Re-crime dynamic multi-factor dissociation analysis method and device

Abstract

The invention discloses a re-crime dynamic multi-factor dissociation analysis method and device, and the method comprises the steps: setting an adversarial learning framework which comprises a static representation generator used for generating a static representation, an evolutionary representation generator used for generating an evolutionary representation, and a discriminator, inputting the dynamic complex crime network into a static representation generator and an evolutionary representation generator for feature learning, generating static representations and evolutionary representations of shadow line re-crimes at different time points, and constructing a training data set; constructing a loss function to perform adversarial decoupling training on the adversarial learning framework, and generating static representation and evolution representation corresponding to the whole dynamic complex crime network by the trained adversarial learning framework; constructing a solution ion space and carrying out subspace mapping decoupling; according to a decoupling result, key factors influencing re-crime are identified; the method has the advantages that support is provided for formulating targeted intervention measures, and the probability of heavy crimes is effectively reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of recidivism behavior analysis, and particularly relates to a dynamic multi-factor dissociation analysis method and device for recidivism. Background Art

[0002] With the increasing complexity of society and the intensification of population mobility, criminal behaviors have shown more diverse and complex characteristics. Identifying the key influencing factors of recidivism has become an important task in judicial and social governance. Currently, the existing analysis of influencing factors of recidivism mainly relies on statistical analysis, simple machine learning models, or identification methods of single factors. However, criminal behavior is not an isolated event but the result of the combined action of multiple factors. There is a high degree of correlation between the previous and subsequent criminal behaviors of criminals, which reflects potential behavior patterns or incentives and has important reference value for the analysis of influencing factors of future crimes.

[0003] Chinese Patent Publication No. CN108596386A discloses a method for predicting the probability of a prisoner's recidivism, including the following steps: (1) extracting the prisoner's data from a database, where the data includes the prisoner's personal file, criminal record, and influencing factors associated with recidivism; (2) cleaning the extracted data to obtain valid data and existing feature variables; (3) constructing features based on the valid data to obtain constructed feature variables; (4) selecting a sample set with data balance; (5) screening the existing feature variables and constructed feature variables of the sample set to obtain significant feature variables; (6) substituting the significant feature variables and valid data of the sample set into a classifier algorithm for fitting to obtain a classifier model; (7) inputting the prisoner's data into the classifier model to calculate the probability of the prisoner's recidivism. It does not fully explore the factor mechanism of recidivism, does not analyze the factors of recidivism and the attributes of these factors, so it is difficult to know the reasons and attributes of recidivism and difficult to formulate targeted intervention measures.

[0004] In summary, the existing methods fail to fully explore the factor mechanism of recidivism and ignore the coupling relationship between multiple factors such as previous and subsequent crimes, resulting in ineffective extraction of key information, especially insufficient in the dissociation of the factor mechanism of recidivism, making it difficult to deeply understand criminal behavior, difficult to capture the influence intensity of key attributes on individual behavior tendencies, and thus difficult to provide support for the formulation of targeted intervention measures and unable to effectively reduce the probability of recidivism. Summary of the Invention

[0005] The technical problem to be solved by the present invention is that the existing dynamic multi-factor dissociation analysis method for recidivism is difficult to capture the influence intensity of key attributes on individual behavior tendencies, and thus difficult to provide support for the formulation of targeted intervention measures and unable to effectively reduce the probability of recidivism.

[0006] The present invention solves the above technical problems through the following technical means: A dynamic multi-factor dissociation analysis method for recidivism, including:

[0007] S1. Set up an adversarial learning framework, which includes a static representation generator for generating static representations, an evolutionary representation generator for generating evolutionary representations, and a discriminator for discriminating whether the results generated by the static representation generator and the evolutionary representation generator are real and whether there is overlapping information between the two; input the dynamic complex crime network into the static representation generator and the evolutionary representation generator for feature learning to generate static representations and evolutionary representations of recidivism with hatching at different time points, and construct a training data set;

[0008] S2. Use the training data set to construct a loss function to perform adversarial decoupling training on the adversarial learning framework, stop training when the loss function value is the smallest, and obtain a trained adversarial learning framework. The trained adversarial learning framework generates static representations and evolutionary representations corresponding to the entire dynamic complex crime network;

[0009] S3. Construct a dissociation subspace and perform subspace mapping. Subspace mapping refers to decoupling the static representation and the evolutionary representation corresponding to the entire dynamic complex crime network;

[0010] S4. According to the decoupling results, identify the key factors affecting recidivism.

[0011] The present invention can effectively extract two types of multi-factor representations, static and evolutionary, from a complex crime network, and perform dissociation analysis to achieve the dissociation of different subspaces, enabling the model to independently identify and analyze the influence of various factors on behavior, helping to understand the influence intensity of key attributes on individual behavior tendencies, and finally, according to the decoupling results, identify the key factors affecting recidivism, thereby providing support for the formulation of targeted intervention measures and effectively reducing the probability of reoffending.

[0012] Further, S1 includes:

[0013] Input the dynamic complex crime network {G 1 , G 2 , …, G T}, where T is the total number of time segments, and G T is the crime network for the T time period. The crime network for each time period includes multiple nodes, and each node includes static information and dynamic factors. The static information includes basic information, criminal history, and criminal records, and the dynamic factors include psychological changes, social interactions, social support, and economic status;

[0014] Input the dynamic complex crime network into the static representation generator and the evolutionary representation generator for feature learning, and extract the first positive pair corresponding to time slice C1 and time slice C2 and the first negative pair represents the static representation of node v in time slice C1, represents the static representation of node v in time slice C2, represents the static representation of node u in time slice C1; extract the second positive pair and the second negative pair represents the node representation of node v in time period t and S v represents the static representation of node v, represents the evolving representation of node v in time period t. Since the static representation of a node is the same at any moment, while the evolving representation changes with time, the static representation here has no superscript t, and the evolving representation has superscript t, represents the node representation of node w adjacent to node v in time period t, represents the node representation of node u not adjacent to node v in time period t. Use the first positive pair, the first negative pair, the second positive pair, and the second negative pair to construct the training dataset.

