Bayesian network structure optimization method and device based on BRKGA algorithm

Abstract

The present application relates to a Bayesian network structure optimization method and device based on the BRKGA algorithm. The method comprises: generating a random key vector composed of multiple random numbers between [0,1]; using the NO TEARS algorithm as a decoder to decode the random key vector to obtain an initial solution to the Bayesian network structure optimization problem; calculating the fitness value of each initial solution according to the initial solution and a pre-set loss function; sorting the fitness values ​​to obtain an elite individual set; using a genetic algorithm to perform crossover and mutation operations according to the elite individual set, and outputting a global approximate solution to the Bayesian network structure optimization problem; the global approximate solution is a continuous value matrix; discretizing the continuous value matrix to obtain a global approximate adjacency matrix, and obtaining an optimized Bayesian network structure according to the global approximate adjacency matrix. The present method can improve the accuracy of Bayesian network structure learning.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular to a Bayesian network structure optimization method, device, computer equipment and storage medium based on a BRKGA algorithm. Background Art

[0002] In the existing medical diagnosis system and recommendation system, there is a lack of big data support when diagnosing diseases and making preference recommendations, and the accuracy of diagnosis and recommendation is low. With the development of technology in the field of machine learning, Bayesian networks have emerged. Bayesian networks are a general probabilistic graphical model. Its structure is represented by a directed acyclic graph (DAG), which can represent the correlation between random variables. It is a graphical representation of the joint probability distribution of random variables, so it has strong interpretability. In the medical diagnosis system, random variables can be data such as diseases and patient personal information. In the preference recommendation system, random variables can be data such as user access types and preferences. Bayesian networks can be used in medical diagnosis systems and recommendation systems because they can represent uncertain knowledge and perform reasoning and calculation expression. They can assist in disease diagnosis and preference recommendation.

[0003] The current use of Bayesian networks in assisting disease diagnosis and preference recommendations has low accuracy. Summary of the invention

[0004] Based on this, it is necessary to provide a Bayesian network structure optimization method and device based on the BRKGA algorithm that can improve the accuracy of disease diagnosis in response to the above technical problems.

[0005] A Bayesian network structure optimization method based on BRKGA algorithm, the method comprising:

[0006] When it is detected that the prediction accuracy of the Bayesian network in the task to be predicted is lower than a preset value, a Bayesian network structure optimization instruction is generated; the task to be predicted includes a disease prediction task;

[0007] According to the Bayesian network structure optimization instruction, the Bayesian network structure is obtained; the Bayesian network structure is a DAG graph; the DAG graph includes: a node set and an edge set; the nodes in the node set are random variables in the task to be predicted;

[0008] Generate a random key vector consisting of multiple random numbers between [0,1];

[0009] The NO TEARS algorithm is used as a decoder to decode the random key vector and obtain the initial solution to the Bayesian network structure optimization problem;

[0010] According to the initial solution and the pre-set loss function, calculate the fitness value of each initial solution;

[0011] Sort the fitness values ​​to obtain the elite individual set;

[0012] According to the elite individual set, a genetic algorithm is used to perform crossover and mutation operations, and a global approximate solution to the Bayesian network structure optimization problem is output; the global approximate solution is a continuous value matrix;

[0013] Discretize the continuous value matrix to obtain a global approximate adjacency matrix, and obtain the optimized Bayesian network structure based on the global approximate adjacency matrix;

[0014] The optimized Bayesian network structure is used to assist in disease prediction.

[0015] In one embodiment, the tasks to be predicted include a disease prediction task and a user preference recommendation task; the random variables refer to nodes in the Bayesian network, including disease and patient information; the disease and patient information are extracted from the medical system; the NO TEARS algorithm is used as a decoder to decode the random key vector to obtain an initial solution to the Bayesian network structure optimization problem, including:

[0016] According to the NO TEARS algorithm, the first optimization model is constructed as follows:

[0017]

[0018] h(W)=tr(e WοW )-d

[0019] Where l represents the loss function, W represents the real number matrix, X represents the data matrix, the data in the data matrix represents the data in the medical system, including patient information, n is the size of the random key vector; WοW represents the Hardmard product, e WοW represents the matrix index, h(W) represents the acyclic constraint in the directed acyclic graph, and d represents the data dimension;

[0020] The first optimization model is optimized by the augmented Lagrangian method with a quadratic penalty term, and the second optimization model is obtained as follows:

[0021]

[0022] According to the dual expression of the second optimization model, the decoder is obtained as:

[0023] D(α)=minL ρ (W,α)

[0024]

[0025]

[0026] in, is the initial solution, α is the Lagrange multiplier, and ρ is the step size;

[0027] The random key vector is input into the decoder as a real number matrix for decoding, and an initial solution to the Bayesian network structure optimization problem is obtained; the initial solution represents multiple Bayesian network structures for prediction.

[0028] In one of the embodiments, the fitness values ​​are sorted to obtain an elite individual set, including: sorting the fitness values ​​from small to large, and selecting the initial solutions corresponding to the top 20% after sorting to construct the elite individual set; the fitness value represents the matching success value between the Bayesian network structure in the initial solution and the data in the current task to be predicted.

[0029] In one of the embodiments, a genetic algorithm is used to perform crossover and mutation operations based on an elite individual set to output a global approximate solution to a Bayesian network structure optimization problem, including: a genetic algorithm is used to perform crossover and mutation operations based on an elite individual set, mutant individuals are generated through a mutation operator, the elite individual set is directly copied to the next generation and a crossover operation is performed to obtain crossover individuals, the mutant individuals and the crossover individuals are merged to form a new generation of individual sets, and the fitness value of each individual is calculated, and when a preset iteration condition is met, a global approximate solution to the Bayesian network structure optimization problem is obtained; the global approximate solution is a continuous value matrix; the continuous value matrix is ​​discretized to obtain a global approximate adjacency matrix, and the global approximate adjacency matrix represents a Bayesian network structure that can describe the relationship between patient information and diseases.

