboring node corresponding to each node, calculate the proportion of the weight of the connection between each neighboring node and the corresponding node to the total weight of all connections of the corresponding node, and the proportion of the weight of the connection between each neighboring node and all other neighboring nodes of the corresponding node to the total weight of all connections of the corresponding node. Multiply the two obtained proportions of all neighboring nodes corresponding to each node and add them one by one to obtain the constraint degree of each node. The constraint degree measures the degree to which the node is restricted by its neighboring nodes. The control capability is based on the ratio of the number of shortest paths between any two nodes to the total number of shortest paths, where the shortest path is the path with the fewest edges connecting two nodes.

[0010] Preferably, step S2 divides the time range formed by the timestamp into multiple time windows of equal length. Based on each time window, it obtains the total resource input, total resource output, resource conversion coefficient, functional role type, effective scale, connection efficiency, constraint degree and control capability of each node in each time window, and generates a change sequence of the functional role type of each node in all time windows. It should be noted that each element in the change sequence represents whether the functional role changes between two adjacent time windows. The LZ complexity of the functional role type change sequence corresponding to each node is calculated based on the change sequence, and used as the functional role change frequency. The resource flow fluctuation coefficient is obtained by multiplying the coefficient of variation of the total resource input, the total resource output, and the resource conversion coefficient of each node in all time windows. The structural feature change rate is obtained by multiplying the coefficients of variation of the effective scale, connection efficiency, constraint degree, and control capability of each node in all time windows. After performing a hyperbolic tangent transform on the obtained functional role change frequency, resource flow fluctuation coefficient and structural feature change rate, the resulting value is used as the functional stability index of each node. All quantified values ​​are combined with a dynamic semantic association graph to form a multidimensional functional graph of the crime network.

[0011] Preferably, S3 is based on the multidimensional functional graph of the crime network, and divides the time range formed by the timestamp into multiple time steps of equal length, for example, each time step is one week. The sum of the weights of all edges of each node in each time step is calculated and used as the activity intensity to obtain the activity intensity time series of each node. Wavelet transform is used to smooth the activity intensity time series of each node. The smoothed activity intensity time series of every two nodes are combined into a two-dimensional time series matrix. The rows of the two-dimensional time series matrix represent the time step, and the columns of the two-dimensional time series matrix represent the activity intensity of the two nodes. Based on each node pair, the Bayesian information criterion value of the model under each lag order is calculated. The lag order with the smallest Bayesian information criterion value is selected as the optimal lag order to complete the causal relationship test. In this application, since the Bayesian information criterion can balance the model fit and model complexity, the optimal lag order is selected to avoid overfitting or underfitting. Based on the activity intensity data of one of the nodes in each node pair over several time steps, the first prediction model predicts the activity intensity at the current time step, and the second prediction model predicts the activity intensity of one of the nodes in the node pair over several time steps based on the activity intensity data of the two nodes in the node pair over several time steps. The numerator is the ratio of the predicted difference in activity intensity between two current time steps to the optimal lag order. The denominator is the result of dividing the predicted activity intensity of a node at the current time step by the total length of the time series minus twice the optimal lag order minus one, and the statistic value is obtained to complete the causal relationship test. In S3, the nodes are used as vertices, and the causal relationships obtained from the test are used as directed edges. The weight of the directed edge is equal to the corresponding statistical value. The optimal lag order is added as a time delay attribute to each directed edge, that is, the activity change of node A will have the greatest impact on the activity of node B after a certain number of time steps. Based on the causal association network, all directed paths with a length greater than or equal to two are extracted. Each directed path is treated as a causal chain. The total strength is calculated by multiplying the weights of all edges on the path in each causal chain. Causal chains with a total strength greater than the average strength are retained. It should be noted that the average strength is a preset threshold, which can be defined by the user according to the actual situation in the actual evaluation.

[0012] Preferably, S3 calculates the ratio of the product of the activity intensity of each vertex and the control capability of the corresponding vertex in the causal relationship network to the functional stability index, and then calculates the ratio of the functional coverage breadth of each vertex to the constraint degree. The two ratios are then used as the basic risk value. The basic risk value comprehensively considers the current activity intensity, mediation control capability, functional stability, functional coverage breadth and constraint degree of the vertex, and can fully reflect the basic risk level of the vertex. Based on the similarity between the current causal network and all causal networks, the similarity is calculated as the average of the point mutual information similarity of the feature vectors of corresponding vertices in the two networks. Then, the top several historical scenarios with the highest similarity are found. Finally, the risk values ​​of the corresponding vertices in these historical scenarios are multiplied together. As a historical correction coefficient, the base risk value is multiplied by the historical correction coefficient to obtain the real-time risk value of each vertex. Collect real-time risk values ​​for all vertices, construct a risk value distribution histogram, and classify the risk values ​​as follows: Arrange all risk values ​​in ascending order; initialize the grading scheme by dividing all risk values ​​into one level; further divide each level in the current grading scheme into two sub-levels, and calculate the sum of the within-group variances for each division method; select the division method with the smallest sum of within-group variances as the new grading scheme; repeat the steps until the preset number of gradings is reached, and divide the corresponding risk levels according to the number of gradings. The sum of variances within a group is calculated as follows: for each sub-level, calculate the average of all risk values ​​within that sub-level; then calculate the sum of squares of the differences between each risk value and the average of that sub-level; finally, sum the sums of squares for all sub-levels. Based on the geographical location information of the vertices, all vertices are mapped to a geographic grid, and the regional risk index of each geographic grid is calculated. First, the real-time risk values ​​of all vertices in each geographic grid are multiplied together, and then Gaussian spatial smoothing is performed to transform the real-time risk values ​​of vertices in the same region in the geographic grid into regional risks. Based on the entities associated with the vertices, i.e. the entities extracted in S1, vertices related to endangered animal species are identified. The real-time risk value is converted into species risk by integrating the relevant vertices. The regional risk and species risk are combined by constructing a risk value distribution histogram and classifying the risk values ​​to determine the corresponding regional risk level and species risk level, thus completing the overall assessment.

