Improved DS evidence theory method based on liver disease consultation and discussion model

Abstract

The application relates to an improved DS evidence theory method based on a liver disease consultation research model, which comprises the following steps: first, establishing a dispute table and a dialogue tree model for consultation, creating a recognition framework, and obtaining initial evaluation values of each node and rule relations and corresponding strengths between the nodes; obtaining evaluation values of each evidence node for a conclusion node according to a CEM matrix; in a fusion process, calculating Pignistic probability distances and conflict coefficients and synthesizing the conflict distances, and scaling the distances by considering conflict consistency degrees between evidence sources during the synthesis; second, calculating the credibility of each evidence source; and finally, weighting and averaging the evidence sources according to the credibility of each node to correct and fuse the evidence sources by using a traditional DS evidence combination rule to obtain final evaluation values of the nodes. In an example, the method can effectively solve the conflict problem without changing the DS combination rule, and has good anti-interference ability and convergence.

Description

Technical Field

[0001] This invention relates to an improved DS evidence theory method based on a liver disease consultation and discussion model, belonging to the technical field of debate and discussion models. Background Technology

[0002] With the continuous development of science and technology, debate has become an important research direction in the field of artificial intelligence. In medical liver disease consultation systems, debates are usually conducted in a dialogue format, where various doctors present their opinions on a particular diagnosis or treatment plan. Through natural language understanding, this can be transformed into a dispute table, which includes evidence, conclusions, initial evaluation values ​​of the evidence, and the rule relationships and response strengths between the evidence and the conclusion. By treating the evidence in the dispute as child nodes and the conclusion as the parent node, the entire debate scenario can be transformed into a dialogue tree. The credibility of each node in the dialogue tree depends on its initial evaluation value and the evaluations of the node by other evidence nodes. Therefore, it is necessary to integrate the node's initial evaluation with the objective evaluations of other nodes to obtain a comprehensive evaluation value of all doctors' opinions, which is used to judge the credibility of the node. In actual debates, due to the uncertainty and incompleteness of information, the statements in the dispute and the relationships between the disputes are uncertain. DS evidence theory itself has the ability to handle uncertainties, therefore, this theory can be well applied in debate discussion models.

[0003] While the DS evidence theory is capable of handling uncertainties, when there is significant conflict among various sources of evidence, the traditional DS evidence combination rules often yield illogical results. In some extreme cases, such as the Zadeh counterexample, the result of evidence fusion can even contradict intuitive judgment. Therefore, to address the limitation of the traditional DS evidence theory in handling highly conflicting evidence, many scholars both domestically and internationally have made improvements. These improvements mainly focus on modifying the combination rules and the sources of evidence. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing methods by providing an improved DS evidence theory method based on conflicting evidence weighting correction, which modifies the evidence source. This method involves obtaining the self-evaluation value of a node in the dialogue tree of a liver disease consultation system, as well as the evaluation values ​​of that node from other relevant child nodes. The identification framework consists of a triplet, including trust, distrust, and ignorance propositions. Each proposition has evaluation values ​​for the node itself and for each of its other relevant child nodes. All these evaluation values ​​are used as basic probability allocation values ​​and fused using the improved DS evidence theory method of this invention. Finally, the credibility of the node is judged based on the fusion result.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] An improved DS evidence theory method based on a liver disease consultation and discussion model includes the following steps:

[0007] Step 1: Create a liver disease consultation and discussion model, determine the recognition framework, transform the consultation scenario into a dispute table, and create a dialogue tree. The specific details are as follows:

[0008] Step 1.1, determine the recognition frame Θ, the power set 2 of the recognition frame used in this invention. Θ ={Belief, Disbelief, Ignorance}, comprising three propositions: trust, distrust, and ignorance, representing the doctor's degree of trust, distrust, and misunderstanding of the statement, respectively. The evaluation value of each node is represented by the vector w = (b, d, i).