[0030] In one of the embodiments, the preset iteration condition is a maximum number of iterations.

[0031] In one embodiment, the continuous value matrix is ​​discretized to obtain a global approximate adjacency matrix, including: if the element in the continuous value matrix is ​​greater than a threshold, the element is set to 1, if the element in the continuous value matrix is ​​less than the threshold, the element is set to 0, and a global approximate adjacency matrix is ​​obtained; the global approximate adjacency matrix represents a Bayesian network structure that matches the data in the task to be predicted and can describe the relationship between patient information and disease.

[0032] In one embodiment, an optimized Bayesian network structure is obtained according to a global approximate adjacency matrix, including: calculating the trace in the adjacency matrix according to the global approximate adjacency matrix, if the trace in the adjacency matrix is ​​not equal to zero, randomly selecting two nodes corresponding to the non-zero elements of the adjacency matrix, deleting the edges corresponding to the nodes or reversing the edges between the two nodes until the trace of the adjacency matrix is ​​equal to zero; when the trace of the current matrix is ​​zero, calculating the nth power of the matrix, using the above method to continue to determine whether the trace of the matrix is ​​zero and make adjustments, recursively repeating the above process until n is equal to the dimension of the matrix, and obtaining the optimized Bayesian network structure; the optimized Bayesian network structure is the Bayesian network structure with the best prediction effect.

[0033] A Bayesian network structure optimization device based on BRKGA algorithm, the device comprising:

[0034] The detection module is used to generate a Bayesian network structure optimization instruction when it is detected that the Bayesian network prediction accuracy in the task to be predicted is lower than a preset value; the task to be predicted includes a disease prediction task; according to the Bayesian network structure optimization instruction, the Bayesian network structure is obtained; the Bayesian network structure is a DAG graph; the DAG graph includes: a node set and an edge set; the nodes in the node set are random variables in the task to be predicted;

[0035] A data acquisition module is used to generate a random key vector consisting of multiple random numbers between [0,1];

[0036] A decoding module is used to decode the random key vector using the NO TEARS algorithm as a decoder to obtain an initial solution to the Bayesian network structure optimization problem;

[0037] A fitness value acquisition module is used to calculate the fitness value of each initial solution according to the initial solution and a pre-set loss function; sort the fitness values ​​to obtain an elite individual set and a non-elite individual set;

[0038] The crossover and mutation module is used to perform crossover and mutation operations based on the elite individual set using a genetic algorithm to output a global approximate solution to the Bayesian network structure optimization problem; the global approximate solution is a continuous value matrix;

[0039] The discrete module is used to discretize the continuous value matrix to obtain the global approximate adjacency matrix, and obtain the optimized Bayesian network structure based on the global approximate adjacency matrix; the optimized Bayesian network structure is used to assist in disease prediction.

[0040] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0041] When it is detected that the prediction accuracy of the Bayesian network in the task to be predicted is lower than a preset value, a Bayesian network structure optimization instruction is generated; the task to be predicted includes a disease prediction task;

[0042] According to the Bayesian network structure optimization instruction, the Bayesian network structure is obtained; the Bayesian network structure is a DAG graph; the DAG graph includes: a node set and an edge set; the nodes in the node set are random variables in the task to be predicted;

[0043] Generate a random key vector consisting of multiple random numbers between [0,1];

[0044] The NO TEARS algorithm is used as a decoder to decode the random key vector and obtain the initial solution to the Bayesian network structure optimization problem;

[0045] According to the initial solution and the pre-set loss function, calculate the fitness value of each initial solution;

[0046] Sort the fitness values ​​to obtain the elite individual set;

[0047] According to the elite individual set, a genetic algorithm is used to perform crossover and mutation operations, and a global approximate solution to the Bayesian network structure optimization problem is output; the global approximate solution is a continuous value matrix;

[0048] Discretize the continuous value matrix to obtain the global approximate adjacency matrix, and obtain the optimized Bayesian network structure based on the global approximate adjacency matrix;

[0049] The optimized Bayesian network structure is used to assist in disease prediction.