[0013] This invention also provides a big data-based endangered animal crime risk assessment system, the system comprising: The data processing and modeling module is used to unify multi-source heterogeneous data, complete the standardization transformation of text segmentation and non-text data, extract candidate features and construct a feature co-occurrence matrix, generate entities through point mutual information similarity calculation and hierarchical clustering, mine effective relationships between entities and calculate confidence, and finally construct a dynamic semantic association graph with complete timestamp attributes. The multidimensional functional graph construction module is used to classify edge types based on dynamic semantic association graphs and construct resource flow matrices, calculate the total resource input, total resource output and resource conversion coefficient of nodes, classify node functional roles by combining three-dimensional functional space and density clustering with niche differentiation strategy, analyze network topology, calculate the effective scale, connection efficiency, constraint degree and control capability of nodes, generate functional stability index by combining time window analysis, and generate multidimensional functional graph of criminal network. The assessment module is used to construct a time series of node activity intensity based on a multidimensional functional map of a crime network, perform causal relationship testing on nodes after smoothing by wavelet transform, construct a causal association network with time delay attributes, calculate the basic risk value by comprehensively considering multidimensional indicators, obtain the real-time risk value, classify the risk level of nodes, and further complete the spatial smoothing calculation of regional risk and the association assessment of the risk of endangered animal species.

[0014] Beneficial effects This invention provides a method and system for risk assessment of endangered animal crimes based on big data, which has the following beneficial effects: By constructing a resource flow matrix to quantify the total resource input, total resource output, and resource conversion coefficient of nodes, the nodes are mapped to a three-dimensional functional space to form a three-dimensional point cloud. A density clustering algorithm combined with a niche differentiation strategy is used to allow functional roles in the crime network to emerge naturally from the data, avoiding the limitations of predefined roles. This can effectively identify emerging functional roles and hidden division of labor patterns that traditional methods cannot discover. On this basis, the effective scale, connection efficiency, degree of constraint, and control ability of nodes are further quantified. The functional stability index of nodes is calculated through multi-time window analysis to characterize the structural position, functional role, and dynamic evolution characteristics of nodes in the crime network. By constructing a time series of node activity intensity and smoothing it with wavelet transform, a dual prediction model combined with the Bayesian information criterion is used to conduct a rigorous causal relationship test. A causal association network with time delay attributes is constructed and high-influence causal chains are extracted, which effectively distinguishes causal relationships, reduces the false positive and false negative rates of risk assessment, calculates the basic risk value by comprehensively considering multi-dimensional indicators, and generates historical correction coefficients by combining historical similar scenarios, thus obtaining real-time updated node risk values ​​and improving the accuracy and dynamism of risk assessment. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 This invention provides a method and system for assessing the risk of endangered animal crimes based on big data, comprising the following steps: S1. Based on the multi-source data obtained from big data, perform word segmentation and feature extraction, and construct a feature co-occurrence matrix with all extracted candidate features. Based on the feature co-occurrence matrix, perform feature clustering to generate entities, extract all entities containing at least two different entities, and perform relationship discovery. Combine the obtained confidence scores with entities and automatically discovered effective relationships to construct a dynamic semantic association graph. S1 performs sentence-by-sentence word segmentation on all text data in the multi-source data, extracts all nouns, verbs and quantifiers as candidate features, converts all field values ​​in all non-text data in the multi-source data into candidate features in string form, assigns a unique identifier to each candidate feature, and records the total number of times each candidate feature appears in all data sources, the number of data sources in which each candidate feature appears, and the content of the context window for each appearance of each candidate feature. Based on the multi-source data obtained from big data, including customs seizure records, publicly available data from social media, cross-border logistics trajectory data, financial transaction flow data, animal habitat monitoring data, etc.; the above operations will uniformly convert the raw data of different formats and sources into a computable feature representation, retaining all elements of information that the data may contain. A two-dimensional co-occurrence matrix is ​​constructed using all candidate features as rows and columns. The total number of times any two candidate features co-occur within the same context window is used as the element value corresponding to any two candidate features in the two-dimensional co-occurrence matrix. Specifically, the size of the context window is set to a single sentence or a single data record. That is, if two features appear in the same action record, the same social media post, the same logistics trajectory, or the same financial transaction, it is considered as one co-occurrence. Based on the feature co-occurrence matrix, the point mutual information similarity between any two candidate features is calculated as follows: the ratio of the total number of occurrences of each of the two candidate features to the total number of occurrences of all candidate features is calculated as the probability of each of the two candidate features occurring; the ratio of the total number of co-occurrences of the two candidate features to the total number of occurrences of all candidate features is calculated as the probability of the two candidate features co-occurring; the natural logarithm of the result of dividing the probability of co-occurrence of the two features by the product of the probabilities of occurrence of the two features individually is taken to obtain the point mutual information similarity value. Each candidate feature is treated as an independent cluster, and a hierarchical clustering algorithm is used to cluster all candidate features based on the point mutual information similarity value. The two clusters with the highest similarity are merged into a new cluster. After repeated merging, the feature clustering is completed, and each final cluster is treated as an entity.

[0018] It should be noted that the specific steps of the hierarchical clustering algorithm are as follows: each candidate feature is regarded as an independent cluster, and then the average similarity between all clusters is calculated. The average similarity between all clusters is calculated by taking the arithmetic mean of the point mutual information similarity values ​​of all features in the first cluster and all features in the second cluster, and merging the two clusters with the highest similarity into a new cluster.