[0009] Step 1.2: Through natural language understanding, all doctors' opinions in the consultation scenario are transformed into a dispute table. Each dispute contains four parts: evidence, conclusion, initial evaluation value of the evidence, rule relationship between the evidence and the conclusion, and response strength. The disputed evidence is treated as child nodes, and the conclusion as the parent node, forming a dialogue tree. The rule relationship between evidence and conclusion is divided into four types:

[0010]

[0011] Where p represents evidence and h represents conclusion. The former two represent evidence or evidence that supports the conclusion, while the latter two represent evidence or evidence that opposes the conclusion.

[0012] The response strength of evidence and conclusion is expressed as

[0013]

[0014] Its value range is 0-1.

[0015] Step two: Based on the initial evaluation values ​​of each disputed piece of evidence, the rule-based relationship between evidence and conclusion, and the response strength, calculate the evaluation value of that evidence node relative to the conclusion node. The specific steps are as follows:

[0016] Step 2.1: For a dispute A, based on the rule relationship between evidence and conclusion and the response strength, obtain the complete evidence mapping matrix CEM(A).

[0017]

[0018] Step 2.2: Calculate the evaluation value w(h) of the conclusion based on the initial evaluation value w(p) of evidence p in dispute A and the CEM matrix.

[0019] w(h) = w(p) × CEM(A)

[0020] Step 3: Calculate the Pengistic probability distance matrix P between the n nodes, as detailed below.

[0021] Step 3.1: For a given node in the dialogue tree, the evaluation values ​​of that node's child nodes can be obtained through the above steps. Assuming that the node has n-1 related child nodes (n≥1), plus the node itself, a total of n evaluation value vectors w1, ..., w2 can be obtained. n (n≥1), use m respectively i (b), m i (d), m i (a) represents the probability assignment function m of propositions b, d, a at the i-th node. i Assign a trust level value.

[0022] Step 3.2, calculate the Pignistic probability function of the evaluation value vector of n nodes. The formula for calculating the Pignistic probability function of the i-th node is as follows:

[0023]

[0024] Where A represents a proposition in the recognition framework, and |A| represents the cardinality of a subset A, i.e., the number of elements contained in set A. For this recognition framework, |Belief| = 1, |Disbelief| = 1, and |Ignorance| = 1.

[0025] Step 3.3, calculate the Pignistic probability distance matrix P between the n nodes.

[0026]

[0027] in The distance between node i and node j is represented by the following formula:

[0028]

[0029] Step 4: Calculate the conflict coefficient matrix K among the n nodes.

[0030]

[0031] Where k ij The collision coefficient between node i and node j is represented by the following formula:

[0032]

[0033] Where m i Let A represent the probability assignment function for the i-th node. i and B jThese represent the propositions for node i and node j within this recognition framework, respectively.

[0034] Step 5: Calculate the conflict distance matrix D between the n nodes, as detailed below.

[0035] Step 5.1, calculate the evidence conflict / consistency G among n nodes. ch (E i E j )

[0036]

[0037] G ch (E i E j ) represents the degree of evidence conflict / consistency between node i and node j, where C(E) i E j ) and H(E i E j ) represent the number of evidence conflicts and the number of evidence consistency for nodes i and j, respectively. The specific formulas are as follows:

[0038]

[0039]

[0040] Step 5.2, calculate the conflict distance matrix D between the n nodes.

[0041]

[0042] Where d ij The collision distance between node i and node j is represented by the following formula:

[0043]

[0044] Step 6: Calculate the credibility Crd(i) of the n nodes, as detailed below.

[0045] Step 6.1, calculate the similarity matrix S between the n nodes.

[0046]

[0047] Where sim(m) i ,m j The similarity between node i and node j is represented by ). Similarity and distance are opposite concepts, therefore similarity can be represented as .

[0048] sim(m i ,m j )=1-d ij

[0049] Step 6.2, calculate the support Sup(i) of the n nodes.

[0050]

[0051] This involves summing all elements in each row of the similarity matrix except for its own similarity score. This method can reflect the degree to which a node is supported by all other related nodes.

[0052] Step 6.3: Normalize the support of the n nodes to obtain the confidence level Crd(i).