[0050] The above-mentioned Bayesian network structure optimization method and device based on the BRKGA algorithm, BRKGA is a general framework for solving combinatorial optimization problems. First, when it is detected that the prediction accuracy of the Bayesian network in the task to be predicted is lower than the preset value, a Bayesian network structure optimization instruction is generated; the task to be predicted includes the disease prediction task and the user preference recommendation task, and the Bayesian network in the task to be predicted is used for Biqing prediction and user preference recommendation, wherein the nodes in the Bayesian network represent the disease and patient information; the disease and patient information can be extracted from the medical system; according to the Bayesian network structure optimization instruction, the Bayesian network structure is obtained; a random key vector composed of multiple random numbers between [0,1] is generated; the random key vector is the input of the decoder, and NO is added in BRKGA The TEARS algorithm is used as a decoder to decode the random key vector to obtain the initial solution to the Bayesian network structure optimization problem, which is also the local optimal solution. The initial solution represents multiple Bayesian network structures for prediction. The multiple Bayesian network structures for prediction are first calculated, and then the fitness value of each initial solution is calculated according to the pre-set loss function; all fitness values ​​are sorted from small to large, and the initial solutions corresponding to the top 20% after sorting are selected to construct an elite individual set. The fitness value represents the matching success value between the Bayesian network structure in the initial solution and the data in the current task to be predicted. By calculating the fitness value of the initial solution, a network that matches the data in the current task to be predicted is selected. The prediction accuracy can be improved by using the network that matches the data more closely in subsequent predictions; then, mutation and crossover operators are introduced in BRKGA to perform mutation and crossover operations on the elite individuals, and the Bayesian network structure optimization problem is output. A global approximate solution to the problem; thereby improving the traditional genetic algorithm, wherein the global approximate solution is a continuous value matrix; the continuous value matrix is ​​discretized to obtain a global approximate adjacency matrix, the global approximate adjacency matrix represents a Bayesian network structure that matches the data in the task to be predicted and can describe the relationship between patient information and disease, the global approximate adjacency matrix is ​​de-ringed to obtain an optimized Bayesian network structure, the de-ring operation is used to delete the rings in the global approximate matrix that do not match the data highly, and the obtained Bayesian network structure is the network structure that best matches the data in the prediction task and has the best prediction effect, the optimized Bayesian network structure is used to reflect the correlation of the data in the task to be predicted, so that when the optimized Bayesian network is used to assist doctors in making disease judgments, the disease and patient information are input into a computer, and according to the correlation between the disease and the patient information, the disease prediction can be better assisted. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A schematic diagram of a flow chart of a Bayesian network structure optimization method based on a BRKGA algorithm in one embodiment;

[0052] Figure 2is a structural block diagram of a Bayesian network structure optimization device based on the BRKGA algorithm in one embodiment;

[0053] Figure 3 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0055] In one embodiment, Figure 1 As shown, a Bayesian network structure optimization method based on the BRKGA algorithm is provided, comprising the following steps:

[0056] Step 102, when it is detected that the prediction accuracy of the Bayesian network in the task to be predicted is lower than a preset value, a Bayesian network structure optimization instruction is generated; the task to be predicted includes a disease prediction task; and the Bayesian network structure is obtained according to the Bayesian network structure optimization instruction.

[0057] The Bayesian network structure is a DAG graph; the DAG graph includes: a node set and an edge set; the nodes in the node set are random variables in the task to be predicted.

[0058] There can be many scenarios for the task to be predicted, for example, in the field of medical diagnosis, predicting the condition based on the patient's symptoms; or in the recommendation system, predicting the user's preferences based on the user's access type, etc.; the elements that can reflect the prediction task can be used as random variables in the prediction task. For example, in the field of medical diagnosis, random variables are some symptoms or patient information (such as cough, back pain, age, gender, etc.), DAG graph is a directed acyclic graph, the structure of the Bayesian network is represented by a directed acyclic graph, which is composed of nodes representing variables and directed edges connecting these nodes, that is, node sets and edge sets. Nodes represent random variables, and directed edges between nodes represent the dependency relationship between nodes (from parent nodes to their child nodes). Conditional probability is used to express the strength of the relationship, and if there is no node, prior probability is used to express information.

[0059] In a system where the task to be predicted is implemented using a Bayesian network, as time goes by or when the scenario changes significantly, the prediction accuracy of the Bayesian network stored in the system may be greatly reduced and cannot meet actual needs. Therefore, it is necessary to set a preset value in the system, that is, when it is detected that the prediction accuracy of the Bayesian network in the task to be predicted is lower than the preset value, the Bayesian network structure is optimized. For the system, a Bayesian network structure optimization instruction can be generated, and then the structure is optimized according to the Bayesian network structure optimization instruction.

[0060] Step 104: Generate a random key vector consisting of a plurality of random numbers between [0, 1].

[0061] The random key vector is a vector composed of multiple random numbers. The random number is any value between [0,1]. For example, the random number can be 0.3, 0.5, 0.8, 1, etc. There is no specific restriction here, and the random number generation rule shall prevail.

[0062] Step 106: Use the NO TEARS algorithm as a decoder to decode the random key vector to obtain an initial solution to the Bayesian network structure optimization problem.

[0063] The NO TEARS algorithm provides a method to re-examine the Bayesian network structure problem. This method first defines the loss function and proposes a new method to define acyclic constraints. According to a specific optimization problem, the NO TEARS algorithm is used as a decoder to decode the random key vector, and the decoded initial solution is used as the initial solution to the Bayesian network structure optimization problem. The initial solution represents multiple Bayesian network structures for prediction, and the Bayesian network structure contains the causal relationship between the various features in the medical system data.

[0064] Step 108: Calculate the fitness value of each initial solution based on the initial solution and the preset loss function, sort the fitness values, and obtain the elite individual set.

[0065] In the present invention, the fitness value is determined by calculating the value of the loss function of the initial solution, and the fitness values ​​are sorted from small to large, and a part of the initial solutions are selected as the elite individual set. Elite individuals refer to the individuals with the highest fitness searched so far in the evolution of the population. The fitness value represents the matching success value between the Bayesian network structure in the initial solution and the data in the current task to be predicted. By calculating the fitness value of the initial solution, a network with a relatively high match with the data in the current task to be predicted is selected, and the prediction accuracy can be improved in subsequent predictions by using a network with a relatively high match with the data.

[0066] Step 110, based on the elite individual set, a genetic algorithm is used to perform crossover and mutation operations, and a global approximate solution to the Bayesian network structure optimization problem is output.

[0067] The global approximate solution is a continuous-valued matrix.