[0019] S1 extracts all context windows containing at least two different entities, counts the frequency of verb features connecting the two entities in each context window, takes the three most frequent verb features in each pair of entities as candidate relations between each pair of entities, calculates the point mutual information confidence of each candidate relation, and takes the point mutual information confidence of each candidate relation as the effective relation between each pair of entities. The point mutual information confidence score is calculated as follows: First, the probability of the verb feature and the pair of entities appearing together is calculated. Then, the product of the probability of the verb feature appearing together and the probability of the pair of entities appearing together is calculated. Next, the probability of the verb feature and the pair of entities appearing together is divided by the product of the probability of the verb feature appearing together and the probability of the pair of entities appearing together. Finally, the natural logarithm of the result of the product of the probability of the verb feature and the pair of entities appearing together and the probability of the pair of entities appearing together is taken to obtain the point mutual information confidence score. Then, with entities as nodes and effective relationships as edges, the hyperbolic tangent transform value of the point mutual information confidence of each effective relationship is used as the weight of the edge. A timestamp attribute is added to each node and each edge to record the time of the first appearance of each node and / or each edge, the time of the last appearance of each node and / or each edge, and the time point of each appearance of each node and / or each edge, forming a dynamic semantic association graph with a complete time dimension. It should be noted that the hyperbolic tangent transform can map the point mutual information confidence values ​​of any real number range to a range between -1 and 1, compressing the influence of extreme values ​​and making the edge weight distribution more reasonable. In practice, the specific steps of the hyperbolic tangent transform are as follows: First, calculate twice the point mutual information confidence value of each effective relation; then calculate twice the power of the point mutual information confidence value of each effective relation minus the power of negative twice the point mutual information confidence value of each effective relation for the natural exponent; finally, calculate the natural exponent... The hyperbolic tangent transform value is obtained by adding twice the confidence level of the point-to-point mutual information of each effective relation to twice the natural exponent raised to the power of the confidence level of the point-to-point mutual information of each effective relation; finally, the result of subtracting twice the confidence level of the point-to-point mutual information of each effective relation from twice the confidence level of the natural exponent raised to the power of the confidence level of the point-to-point mutual information of each effective relation and dividing the result by adding twice the confidence level of the point-to-point mutual information of each effective relation to twice the confidence level of the natural exponent raised to the power of the confidence level of the point-to-point mutual information of each effective relation is divided by the result of adding twice the confidence level of the point-to-point mutual information of each effective relation to twice the confidence level of the natural exponent raised to the power of the confidence level of the point-to-point mutual information of each effective relation.

[0020] S2. Based on the dynamic semantic association graph, extract the semantics of all edges and the corresponding relational verbs. After classifying the edge types, construct a node resource flow matrix and calculate the total resource input and total resource output of each node. Calculate the resource conversion coefficient of each node based on the total resource output and total resource input. According to the niche differentiation strategy, establish a three-dimensional point cloud with the total resource input, total resource output, and resource conversion coefficient of each node. Based on the standardization processing and clustering calculation of the three-dimensional point cloud, quantify the functional coverage breadth and functional similarity of each node based on the obtained clusters. Then find the neighbor nodes of each node. Calculate the effective scale, connection efficiency, constraint degree, and control capability of each node based on the formed neighbor set to complete the stability assessment of each node. Combine all the quantified values ​​with the dynamic semantic association graph to form a multi-dimensional functional graph of the crime network. S2 extracts all edges based on the dynamic semantic association graph. According to the semantics of the verbs corresponding to all extracted edges, the extracted edges are divided into types including input edges, output edges, and bidirectional edges. A two-dimensional resource flow matrix is ​​constructed with all nodes as rows and columns. The element value of any position in the two-dimensional resource flow matrix is ​​used as the sum of the weights of all output edges flowing from the corresponding row node to the corresponding column node. In some other specific implementations, for bidirectional edges, the weights of the bidirectional edges are evenly distributed to the corresponding positions in the two directions. Input edges represent relationships where resources flow from other nodes to the corresponding node, such as purchasing, receiving, and importing; output edges represent relationships where resources flow from the corresponding node to other nodes, such as selling, sending, and exporting; bidirectional edges represent relationships where resources flow in both directions, such as exchanging and cooperating. Based on a two-dimensional resource flow matrix, the sum of all element values ​​in the corresponding column of each node yields the total amount of resources flowing from all other nodes to each node, which is the total resource input. Similarly, the sum of all element values ​​in the corresponding row of each node yields the total amount of resources flowing from each node to all other nodes, which is the total resource output. The total resource input refers to a node's ability to acquire resources from other nodes, serving as a quantitative indicator of a node's resource acquisition position in the network. The total resource output refers to a node's ability to provide resources to other nodes, serving as a quantitative indicator of a node's resource supply position in the network. Based on the ratio of total resource output to total resource input, a hyperbolic tangent transform is performed on the ratio. The hyperbolic tangent transform value is then normalized to obtain the resource conversion coefficient. There exists a situation where the total resource input is zero and the resource conversion coefficient is one when the total resource output is zero. In the above, the hyperbolic tangent transform of the point mutual information confidence value maps the point mutual information confidence value of any real number range to a range between -1 and 1. The core of this step remains the same. To further understand the technical solution disclosed in this application, the specific steps of the hyperbolic tangent transform here include: calculating twice the ratio of total resource output to total resource input; calculating the natural index raised to the power of twice the ratio of total resource output to total resource input, minus the natural index raised to the power of negative twice the ratio of total resource output to total resource input; calculating the natural index raised to the power of twice the ratio of total resource output to total resource input, plus the natural index raised to the power of negative twice the ratio of total resource output to total resource input; and dividing the result of subtracting the natural index raised to the power of twice the ratio of total resource output to total resource input, by the natural index raised to the power of twice the ratio of total resource output to total resource input, plus the natural index raised to the power of negative twice the ratio of total resource output to total resource input, to obtain the hyperbolic tangent transform value. In summary, the resource conversion coefficients obtained through hyperbolic tangent transform reflect a node's ability to transform input resources into output resources, which facilitates subsequent cluster analysis. S2 constructs a three-dimensional functional space using the total resource input, total resource output, and resource conversion coefficient of each node as three coordinate axes. Each node is mapped to the three-dimensional functional space constructed by the total resource input, total resource output, and resource conversion coefficient of each node, forming a three-dimensional point cloud. The three-dimensional functional space refers to the ability to integrate the three core resource features of a node into a unified spatial representation, so that nodes with different functional roles can naturally form different clusters in the space. After standardizing the 3D point cloud, clustering calculations are performed on the standardized 3D point cloud to obtain clusters.

