Measurement and attribute reduction method adaptive to knowledge graph concept cognition unique features

Abstract

The invention discloses a measurement and attribute reduction method adaptive to knowledge graph concept cognition unique features, and relates to the technical field of measurement and attribute reduction of knowledge graphs. In the multi-concept scene, the invention provides a novel attribute reduction method, which is different from the existing attribute reduction method, can effectively process the multi-concept scene under the concept cognition background on the knowledge graph, and aims to find a group of attribute sets which are closer to decision attributes but not the whole condition attribute set; in order to solve the problem that an existing measurement method cannot distinguish a plurality of reduction corresponding to equivalent division and fine division, the invention provides a division closeness degree with high distinction degree so as to avoid refinement and reduce interlacing; therefore, the reduction method provided by the invention can find better attribute reduction from the attribute set which only has differences in equivalent division and fine division.

Description

Technical Field

[0001] The present invention relates to the field of knowledge graphs and attribute reduction technologies, and specifically to a measurement and attribute reduction method that adapts to the unique cognitive characteristics of knowledge graph concepts. Background Art

[0002] Attribute reduction is essential for rapidly achieving concept recognition in knowledge graphs. To provide prior knowledge for machine understanding and thinking, knowledge graph concept recognition aims to identify the typical characteristics of concepts from the information associated with the concept in the knowledge graph. These characteristics include: 1) the significance of the concept's related attributes; 2) the domain to which the concept belongs; 3) the concept's decision rules; and other features. Knowledge graphs can provide data support for concept recognition because they contain relevant information. For example, a knowledge graph contains the concept "famous director," instances "Person 1 / Person 2 / Person 3...", and triples such as <Person 1, Award, Oscar for Best Director>, <Person 2, Award, Oscar for Best Director>, and <Person 1, Father, Person X>. Based on this data, we can derive cognitive knowledge. For example, "Oscar for Best Director" can confirm that a person belongs to the concept "famous director." However, the attribute "father" and the triple <Person 1, Father, Person X> do not contribute to concept recognition in the knowledge graph, and their presence significantly hinders the acquisition of such cognitive knowledge.

[0003] Attribute reduction for knowledge graph concept cognition has two unique features: 1) Concept cognition on knowledge graphs is based on the relationship between conditional attribute data and decision attribute data. Therefore, attribute reduction for knowledge graph concept cognition aims to find a set of attributes that is closer to the decision attributes. 2) A notable feature of the raw data for knowledge graph concept cognition is that an instance may belong to multiple concepts. Therefore, attribute reduction methods for knowledge graph concept cognition need to be able to handle this one-to-many relationship. Existing attribute reduction methods include primitive rough set attribute reduction methods and rough set-based multi-label attribute reduction methods. Primitive rough set attribute reduction methods are mainly based on complement conditional entropy, Shannon conditional entropy, combined conditional entropy, or inconsistency rate. The multiple labels in multi-label attribute reduction methods are similar to multiple concepts, while other multi-label methods are based on fuzzy rough sets, whose data is fuzzy data represented by probabilities. However, both attribute reduction methods suffer from two problems: 1. They fail to distinguish between multiple reductions corresponding to equivalent partitions and finer partitions, resulting in all attribute subsets that satisfy the conditions being considered reduction results; 2. They aim to find a set of attributes that is closer to the entire set of conditional attributes, rather than the decision attributes. In the above example, existing attribute reduction methods may retain irrelevant attributes such as "father." Treating these as reduction results would severely slow the acquisition of cognitive knowledge about "famous director."

[0004] In summary, existing attribute reduction methods are powerless in the face of the above two unique features, so a new solution is needed to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to provide a measurement and attribute reduction method that adapts to the unique cognitive characteristics of knowledge graph concepts to solve the technical problems raised in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a measurement method adapted to the unique cognitive characteristics of knowledge graph concepts, comprising at least the following steps:

[0007] S1: During attribute reduction, the positive threshold is used to determine the correspondence between instance partitions before and after attribute reduction, that is, the correspondence between the instance partitions of the conditional attribute subset and the instance partitions of the decision attribute. The positive threshold is used as a screening mechanism to help determine which instances are correctly partitioned.