[0053]

[0054] Step 7: Modify the evidence sources using credibility as the weight, and perform traditional DS evidence combination on the modified results. The specific steps are as follows:

[0055] Step 7.1: Modify the evidence source by using the credibility Crd(i) of each node as a weight to perform a weighted average correction.

[0056]

[0057] Step 7.2: Use the traditional DS evidence combination rules to synthesize the modified n nodes n-1 times. The synthesis formula is as follows:

[0058]

[0059] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0060] This invention first establishes a dispute table and a dialogue tree model for consultations, creates an identification framework, and obtains the initial evaluation values ​​of each node, as well as the rule relationships and corresponding strengths between nodes. Then, based on the CEM matrix, it obtains the evaluation values ​​of each evidence node for the conclusion node. Next, for nodes with multiple pieces of evidence in the dialogue tree, fusion is performed starting from the leaf nodes of the dialogue tree. During the fusion process, the pignistic probability distance and conflict coefficient are calculated and synthesized to obtain the conflict distance. Simultaneously, the consistency of conflict between evidence sources is considered to scale the distance. Third, the credibility of each evidence source is calculated. Finally, the evidence sources are weighted and averaged according to the credibility of each node, and fusion is performed using traditional DS evidence combination rules to obtain the final evaluation value of the node. The numerical examples show that this method can effectively handle conflict problems without changing the DS combination rules, and has good anti-interference ability and convergence.

[0061] This invention applies an improved DS evidence theory to liver disease consultations. By modifying the evidence sources, evidence fusion avoids the "zero confidence problem" inherent in traditional DS methods. The Pignistic probability function and conflict coefficient are used as the evidence weights for modifying the evidence sources; when the opinions of two doctors differ significantly, the corresponding weights are relatively reduced, thus effectively improving the resistance to interference. The conflict distance is appropriately scaled based on the degree of conflict / consistency between the evidence; when the degree of conflict is high, the conflict distance is appropriately increased, and when the degree of consistency is high, the conflict distance is appropriately decreased, greatly improving the convergence of evidence fusion. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a flowchart of the improved DS evidence theory method based on the liver disease consultation and discussion model of the present invention;

[0064] Figure 2 This is an example diagram of a dispute table for the improved DS evidence theory method based on the liver disease consultation and discussion model of the present invention. It contains 6 disputes, each of which includes evidence, conclusion, and initial evaluation value of the evidence. It also gives the rule relationship between the evidence and the conclusion and the response strength.

[0065] Figure 3 This invention relates to an improved DS evidence theory method based on a liver disease consultation and discussion model. Figure 2 The corresponding dialogue tree shows the direction of the arrows, indicating the direction from the disputed evidence to the conclusion. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] In real-world consultation scenarios, debates often take place over a target statement, such as a diagnosis or treatment plan. Doctors engage in a series of debates before reaching a consensus on whether to trust the statement. Each doctor's viewpoint can be considered a point of contention. The entire consultation model's improved DS evidence theory processing flow is as follows: Figure 1As shown.

[0068] The specific embodiments of the present invention are as follows:

[0069] An improved DS evidence theory method based on a liver disease consultation and discussion model includes the following steps:

[0070] Step 1: Create a liver disease consultation and discussion model, determine the recognition framework, transform the consultation scenario into a dispute table, and create a dialogue tree. The specific details are as follows:

[0071] Step 1.1, determine the recognition frame Θ, the power set 2 of the recognition frame used in this invention. Θ ={Belief, Disbelief, Ignorance}, comprising three propositions: trust, distrust, and ignorance, representing the doctor's degree of trust, distrust, and misunderstanding of the statement, respectively. The evaluation value of each node is represented by the vector w = (b, d, i).

[0072] Step 1.2: Through natural language understanding, all doctors' opinions in the consultation scenario are transformed into a dispute table. Each dispute contains four parts: evidence, conclusion, initial evaluation value of the evidence, rule relationship between the evidence and the conclusion, and response strength. The disputed evidence is treated as child nodes, and the conclusion as the parent node, forming a dialogue tree. The rule relationship between evidence and conclusion is divided into four types:

[0073]

[0074] Where p represents evidence and h represents conclusion. The former two represent evidence or evidence that supports the conclusion, while the latter two represent evidence or evidence that opposes the conclusion.