[0015] Furthermore, S2 includes:

[0016] The loss of the static representation generator is where sim is a function for calculating the similarity between node representations, and τ is a coefficient for controlling the smoothness of the distribution, represents the expected value of the static representation of node v, and V is the set of nodes not adjacent to node v;

[0017] The loss of the evolving representation generator is where two nodes are connected to form an edge, and ε t represents the set of edges at time t, and U represents the set of other nodes that are not node v and its neighbor nodes;

[0018] The loss function of the discriminator is used to dissociate the overlapping information between the static representation and the evolving representation, and the formula is as follows

[0019]

[0020] where H represents the generator and d represents the discriminator, represents the expected value for all time series, represents the combination of the static and evolving representations of the true samples randomly sampled from the same node at time t, represents the feature representation of the fake samples randomly generated from different nodes at time t, represents the discriminator's pair of discrimination result;

[0021] The final loss function is where λ1, λ2, and λ3 are all set hyperparameters, is a regularization term, and v(H,D) represents a function of the interaction between the generator and the discriminator during the adversarial training process, that is, the loss function of the discriminator; when the loss function value converges or reaches a preset stopping criterion, the training is stopped to obtain a trained adversarial learning framework.

[0022] Furthermore, S3 includes:

[0023] Set the basic attributes, time-evolution attributes, psychological attributes, environmental attributes, and social attributes as different decoupled ion spaces, decouple the static representation and evolutionary representation corresponding to the entire dynamic complex crime network to obtain a decoupling result. Among them, the decoupling result of node v is:

[0024] Among them, S 基本属性 represents the static representation of node v belonging to the basic attributes, represents the evolutionary representation of node v belonging to the basic attributes at time t.

[0025] Furthermore, the basic attributes include the basic information of an individual, such as gender and native place; the time-evolution attributes reflect the characteristics of an individual changing over time, such as physical condition, employment status, and income level; the psychological attributes include emotional characteristics such as anxiety, depression, and impulse control; the environmental attributes include the physical and social environmental factors where an individual is located, such as family environment, safety of the residential community, educational resources, and social services; the social attributes describe the social network and relationship network of an individual, such as friends, family members, and work partners.

[0026] Furthermore, S4 includes:

[0027] Map the decoupling result to the probability space using the activation function Softmax to output the probability of each class. Each probability value corresponds to the recidivism probability under this decoupled ion space. The probability formula corresponding to the decoupling result of node v is as follows:

[0028]

[0029] Among them, represents the output probability of the static representation of node v belonging to the basic attributes, represents the output probability of the evolutionary representation of node v belonging to the basic attributes at time t;

[0030] Set a threshold for the output probability. If the probability of a certain attribute of the decoupling result of the calculated node is higher than this threshold, then this attribute is a key factor affecting recidivism.

[0031] The present invention also provides a recidivism dynamic multi-factor dissociation analysis device, including:

[0032] The dataset acquisition module is used to set up an adversarial learning framework. The adversarial learning framework includes a static representation generator for generating static representations, an evolutionary representation generator for generating evolutionary representations, and a discriminator. The discriminator is used to determine whether the results generated by the static representation generator and the evolutionary representation generator are real and whether there is overlapping information between the two. The dynamic complex crime network is input into the static representation generator and the evolutionary representation generator for feature learning to generate static representations and evolutionary representations of repeat offenses at different time points, and a training dataset is constructed.

[0033] The adversarial decoupling training module is used to use the training dataset to construct a loss function to perform adversarial decoupling training on the adversarial learning framework. When the value of the loss function is minimized, the training stops, and a trained adversarial learning framework is obtained. The trained adversarial learning framework generates static representations and evolutionary representations corresponding to the entire dynamic complex crime network.

[0034] The decoupling module is used to construct a decoupled subspace and perform subspace mapping. Subspace mapping refers to decoupling the static representations and evolutionary representations corresponding to the entire dynamic complex crime network.

[0035] The key factor acquisition module is used to identify key factors affecting repeat offenses according to the decoupling results.

[0036] Furthermore, the dataset acquisition module is also used for:

[0037] Input the dynamic complex crime network {G 1 ,G 2 ,…,G T}, where T is the total number of time segments, and G T is the crime network for the T time period. The crime network for each time period includes multiple nodes, and each node includes static information and dynamic factors. The static information includes basic information, criminal history, and criminal records, and the dynamic factors include psychological changes, social interactions, social support, and economic status.

[0038] The dynamic complex crime network is input into the static representation generator and the evolutionary representation generator for feature learning to extract the first positive pair corresponding to time slice C1 and time slice C2 and the first negative pair represents the static representation of node v at time slice C1, represents the static representation of node v at time slice C2, represents the static representation of node u at time slice C1; extract the second positive pair and the second negative pair represents the node representation of node v at time t and s v represents the static representation of node v, represents the evolutionary representation of node v at time t, The node representation of node w adjacent to node v at time period t The node representation of node u not adjacent to node v at time period t. A training data set is constructed using the first positive pair, the first negative pair, the second positive pair, and the second negative pair.

[0039] Furthermore, the adversarial decoupling training module is also used for:[[]]

[0040] The loss of the static representation generator is where sim is a function for calculating the similarity between node representations, and τ is a coefficient for controlling the smoothness of the distribution. denotes the expected value of the static representation of node v, and V is the set of nodes not adjacent to node v;

[0041] The loss of the evolutionary representation generator is where two nodes are connected to form an edge, and ε t denotes the set of edges at time t, and U denotes the set of other nodes that are not node v and its neighbor nodes;

[0042] The loss function of the discriminator is used to dissociate the overlapping information between the static representation and the evolutionary representation, and the formula is as follows

[0043]

[0044] where H represents the generator and D represents the discriminator. denotes the expected value for all time series. represents the combination of the static and evolutionary representations of the real samples randomly sampled from the same node at time t. represents the feature representation of the fake samples randomly generated from different nodes at time t. represents the discriminator's result for ;

[0045] The final loss function is where λ1, λ2, and λ3 are all set hyperparameters. is the regularization term, and V(H, D) represents the function of the interaction between the generator and the discriminator during the adversarial training process, that is, the loss function of the discriminator; when the loss function value converges or reaches the preset stopping criterion, the training is stopped, and the trained adversarial learning framework is obtained.