[0008] Define the positive domain;

[0009] Let [x] P ={X1,X2,…,X m} and [x] D ={Y1,Y2,…,Y n} are two partitions of the decision information table S = (U, A = C∪D, V, f), then the positive domain of P with respect to D is defined as follows:

[0010]

[0011] Among them, X i is [x] P The elements in (also a set), Y j is [x] D The elements in (also a set); U is the sample (or instance) set; C is the condition attribute set; D is the decision attribute set; V is the attribute value set; f is the mapping function between attributes and values;

[0012] S2:

[0013] In the process of attribute reduction, the similarity between two partitions is evaluated by partition similarity, thereby providing support for partition closeness. The partition similarity is as follows:

[0014] Let [x] P ={X1,X2,…,X m} and [x] D ={Y1,Y2,…,Y n} are two partitions, [x] P and [x] D The partition similarity between is calculated as follows:

[0015]

[0016] The formula of S2 divides the similarity to calculate all X i and Y j The weighted sum of the intersections between them, and the ratio of the square of the intersection to the union is used to measure the similarity between them. The larger the ratio, the more similar the two partitions are.

[0017] The final result is divided by the size of the entire set U to ensure a standardized similarity value;

[0018] S3: calculation of partition proximity;

[0019] Let [x] P ={X1,X2,…,X m} and [x] D ={Y1,Y2,…,Y n} are two partitions, then the closeness of P to D is calculated as follows:

[0020]

[0021] Among them, PS([x] P ,[x] D ) is the similarity between the two partitions, which measures the [x] P and [x] D The similarity of POS P (D) is the positive domain of P relative to D, that is, the degree to which the instance is correctly divided;

[0022] In the attribute reduction process, partition closeness is used to evaluate the "closeness" between the partition of the instance after reduction and the partition of the instance by the decision attribute, because partition closeness can quantify and understand the strength of the association between the two partitions in practical applications.

[0023] The attribute reduction method adapted to the unique cognitive characteristics of knowledge graph concepts includes at least the following steps:

[0024] S1: Define the multi-concept decision information table;

[0025] Based on the measurement method adapted to the unique cognitive characteristics of knowledge graph concepts described in claim 1, a corresponding decision information table is adopted, that is, S = (U, A = C∪D, V, f). If a certain attribute value in the decision attribute set D corresponds to multiple different classification concepts, then S is regarded as a multi-concept decision information table;

[0026] In machine learning or pattern recognition of knowledge graphs, identifying multi-concept decision information tables helps optimize the decision process and avoid oversimplification or misclassification;

[0027] S2: Define the alternative conditional attribute set:

[0028] When S is a multi - concept decision information table, is a conditional attribute set, [x] P , [x] C and [x] D are respectively the partitions of U induced by P, C, and D;

[0029] If PC(D|P) ≥ PC(D|C), then P is an alternative conditional attribute set of C;

[0030] In a knowledge graph, the adoption of an alternative conditional attribute set can help identify which conditional attributes can replace the full set of conditional attributes, thus reducing the computational complexity in the decision - making process;

[0031] S3: Define the main conditional attribute set;

[0032] Let P be an alternative conditional attribute set of a multi - concept decision information table;

[0033] If for any attribute PC(D|P∪{a}) < PC(D|P) and PC(D|P - {a}) < PC(D|P), then P is a main conditional attribute set of S, and the explanation is as follows:

[0034] When the conditional attribute set P is an alternative conditional attribute set, further examine whether it is a main conditional attribute set. If for any attribute a, the conditional information gain becomes smaller after adding or removing this attribute, then P is considered a main conditional attribute set;

[0035] It is crucial for constructing a decision - making system with a concise and efficient knowledge graph concept cognition because it can help identify the key attributes that are indispensable in the decision - making process;

[0036] S4: Define the attribute reduction by synthesizing S1 - S3;

[0037] When S is a multi - concept decision information table, is a main conditional attribute set, and (P′≠P), PC(D|P′) ≤ PC(D|P), then P is an attribute reduction of C, and the explanation is as follows:

[0038] There may be multiple main conditional attribute sets, and select a conditional attribute set with the largest partition proximity as the attribute reduction;

[0039] Through attribute reduction, those unimportant or redundant attributes in the decision - making process can be removed, thus improving the efficiency and accuracy of the decision - making process.