[0021] In some other specific implementations, the standardization process can refer to: calculating the average value and standard deviation of all points on the three coordinate axes, then subtracting the average value of all points on the corresponding coordinate axis from the value of each point on each coordinate axis, and then dividing by the standard deviation of all points on the corresponding coordinate axis, so that the three coordinate axes have the same scale.

[0022] The clustering calculation for S2 is as follows: Based on each point in the three-dimensional functional space, calculate the number of other points contained within a preset radius centered on each point, and use this as the density of each point; Mark all points with a density greater than the average density as core points; Randomly select unvisited core points, mark the randomly selected unvisited core points as visited, and create new clusters; Add all points within a preset radius of randomly selected unvisited core points to the queue to be processed; If a current processing point in the queue has not been visited, mark it as visited. If a current processing point in the queue is a core point, add all unvisited points within a preset radius of the current processing point to the queue. Add the current processing point to the current cluster. Process each point in the queue sequentially until the queue is empty, thus completing the construction of a cluster. Repeat the steps until all core points have been visited. Mark all points that are not assigned to any cluster as noise points; S2 assigns each cluster to a functional role type based on its central position in the three-dimensional functional space and the niche differentiation strategy. As mentioned above, this application uses total resource input, total resource output, and resource conversion coefficient as coordinate axes. Therefore, the differences in spatial position also determine the differences in the above three quantitative indicators. That is, clusters with high total resource input and low total resource output correspond to resource consumption roles, clusters with low total resource input and high total resource output correspond to resource production roles, clusters with both high total resource input and high total resource output and moderate resource conversion coefficients correspond to resource transformation roles, and clusters with extremely high resource conversion coefficients correspond to resource processing roles. This reflects the specific implementation of the niche differentiation strategy in this application. Then, based on the Euclidean distance from each node to the center of all clusters, the functional coverage breadth of each node is obtained after normalization. Based on the Chebyshev distance between corresponding points of each pair of nodes in the three-dimensional functional space, the functional similarity between each pair of nodes is calculated. The Chebyshev distance is the maximum value of the absolute values ​​of the differences on the three coordinate axes. The functional similarity value is obtained by taking the reciprocal of the natural exponent of the Chebyshev distance between corresponding points of each pair of nodes in the three-dimensional functional space, because the Chebyshev distance can effectively measure the maximum difference between two nodes in all dimensions and reflects the essential difference of the functional roles of nodes better than the Euclidean distance.

[0023] S2 forms a neighbor set based on all directly connected neighbor nodes of each node. The neighbor set is the basis for analyzing the structural position of a node in the network and directly reflects the local connectivity of the node. Based on any two neighbor nodes in the neighbor set corresponding to each node, a directly connected edge is found. If a directly connected edge exists, any two neighbor nodes with a directly connected edge are considered as redundant connections. This reflects the tightness of the connection between the node's neighbors. The more redundant connections there are, the less obvious the structural hole advantage of the node is. Based on the total number of neighbors of each node, the number of all redundant connections in the neighbor set corresponding to each node is calculated. The result of subtracting the natural logarithm of the number of all redundant connections in the neighbor set corresponding to each node from the total number of neighbors of each node is subjected to hyperbolic tangent transformation to obtain the effective size of each node. The larger the effective size, the more different functional groups connected to each node. The hyperbolic tangent transformation follows the same calculation logic as the above scheme, which will not be repeated here. The connection efficiency is obtained by taking the natural logarithm of the ratio of the effective size of each node to the total number of neighbors of each node. Based on each neighboring node corresponding to each node, calculate the proportion of the weight of the connection between each neighboring node and the corresponding node to the total weight of all connections of the corresponding node, and the proportion of the weight of the connection between each neighboring node and all other neighboring nodes of the corresponding node to the total weight of all connections of the corresponding node. Multiply the two obtained proportions of all neighboring nodes corresponding to each node and add them one by one to obtain the constraint degree of each node. The constraint degree measures the degree to which the node is restricted by its neighboring nodes. The control capability is based on the ratio of the number of shortest paths between any two nodes to the total number of shortest paths, where the shortest path is the path with the fewest edges connecting two nodes.