[0068] Crossover and mutation operators are important components of genetic algorithms, and their purpose is to search the solution space and avoid the search process from falling into a local minimum. The basic principle of the crossover operator is to exchange several genes between two chromosomes according to certain strategies. The basic principle of the mutation operator is to randomly select a chromosome and change its value. In the present invention, crossover and mutation operations are performed on elite individuals to generate a set of individuals with higher fitness, which can be represented as a global approximate solution of the Bayesian network.

[0069] Step 112, discretize the continuous value matrix to obtain a global approximate adjacency matrix, and obtain an optimized Bayesian network structure based on the global approximate adjacency matrix; use the optimized Bayesian network structure to assist in disease prediction.

[0070] Since the result returned by the genetic algorithm is a continuous value matrix, not 0 or 1, it is necessary to convert the continuous value into a discrete value. Through discretization processing, the obtained matrix is ​​a global approximate adjacency matrix. The adjacency matrix cannot ensure the acyclicity of the directed graph. After de-ringing the global approximate adjacency matrix, an optimized Bayesian network structure can be obtained; the global approximate adjacency matrix represents a Bayesian network structure that matches the data in the task to be predicted and can describe the relationship between patient information and disease. The global approximate adjacency matrix is ​​de-ringed to obtain an optimized Bayesian network structure. The de-ringing operation is used to delete the rings in the global approximate matrix that do not match the data highly. The obtained Bayesian network structure is the network structure that best matches the data in the prediction task and has the best prediction effect. The optimized Bayesian network structure is used to reflect the correlation of the data in the task to be predicted, so that when the method is used to assist doctors in diagnosing symptoms, the symptoms and patient information are input into the computer. According to the correlation between the symptoms and patient information, the disease prediction can be better assisted.

[0071] The above-mentioned Bayesian network structure optimization method based on the BRKGA algorithm, BRKGA is a general framework for solving combinatorial optimization problems. First, when it is detected that the prediction accuracy of the Bayesian network in the task to be predicted is lower than the preset value, a Bayesian network structure optimization instruction is generated; the task to be predicted includes the disease prediction task and the user preference recommendation task, and the Bayesian network in the task to be predicted is used for Biqing prediction and user preference recommendation, wherein the nodes in the Bayesian network represent the disease and patient information; the disease and patient information can be extracted from the medical system; according to the Bayesian network structure optimization instruction, the Bayesian network structure is obtained; a random key vector composed of multiple random numbers between [0,1] is generated; the random key vector is the input of the decoder, and NO is added in BRKGA The TEARS algorithm is used as a decoder to decode the random key vector to obtain the initial solution to the Bayesian network structure optimization problem, which is also the local optimal solution. The initial solution represents multiple Bayesian network structures for prediction. The multiple Bayesian network structures for prediction are first calculated, and then the fitness value of each initial solution is calculated according to the pre-set loss function; all fitness values ​​are sorted from small to large, and the initial solutions corresponding to the top 20% after sorting are selected to construct an elite individual set. The fitness value represents the matching success value between the Bayesian network structure in the initial solution and the data in the current task to be predicted. By calculating the fitness value of the initial solution, a network that matches the data in the current task to be predicted is selected. The prediction accuracy can be improved by using the network that matches the data more closely in subsequent predictions; then, mutation and crossover operators are introduced in BRKGA to perform mutation and crossover operations on the elite individuals, and the Bayesian network structure optimization problem is output. A global approximate solution to the problem; thereby improving the traditional genetic algorithm, wherein the global approximate solution is a continuous value matrix; the continuous value matrix is ​​discretized to obtain a global approximate adjacency matrix, the global approximate adjacency matrix represents a Bayesian network structure that matches the data in the task to be predicted and can describe the relationship between patient information and disease, the global approximate adjacency matrix is ​​de-ringed to obtain an optimized Bayesian network structure, the de-ring operation is used to delete the rings in the global approximate matrix that do not match the data highly, and the obtained Bayesian network structure is the network structure that best matches the data in the prediction task and has the best prediction effect, the optimized Bayesian network structure is used to reflect the correlation of the data in the task to be predicted, so that when the optimized Bayesian network is used to assist doctors in making disease judgments, the disease and patient information are input into a computer, and according to the correlation between the disease and the patient information, the disease prediction can be better assisted.

[0072] In one embodiment, the tasks to be predicted include a disease prediction task and a user preference recommendation task; the random variables refer to nodes in the Bayesian network, including disease and patient information; the disease and patient information are extracted from the medical system; the NO TEARS algorithm is used as a decoder to decode the random key vector to obtain an initial solution to the Bayesian network structure optimization problem, including:

[0073] According to the NO TEARS algorithm, the first optimization model is constructed as follows:

[0074]

[0075] h(W)=tr(e WοW )-d

[0076] Where l represents the loss function, W represents the real number matrix, X represents the data matrix, the data in the data matrix represents the data in the medical system, including patient information, n is the size of the random key vector; WοW represents the Hardmard product, e WοW represents the matrix index, h(W) represents the acyclic constraint in the directed acyclic graph, and d represents the data dimension;

[0077] The first optimization model is optimized by the augmented Lagrangian method with a quadratic penalty term, and the second optimization model is obtained as follows:

[0078]

[0079] According to the dual expression of the second optimization model, the decoder is obtained as:

[0080] D(α)=minL ρ (W,α)

[0081]

[0082]

[0083] in, is the initial solution, α is the Lagrange multiplier, and ρ is the step size;

[0084] Among them, the real matrix means that all elements in the matrix are real numbers, the data matrix refers to the matrix composed of all patient information, and the real matrix W can be used as the adjacency matrix A(W)∈{0,1} d×d The specific meaning of this function is: if the elements in W are not 0, then set the elements in A(W) to 1; otherwise, set the elements in A(W) to 0.