[0040] A method for obtaining attribute reduction by a heuristic algorithm comprises at least the following steps:

[0041] S1: Input a concept into the computer, obtain all instances of the concept from the knowledge graph, and obtain all attribute features associated with these instances, including attributes, attribute values, and attribute value types;

[0042] S2: Generate a reverse golden section point n based on the size of the entire conditional attribute set C;

[0043] S3: Randomly generate 10 initial attribute sets P of size n;

[0044] S4: The key-value pair that combines the initial attribute set P and its partition closeness PC(D|P)<P,PC(D|P)> Put into candidate set δ can middle;

[0045] S5: Execute the iterative algorithm within the predefined maximum number of executions;

[0046] S6: In the candidate set δ can The attribute set with the maximum partition closeness is retained;

[0047] S7: Generate attribute reduction set δ by taking the attribute set that meets the definition as attribute reduction reduction .

[0048] Furthermore, the S5 at least includes the following steps:

[0049] S51: If the current execution count of the algorithm is a multiple of interval, that is, a multiple of the interval between two optimization operations, then in the candidate set δ can Select all attribute sets P with the maximum partition closeness and put them into the attribute set set δ P middle;

[0050] S52: If it is not a multiple, randomly select an attribute set P and put it into the candidate set set δ P middle;

[0051] S53: Loop through attribute set δ P The attribute set P in .

[0052] Furthermore, the S53 at least includes the following steps:

[0053] S531: Expand each selected attribute set P multiple times. Each time, select an attribute from the conditional attribute set C that is not repeated in the attribute set P, merge it with P to form a new attribute set P′, and put it into the deformation set set δ P′ middle;

[0054] S532: Perform multiple contractions on each selected attribute set P, removing one attribute from P each time to form a new attribute set P′ and put it into the deformation set set δ P′ middle;

[0055] S533: From the deformation set set δ P′ Remove the candidate set δ can The set of attributes that have already appeared in ;

[0056] S534: Combine the deformation set δ P′ The key-value pairs of the remaining attribute set and its partition closeness PC(D|P′)<P′,PC(D|P′)> Put into candidate set δ can middle.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] 1. In multi-concept scenarios, this paper proposes a novel attribute reduction method. Unlike existing attribute reduction methods, it can effectively handle multi-concept scenarios in the context of concept cognition on knowledge graphs. Its goal is to find a set of attributes that is closer to the decision attribute rather than the entire conditional attribute set.

[0059] 2. To address the problem that existing metrics cannot distinguish between multiple reductions corresponding to equivalence partitions and finer partitions, this paper proposes a highly discriminative partition closeness to avoid refinement and reduce overlap. This enables the reduction method of the present invention to find better attribute reductions from attribute sets that differ only in equivalence partitions and finer partitions.

[0060] 3. The present invention also proposes a heuristic algorithm to generate attribute reduction. The heuristic algorithm can obtain the optimal reduction result when executed a sufficient number of times. It has the advantages of accuracy and speed and can improve the efficiency of data analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0062] Figure 1 Classification performance graphs for attribute reduction generated by different metric methods of the present invention;

[0063] Figure 2 Graph showing the classification performance before and after attribute reduction of the present invention;

[0064] Figure 3 This is the classification time diagram before and after attribute simplification of the present invention;

[0065] Figure 4 This is a flow chart of the heuristic algorithm of the present invention. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0067] Example 1:

[0068] The measurement method adapted to the unique cognitive characteristics of knowledge graph concepts includes at least the following steps:

[0069] S1: During attribute reduction, the positive threshold is used to determine the correspondence between instance partitions before and after attribute reduction, that is, the correspondence between the instance partitions of the conditional attribute subset and the instance partitions of the decision attribute. The positive threshold is used as a screening mechanism to help determine which instances are correctly partitioned.