[0024] S2 divides the time range formed by the timestamp into multiple time windows of equal length. Based on each time window, it obtains the total amount of resource input, total amount of resource output, resource conversion coefficient, functional role type, effective scale, connection efficiency, constraint degree and control capability of each node in each time window. It generates a sequence of changes in the functional role type of each node in all time windows. It should be noted that each element in the change sequence represents whether the functional role changes between two adjacent time windows. The LZ complexity of the functional role type change sequence corresponding to each node is calculated based on the change sequence, and used as the functional role change frequency. The resource flow fluctuation coefficient is obtained by multiplying the coefficient of variation of the total resource input, the total resource output, and the resource conversion coefficient of each node in all time windows. The structural feature change rate is obtained by multiplying the coefficients of variation of the effective scale, connection efficiency, constraint degree, and control capability of each node in all time windows. In some other specific implementations, the LZ complexity is calculated in this application as follows: starting from the first element of the functional role type change sequence corresponding to each node, subsequent elements are added to the current substring in turn until the current substring has not appeared in the previous part of the functional role type change sequence corresponding to each node. At this time, a new substring is recorded, and then the above process is repeated starting from the next element. The number of substrings obtained is the LZ complexity of the functional role type change sequence corresponding to each node. The coefficient of variation mentioned above refers to the standard deviation of the total resource input of each node across all time windows divided by the average of the total resource input of each node across all time windows. This is because the coefficient of variation can eliminate the differences in dimensions and scales between different indicators and measure the relative fluctuation of the indicators. Taking the natural logarithm after multiplication can integrate the fluctuation of the three indicators and compress the influence of extreme values. After performing a hyperbolic tangent transform on the obtained functional role change frequency, resource flow fluctuation coefficient and structural feature change rate, the resulting value is used as the functional stability index of each node. By combining all quantitative values ​​with a dynamic semantic association graph, a multidimensional functional map of the crime network is formed. The multidimensional functional map of the crime network is labeled with total resource input, total resource output, resource conversion coefficient, type of functional role, functional coverage, functional similarity, effective scale, connection efficiency, degree of constraint, control ability, and functional stability index. In summary, by differentiating ecological niches, functional roles emerge naturally from the data, avoiding the limitations of predefined roles, controlling key nodes of resource flow, and correcting the bias of assessing node importance based solely on resource flow, thus providing a comprehensive characterization of the internal structure and functional division of labor in endangered animal crimes.

[0025] S3. Construct node activity time series based on the multidimensional functional graph of the crime network. After smoothing the node activity time series, generate node pair activity sequence combinations and perform causal relationship testing. The first and second prediction models combine the prediction results with the lag order obtained from the causal relationship test to calculate statistics. Construct a causal association network with the obtained statistics, extract the causal chains of the causal association network, perform real-time risk analysis on the nodes retained in the causal chains, and classify the risk levels.

[0026] S3 is based on the multidimensional functional graph of the crime network and divides the time range formed by timestamps into multiple time steps of equal length, such as one week for each time step. The sum of the weights of all edges of each node in each time step is calculated and used as the activity intensity to obtain the activity intensity time series of each node. Wavelet transform is used to smooth the activity intensity time series of each node. The smoothed activity intensity time series of every two nodes are combined into a two-dimensional time series matrix. The rows of the two-dimensional time series matrix represent the time step, and the columns of the two-dimensional time series matrix represent the activity intensity of the two nodes. Based on each node pair, the Bayesian information criterion value of the model under each lag order is calculated. The lag order with the smallest Bayesian information criterion value is selected as the optimal lag order to complete the causal relationship test. In this application, since the Bayesian information criterion can balance the model fit and model complexity, the optimal lag order is selected to avoid overfitting or underfitting. Based on the activity intensity data of one of the nodes in each node pair over several time steps, the first prediction model predicts the activity intensity at the current time step, and the second prediction model predicts the activity intensity of one of the nodes in the node pair over several time steps based on the activity intensity data of the two nodes in the node pair over several time steps. The numerator is the ratio of the predicted difference in activity intensity between the two current time steps to the optimal lag order. The denominator is the result of dividing the predicted activity intensity of one node at the current time step by the total length of the time series minus twice the optimal lag order minus one. This yields a statistical value, which is used to complete the causal relationship test. The existence of a statistical value indicates that there is a causal relationship between the two nodes. The statistical value is a quantitative representation of the strength of the causal relationship. Taking the first prediction model as an example, the parameter fitting process of the linear regression model is as follows: calculate the mean of the independent variables and the mean of the dependent variable for all nodes (node ​​B); then calculate the covariance of each independent variable with respect to the dependent variable, as well as the variance of each independent variable; next, divide the covariance by the variance to obtain the coefficients of the independent variables; finally, subtract the sum of the coefficients of all independent variables multiplied by the corresponding means of the independent variables from the mean of the dependent variable to obtain the intercept term. The first prediction model serves as the baseline model, using only the historical data of node B itself to predict the future activities of node B, reflecting the autocorrelation of node B's activities. The second prediction model incorporates historical data of node A into the baseline model, reflecting the impact of node A's activities on node B's activities. Both prediction models reference commonly used linear fitting functions and are trained using historical data. S3 uses nodes as vertices and the causal relationships obtained from the test as directed edges. The weight of the directed edge is equal to the corresponding statistical value. It constructs a causal relationship network and adds the optimal lag order as a time delay attribute to each directed edge. That is, the change in the activity of node A will have the greatest impact on the activity of node B after a certain number of time steps. Based on the causal association network, all directed paths with a length greater than or equal to two are extracted. Each directed path is treated as a causal chain. The total strength is calculated by multiplying the weights of all edges on the path in each causal chain. Causal chains with a total strength greater than the average strength are retained. It should be noted that the average strength is a preset threshold, which can be defined by the user according to the actual situation in the actual evaluation.