[0046] Furthermore, the decoupling module is also used for:[[]]

[0047] Set the basic attributes, time evolution attributes, psychological attributes, environmental attributes, and social attributes as different decoupling subspaces, and decouple the static representation and the evolutionary representation corresponding to the entire dynamic complex crime network to obtain the decoupling result. Among them, the decoupling result of node v is:[[]]

[0048] Among them, S 基本属性 represents the static representation belonging to the basic attributes in node v, and represents the evolutionary representation belonging to the basic attributes in node v at time t.

[0049] Furthermore, the basic attributes include the basic information of an individual, such as gender and native place; the time-evolution attributes reflect the characteristics of an individual changing over time, such as physical condition, employment status, and income level; the psychological attributes include emotional characteristics such as anxiety, depression, and impulse control; the environmental attributes include the physical and social environmental factors where the individual is located, such as family environment, safety of the residential community, educational resources, and social services; the social attributes describe the social network and relationship network of an individual, such as friends, family members, and work partners.

[0050] Furthermore, the key factor acquisition module is also used for:

[0051] mapping the decoupling result to the probability space using the activation function Softmax to output the probability of each class, and each probability value corresponds to the recidivism probability under this decoupled subspace. The probability formula corresponding to the decoupling result of node v is as follows:

[0052]

[0053] Among them, represents the output probability of the static representation belonging to the basic attributes in node v, and represents the output probability of the evolutionary representation belonging to the basic attributes in node v at time t;

[0054] Set a threshold for the output probability. If the probability of a certain attribute of the decoupling result of the calculated node is higher than this threshold, then this attribute is a key factor affecting recidivism.

[0055] The advantages of the present invention are as follows: The present invention can effectively extract two types of multi-factor representations, static and evolutionary, from a complex crime network, and perform dissociation analysis to achieve the dissociation of different subspaces, enabling the model to independently identify and analyze the influence of various factors on behavior, helping to understand the influence intensity of key attributes on individual behavior tendencies, and finally, based on the decoupling result, identifying the key factors affecting recidivism, thereby providing support for the formulation of targeted intervention measures and effectively reducing the recidivism probability. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a framework diagram of a dynamic multi-factor dissociation analysis method for recidivism disclosed in Embodiment 1 of the present invention;

[0057] Figure 2Flowchart of a method for dynamic multi-factor dissociation analysis of recidivism disclosed in Embodiment 1 of the present invention;

[0058] Figure 3 Structural schematic diagram of a device for dynamic multi-factor dissociation analysis of recidivism disclosed in Embodiment 2 of the present invention. Detailed implementation manners

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0060] Embodiment 1

[0061] As Figure 1 and Figure 2 shown, Embodiment 1 of the present invention provides a method for dynamic multi-factor dissociation analysis of recidivism, innovatively proposing a discrete-time dynamic graph dissociation learning framework combined with a causal graph. For the input dynamic complex crime network G(t), based on the adversarial learning framework, two types of representation generators, namely static and evolutionary, and a dissociation discriminator are proposed, and the causal graph is used to guide the dissociation learning, ultimately obtaining static representations and representations evolving over time to achieve factor dissociation analysis of the previous and subsequent crimes. This method mainly includes four steps. First, through temporal modeling and feature extraction of the dynamic complex crime network, multi-factor representations that can capture the factors influencing recidivism at different time nodes are generated. Second, dissociation techniques are applied to analyze multiple factors. Then, a dissociation subspace is constructed to divide the identified influencing factors according to their independence, so as to more clearly express and understand the roles of different factors in the analysis; a causal graph model is established to deeply understand the mechanism influencing recidivism by analyzing the causal relationships between factors. The relevant descriptions of the specific steps are as follows:

[0062] S1. Set up an adversarial learning framework, which includes a static representation generator for generating static representations, an evolutionary representation generator for generating evolutionary representations, and a discriminator. The discriminator is used to determine whether the results generated by the static representation generator and the evolutionary representation generator are real and whether there is overlapping information between the two; input the dynamic complex crime network into the static representation generator and the evolutionary representation generator for feature learning, generate static representations and evolutionary representations of the shadow recidivism at different time points, and construct a training dataset; the specific process is as follows:

[0063] By inputting the dynamic complex crime network {G 1 ,G2 , …, G T}, feature learning can be performed to generate multi-factor representations that affect recidivism at different time nodes. Considering that the factors affecting recidivism can be divided into static factors and evolutionary factors, the present invention designs two representation generators: a static representation generator G s and an evolutionary representation generator G e . The static representation generator G s is designed to identify the stable features inherent in an individual, which should remain consistent in any arbitrary local time segment, such as an individual's family background and educational environment, etc. To achieve this goal, the present invention designs an auxiliary task based on time-slice contrast learning, and performs feature generation through bidirectional Bernoulli sampling and structural time modeling, optimizing the representations of node v in time slice pairs C1 and C2, thereby ensuring the consistency of the same attributes. In this process, two time segments C1 and C2 of node v obtained through bidirectional Bernoulli sampling, bidirectional Bernoulli sampling selects samples with a certain probability, making the sample selection of each node more flexible. Compared with random sampling, it can more effectively capture the relationships and evolution characteristics between nodes, and make full use of the information in the graph. The representation of the node is learned through GCN In this embodiment, GCN learning refers to inputting the data corresponding to the time slice into the generator to obtain the corresponding node representation. and are regarded as the positive pairs in contrast learning while the representations of two different nodes v and u in time slice C1 are used as negative pairs in contrast learning

[0064] In the evolutionary representation generator G e , multiple dynamic factors affecting recidivism are modeled, including social support, economic status, employment situation, and behavior patterns, etc., which change over time. For this reason, the model adopts a structure with non-shared parameters but similar to the static generator, so as to learn static features and dynamic changes separately. The final node representation can be combined by the static representation and the dynamic representation, expressed as The present invention designs structural proximity contrast learning based on the graph structure, and regards the representations of node v and its adjacent node w and as the positive pairs in contrast learning while the representations of node v and its non-adjacent node u and are regarded as negative pairs