[0070] Define the positive domain;

[0071] Let [x] P ={X1,X2,…,X m} and [x] D ={Y1,Y2,…,Y n} are two partitions of the decision information table S = (U, S = C∪D, V, f), then the positive domain of P with respect to D is defined as follows:

[0072]

[0073] Among them, X i is [x] P The subset elements in (also a set), Y j is [x] D The elements in (also a set); U is the sample (or instance) set; C is the condition attribute set; D is the decision attribute set; V is the attribute value set; f is the mapping function between attributes and values;

[0074] S2: During attribute reduction, the similarity between two partitions is evaluated by partition similarity, thereby providing support for partition closeness. The partition similarity is as follows:

[0075] Let [x] P ={X1,X2,…,X m} and [x] D ={Y1,Y2,…,Y n} are two partitions, [x]P and [x] D The partition similarity between is calculated as follows:

[0076]

[0077] The formula of S2 divides the similarity to calculate all X i and Y j The weighted sum of the intersections between them, and the ratio of the square of the intersection to the union is used to measure the similarity between them. The larger the ratio, the more similar the two partitions are.

[0078] The final result is divided by the size of the entire set U to ensure a standardized similarity value;

[0079] S3: calculation of partition proximity;

[0080] Let [x] P ={X1,X2,…,X m} and [x] D ={Y1,Y2,…,Y n} are two partitions, then the closeness of P to D is calculated as follows:

[0081]

[0082] Among them, PS([x] P ,[x] D ) is the similarity between the two partitions, which measures the [x] P and [x] D ; similarity; POS P (D) is the positive domain of P relative to D, that is, the degree to which the instance is correctly divided;

[0083] In the attribute reduction process, partition closeness is used to evaluate the "closeness" between the partition of the instance after reduction and the partition of the instance by the decision attribute, because partition closeness can quantify and understand the strength of the association between the two partitions in practical applications.

[0084] Example 2:

[0085] The attribute reduction method adapted to the unique cognitive characteristics of knowledge graph concepts includes at least the following steps:

[0086] S1: Define the multi-concept decision information table;

[0087] The measurement method for adapting to the unique cognitive characteristics of knowledge graph concepts based on claim 1 adopts a corresponding decision information table, that is, S = (U, A = C∪D, V, f). If a certain attribute value in the decision attribute set D corresponds to multiple different classification concepts, then S is regarded as a multi-concept decision information table;

[0088] In machine learning or pattern recognition of knowledge graphs, identifying multi - concept decision information tables helps optimize the decision - making process and avoid over - simplification or misclassification;

[0089] S2: Define the alternative conditional attribute set:

[0090] When S is a multi - concept decision information table, is a conditional attribute set, [x] P 、[x] C and [x] D are partitions of U induced by P, C, and D respectively;

[0091] If PC(D|P)≥PC(D|C), then P is an alternative conditional attribute set of C;

[0092] In a knowledge graph, the adoption of an alternative conditional attribute set can help identify which conditional attributes can replace the full set of conditional attributes, thus reducing the computational complexity in the decision - making process;

[0093] S3: Define the main conditional attribute set;

[0094] Let P be an alternative conditional attribute set of a multi - concept decision information table;

[0095] If for any attribute PC(D|P∪{a})<PC(D|P) and PC(D|P - {a})<PC(D|P), then P is a main conditional attribute set of S, and the explanation is as follows:

[0096] When the conditional attribute set P is an alternative conditional attribute set, further examine whether it is a main conditional attribute set. If for any attribute a, the conditional information gain becomes smaller after adding or removing the attribute, then P is considered a main conditional attribute set;

[0097] It is crucial for constructing a decision - making system for concise and efficient knowledge graph concept cognition because it can help identify the key attributes that are indispensable in the decision - making process;

[0098] S4: Define attribute reduction by integrating S1 - S3;

[0099] When S is a multi - concept decision information table, is a main conditional attribute set, and (P′≠P), PC(D|P′)≤PC(D|P), then P is an attribute reduction of C, and the explanation is as follows:

[0100] There may be multiple main conditional attribute sets. Select a conditional attribute set with the maximum partition closeness as the attribute reduction;

[0101] Through attribute reduction, those attributes that are unimportant or redundant in the decision-making process can be removed, thereby improving the efficiency and accuracy of the decision-making process.

[0102] Example 3:

[0103] See Figure 4 , a method for obtaining attribute reduction by a heuristic algorithm, comprising at least the following steps:

[0104] S1: Input a concept into the computer, obtain all instances of the concept from the knowledge graph, and obtain all attribute features associated with these instances. Attribute features include attributes, attribute values, and attribute value types.