[0085] The random key vector is input as a real number matrix into the decoder for decoding, and the initial solution of the Bayesian network structure optimization problem is obtained. The NO TEARS algorithm is an optimization algorithm, but the Bayesian structure optimization problem is a non-convex optimization problem, so the exact solution cannot be found through the above process. Therefore, the BRKGA algorithm is needed to find more possible solutions. The above process is regarded as a decoder; the initial solution represents multiple Bayesian network structures for prediction.

[0086] The obtained initial solution is input into the preset loss function to calculate the fitness value of the initial solution, and the elite individual set and the non-elite individual set are obtained according to the size of the fitness value.

[0087] In one of the embodiments, the fitness values ​​are sorted to obtain an elite individual set, including: sorting the fitness values ​​from small to large, and selecting the initial solutions corresponding to the top 20% after sorting to construct an elite individual set. The elite individuals can be used as initial data in a genetic algorithm and as input to the genetic algorithm, and new offspring will also have better fitness values ​​when inheriting elements in the elite individual set. Specifically, when setting the number of mutant individuals, the number of mutant individuals is set to be the same as the number of elite individuals in this embodiment; the fitness value represents the matching success value of the Bayesian network structure in the initial solution and the data in the current task to be predicted. When the Bayesian network structure with a higher fitness value is used for disease prediction, the data matching degree of patient information and symptoms is higher, which can improve the accuracy of disease prediction.

[0088] In one of the embodiments, a genetic algorithm is used to perform crossover and mutation operations based on an elite individual set to output a global approximate solution to a Bayesian network structure optimization problem, including: a genetic algorithm is used to perform crossover and mutation operations based on an elite individual set, mutant individuals are generated through a mutation operator, the elite individual set is directly copied to the next generation and a crossover operation is performed to obtain crossover individuals, the mutant individuals and the crossover individuals are merged to form a new generation of individual sets, and the fitness value of each individual is calculated, and when a preset iteration condition is met, a global approximate solution to the Bayesian network structure optimization problem is obtained; the global approximate solution is a continuous value matrix; the continuous value matrix is ​​discretized to obtain a global approximate adjacency matrix, and the global approximate adjacency matrix represents a Bayesian network structure that can describe the relationship between patient information and diseases.

[0089] A genetic algorithm is used to perform crossover and mutation operations on the elite individual set. The fitness value of the individual obtained in each operation is better than the previous one. This process is iterated continuously. When the preset iteration conditions are reached, the set of individuals with the best fitness is obtained. These individuals can be used as the global approximate solution to the Bayesian network structure optimization problem.

[0090] In one embodiment, the preset iteration condition is a maximum number of iterations, which is a value predefined by a user.

[0091] In one embodiment, the continuous value matrix is ​​discretized to obtain a global approximate adjacency matrix, including: if the element in the continuous value matrix is ​​greater than a threshold, the element is set to 1, if the element in the continuous value matrix is ​​less than the threshold, the element is set to 0, to obtain a global approximate adjacency matrix.

[0092] The continuous value matrix is ​​the global approximate solution of the Bayesian network. The continuous value matrix is ​​discretized and the continuous values ​​are converted into discrete values. The obtained discrete values ​​are the global approximate adjacency matrix. The global approximate adjacency matrix represents a Bayesian network structure that matches the data in the task to be predicted and can describe the relationship between patient information and disease.

[0093] In one embodiment, an optimized Bayesian network structure is obtained according to the global approximate adjacency matrix, including:

[0094] According to the global approximate adjacency matrix, the trace in the adjacency matrix is ​​calculated. If the trace in the adjacency matrix is ​​not equal to zero, two nodes corresponding to the non-zero elements of the adjacency matrix are randomly selected, and the edges corresponding to the nodes are deleted or the edges between the two nodes are reversed until the trace of the adjacency matrix is ​​equal to zero; when the trace of the current matrix is ​​zero, the nth power of the matrix is ​​calculated, and the above method is used to continue to determine whether the trace of the matrix is ​​zero and make adjustments, and the above process is recursively repeated until n is equal to the dimension of the matrix to obtain the optimized Bayesian network structure; the optimized Bayesian network structure is the Bayesian network structure with the best prediction effect.

[0095] The elements in the adjacency matrix obtained after discretization are either 0 or 1. However, this still cannot ensure the acyclicity of the directed graph. Therefore, it is necessary to design an algorithm to delete the loops contained in the graph. In order to ensure the acyclicity of the Bayesian network, an important property of DAG is used here: the adjacency matrix a ij The nth power means that it takes n steps to go from i to j. The first step in designing the algorithm in this embodiment is to calculate the trace of the adjacency matrix. If the trace of the matrix is ​​not equal to 0, two nodes corresponding to the non-zero elements of the adjacency matrix are randomly selected, and the edges corresponding to the nodes are deleted or the edges between the two nodes are reversed until the trace of the adjacency matrix is ​​equal to zero; when the trace of the current matrix is ​​zero, the nth power of the matrix is ​​calculated, and the above method is used to continue to determine whether the trace of the matrix is ​​zero and make adjustments, and the above process is recursively repeated until n is equal to the dimension of the matrix. The beneficial effects of the above embodiments of the present invention are described below through a specific experimental process:

[0096] The experiments in this paper are implemented in R language, and the networks used in the experiments are downloaded from the Bayesian network repository. The detailed information of the networks used in this paper is listed in Table 1. Samples are taken from these networks with sample sizes of 100, 500, and 1000. The population size p is set to 50, and the maximum number of iterations is 100. The sizes of elite individuals and mutant individuals are pm = 0.2*p and p respectively. e =0.2*p,ρ e Set to 0.7. All results are the average of 10 experiments. SHD (structural Hamming distance) is used to evaluate the performance of the algorithm. The smaller the SHD value, the more similar the two directed acyclic graphs are. First, the experimental results of the algorithms performed on these Bayesian networks are listed in Table 1, with sample sizes of 100, 500 and 1000, respectively. The BRKGA algorithm is compared with the original NOTEARS algorithm and the standard genetic algorithm. For NOT TEARS, the maximum iteration t=100, the tolerance ∈=1e-8, and the minimum forward rate c=0.25 are set. The population size of the genetic algorithm is 50, the crossover probability is 0.9, and the mutation probability is 0.1. The experimental results can be seen in Table 1, and the smallest SHD value in the table is bolded. It can be seen from Table 1 that SHD increases with the number of edges in the network. Except for the CANCER and EARTHQUAKE networks, the BRKGA algorithm in the present invention achieves a smaller SHD value than other algorithms, indicating that the DAG obtained by the BRKGA algorithm in the present invention is closer to the real network structure than the two algorithms. The table shows that NO TEARS obtains smaller SHD values ​​on the CANCER network and the EARTHQUAKE network than the standard GA and BRKGA. The table also shows that NO TEARS performs better than GA on the CANCER, EARTHQUAKE, ASIA, CHILD, and ALARM networks. Similarly, NO TEARS's SHD values ​​are better than BRKGA on the CANCER network, the EARTHQUAKE network, and the HAILFINDER network with a sample size of 1000. Compared with the standard GA, the BRKGA algorithm in the present invention has better performance on all networks. When the network is small, NO TEARS can achieve better SHD values ​​than the algorithm of the present invention because NO TEARS has found a DAG close to the global exact solution. However, as the number of nodes increases, the solution obtained by NO TEARS is increasingly different from the solution obtained by the algorithm of the present invention. The BRKGA algorithm in the present invention can obtain a continuous solution and discretize it through an indicator function. Therefore, even if the optimal solution obtained by the BRKGA algorithm in the present invention achieves a smaller loss value, its adjacency matrix may be the same. Although BRKGA only adds a decoder and generates random key vectors, it significantly enhances the performance of standard GA.

[0097] Table 1 Experimental results

[0098]

[0099] In one embodiment, a medical diagnosis system is constructed using a Bayesian network, a patient information set X is input into the medical diagnosis system, a first optimization model containing the patient information is constructed, a second optimization model containing the patient information is obtained by optimizing the first optimization model, a decoder is constructed according to the second optimization model, and decoding is performed using a random key vector to obtain an initial solution to the Bayesian network structure optimization problem containing the patient information, the initial solution represents multiple Bayesian network structures for prediction, the Bayesian network structure includes the causal relationship between the features in the medical system data, and the fitness value of each initial solution is calculated according to the initial solution and a preset loss function; the fitness value represents the matching success value of the Bayesian network structure in the initial solution with the data in the current task to be predicted, and by calculating the fitness value of the initial solution, a network that matches the data in the current task to be predicted is selected, and the network that matches the data more closely can be used to improve the prediction accuracy in subsequent predictions. The fitness values ​​are sorted to obtain an elite individual set; based on the elite individual set, a genetic algorithm is used to perform crossover and mutation operations to output a global approximate solution to the Bayesian network structure optimization problem; the global approximate solution is a continuous value matrix; the continuous value matrix is ​​discretized to obtain a global approximate adjacency matrix, the global approximate adjacency matrix represents a Bayesian network structure that matches the data in the task to be predicted and can describe the relationship between patient information and disease, the global approximate adjacency matrix is ​​de-ringed to obtain an optimized Bayesian network structure, the de-ringing operation is used to delete the rings in the global approximate matrix that do not match the data highly, and the obtained Bayesian network structure is the network structure that best matches the data in the prediction task and has the best prediction effect, the optimized Bayesian network structure is used to reflect the correlation of the data in the task to be predicted, so that when the method is used to assist doctors in making disease judgments, the disease and patient information are input into the computer, and according to the correlation between the disease and the patient information, the disease prediction can be better assisted.

[0100] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but the steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of the steps, and the steps can be executed in other orders. Moreover, Figure 1At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0101] In one embodiment, Figure 2 As shown, a Bayesian network structure optimization device based on the BRKGA algorithm is provided, comprising: a detection module 201, a data acquisition module 202, a decoding module 203, a fitness value acquisition module 204, a crossover and mutation module 205 and a discrete module 206, wherein:

[0102] The detection module 201 is used to generate a Bayesian network structure optimization instruction when it is detected that the Bayesian network prediction accuracy in the task to be predicted is lower than a preset value; the task to be predicted includes a disease prediction task; according to the Bayesian network structure optimization instruction, the Bayesian network structure is obtained; the Bayesian network structure is a DAG graph; the DAG graph includes: a node set and an edge set; the nodes in the node set are random variables in the task to be predicted.

[0103] The data acquisition module 202 is used to generate a random key vector composed of a plurality of random numbers between [0, 1].

[0104] The decoding module 203 is used to decode the random key vector using the NO TEARS algorithm as a decoder to obtain an initial solution to the Bayesian network structure optimization problem.

[0105] The fitness value acquisition module 204 is used to calculate the fitness value of each initial solution according to the initial solution and a preset loss function; sort the fitness values ​​to obtain an elite individual set.

[0106] The crossover and mutation module 205 is used to perform crossover and mutation operations using a genetic algorithm based on the elite individual set, and output a global approximate solution to the Bayesian network structure optimization problem; the global approximate solution is a continuous value matrix.