[0065] Based on the above principle, input the dynamic complex crime network {G 1 , G 2 , …, G T}, T is the total number of time segments, GT The criminal network for time period T, where the criminal network for each time period includes multiple nodes, and each node includes static information and dynamic factors. The static information includes basic data, criminal history, and criminal records, and the dynamic factors include psychological changes, social interactions, social support, and economic status; the dynamic complex criminal network is input into the static representation generator and the evolutionary representation generator for feature learning to extract the first positive pair corresponding to time slices C1 and C2 and the first negative pair represents the static representation of node v in time slice C1, represents the static representation of node v in time slice C2, represents the static representation of node u in time slice C1; extract the second positive pair and the second negative pair represents the node representation of node v in time period t and s v represents the static representation of node v, represents the evolutionary representation of node v in time period t, represents the node representation of node w adjacent to node v in time period t, represents the node representation of node u not adjacent to node v in time period t, and construct a training data set using the first positive pair, the first negative pair, the second positive pair, and the second negative pair.

[0066] S2. Using the training data set, construct a loss function to perform adversarial decoupling training on the adversarial learning framework, and stop training when the loss function value is minimized to obtain a trained adversarial learning framework. The trained adversarial learning framework generates the static representation and the evolutionary representation corresponding to the entire dynamic complex criminal network; the specific process is as follows:

[0067] The loss of the static representation generator is where sim is a function for calculating the similarity between node representations, and τ is a coefficient for controlling the smoothness of the distribution, represents the expected value of the static representation of node v; this function aims to learn representations where the positive pairs are closer and the negative pairs are farther apart. V is the set of nodes not adjacent to node v.

[0068] The loss of the evolutionary representation generator is where two nodes are connected to form an edge, and ε t represents the set of edges at time t, and U represents the set of other nodes that are not node v and its neighbor nodes; through the above method, the present invention effectively combines static and dynamic features and provides a more comprehensive representation learning mechanism for the analysis of recidivism influencing factors.

[0069] To effectively dissociate the overlapping information between the static representation and the evolutionary representation, the goal is to minimize the mutual information between them. This process can be expressed as:

[0070]

[0071] Among them, H represents a set consisting of two generators. I represents the mutual information of nodes. S represents the distribution of static representations of all nodes, while E represents the distribution of evolved representations of all nodes. Minimizing the mutual information is equivalent to minimizing the KL divergence between the joint distribution p(S, E) and the marginal distributions p(S)p(E). Therefore, the true data distribution p d = p(S, E) and the generated data distribution p g = p(S)p(E) are constructed, and the KL divergence between the two is minimized within the framework of a generative adversarial network. That is, the loss function formula of the discriminator is as follows:

[0072]

[0073] Among them, H represents the generator, which attempts to generate static and dynamic representations to make the generated representations as close as possible to the true representations of the nodes, that is, to make the generated samples look like they come from the same node The discriminator D is a multi-layer perceptron with dissociative perception ability and is responsible for judging whether these samples come from the true distribution. The discriminator not only judges the authenticity of the samples but also provides feedback on the overlapping information between the two parts of the representations of the generator, enhancing the dissociation between the static and evolved representations. denotes the expected value over all time series, denotes the combination of the static and evolved representations of true samples randomly sampled from the same node at time t, denotes the feature representation of fake samples randomly sampled from different nodes at time t, denotes the discriminant result of the discriminator for ;

[0074] The loss function of the discriminator is simplified to Loss(D) = -V(H, D), where V(H, D) represents the function of the interaction between the generator and the discriminator during the adversarial training process, measuring the performance of the generator and the discriminant ability of the discriminator, that is, the loss function formula of the discriminator recorded above.

[0075] The final loss function of the model is Among them, λ1, λ2, and λ3 are all set hyperparameters, is the regularization term used to prevent overfitting, is an L2 regularization term, where ω iDenote the \(i\)-th parameter in the model, and sum the squared values of each parameter in the model through a regularization term. When the loss function value converges or reaches a preset stopping criterion, stop the training to obtain a trained adversarial learning framework. By introducing an adversarial strategy, it aims to capture the evolution influencing factors of individuals at different times, enhance the understanding of the mechanism of criminal behavior, and thus provide theoretical support for formulating more effective prevention and rehabilitation measures.

[0076] S3. Construct a decoupled subspace and perform subspace mapping. Subspace mapping refers to decoupling the static representation and evolutionary representation corresponding to the entire dynamic complex crime network. The specific process is as follows:

[0077] The subspace of multi-factor dissociation is used to classify and independently process various influencing attributes, thereby avoiding interference between different attributes and facilitating the analysis of the independent effects and interaction effects of each attribute in different situations. The decoupled GCN model DisenGCN is adopted to ensure that the generated representations are composed of different decoupled subspaces, and the routing mechanism of the model can capture relevant factors in different aspects of nodes. For example, the DisenGCN model can decouple a general representation \(R\) into \(K\) representations closely related to \(K\) factors, \(R = [R_1, R_2, \ldots, R K . Specifically, the following main attributes are designed as different decoupled subspaces:

[0078] Basic attributes: Include the core basic information of individuals, such as gender and native place, etc. Such attributes are relatively stable and are usually not significantly affected by the external environment.

[0079] Temporal evolution: Reflect the characteristics of individuals changing over time, such as physical condition, employment status, and income level, etc. These attributes change significantly in the time series and directly affect the life stability and social integration degree of individuals.

[0080] Psychological attributes: Such as emotional characteristics like anxiety, depression, and impulse control. Such characteristics may change over time and are easily affected by environmental and situational factors.

[0081] Environmental attributes: Include the physical and social environmental factors where individuals are located, such as family environment, safety of the residential community, educational resources, and social services, etc., which affect the daily decisions and behavior choices of individuals.