[0105] S2: Generate a reverse golden section point n based on the size of the entire conditional attribute set C;

[0106] S3: Randomly generate 10 initial attribute sets P of size n;

[0107] S4: The key-value pair that combines the initial attribute set P and its partition closeness PC(D|P)<P,PC(D|P)> Put into candidate set δ can middle;

[0108] S5: Execute the iterative algorithm within the predefined maximum number of executions;

[0109] S6: In the candidate set δ can The attribute set with the maximum partition closeness is retained;

[0110] S7: Generate attribute reduction set δ by taking the attribute set that meets the definition as attribute reduction reduction .

[0111] S5 includes at least the following steps:

[0112] S51: If the current execution count of the algorithm is a multiple of interval, that is, a multiple of the interval between two optimization operations, then in the candidate set δ can Select all attribute sets P with the maximum partition closeness and put them into the attribute set set δ P middle;

[0113] S52: If it is not a multiple, randomly select an attribute set P and put it into the candidate set set δ P middle;

[0114] S53: Loop through attribute set δ P The attribute set P in .

[0115] S53 at least includes the following steps:

[0116] S531: Expand each selected attribute set P multiple times. Each time, select an attribute from the conditional attribute set C that is not repeated in the attribute set P, merge it with P to form a new attribute set P′, and put it into the deformation set set δ P′ middle;

[0117] S532: Perform multiple contractions on each selected attribute set P, removing one attribute from P each time to form a new attribute set P′ and put it into the deformation set set δ P′ middle;

[0118] S533: From the deformation set set δ P′ Remove the candidate set δ can The set of attributes that have already appeared in ;

[0119] S534: Combine the deformation set δ P′ The key-value pairs of the remaining attribute set and its partition closeness PC(D|P′)<P′,PC(D|P′)> Put into candidate set δ can middle.

[0120] The following experimental demonstration supports the above examples: The experimental data comes from YAGO (https: / / yago-knowledge.org / ), one of the three most widely used knowledge graphs (YAGO, DBpedia, and Freebase). YAGO contains over 10,000 concepts, each with varying numbers of instances, making it easier to select the concepts needed for the experiment.

[0121] If experiments on randomly selected concepts yield good results, it demonstrates that our algorithm is universal and works in the vast majority of cases. Furthermore, since running experiments over a shorter timeframe better demonstrates the results across the temporal dimension, we randomly selected four concepts from a total of 2056 concepts with instances between 600 and 700: "wikicat_Machine_guns," "wikicat_Baltic_islands," "wikicat_Tunisian_footballers," and "wikicat_Painters_of_Emilia__Romagna." For aesthetic reasons, these four concepts were replaced with "Concept 1," "Concept 2," "Concept 3," and "Concept 4" in subsequent experiments. Their statistical information is shown in Table 1. As can be seen from Table 1, the number of instances belonging to multiple concepts varies, ranging from 0.59% to 35.05%.

[0122] Table 1: Dataset statistics

[0123]

[0124]

[0125] 1. Experiment on the effectiveness of dividing proximity

[0126] Different metrics may produce different attribute reductions, but each metric consistently outperforms the others. Since attribute reductions often yield more than one result, Tables 2, 3, 4, and 5 show partial attribute reductions with significant overlap for ease of comparison. E, H, K, φ, and PC represent the attribute reductions for the entire set of conditional attributes from metrics E(D|P), H(D|P), K(D|P), φ(D|P), and PC(D|P), respectively. These tables show that for Concept 1, Concept 2, and Concept 4, the attribute reductions of E(D|P), H(D|P), and K(D|P) are highly consistent. For Concept 3, the attribute reductions of K(D|P) and φ(D|P) are empty. Furthermore, judging by the metric values ​​in the tables, each metric consistently outperforms the attribute reductions obtained by the other methods. It should be noted that the results of the existing measurement method and partitioning proximity represent distance and closeness respectively, so the former and the latter prefer the minimum and maximum values ​​respectively, and when the measurement values ​​are the same, they prefer fewer attributes.