[0075] The response strength of evidence and conclusion is expressed as

[0076]

[0077] Its value range is 0-1.

[0078] Step two: Based on the initial evaluation values ​​of each disputed piece of evidence, the rule-based relationship between evidence and conclusion, and the response strength, calculate the evaluation value of that evidence node relative to the conclusion node. The specific steps are as follows:

[0079] Step 2.1: For a dispute A, based on the rule relationship between evidence and conclusion and the response strength, obtain the complete evidence mapping matrix CEM(A).

[0080]

[0081] Step 2.2: Calculate the evaluation value w(h) of the conclusion based on the initial evaluation value w(p) of evidence p in dispute A and the CEM matrix.

[0082] w(h) = w(p) × CEM(A)

[0083] Step 3: Calculate the Pengistic probability distance matrix P between the n nodes, as detailed below.

[0084] Step 3.1: For a given node in the dialogue tree, the evaluation values ​​of that node's child nodes can be obtained through the above steps. Assuming that the node has n-1 related child nodes (n≥1), plus the node itself, a total of n evaluation value vectors w1, ..., w2 can be obtained. n (n≥1), use m respectively i (b), m i (d), m i (a) represents the probability assignment function m of propositions b, d, a at the i-th node. i Assign a trust level value.

[0085] Step 3.2, calculate the Pignistic probability function of the evaluation value vector of n nodes. The formula for calculating the Pignistic probability function of the i-th node is as follows:

[0086]

[0087] Where A represents a proposition in the recognition framework, and |A| represents the cardinality of a subset A, i.e., the number of elements contained in set A. For this recognition framework, |Belief| = 1, |Disbelief| = 1, and |Ignorance| = 1.

[0088] Step 3.3, calculate the Pignistic probability distance matrix P between the n nodes.

[0089]

[0090] in The distance between node i and node j is represented by the following formula:

[0091]

[0092] Step 4: Calculate the conflict coefficient matrix K among the n nodes.

[0093]

[0094] Where k ij The collision coefficient between node i and node j is represented by the following formula:

[0095]

[0096] Where m iLet A represent the probability assignment function for the i-th node. i and B j These represent the propositions for node i and node j within this recognition framework, respectively.

[0097] Step 5: Calculate the conflict distance matrix D between the n nodes, as detailed below.

[0098] Step 5.1, calculate the evidence conflict / consistency G among the n nodes. ch (E i E j )

[0099]

[0100] G ch (E i E j ) represents the degree of evidence conflict / consistency between node i and node j, where C(E) i E j ) and H(E i E j ) represent the number of evidence conflicts and the number of evidence consistency for nodes i and j, respectively. The specific formulas are as follows:

[0101]

[0102]

[0103] Step 5.2, calculate the conflict distance matrix D between the n nodes.

[0104]

[0105] Where d ij The collision distance between node i and node j is represented by the following formula:

[0106]

[0107] Step 6: Calculate the credibility Crd(i) of the n nodes, as detailed below.

[0108] Step 6.1, calculate the similarity matrix S between the n nodes.

[0109]

[0110] Where sim(m) i ,m j The similarity between node i and node j is represented by ). Similarity and distance are opposite concepts, therefore similarity can be represented as .

[0111] sim(m i ,m j)=1-d ij

[0112] Step 6.2, calculate the support Sup(i) of the n nodes.

[0113]

[0114] This involves summing all elements in each row of the similarity matrix except for its own similarity score. This method can reflect the degree to which a node is supported by all other related nodes.

[0115] Step 6.3: Normalize the support of the n nodes to obtain the confidence level Crd(i).

[0116]

[0117] Step 7: Modify the evidence sources using credibility as the weight, and perform traditional DS evidence combination on the modified results. The specific steps are as follows:

[0118] Step 7.1: Modify the evidence source by using the credibility Crd(i) of each node as a weight to perform a weighted average correction.