[0082] Social attributes: Describe the social network and relationship network of individuals, such as friends, family members, and work partners, etc., which may affect the behavior choices of individuals and may evolve with social interactions.

[0083] Based on the construction of the decoupled subspace, decouple the static representation and evolutionary representation corresponding to the entire dynamic complex crime network to obtain a decoupling result. Among them, the decoupling result of node \(v\) is:

[0084] Among them, S 基本属性 represents the static representation of the basic attributes in node v, and represents the evolutionary representation of the basic attributes in node v at time t.

[0085] By dissociating different subspaces, the model can independently identify and analyze the impacts of various factors on behavior, which helps to understand the influence intensity of key attributes on individual behavior tendencies, and further provides support for the formulation of targeted intervention measures. For example, if it is found that psychological attributes have a significant impact on behavior tendencies, psychological counseling and grooming can be considered as priorities.

[0086] S4. According to the decoupling results, identify the key factors affecting recidivism. The specific process is as follows:

[0087] In the dynamic dissociation process, the causal graph plays a key role. By clarifying the causal relationships between different factors, the model can identify the key factors and potential interfering factors affecting recidivism. The causal graph details the various factors and their interactions in the tasks related to recidivism. The specific analysis is as follows:

[0088] Graph data G: The dataset that forms the basis of the study, containing information about individuals and their social environment, including demographic data, psychological data, individual criminal records, family situations, and community statistics, etc.

[0089] Causal factor C: The main factors directly leading to recidivism, which may include physical conditions, family structures, social relationships, and community support, etc.

[0090] Irrelevant factor T: Factors that, although present in the system, do not have a direct impact on recidivism. These are usually interference items or noise and need to be excluded during the modeling process.

[0091] Evolutionary factor E: Factors that change over time and may affect the probability of recidivism, such as income levels, social status, and family health conditions after release from prison, etc.

[0092] Static factor S: Factors that are relatively stable and do not change with time or environment, such as information about gender, native place, and family background, etc.

[0093] Representation towards R: Contains static and dynamic information about individuals and is used to identify and analyze criminal factors.

[0094] Key factor Y: The important factors identified that affect an individual's re - offending.

[0095] The following relationships exist among the elements in the causal graph: ① The graph data G affects the causal factor C and the irrelevant factor T; ② The causal factor C affects the static factor S and the evolutionary factor E; ③ The static factor S and the evolutionary factor E together determine the representation vector R; ④ The representation vector R determines the key factor Y, which is the main factor affecting reoffending. Based on the above relationships, the present invention identifies the causal factor C by collecting and analyzing the graph data G, and eliminates the influence of the factor T as much as possible. Then, the representation vector R is constructed using the static factor S and the evolutionary factor E, and the key factor Y is analyzed through the vector. This method aims to more accurately identify and analyze the factors affecting reoffending.

[0096] Specifically, the decoupling result is mapped to the probability space using the activation function Softmax to output the probability of each class. Each probability value corresponds to the reoffending probability in this decoupling space. The probability formula corresponding to the decoupling result of node v is as follows:

[0097]

[0098] in, represents the output probability of the static representation of the basic attributes in node v, represents the output probability of the evolution representation of the basic attribute in the node v at time t;

[0099] The output probability is set to a threshold (such as 0.5). If the probability of a certain attribute of the calculated node decoupling result is higher than the threshold, it can be considered that the attribute has a greater impact on reoffending, and the attribute is the key factor affecting reoffending. The corresponding static factor is the static factor S in the causal graph, and the corresponding evolutionary factor is the evolutionary factor E in the causal graph.

[0100] Based on the above description, for ease of understanding, the entire method process of the present invention is summarized as follows:

[0101] Enter the dynamic complex criminal network {G 1 , G 2 , ..., G T}, the static information (such as basic information, criminal history and criminal record) and dynamic factors (such as psychological changes, social interactions, social support and economic status) of each node in the network are represented as graph data. The time dimension t is the main axis.

[0102] Two time segments C1 and C2 of node v are obtained through two-way Bernoulli sampling. The model generates the static representation S and the evolutionary representation E of each node under each time slice through contrastive learning. The static representation ensures through time slice contrast that the learned representation remains consistent for the same node in different time slices and does not change over time. The evolutionary representation learns time-related dynamic information, such as income and employment status, through structural proximity contrastive learning. For example, the static representation of node 1 in the first time slice is (corresponding to family background and educational environment), and the evolutionary representation is (corresponding to income and employment status).

[0103] Under the adversarial learning framework, the generator H is composed of a static generator Gs and a dynamic generator Ge. The goal is to make the generated representation as close as possible to the true representation of the node. The dissociation discriminator D judges whether the generated representation really comes from the same node, tells the generator whether there is overlapping information in the two parts of the representation, and enhances its dissociation ability to ensure that the static and dynamic representations do not interfere with each other.

[0104] During the training process, the KL divergence between the static representation and the evolutionary representation gradually decreases, reflecting the improvement of the model's dissociation ability. The static generator, the evolutionary generator, and the discriminator are optimized through the loss function Loss. As the Loss function decreases, the model gradually converges. The positive pairs in the contrastively learned representation get closer, and the negative pairs are pulled apart, ensuring that the generated representation can effectively distinguish different nodes and learn a representation close enough to the true one, and the generated representation reduces the mutual information between the static and evolutionary information, enhancing the dissociation of the static and evolutionary factors. Finally, the representation R of the model is obtained as (S, E T ).

[0105] Combined with the actual situation, five dissociation subspaces are designed. The dissociation GCN model DisenGCN constructs dissociation subspaces for the static and evolutionary representations respectively, and the neighbor routing mechanism of this model ensures that the dissociated different subspaces are closely related to the relationships between factors under each subspace.

[0106] For each node representation R learned, the activation function Softmax is used to map the node representation to the probability space, and the probability under each subspace is output Each probability value corresponds to the recidivism probability under this dissociation subspace.