[0127] Table 2: Comparison of attribute reduction of “Concept 1”

[0128]

[0129] Table 3: Comparison of attribute reduction of “Concept 2”

[0130]

[0131]

[0132] Table 4: Comparison of attribute reduction of “Concept 3”

[0133]

[0134] Table 5: Comparison of attribute reduction of “Concept 4”

[0135]

[0136] From the classification performance of attribute reductions generated by different measurement methods, no measurement method maintains an absolute advantage in every concept. In order to objectively compare which measurement method generates better attribute reductions, we conducted a classification experiment on attribute reductions and determined the quality of attribute reductions based on the classification results, such as Figure 1As shown in Figure 2, the four classification algorithms conduct classification experiments on each concept separately, and the test instances of each concept come from the first two sub-concepts with the largest number of instances. Figure 1 The numbers on the X-axis represent the number of original instances, for example, 10 means that the four classification methods classify 10 instances simultaneously, C represents the entire conditional attribute set, and EHK, EH and Kφ represent the same attribute reduction from different metric methods. Figure 1 The left Y-axis and right Y-axis in the graph represent the number and percentage of correctly classified instances, respectively. In order to avoid the differences caused by the classification algorithms, we comprehensively evaluate the classification results of the four classification algorithms, that is, the classification results of the four algorithms constitute a stacked bar chart. Figure 1 It can be seen that for different concepts, no measurement method maintains an absolute advantage in every concept.

[0137] However, based on the above classification performance, PC is the best metric through comprehensive evaluation. The classification results of the five metrics on each instance set of each concept can be ranked. For example, for “Concept 1” and Figure 1 The number 10 on the x-axis corresponds to percentages of 64, 64, 58, and 61 for C, EHK, φ, and PC, respectively. Therefore, the classification performance rankings for E, H, K, φ, and PC are 1, 1, 3, and 2, respectively. This corresponds to the first number in each row of the "Concept 1" column in Table 2. Furthermore, the scale of attribute reduction must be considered. Because the attribute reduction obtained by a metric for a concept is consistent and applied five times (i.e., the numbers 10, 20, 30, 40, and 50 for each concept), the attribute reduction ranking for each concept is multiplied by 5 to simplify the calculation. For example, as shown in Table 2, the attribute reductions for EHK, φ, and PC are {17, 29, 51, 69, 73, 74}, {17, 51, 74}, and {17, 51, 69, 73, 74}, respectively. Therefore, the attribute reduction rankings for E, H, K, φ, and PC are 3, 3, 1, and 2, respectively. This corresponds to the first number in each row of the "P1" column in Table 6. As can be seen from Table 6, by calculating the comprehensive ranking, it can be concluded that PC is the optimal measurement method.

[0138] Table 6: Ranking of attribute reduction classification performance

[0139]

[0140] 2. Validity Experiment of Attribute Reduction

[0141] Attribute reduction is the process of simplifying a dataset by identifying and removing redundant or irrelevant attributes while maintaining classification or decision-making capabilities. Therefore, we compared classification performance before and after attribute reduction. The experimental data combined all data from four concepts (i.e., "Concept 1," "Concept 2," "Concept 3," and "Concept 4"). Figure 2 The numbers on the X-axis and “before / after” represent the number of original instances of each concept and the situation before and after attribute reduction, respectively. Figure 2 The left and right Y axes in represent the number and percentage of correctly classified instances, respectively. Figure 2 The numbers on the stacked bar chart in represent the number of correctly classified instances for each concept. Figure 2 It can be seen that for the four machine learning algorithms (i.e., decision tree, naive Bayes, random forest, and support vector machine), the difference in classification performance before and after attribute reduction is very small, which means that attribute reduction effectively preserves the original characteristics of the data.

[0142] The value of attribute reduction lies in its ability to significantly reduce the complexity of data processing and improve the efficiency of data analysis. Therefore, we compared the classification time before and after attribute reduction. The experimental data combined all the data from four concepts (i.e., "Concept 1", "Concept 2", "Concept 3", and "Concept 4"). Figure 3 The X-axis, left Y-axis, and right Y-axis represent the number of all original instances of the four concepts, the classification time, and the percentage of classification time before and after attribute reduction. Figure 3 It can be seen that for the four machine learning algorithms (i.e., decision tree, naive Bayes, random forest, and support vector machine), there is a significant difference in classification time before and after attribute reduction, which means that attribute reduction can significantly reduce the complexity of data processing and improve the efficiency of data analysis.