[0119]

[0120] Step 7.2: Use the traditional DS evidence combination rules to synthesize the modified n nodes n-1 times. The synthesis formula is as follows:

[0121]

[0122] Example

[0123] To further verify the effectiveness of this improved DS evidence theory method, this method is compared and verified with other classic algorithms using specific numerical examples.

[0124] Example: In the recognition framework Θ = {A, B, C}, the probability allocation functions of the three propositions are m1, m2, m3, m4, and m5, respectively. The fused probability allocation function is m. The basic probability allocation in this example is shown in Table 2. Among them, evidence source 2 is interference evidence, and the basic probability allocation value of proposition A in this evidence is much smaller than that of other evidence sources.

[0125] Table 2 shows the basic probability distribution of each piece of evidence in the example.

[0126] evidence A B C <![CDATA[m1]]> 0.5 0.2 0.3 <![CDATA[m2]]> 0 0.9 0.1 <![CDATA[m3]]> 0.55 0.1 0.35 <![CDATA[m4]]> 0.55 0.1 0.35 <![CDATA[m5]]> 0.55 0.1 0.35

[0127] The five evidence bodies in Table 2 were divided into four groups for experiments: m1, m2; m1, m2, m3; m1, m2, m3, m4; and m1, m2, m3, m4, m5. The method of this invention is now compared with some classic methods such as the traditional DS method, Yager rule method, Murphy method, and Jousselme distance method. The final fusion results of various methods are shown in Table 3.

[0128] Table 3 Comparison of fusion results under various methods

[0129]

[0130] As shown in Table 3, the second evidence source assigns a value of 0 to proposition A. Due to the "zero confidence problem" in the traditional DS method, the fusion result of the traditional DS method always assigns a value of 0 to proposition A, regardless of how many evidence sources are fused. However, in this example, except for evidence source 2, the probability assignment value of proposition A given by each evidence source is relatively high. Theoretically, the final fusion result should be that the probability of proposition A is much greater than that of propositions B and C. Therefore, the fusion fails.

[0131] The Yager rule method allocates the conflict probability k entirely to the identification frame during the fusion process, resulting in low probabilities for each proposition after fusion. As the number of fused evidence sources increases, the probability eventually approaches 0, making it difficult to use to determine the result. Therefore, this method is not very meaningful in multi-evidence fusion.

[0132] Although the Murphy method generally follows the actual situation as the number of fused evidence sources increases, it cannot quickly eliminate the interference of evidence source 2 when there are few evidence sources. As can be seen from the change from combination 1 to combination 2, the probability of proposition A does not increase very quickly.

[0133] Compared to the Murphy method, the Jousselme distance method considers the interrelationships between evidence sources. Therefore, the convergence speed of proposition A is effectively improved from combination 1 to combination 2, and the interference from evidence source 2 is greatly reduced.

[0134] The method of this invention effectively combines the Pignistic probability distance and the conflict coefficient into a conflict distance when processing the interrelationships between evidence sources. It also considers the impact of evidence conflict / consistency, increasing the conflict distance when the degree of evidence conflict is high and decreasing it when the degree of consistency is high by setting a threshold, thus considering more comprehensive factors. Therefore, as shown in Combination 1, this method has better anti-interference capabilities than other methods, and its convergence speed is also relatively fast.