[0107] The constructed causal graph is used to clarify the causal relationships between different factors, enabling the model to identify the key factors and potential interfering factors affecting recidivism. Irrelevant factors are excluded through the dissociation process. When identifying key factors, a threshold (e.g., 0.5) is set for the output probability of each feature. If the probability of a certain feature is higher than this threshold, it can be considered that this feature has a greater impact on recidivism. For example: if the probability is higher than the threshold, those belonging to the static factor channel are classified as static factors, and those belonging to the evolutionary factor channel are classified as evolutionary factors.

[0108] Combined with the key factors of recidivism, a real-time recidivism risk assessment can be generated through the model. For individuals evaluated as high-risk, we can carry out personalized interventions based on specific key factors. For example: if it is shown that environmental attributes have a significant impact on individual behavior, then in the intervention measures, the family and community support mechanisms can be preferentially enhanced to ensure that individuals can obtain positive environmental impacts. Through the implementation of intervention strategies, the likelihood of an individual's recidivism can be effectively reduced, and the social security level can be improved.

[0109] The pseudocode of the entire method process is as follows:

[0110] Input: Complex crime network sequence G(t) = {G 1 , G 2 , …, G T}, where T is the total number of time segments: t = 1, …, T

[0111] Output: Static representation of crime Evolutionary representation of crime

[0112] 1: While not converged do

[0113] 2: Draw L from the uniform distribution U(1, T), and draw t from U(1, T - L) i , t j = t i + L - 1

[0114] 3: Calculate the sampling vector k 1 , k 2 , where the calculation formula for the sampling vector when the sampling step is m and the sampling time is t is with the sampling step m = 1, 2, …, L. p m = f(m; p; L) is the calculated probability of m, p ∈ (0, 1) is the success probability of each Bernoulli trial. g m is the perturbation term drawn from the Gumbel distribution Gumbel(0, 1), and τg is the temperature parameter. Through this formula, the model hopes to generate a p m and the perturbation term g mThe probability sampling vector. The temperature parameter τg controls the smoothness of the selection probability. When it is large, the sampling tends to be uniform.

[0115] 4: Calculate the snapshot index vector The index vector is the integration of the sampled vectors calculated above. The subscript j is the j-th sampling step, and the superscript i is the i-th sampling moment.

[0116] 5: Time slice

[0117] 6: Static multi-factor representation of crime: S 1 = GCN(C1), S 2 = GCN(C2). The capital S represents the set of static representations, and the superscript represents the corresponding time slice.

[0118] 7: Multi-factor representation of crime evolution: Here, the capital E also represents a set, which is the set of evolutionary representations, and the superscript represents the corresponding moment.

[0119] 8: Calculate and minimize the loss of the generator H, where

[0120] 9: Perform the following operations for the number of Epochs of the Discriminator

[0121] 10: Calculate and minimize the loss of the discriminator D,

[0122] 11: end while

[0123] 12: Final static representation:

[0124] 13: Calculate the disentangled subspaces DisenGCN(S) and DisenGCN(E 1 , E 2 , …, E T of the static representation S and the evolutionary representations. Obtain the representation of node v at each time slice t ).

[0125] 14: Map the node representation R in each subspace to the probability space, and use the activation function Softmax to output the probability of each subspace

[0126] 15: Causal graph-guided judgment of key factors for recidivism: Try to eliminate the influence of irrelevant factors through the disentanglement model. In the output results, if the probability is high, it indicates a key influencing factor for recidivism.

[0127] Through the above technical solutions, the present invention proposes a dynamic multi-factor dissociation analysis method for recidivism, which can solve the innovative technology of the dissociation analysis method for dynamic influencing factors of recidivism and improve the explanatory power of the model for the complex relationships among multiple factors. At the same time, through adversarial learning and contrastive learning strategies, the model is not only applicable to static scenarios but also can handle the dynamic changes of individual behaviors. The dynamic complex network model can comprehensively capture the multi-dimensional information of crime-related data, providing a scientific basis for judicial institutions to formulate intervention measures. It aims to capture the evolutionary influencing factors of individuals at different times, enhance the understanding of the mechanism of criminal behavior, and thus provide theoretical support for formulating more effective prevention and rehabilitation measures. By dissociating different subspaces, the model can independently identify and analyze the influence of various factors on behavior, helping to understand the influence intensity of key attributes on individual behavior tendencies, and further providing support for the formulation of targeted intervention measures.

[0128] Example 2

[0129] As Figure 3 shown, based on Example 1, the second embodiment of the present invention further provides a dynamic multi-factor dissociation analysis device for recidivism, including:

[0130] A dataset acquisition module, configured to set an adversarial learning framework. The adversarial learning framework includes a static representation generator for generating static representations, an evolutionary representation generator for generating evolutionary representations, and a discriminator. The discriminator is used to determine whether the results generated by the static representation generator and the evolutionary representation generator are real and whether there is overlapping information between the two; input the dynamic complex crime network into the static representation generator and the evolutionary representation generator for feature learning, generate static representations and evolutionary representations of shadow recidivism at different time points, and construct a training dataset;

[0131] An adversarial decoupling training module, configured to use the training dataset to construct a loss function to perform adversarial decoupling training on the adversarial learning framework, and stop training when the loss function value is the smallest to obtain a trained adversarial learning framework. The trained adversarial learning framework generates static representations and evolutionary representations corresponding to the entire dynamic complex crime network;

[0132] A decoupling module, configured to construct a decoupled subspace and perform subspace mapping. Subspace mapping refers to decoupling the static representations and evolutionary representations corresponding to the entire dynamic complex crime network;

[0133] A key factor acquisition module, configured to identify key factors affecting recidivism according to the decoupling results.

[0134] Specifically, the dataset acquisition module is further configured to:

[0135] Input the dynamic complex crime network {G 1 ,G2 , …, G T , where T is the total number of time segments, and G T is the crime network for the T time period. The crime network for each time period includes multiple nodes, and each node includes static information and dynamic factors. The static information includes basic data, criminal history, and criminal records, and the dynamic factors include psychological changes, social interactions, social support, and economic status;

[0136] Input the dynamic complex crime network into the static representation generator and the evolutionary representation generator for feature learning, and extract the first positive pair corresponding to time slices C1 and C2 and the first negative pair represents the static representation of node v in time slice C1, represents the static representation of node v in time slice C2, represents the static representation of node u in time slice C1; extract the second positive pair and the second negative pair represents the node representation of node v at time t and s v represents the static representation of node v, represents the evolutionary representation of node v at time t, represents the node representation of node w adjacent to node v at time t, represents the node representation of node u not adjacent to node v at time t. Use the first positive pair, the first negative pair, the second positive pair, and the second negative pair to construct a training data set.