[0143] 3. Experiment on the effectiveness of heuristic algorithms

[0144] Under what circumstances can the heuristic algorithm HAR guarantee optimal results? HAR sometimes achieves optimal results and sometimes suboptimal results. Therefore, we show the proportion of optimal results after 10 runs, as shown in Tables 7, 8, 9, and 10, where the parameter interval varies from 10 to 50 and the number of executions varies from 1000 to 5000. These tables show that for "Concept 1" and "Concept 2," HAR consistently achieves optimal results as long as the number of executions exceeds 400. For "Concept 3" and "Concept 4," the execution thresholds are 2000 and 4000, respectively. Experimental results show that increasing the number of executions ensures that HAR achieves both accuracy and speed.

[0145] Table 7: Proportion of HAR achieving the best results for “Concept 1”

[0146]

[0147]

[0148] Table 8: Proportion of HAR achieving the best results for “Concept 2”

[0149]

[0150] Table 9: Proportion of HAR achieving the best results for “Concept 3”

[0151]

[0152] Table 10: Proportion of HAR achieving the best results for “Concept 4”

[0153]

[0154] In summary, the present invention has the following key innovations that cannot be ignored:

[0155] 1. We propose a highly discriminative partition closeness measure to avoid refinement and reduce overlap. Unlike existing metrics, it can distinguish attribute sets that differ only in equivalent partitions and fine partitions.

[0156] 2. A novel attribute reduction method is proposed to find a set of attributes that are closer to the decision attributes than the entire set of conditional attributes in a multi-concept scenario.

[0157] 3. We propose a heuristic algorithm to generate attribute reductions. In the attribute reduction scenario in the context of knowledge graph concept cognition, we increase the number of executions to ensure that the algorithm has both accuracy and speed advantages.

[0158] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A measurement method adapted to the unique cognitive characteristics of knowledge graph concepts, characterized by: At least the following steps are included: S1: During attribute reduction, the positive threshold is used to determine the correspondence between instance partitions before and after attribute reduction, that is, the correspondence between the instance partitions of the conditional attribute subset and the instance partitions of the decision attribute. The positive threshold is used as a screening mechanism to help determine which instances are correctly partitioned. Define the positive domain; Let [x] P ={X1,X2,…,X m } and [x] D ={Y1,Y2,…,Y n } are two partitions of the decision information table S = (U, A = C∪D, V, f), then the positive domain of P with respect to D is defined as follows: Among them, X i is [x] P The elements in is also a set), Y j is [x] D The elements in is also a set; U is the set of samples or instances; C is the condition attribute set; D is the decision attribute set; V is the attribute value set; f is the mapping function between attributes and values; S2: During attribute reduction, the similarity between two partitions is evaluated by partition similarity, thereby providing support for partition closeness. The partition similarity is as follows: Let [x] P ={X1,X2,…,X m } and [x] D ={Y1,Y2,…,Y n } are two partitions, [x] P and [x] D The partition similarity between is calculated as follows: The formula of S2 divides the similarity to calculate all X i and Y j The weighted sum of the intersections between them, and the ratio of the square of the intersection to the union is used to measure the similarity between them. The larger the ratio, the more similar the two partitions are. The final result is divided by the size of the entire set U to ensure a standardized similarity value; S3: calculation of partition proximity; Let [x] P ={X1,X2,…,X m } and [x] D ={Y1,Y2,…,Y n } are two partitions, then the closeness of P to D is calculated as follows: Among them, PS([x] P ,[x] D ) is the similarity between the two partitions, which measures the [x] P and [x] D The similarity of POS P (D) is the positive domain of P relative to D, that is, the degree to which the instance is correctly divided; During attribute reduction, partition closeness is used to evaluate the closeness between the partition of the reduced instance and the partition of the instance by the decision attribute, because partition closeness can quantify and understand the strength of the association between the two partitions in practical applications.