[0135] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. An improved DS evidence theory method based on a liver disease consultation and discussion model, characterized in that: Includes the following steps: Step 1: Create a liver disease consultation and discussion model, determine the recognition framework, transform the consultation scenario into a dispute table, and create a dialogue tree; Step 2: Calculate the evaluation value of each piece of evidence for the conclusion node based on the initial evaluation value of each piece of evidence, the rule relationship between the evidence and the conclusion, and the response strength. Step one specifically includes: Step 1.1: Determine the recognition framework Identify the idempotent set of the framework It includes three propositions: trust, distrust, and ignorance, which represent the doctor's level of trust, distrust, and ambiguity regarding the statement, respectively. The evaluation value of each node is represented by the vector w=(b,d,i). Step 1.2: Through natural language understanding, all doctors' opinions in the consultation scenario are transformed into a dispute table. Each dispute contains four parts: evidence, conclusion, initial evaluation value of the evidence, rule relationship between the evidence and the conclusion, and response strength. The evidence in the dispute is used as child nodes, and the conclusion is used as the parent node to form a dialogue tree. The rule relationship between the evidence and the conclusion is divided into four types: ； Where p represents evidence and h represents conclusion. The former two represent evidence or evidence that supports the conclusion, while the latter two represent evidence or evidence that opposes the conclusion. The strength of the response between evidence and conclusion is expressed as follows: ； Its value range is 0-1; Step 3: Calculate the Pengistic probability distance matrix P between the n nodes; Step 4: Calculate the conflict coefficient matrix K between the n nodes; Step four specifically includes: Calculate the conflict coefficient matrix K between n nodes: ； in The conflict coefficient between node i and node j is represented by the following formula: ； in Describes the probability assignment function for the i-th node. and These represent the propositions for node i and node j under this recognition framework, respectively. Step 5: Calculate the conflict distance matrix D between the n nodes; Step five specifically includes: Step 5.1: Calculate the evidence conflict / consistency among n nodes. : ； This represents the degree of evidence conflict / consistency between node i and node j, where and Let i and j represent the evidence conflict and consistency rates for nodes i and j, respectively, using the following formulas: ； ； Step 5.2: Calculate the conflict distance matrix D between the n nodes: ； in The collision distance between node i and node j is represented by the following formula: ； in This represents the distance between node i and node j; Step 6: Calculate the credibility Crd(i) of the n nodes; Step 7: Modify the evidence source with credibility as the weight, and perform traditional DS evidence combination on the modified result.

2. The improved DS evidence theory method based on a liver disease consultation and discussion model according to claim 1, characterized in that: Step two specifically includes: Step 2.1: For a dispute A, based on the rule relationship between evidence and conclusion and the response strength, obtain the complete evidence mapping matrix CEM(A). ； Step 2.2: Calculate the evaluation value w(h) of the conclusion based on the initial evaluation value w(p) of evidence p in dispute A and the CEM matrix. 。 3. The improved DS evidence theory method based on a liver disease consultation and discussion model according to claim 1, characterized in that: Step three specifically includes: Step 3.1: For a node in the dialogue tree, the evaluation values ​​of its child nodes can be obtained through the above steps. Assuming that the node has n-1 related child nodes (n≥1), plus the node itself, a total of n evaluation value vectors w1, ..., w2 can be obtained. n (n≥1), use m respectively i (b), m i (d), m i (a) represents the probability assignment function m of propositions b, d, a at the i-th node. i Assign a trust level value; Step 3.2: Calculate the Pignistic probability function of the evaluation value vector of n nodes. The formula for calculating the Pignistic probability function of the i-th node is: ； Where |B| represents the cardinality of the subset B, that is, the number of elements contained in the set B; Step 3.3: Calculate the Pengistic probability distance matrix P between the n nodes; ； in The distance between node i and node j is represented by the following formula: 。 4. The improved DS evidence theory method based on a liver disease consultation and discussion model according to claim 1, characterized in that: Step six specifically includes: Step 6.1: Calculate the similarity matrix S between the n nodes. ； in Let represent the similarity between node i and node j. Similarity and distance are opposite concepts, therefore similarity can be represented as: ； Step 6.2, calculate the support Sup(i) of the n nodes: ； Summing all elements in each row of the similarity matrix except for its own similarity score reveals the degree to which a node is supported by all other related nodes. Step 6.3: Normalize the support of the n nodes to obtain the confidence level Crd(i): 。 5. The improved DS evidence theory method based on a liver disease consultation and discussion model according to claim 1, characterized in that: Step seven specifically includes: Step 7.1: Modify the evidence source by using the credibility Crd(i) of each node as a weight for weighted averaging correction. ； Step 7.2: Use the traditional DS evidence combination rules to synthesize the modified n nodes n-1 times. The synthesis formula is: 。

Images