[0107] The discretization module 206 is used to discretize the continuous value matrix to obtain a global approximate adjacency matrix, and obtain an optimized Bayesian network structure based on the global approximate adjacency matrix; the optimized Bayesian network structure is used to assist in disease prediction.

[0108] In one embodiment, the decoding module 203 is further used to decode the random key vector using the NO TEARS algorithm as a decoder to obtain an initial solution to the Bayesian network structure optimization problem, including:

[0109] According to the NO TEARS algorithm, the first optimization model is constructed as follows:

[0110]

[0111] h(W)=tr(e WοW )-d

[0112] Where l represents the loss function, W represents the real number matrix, X represents the data matrix, the data in the data matrix represents the data in the medical system, including patient information, n is the size of the random key vector; WοW represents the Hardmard product, e WοW represents the matrix index, h(W) represents the acyclic constraint in the directed acyclic graph, and d represents the data dimension;

[0113] The first optimization model is optimized by the augmented Lagrangian method with a quadratic penalty term, and the second optimization model is obtained as follows:

[0114]

[0115] According to the dual expression of the second optimization model, the decoder is obtained as:

[0116] D(α)=minL ρ (W,α)

[0117]

[0118]

[0119] in, is the initial solution, α is the Lagrange multiplier, and ρ is the step size;

[0120] The random key vector is input into the decoder as a real number matrix for decoding, and an initial solution to the Bayesian network structure optimization problem is obtained; the initial solution represents multiple Bayesian network structures for prediction.

[0121] In one embodiment, the fitness value acquisition module 204 is also used to sort the fitness values ​​to obtain an elite individual set, including: sorting the fitness values ​​from small to large, and selecting the initial solutions corresponding to the top 20% after sorting to construct an elite individual set; the fitness value represents the matching success value between the Bayesian network structure in the initial solution and the data in the current task to be predicted.

[0122] In one embodiment, the crossover and mutation module 205 is also used to perform crossover and mutation operations based on the elite individual set using a genetic algorithm to output a global approximate solution to the Bayesian network structure optimization problem, including: performing crossover and mutation operations based on the elite individual set using a genetic algorithm, generating mutant individuals through a mutation operator, the elite individual set is directly copied to the next generation and a crossover operation is performed to obtain crossover individuals, the mutant individuals and the crossover individuals are merged to form a new generation of individual sets, and the fitness value of each individual is calculated, and when the preset iteration condition is met, a global approximate solution to the Bayesian network structure optimization problem is obtained; the global approximate solution is a continuous value matrix; the continuous value matrix is ​​discretized to obtain a global approximate adjacency matrix, and the global approximate adjacency matrix represents a Bayesian network structure that can describe the relationship between patient information and diseases.

[0123] In another embodiment, the crossover and mutation module 205 is further configured to preset an iteration condition as a maximum number of iterations.

[0124] In one embodiment, the discretization module 206 is also used to discretize the continuous value matrix to obtain a global approximate adjacency matrix, including: if the element in the continuous value matrix is ​​greater than a threshold, the element is set to 1, if the element in the continuous value matrix is ​​less than the threshold, the element is set to 0, and a global approximate adjacency matrix is ​​obtained; the global approximate adjacency matrix represents a Bayesian network structure that matches the data in the task to be predicted and can describe the relationship between patient information and disease.

[0125] In another embodiment, the discrete module 206 is also used to obtain an optimized Bayesian network structure based on the global approximate adjacency matrix, including: calculating the trace in the adjacency matrix based on the global approximate adjacency matrix, if the trace in the adjacency matrix is ​​not equal to zero, randomly selecting two nodes corresponding to the non-zero elements of the adjacency matrix, deleting the edges corresponding to the nodes or reversing the edges between the two nodes until the trace of the adjacency matrix is ​​equal to zero; when the trace of the current matrix is ​​zero, calculating the nth power of the matrix, using the above method to continue to determine whether the trace of the matrix is ​​zero and make adjustments, recursively repeating the above process until n is equal to the dimension of the matrix, and obtaining the optimized Bayesian network structure; the optimized Bayesian network structure is the Bayesian network structure with the best prediction effect.

[0126] For the specific definition of the Bayesian network structure optimization device based on the BRKGA algorithm, please refer to the definition of the Bayesian network structure optimization method based on the BRKGA algorithm above, which will not be repeated here. Each module in the above-mentioned Bayesian network structure optimization device based on the BRKGA algorithm can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0127] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a Bayesian network structure optimization method based on the BRKGA algorithm is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a key, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0128] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0129] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method in the above embodiment when executing the computer program.

[0130] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method in the above embodiment are implemented.

[0131] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0132] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0133] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A Bayesian network structure optimization method based on BRKGA algorithm, characterized in that: The method comprises: When it is detected that the prediction accuracy of the Bayesian network in the task to be predicted is lower than a preset value, a Bayesian network structure optimization instruction is generated; the task to be predicted includes a disease prediction task; According to the Bayesian network structure optimization instruction, a Bayesian network structure is obtained; the Bayesian network structure is a DAG graph; the DAG graph includes: a node set and an edge set; the nodes in the node set are random variables in the task to be predicted; the random variables refer to nodes in the Bayesian network, including symptoms and patient information; the symptoms and patient information are extracted from a medical system; Generate a random key vector consisting of multiple random numbers between [0,1]; Using the NO TEARS algorithm as a decoder to decode the random key vector, and obtaining an initial solution to the Bayesian network structure optimization problem; Calculating the fitness value of each of the initial solutions according to the initial solutions and a preset loss function; Sorting the fitness values ​​to obtain an elite individual set; According to the elite individual set, a genetic algorithm is used to perform crossover and mutation operations, and a global approximate solution to the Bayesian network structure optimization problem is output; the global approximate solution is a continuous value matrix; Discretizing the continuous value matrix to obtain a global approximate adjacency matrix, and obtaining an optimized Bayesian network structure according to the global approximate adjacency matrix; the global approximate adjacency matrix represents a Bayesian network structure capable of describing the relationship between patient information and diseases; The optimized Bayesian network structure is used to assist in disease prediction.