[0027] S3 calculates the ratio of the product of the activity intensity and the control capability of each vertex in the causal network to the functional stability index, and then calculates the ratio of the functional coverage breadth and constraint degree of each vertex. The two ratios are then used as the basic risk value. The basic risk value comprehensively considers the current activity intensity, mediation control capability, functional stability, functional coverage breadth and constraint degree of the vertex, and can fully reflect the basic risk level of the vertex. Based on the similarity between the current causal network and all causal networks, the similarity is calculated as the average of the point mutual information similarity of the feature vectors of corresponding vertices in the two networks. Then, the top several historical scenarios with the highest similarity are found. Finally, the risk values ​​of the corresponding vertices in these historical scenarios are multiplied together. As a historical correction coefficient, the base risk value is multiplied by the historical correction coefficient to obtain the real-time risk value of each vertex. Collect real-time risk values ​​for all vertices, construct a risk value distribution histogram, and classify the risk values ​​as follows: Arrange all risk values ​​in ascending order; initialize the grading scheme by dividing all risk values ​​into one level; further divide each level in the current grading scheme into two sub-levels, and calculate the sum of the within-group variances for each division method; select the division method with the smallest sum of within-group variances as the new grading scheme; repeat the steps until the preset number of gradings is reached, and divide the corresponding risk levels according to the number of gradings. The sum of variances within a group is calculated as follows: for each sub-level, calculate the average of all risk values ​​within that sub-level; then calculate the sum of squares of the differences between each risk value and the average of that sub-level; finally, sum the sums of squares for all sub-levels. Based on the geographical location information of the vertices, all vertices are mapped to a geographic grid, and the regional risk index of each geographic grid is calculated. First, the real-time risk values ​​of all vertices in each geographic grid are multiplied together, and then Gaussian spatial smoothing is performed to transform the real-time risk values ​​of vertices in the same region in the geographic grid into regional risks. Based on the entities associated with the vertices, i.e. the entities extracted in S1, vertices related to endangered animal species are identified. The real-time risk value is converted into species risk by integrating the relevant vertices. The regional risk and species risk are combined by constructing a risk value distribution histogram and classifying the risk values ​​to determine the corresponding regional risk level and species risk level, thus completing the overall assessment.

[0028] For example, this application divides vertices into four risk levels: extremely high risk, high risk, medium risk, and low risk. Different countermeasures are taken according to the risk level. Preferably, this application can also simulate the effect of implementing different intensity perturbations on each high-risk vertex, calculate the decrease magnitude and rate of decrease of the overall network risk index under each perturbation intensity, and then comprehensively consider the action cost and intervention effect to select the intervention scheme with the highest cost-effectiveness as the recommended action measure.

[0029] This invention also provides a big data-based endangered animal crime risk assessment system, comprising: The data processing and modeling module is used to unify multi-source heterogeneous data, complete the standardization transformation of text segmentation and non-text data, extract candidate features and construct a feature co-occurrence matrix, generate entities through point mutual information similarity calculation and hierarchical clustering, mine effective relationships between entities and calculate confidence, and finally construct a dynamic semantic association graph with complete timestamp attributes. The multidimensional functional graph construction module is used to classify edge types based on dynamic semantic association graphs and construct resource flow matrices, calculate the total resource input, total resource output and resource conversion coefficient of nodes, classify node functional roles by combining three-dimensional functional space and density clustering with niche differentiation strategy, analyze network topology, calculate the effective scale, connection efficiency, constraint degree and control capability of nodes, generate functional stability index by combining time window analysis, and generate multidimensional functional graph of criminal network. The assessment module is used to construct a time series of node activity intensity based on a multidimensional functional map of a crime network, perform causal relationship testing on nodes after smoothing by wavelet transform, construct a causal association network with time delay attributes, calculate the basic risk value by comprehensively considering multidimensional indicators, obtain the real-time risk value, classify the risk level of nodes, and further complete the spatial smoothing calculation of regional risk and the association assessment of the risk of endangered animal species.

[0030] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0031] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0032] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for assessing the risk of endangered animal crime based on big data, characterized in that: The specific steps include the following: S1. Based on the multi-source data obtained from big data, perform word segmentation and feature extraction, and construct a feature co-occurrence matrix with all extracted candidate features. Based on the feature co-occurrence matrix, perform feature clustering to generate entities, extract all entities containing at least two different entities, and perform relationship discovery. Combine the obtained confidence scores with entities and automatically discovered effective relationships to construct a dynamic semantic association graph. S1 performs sentence-by-sentence word segmentation on all text data in the multi-source data, extracts all nouns, verbs and quantifiers as candidate features, converts all field values ​​in all non-text data in the multi-source data into candidate features in string form, assigns a unique identifier to each candidate feature, and records the total number of times each candidate feature appears in all data sources, the number of data sources in which each candidate feature appears, and the content of the context window for each appearance of each candidate feature. Construct a two-dimensional co-occurrence matrix with all candidate features as rows and columns. The total number of times any two candidate features co-occur within the same context window is used as the element value of any two candidate features in the two-dimensional co-occurrence matrix. Based on the feature co-occurrence matrix, the point mutual information similarity between any two candidate features is calculated as follows: the ratio of the total number of occurrences of each of the two candidate features to the total number of occurrences of all candidate features is calculated as the probability of each of the two candidate features occurring; the ratio of the total number of co-occurrences of the two candidate features to the total number of occurrences of all candidate features is calculated as the probability of the two candidate features co-occurring; the natural logarithm of the result of dividing the probability of co-occurrence of the two features by the product of the probabilities of occurrence of the two features individually is taken to obtain the point mutual information similarity value. Each candidate feature is treated as an independent cluster, and a hierarchical clustering algorithm is used to cluster all candidate features based on the point mutual information similarity value. The two clusters with the highest similarity are merged into a new cluster. After repeated merging, the feature clustering is completed. Each final cluster is treated as an entity. S2. Based on the dynamic semantic association graph, extract the semantics of all edges and the corresponding relational verbs. After classifying the edge types, construct a node resource flow matrix and calculate the total resource input and total resource output of each node. Calculate the resource conversion coefficient of each node based on the total resource output and total resource input. According to the niche differentiation strategy, establish a three-dimensional point cloud with the total resource input, total resource output, and resource conversion coefficient of each node. Based on the standardization processing and clustering calculation of the three-dimensional point cloud, quantify the functional coverage breadth and functional similarity of each node based on the obtained clusters. Then find the neighbor nodes of each node. Calculate the effective scale, connection efficiency, constraint degree, and control capability of each node based on the formed neighbor set to complete the stability assessment of each node. Combine all the quantified values ​​with the dynamic semantic association graph to form a multi-dimensional functional graph of the crime network. S3. Construct node activity time series based on the multidimensional functional graph of the crime network. After smoothing the node activity time series, generate node pair activity sequence combinations and perform causal relationship testing. The first and second prediction models combine the prediction results with the lag order obtained from the causal relationship test to calculate statistics. Construct a causal association network with the obtained statistics, extract the causal chains of the causal association network, perform real-time risk analysis on the nodes retained in the causal chains, and classify the risk levels.