2. An attribute reduction method that adapts to the unique cognitive characteristics of knowledge graph concepts, characterized by: At least the following steps are included: S1: Define the multi-concept decision information table; Based on the measurement method adapted to the unique cognitive characteristics of knowledge graph concepts described in claim 1, a corresponding decision information table is adopted, that is, S = (U, A = C∪D, V, f). If a certain attribute value in the decision attribute set D corresponds to multiple different classification concepts, then S is regarded as a multi-concept decision information table; In machine learning or pattern recognition of knowledge graphs, identifying multi-concept decision information tables helps optimize the decision process and avoid oversimplification or misclassification; S2: Define the set of alternative conditional attributes: When S is a multi-concept decision information table, is a conditional attribute set, [x] P , [x] C and [x] D are the partitioning of U induced by P, C, and D, respectively; If PC(D|P)≥PC(D|C), then P is an alternative set of conditional attributes of C; In the knowledge graph, the use of alternative conditional attribute sets can help identify which conditional attributes can replace the full set of conditional attributes, thereby reducing the computational complexity of the decision-making process; S3: Define the main condition attribute set; Let P be a set of alternative conditional attributes of a multi-concept decision information table; If any attribute PC(D|P∪{a}) < PC(D|P) and PC(D|P - {a}) < PC(D|P), then P is a main conditional attribute set of S, and the explanation is as follows: When the conditional attribute set P is an alternative conditional attribute set, we further examine whether it is a primary conditional attribute set. If for any attribute a, the conditional information gain becomes smaller after adding or removing the attribute, then P is considered to be a primary conditional attribute set. It is crucial for building a streamlined and efficient decision-making system based on knowledge graph concepts, as it can help identify key attributes that are indispensable in the decision-making process; S4: Define attribute reduction by integrating S1-S3; When S is a multi-concept decision information table, is a set of primary conditional attributes, and PC(D|P′)≤PC(D|P), then P is an attribute reduction of C, which is explained as follows: There may be multiple main condition attribute sets, from which one is selected with the largest partition closeness as attribute reduction; Through attribute reduction, those attributes that are unimportant or redundant in the decision-making process can be removed, thereby improving the efficiency and accuracy of the decision-making process.

3. A method for obtaining attribute reduction by a heuristic algorithm, characterized in that: At least the following steps are included: S1: Input a concept into the computer, obtain all instances of the concept from the knowledge graph, and obtain all attribute features associated with these instances, including attributes, attribute values, and attribute value types; S2: Generate a reverse golden section point n based on the size of the entire conditional attribute set C; S3: Randomly generate 10 initial attribute sets P of size n; S4: The key-value pair that combines the initial attribute set P and its partition closeness PC(D|P)<P,PC(D|P)> Put into candidate set δ can middle; S5: Execute the iterative algorithm within the predefined maximum number of executions; S6: In the candidate set δ can The attribute set with the maximum partition closeness is retained; S7: Generate attribute reduction set δ by taking the attribute set that meets the definition as attribute reduction reduction .

4. The method for obtaining attribute reduction by using a heuristic algorithm according to claim 3, characterized in that: The S5 at least includes the following steps: S51: If the current execution count of the algorithm is a multiple of interval, that is, a multiple of the interval between two optimization operations, then in the candidate set δ can Select all attribute sets P with the maximum partition closeness and put them into the attribute set set δ P middle; S52: If it is not a multiple, randomly select an attribute set P and put it into the candidate set set δ P middle; S53: Loop through attribute set δ P The attribute set P in .

5. The method for obtaining attribute reduction by using a heuristic algorithm according to claim 4, characterized in that: The S53 at least includes the following steps: S531: Expand each selected attribute set P multiple times. Each time, select an attribute from the conditional attribute set C that is not repeated in the attribute set P and merge it with P to form a new attribute set P. ′ , and put it into the deformation set δ P′ middle; S532: Perform multiple contractions on each selected attribute set P, removing one attribute from P each time to form a new attribute set P′ and put it into the deformation set set δ P′ middle; S533: From the deformation set set δ P′ Remove the candidate set δ can The set of attributes that have already appeared in ; S534: Combine the deformation set δ P′ The key-value pairs of the remaining attribute set and its partition closeness PC(D|P′) <P′,PC(D|P ′ )> put into candidate set δ can middle.

Images