2. The method according to claim 1, characterized in that The task to be predicted includes a user preference recommendation task; The NO TEARS algorithm is used as a decoder to decode the random key vector to obtain an initial solution to the Bayesian network structure optimization problem, including: According to the NO TEARS algorithm, the first optimization model is constructed as follows: Where l represents the loss function, W represents the real number matrix, X represents the data matrix, the data in the data matrix represents the data in the medical system, including patient information, and n is the size of the random key vector; represents the Hardmard product, e WοW represents the matrix index, h(W) represents the acyclic constraint in the directed acyclic graph, and d represents the data dimension; The first optimization model is optimized by the augmented Lagrangian method with a quadratic penalty term, and the second optimization model is obtained as follows: According to the dual expression of the second optimization model, the decoder is obtained as: D(a)=minL ρ (W,a) in, is the initial solution, α is the Lagrange multiplier, and ρ is the step size; The random key vector is input as a real number matrix into the decoder for decoding to obtain an initial solution to the Bayesian network structure optimization problem; the initial solution represents multiple Bayesian network structures for prediction.

3. The method according to claim 1, characterized in that The fitness values ​​are sorted to obtain an elite individual set, including: The fitness values ​​are sorted from small to large, and the initial solutions corresponding to the top 20% after sorting are selected to construct an elite individual set; the fitness value represents the matching success value between the Bayesian network structure in the initial solution and the data in the current task to be predicted.

4. The method according to claim 1, characterized in that: According to the elite individual set, a genetic algorithm is used to perform crossover and mutation operations to output a global approximate solution to the Bayesian network structure optimization problem, including: According to the elite individual set, a genetic algorithm is used to perform crossover and mutation operations, and mutant individuals are generated through a mutation operator. The elite individual set is directly copied to the next generation and a crossover operation is performed to obtain crossover individuals. The mutant individuals and the crossover individuals are merged to form a new generation of individual sets, and the fitness value of each individual is calculated. When the preset iteration condition is reached, a global approximate solution to the Bayesian network structure optimization problem is obtained; the global approximate solution is a continuous value matrix; the continuous value matrix is ​​discretized to obtain a global approximate adjacency matrix.

5. The method according to claim 4, characterized in that The preset iteration condition is the maximum number of iterations.

6. The method according to claim 1, characterized in that Discretizing the continuous value matrix to obtain a global approximate adjacency matrix, including: If the element in the continuous value matrix is ​​greater than the threshold, the element is set to 1; if the element in the continuous value matrix is ​​less than the threshold, the element is set to 0 to obtain a global approximate adjacency matrix; the global approximate adjacency matrix represents a Bayesian network structure that matches the data in the task to be predicted and can describe the disease relationship.

7. The method according to claim 1, characterized in that According to the global approximate adjacency matrix, an optimized Bayesian network structure is obtained, including: According to the global approximate adjacency matrix, the trace in the adjacency matrix is ​​calculated. If the trace in the adjacency matrix is ​​not equal to zero, two nodes corresponding to the non-zero elements of the adjacency matrix are randomly selected, and the edges corresponding to the nodes are deleted or the edges between the two nodes are reversed until the trace of the adjacency matrix is ​​equal to zero; when the trace of the current matrix is ​​zero, the nth power of the matrix is ​​calculated, and the above method is used to continue to determine whether the trace of the matrix is ​​zero and make adjustments, and the above process is recursively repeated until n is equal to the dimension of the matrix to obtain an optimized Bayesian network structure; the optimized Bayesian network structure is the Bayesian network structure with the best prediction effect.

8. A Bayesian network structure optimization device based on BRKGA algorithm, characterized in that: The device comprises: A detection module, used for generating a Bayesian network structure optimization instruction when detecting that the prediction accuracy of the Bayesian network in the task to be predicted is lower than a preset value; the task to be predicted includes a disease prediction task; according to the Bayesian network structure optimization instruction, a Bayesian network structure is obtained; the Bayesian network structure is a DAG graph; the DAG graph includes: a node set and an edge set; the nodes in the node set are random variables in the task to be predicted; the random variables refer to nodes in the Bayesian network, including disease and patient information; the disease and patient information are extracted from a medical system; A data acquisition module is used to generate a random key vector consisting of multiple random numbers between [0,1]; A decoding module, used for decoding the random key vector using the NO TEARS algorithm as a decoder to obtain an initial solution to the Bayesian network structure optimization problem; A fitness value acquisition module is used to calculate the fitness value of each of the initial solutions according to the initial solutions and a preset loss function; sort the fitness values ​​to obtain an elite individual set and a non-elite individual set; A crossover and mutation module, used to perform crossover and mutation operations using a genetic algorithm based on the elite individual set, and output a global approximate solution to the Bayesian network structure optimization problem; the global approximate solution is a continuous value matrix; A discretization module is used to discretize the continuous value matrix to obtain a global approximate adjacency matrix, and obtain an optimized Bayesian network structure according to the global approximate adjacency matrix; the global approximate adjacency matrix represents a Bayesian network structure that can describe the relationship between patient information and diseases; The optimized Bayesian network structure is used to assist in disease prediction.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images