2. The method for assessing the risk of endangered animal crime based on big data according to claim 1, characterized in that: S1 extracts all context windows containing at least two different entities, counts the frequency of verb features connecting the two entities in each context window, takes the three most frequent verb features in each pair of entities as candidate relationships between each pair of entities, calculates the point mutual information confidence of each candidate relationship, and takes the point mutual information confidence of each candidate relationship as the effective relationship between each pair of entities. Then, with entities as nodes and effective relationships as edges, the hyperbolic tangent transform value of the point mutual information confidence of each effective relationship is used as the weight of the edge. A timestamp attribute is added to each node and each edge to record the time of the first appearance of each node and / or each edge, the time of the last appearance of each node and / or each edge, and the time point of each appearance of each node and / or each edge, forming a dynamic semantic association graph with a complete time dimension.

3. The method for assessing the risk of endangered animal crime based on big data according to claim 1, characterized in that: S2 extracts all edges based on the dynamic semantic association graph. According to the semantics of the verbs corresponding to all extracted edges, the extracted edges are divided into types including input edges, output edges and bidirectional edges. A two-dimensional resource flow matrix is ​​constructed with all nodes as rows and columns. The element value of any position in the two-dimensional resource flow matrix is ​​used as the sum of the weights of all output edges flowing from the corresponding row node to the corresponding column node. Based on the two-dimensional resource flow matrix, the values ​​of all elements in the corresponding column of each node are added together to obtain the total amount of resources flowing from all other nodes to each node, which is used as the total amount of resource input. Add up all the element values ​​in the corresponding row of each node to get the total amount of resources that flow from each node to all other nodes, which is used as the total amount of resources output. Based on the ratio of total resource output to total resource input, a hyperbolic tangent transform is performed on the ratio. The hyperbolic tangent transform value is then normalized to obtain the resource conversion coefficient. There exists a situation where the total resource input is zero and the resource conversion coefficient is one when the total resource output is zero. The S2 constructs a three-dimensional functional space using the total resource input, total resource output, and resource conversion coefficient of each node as three coordinate axes, mapping each node to the three-dimensional functional space constructed by the total resource input, total resource output, and resource conversion coefficient of each node, thus forming a three-dimensional point cloud; After standardizing the 3D point cloud, clustering calculations are performed on the standardized 3D point cloud to obtain clusters.

4. The method for assessing the risk of endangered animal crime based on big data according to claim 3, characterized in that: The clustering calculation for S2 is specifically as follows: Based on each point in the three-dimensional functional space, calculate the number of other points contained within a preset radius centered on each point, and use this as the density of each point; Mark all points with a density greater than the average density as core points; Randomly select unvisited core points, mark the randomly selected unvisited core points as visited, and create new clusters; Add all points within a preset radius of randomly selected unvisited core points to the queue to be processed; If a current processing point in the queue has not been visited, mark it as visited. If a current processing point in the queue is a core point, add all unvisited points within a preset radius of the current processing point to the queue. Add the current processing point to the current cluster. Process each point in the queue sequentially until the queue is empty, thus completing the construction of a cluster. Repeat the steps until all core points have been visited. Mark all points that are not assigned to any cluster as noise points; S2 assigns each final cluster to a functional role type based on the central position of each cluster in the three-dimensional functional space and the niche differentiation strategy. Then, based on the Euclidean distance from each node to the center of all clusters, the functional coverage breadth of each node is obtained after normalization. Based on the Chebyshev distance between corresponding points of each pair of nodes in the three-dimensional functional space, the functional similarity between each pair of nodes is calculated.

5. The method for assessing the risk of endangered animal crime based on big data according to claim 1, characterized in that: S2 forms a corresponding neighbor set based on all directly connected neighbor nodes of each node, and finds a directly connected edge between any two neighbor nodes in the neighbor set corresponding to each node. If a directly connected edge exists, then any two neighbor nodes with a directly connected edge are treated as redundant connections. Based on the total number of neighbors of each node, calculate the number of all redundant connections in the neighbor set corresponding to each node, and take the hyperbolic tangent transform of the result of subtracting the natural logarithm of the number of all redundant connections in the neighbor set corresponding to each node from the total number of neighbors of each node to obtain the effective size of each node. The connection efficiency is obtained by taking the natural logarithm of the ratio of the effective size of each node to the total number of neighbors of each node. Based on each neighbor node corresponding to each node, calculate the proportion of the weight of the connection between each neighbor node and the corresponding node to the total weight of all connections of the corresponding node, and the proportion of the weight of the connection between each neighbor node and all other neighbor nodes of the corresponding node to the total weight of all connections of the corresponding neighbor node. Multiply the two proportions obtained for all neighbor nodes corresponding to each node and add them one by one to obtain the degree of constraint of each node. The control capability is based on the ratio of the number of shortest paths between any two nodes to the total number of all shortest paths.