[0137] More specifically, the adversarial decoupling training module is also used for:

[0138] The loss of the static representation generator is where sim is a function for calculating the similarity between node representations, and τ is a coefficient for controlling the smoothness of the distribution, represents the expected value of the static representation of node v, and V is the set of nodes not adjacent to node v;

[0139] The loss of the evolutionary representation generator is where two nodes are connected to form an edge, and ε t represents the set of edges at time t, and U represents the set of other nodes that are not node v and its neighbor nodes;

[0140] The loss function of the discriminator is used to dissociate the overlapping information between the static representation and the evolutionary representation, and the formula is as follows

[0141]

[0142] where H represents the generator and D represents the discriminator, represents the expected value for all time series, Combination of static and evolutionary representations of true samples randomly sampled from the same node at time t Feature representation of false samples randomly generated by sampling from different nodes at time t Indicates the discriminator's judgment on The discriminant result;

[0143] The final loss function is where λ1, λ2, and λ3 are all set hyperparameters, is the regularization term, and V(H, D) represents the function of the interaction between the generator and the discriminator during adversarial training, that is, the loss function of the discriminator; when the loss function value converges or reaches the preset stopping criterion, the training is stopped to obtain the trained adversarial learning framework.

[0144] More specifically, the decoupling module is also used for:

[0145] Set the basic attributes, time-evolution attributes, psychological attributes, environmental attributes, and social attributes as different decoupling subspaces to decouple the static representation and evolutionary representation corresponding to the entire dynamic complex crime network, and obtain the decoupling result. Among them, the decoupling result of node v is:

[0146] where S 基本属性 Represents the static representation belonging to the basic attributes in node v, Represents the evolutionary representation belonging to the basic attributes in node v at time t.

[0147] More specifically, the basic attributes include the basic information of the individual, such as gender and native place; the time-evolution attributes reflect the characteristics of the individual changing over time, such as physical condition, employment status, and income level; the psychological attributes include emotional characteristics such as anxiety, depression, and impulse control; the environmental attributes include the physical and social environmental factors where the individual is located, such as family environment, safety of the residential community, educational resources, and social services; the social attributes describe the social network and relationship network of the individual, such as friends, family members, and work partners.

[0148] More specifically, the key factor acquisition module is also used for:

[0149] Map the decoupling result to the probability space using the activation function Softmax to output the probability of each class. Each probability value corresponds to the recidivism probability under this decoupling subspace. The probability formula corresponding to the decoupling result of node v is as follows:

[0150]

[0151] where, Represents the output probability of the static representation belonging to the basic attributes in node v, The output probability of the evolutionary representation belonging to the basic attributes at node v at time t;

[0152] Set a threshold for the output probability. If the probability of a certain attribute of the decoupling result of the calculated node is higher than this threshold, then this attribute is a key factor affecting recidivism.

[0153] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic multi-factor dissociation analysis method for reoffending, characterized in that: include: S1. Setting an adversarial learning framework, the adversarial learning framework includes a static representation generator for generating a static representation, an evolving representation generator for generating an evolving representation, and a discriminator, the discriminator being used to discriminate whether the results generated by the static representation generator and the evolving representation generator are true and whether the two contain overlapping information; The dynamic complex crime network is input into the static representation generator and the evolutionary representation generator for feature learning, and the static representation and evolutionary representation of the lower shadow re-offending at different time points are generated to construct a training data set; S2. Using the training data set, construct a loss function to perform adversarial decoupling training on the adversarial learning framework. When the loss function value is the minimum, stop the training to obtain a trained adversarial learning framework. The trained adversarial learning framework generates a static representation and an evolutionary representation corresponding to the entire dynamic and complex criminal network. S3, construct the solution ion space and perform subspace mapping. Subspace mapping refers to decoupling the static representation and evolutionary representation corresponding to the entire dynamic complex criminal network; S4. Based on the decoupling results, identify the key factors that affect reoffending.

2. A dynamic multi-factor dissociation analysis method for reoffending according to claim 1, characterized in that: S1 includes: Enter the dynamic complex criminal network {G 1 ,G 2 ,…,G T }, T is the total number of time segments, G T The criminal network in each time period is composed of multiple nodes, each of which includes static information and dynamic factors. The static information includes basic information, criminal history and criminal records, and the dynamic factors include psychological changes, social interactions, social support and economic conditions. The dynamic complex crime network is input into the static representation generator and the evolving representation generator for feature learning, and the first positive pair corresponding to time slice C1 and time slice C2 is extracted. and the first negative pair represents the static representation of node v in time slice C1, represents the static representation of node v in time slice C2, Represents the static representation of node u in time slice C1; extracts the second positive pair and the second negative pair The node representation of node v in time period t and s v represents the static representation of node v, represents the evolution of node v in time period t, The node representation of node w adjacent to node v in time period t, The node representation of node u that is not adjacent to node v in time period t is represented, and the training data set is constructed using the first positive pair, the first negative pair, the second positive pair, and the second negative pair.

3. A dynamic multi-factor dissociation analysis method for reoffending according to claim 2, characterized in that S2 include: The loss of the static representation generator is Among them, sim is the function that calculates the similarity between node representations, τ is the coefficient that controls the smoothness of the distribution, represents the expected value of the static representation of node v, V is the set of nodes that are not adjacent to node v; The loss of the evolution representation generator is Among them, two nodes are connected to form an edge, ε t represents the set of edges at time t, and U represents the set of other nodes that are not node v and its neighbor nodes; The loss function of the discriminator is used to disentangle the overlapping information between the static representation and the evolving representation. The formula is as follows Among them, H represents the generator, D represents the discriminator, represents the expected value of all time series, represents the combination of static and evolving representations of real samples randomly sampled from the same node at time t, represents the feature representation of fake samples randomly sampled from different nodes at time t, Represents the discriminator pair The judgment result of The final loss function is Among them, λ1, λ2 and λ3 are all set hyperparameters. is a regularization term, V(H,D) represents the function of the interaction between the generator and the discriminator during adversarial training, that is, the loss function of the discriminator; when the loss function value converges, the training is stopped and a trained adversarial learning framework is obtained.