6. The method for assessing the risk of endangered animal crime based on big data according to claim 5, characterized in that: S2 divides the time range formed by the timestamp into multiple time windows of equal length. Based on each time window, it obtains the total resource input, total resource output, resource conversion coefficient, functional role type, effective scale, connection efficiency, constraint degree and control capability of each node in each time window, and generates a sequence of changes in the functional role type of each node in all time windows. The LZ complexity of the functional role type change sequence corresponding to each node is calculated based on the change sequence, and used as the functional role change frequency. The resource flow fluctuation coefficient is obtained by multiplying the coefficient of variation of the total resource input, the total resource output, and the resource conversion coefficient of each node in all time windows. The structural feature change rate is obtained by multiplying the coefficients of variation of the effective scale, connection efficiency, constraint degree, and control capability of each node in all time windows. After performing a hyperbolic tangent transform on the obtained functional role change frequency, resource flow fluctuation coefficient and structural feature change rate, the resulting value is used as the functional stability index of each node. All quantified values ​​are combined with a dynamic semantic association graph to form a multidimensional functional graph of the crime network.

7. The method for assessing the risk of endangered animal crime based on big data according to claim 1, characterized in that: The S3 is based on the multidimensional functional graph of the crime network, and divides the time range formed by the timestamp into multiple time steps of equal length. The sum of the weights of all edges of each node in each time step is calculated and used as the activity intensity to obtain the activity intensity time series of each node. Wavelet transform is used to smooth the activity intensity time series of each node, and the smoothed activity intensity time series of every two nodes are combined into a two-dimensional time series matrix. Based on each node pair, calculate the Bayesian information criterion value of the model under each lag order, select the lag order with the smallest Bayesian information criterion value as the optimal lag order, and complete the causal relationship test. Based on the activity intensity data of one of the nodes in each node pair over several time steps, the first prediction model predicts the activity intensity at the current time step, and the second prediction model predicts the activity intensity of one of the nodes in the node pair over several time steps based on the activity intensity data of the two nodes in the node pair over several time steps. The numerator is the ratio of the predicted difference in activity intensity at two current time steps to the optimal lag order. The denominator is the result of dividing the predicted activity intensity of a node at the current time step by the total length of the time series minus twice the optimal lag order minus one, and the statistic is obtained to complete the causal relationship test. S3 constructs a causal relationship network with nodes as vertices and the causal relationships obtained from the test as directed edges. The weight of the directed edge is equal to the corresponding statistical value. An optimal lag order is added to each directed edge as a time delay attribute. Based on the causal association network, all directed paths with a length greater than or equal to two are extracted. Each directed path is treated as a causal chain. The total strength is calculated by multiplying the weights of all edges on the path in each causal chain. Causal chains with a total strength greater than the average strength are retained.

8. The method for assessing the risk of endangered animal crime based on big data according to claim 1, characterized in that: S3 calculates the ratio of the product of the activity intensity of each vertex and the control capability of the corresponding vertex in the causal relationship network to the functional stability index, and then calculates the ratio of the functional coverage breadth of each vertex to the constraint degree. The two ratios are then used as the basic risk value. Based on the similarity between the current causal network and all causal networks, the historical correction coefficient is used as the base risk value. The base risk value is multiplied by the historical correction coefficient to obtain the real-time risk value of each vertex. Collect real-time risk values ​​for all vertices, construct a risk value distribution histogram, and classify the risk values ​​as follows: Arrange all risk values ​​in ascending order; initialize the grading scheme by dividing all risk values ​​into one level; further divide each level in the current grading scheme into two sub-levels, and calculate the sum of the within-group variances for each division method; select the division method with the smallest sum of within-group variances as the new grading scheme; repeat the steps until the preset number of gradings is reached, and divide the corresponding risk levels according to the number of gradings. Based on the geographical location information of the vertices, all vertices are mapped to a geographic grid, and the regional risk index of each geographic grid is calculated. First, the real-time risk values ​​of all vertices in each geographic grid are multiplied together, and then Gaussian spatial smoothing is performed to transform the real-time risk values ​​of vertices in the same region in the geographic grid into regional risks. Based on the entities associated with the vertices, vertices related to endangered animal species are identified. The real-time risk value is converted into species risk by integrating the relevant vertices. The regional risk and species risk are combined by constructing a risk value distribution histogram and classifying the risk values ​​to determine the corresponding regional risk level and species risk level, thus completing the overall assessment.

9. A big data-based endangered animal crime risk assessment system, applied to the big data-based endangered animal crime risk assessment method as described in any one of claims 1-8, characterized in that: The system includes: The data processing and modeling module is used to unify multi-source heterogeneous data, complete the standardization transformation of text segmentation and non-text data, extract candidate features and construct a feature co-occurrence matrix, generate entities through point mutual information similarity calculation and hierarchical clustering, mine effective relationships between entities and calculate confidence, and finally construct a dynamic semantic association graph with complete timestamp attributes. The multidimensional functional graph construction module is used to classify edge types based on dynamic semantic association graphs and construct resource flow matrices, calculate the total resource input, total resource output and resource conversion coefficient of nodes, classify node functional roles by combining three-dimensional functional space and density clustering with niche differentiation strategy, analyze network topology, calculate the effective scale, connection efficiency, constraint degree and control capability of nodes, generate functional stability index by combining time window analysis, and generate multidimensional functional graph of criminal network. The assessment module is used to construct a time series of node activity intensity based on a multidimensional functional map of a crime network, perform causal relationship testing on nodes after smoothing by wavelet transform, construct a causal association network with time delay attributes, calculate the basic risk value by comprehensively considering multidimensional indicators, obtain the real-time risk value, classify the risk level of nodes, and further complete the spatial smoothing calculation of regional risk and the association assessment of the risk of endangered animal species.

Images