4. A dynamic multi-factor dissociation analysis method for reoffending according to claim 3, characterized in that: S3 includes: By setting basic attributes, time evolution attributes, psychological attributes, environmental attributes, and social attributes as different decomposition ion spaces, the static representation and evolution representation corresponding to the entire dynamic complex crime network are decoupled to obtain the decoupling result. The decoupling result of node v is: Among them, S 基本属性 Represents the static representation of the basic attributes in node v. Represents the evolution of basic attributes in node v at time t.

5. A dynamic multi-factor dissociation analysis method for reoffending according to claim 4, characterized in that: The basic attributes include basic information of the individual, such as gender and place of origin; time evolution attributes reflect the characteristics of the individual that change over time, such as physical condition, employment status and income level; psychological attributes include emotional characteristics such as anxiety, depression, and impulse control; environmental attributes include physical and social environmental factors in which the individual is located, such as family environment, safety of residential communities, educational resources and social services; social attributes describe the individual's social network and relationship network, such as friends, family members and work partners.

6. The dynamic multi-factor dissociation analysis method for reoffending according to claim 4 is characterized in that S4 include: The decoupling result is mapped to the probability space using the activation function Softmax to output the probability of each class. Each probability value corresponds to the reoffending probability in this decoupling space. The probability formula corresponding to the decoupling result of node v is as follows: in, represents the output probability of the static representation of the basic attributes in node v, represents the output probability of the evolution representation of the basic attribute in the node v at time t; A threshold is set for the output probability. If the probability of a certain attribute of the decoupling result of the calculated node is higher than the threshold, then the attribute is a key factor affecting reoffending.

7. A dynamic multi-factor dissociation analysis device for reoffending, characterized in that: include: A data set acquisition module, used to set up an adversarial learning framework, the adversarial learning framework includes a static representation generator for generating static representation, an evolving representation generator for generating evolving representation, and a discriminator, the discriminator is used to discriminate whether the results generated by the static representation generator and the evolving representation generator are true and whether there is overlapping information between the two; The dynamic complex crime network is input into the static representation generator and the evolutionary representation generator for feature learning, and the static representation and evolutionary representation of the lower shadow re-offending at different time points are generated to construct a training data set; The adversarial decoupling training module is used to use the training data set to construct a loss function to perform adversarial decoupling training on the adversarial learning framework. When the loss function value is minimized, the training is stopped to obtain a trained adversarial learning framework. The trained adversarial learning framework generates a static representation and an evolutionary representation corresponding to the entire dynamic and complex criminal network. The decoupling module is used to construct the decomposition space and perform subspace mapping. Subspace mapping refers to the decoupling of the static representation and the evolutionary representation corresponding to the entire dynamic and complex criminal network; The key factor acquisition module is used to identify the key factors that affect reoffending based on the decoupling results.

8. The dynamic multi-factor dissociation analysis device for reoffending according to claim 7, characterized in that: The dataset acquisition module is also used to: Enter the dynamic complex criminal network {G 1 ,G 2 ,…,G T }, T is the total number of time segments, G T The criminal network in each time period is composed of multiple nodes, each of which includes static information and dynamic factors. The static information includes basic information, criminal history and criminal records, and the dynamic factors include psychological changes, social interactions, social support and economic conditions. The dynamic complex crime network is input into the static representation generator and the evolving representation generator for feature learning, and the first positive pair corresponding to time slice C1 and time slice C2 is extracted. and the first negative pair represents the static representation of node v in time slice C1, represents the static representation of node v in time slice C2, Represents the static representation of node u in time slice C1; extracts the second positive pair and the second negative pair The node representation of node v in time period t and s v represents the static representation of node v, represents the evolution of node v in time period t, The node representation of node w adjacent to node v in time period t, The node representation of node u that is not adjacent to node v in time period t is represented, and the training data set is constructed using the first positive pair, the first negative pair, the second positive pair, and the second negative pair.

9. The dynamic multi-factor dissociation analysis device for reoffending according to claim 8, characterized in that: The adversarial disentangled training module is also used to: The loss of the static representation generator is Among them, sim is the function that calculates the similarity between node representations, τ is the coefficient that controls the smoothness of the distribution, represents the expected value of the static representation of node v, V is the set of nodes that are not adjacent to node v; The loss of the evolution representation generator is Among them, two nodes are connected to form an edge, ε t represents the set of edges at time t, and U represents the set of other nodes that are not node v and its neighbor nodes; The loss function of the discriminator is used to disentangle the overlapping information between the static representation and the evolving representation. The formula is as follows Among them, H represents the generator, D represents the discriminator, represents the expected value of all time series, represents the combination of static and evolving representations of real samples randomly sampled from the same node at time t, represents the feature representation of fake samples randomly sampled from different nodes at time t, Represents the discriminator pair The judgment result of The final loss function is Among them, λ1, λ2 and λ3 are all set hyperparameters. is a regularization term, V(H,D) represents the function of the interaction between the generator and the discriminator during adversarial training, that is, the loss function of the discriminator; when the loss function value converges, the training is stopped and a trained adversarial learning framework is obtained.

10. The dynamic multi-factor dissociation analysis device for reoffending according to claim 9, characterized in that: The decoupling module is also used to: By setting basic attributes, time evolution attributes, psychological attributes, environmental attributes, and social attributes as different decomposition ion spaces, the static representation and evolution representation corresponding to the entire dynamic complex crime network are decoupled to obtain the decoupling result. The decoupling result of node v is: Among them, S 基本属性 Represents the static representation of the basic attributes in node v. Represents the evolution of basic attributes in node v at time t.

Images