A multi-truth discovery method based on attribute fusion

Abstract

The present invention discloses a multi-truth value discovery method based on attribute fusion, which is an unsupervised multi-truth value discovery method. Considering the influence of auxiliary attributes on multi-truth value discovery, the data source expertise and consensus obtained by the auxiliary attributes and the activity of multi-truth attribute values are fused to obtain the support degree of the data source for the attribute value. The multi-truth value discovery problem is converted into a classification problem, and a neural network is used to automatically learn the underlying data features between the data source and the attribute value, so as to finally predict multiple truth values. The accuracy of the experimental results of the multi-truth value discovery method in the field of data integration can be improved. In the face of a multi-source noisy and conflicting data environment, the method can automatically identify, judge and infer complete truth values or data close to the truth values, thereby assisting users in making decisions and improving decision-making efficiency and quality.

Description

Technical Field

[0001] The present invention relates to a multi-truth value discovery method, in particular to a multi-truth value discovery method based on attribute fusion, and belongs to the technical field of data integration. Background Art

[0002] With the development of the internet, the amount of data on the Web has exploded. This massive amount of Web data holds enormous commercial and research value, but not all of it is authentic or usable due to input errors, outdated data, and lost records. Different data sources provide attribute values for several attributes of the same object. Attributes with multiple true values are called multi-truth attributes, and attributes other than multi-truth attributes are called auxiliary attributes. Finding all true values from among the multiple true attribute values provided by a data source for an object is called multi-truth discovery. Currently, multi-truth discovery methods have been successfully applied in many scenarios, such as big data fusion systems, the construction of high-quality knowledge bases, crowd sensing, privacy protection, crowdsourcing aggregation, and the medical field.

[0003] Existing multi-truth discovery methods include: Bo Zhao et al. from the University of Illinois at Urbana, who assumed that the recall and precision of data source quality follow a Beta distribution to construct a probabilistic graphical model (LTM) and used a sampling-based method to infer multiple truth values on multi-truth attributes (Zhao B, Rubinstein BIP, Gemmell J, et al. A bayesian approach to discovering truth from conflicting sources for data integration[J].arXiv preprint arXiv, 2012.1203.058: p.550-561); Xianzhi Wang et al. from the University of Adelaide, who reconstructed the mapping relationship between data sources and multi-truth attribute values, combined the mutually exclusive relationship between attribute values, and proposed a Bayesian method to achieve multi-truth discovery (Wang X, Sheng QZ, Fang XS, et al. An integrated bayesian approach for effective multi-truth discovery[C] / / Proceedings of the 24th ACM International on Conference on Information and Knowledge Management. 2015: p. 493-502); Xueling LIN et al. from the Hong Kong University of Science and Technology considered the domain expertise of the data source and combined the credibility score of multi-truth attribute values to propose a DART model that integrates Bayesian methods (Lin X, Chen L. Domain-aware multi-truth discovery from conflicting sources [J]. Proceedings of the VLDB Endowment, 2018, 11(5): p.635-647); Ma Ruxia et al. from Renmin University of China assumed that the attribute values of objects of different categories should have different credibility and proposed the CTruthFinder algorithm (Ma Ruxia, Meng Xiaofeng. Research on truth value discovery method based on the credibility of data source classification [J]. Computer Research and Development. 2015, 52(09): pp. 1931-1940); Lu Jing et al. from Shanghai University of Technology classified data sources according to the long tail phenomenon of data distribution, and used the correlation of attribute sets to seek the optimal distribution of data source weights to obtain multiple true values (Lu Jing, Hu Cheng, Liu Cong. Research on multiple truth value discovery method based on attribute set correlation and source error [J]. Small and Micro Computer Systems. 2019, 40(03): pp. 601-605).These multi-truth discovery methods focus solely on the multiple truth attributes of an object, failing to consider the impact of auxiliary attributes. This results in unsatisfactory accuracy. Furthermore, these multi-truth discovery methods require manually defined rules, which are difficult to adapt and analyze, and do not accurately reflect the underlying data distribution. Summary of the Invention

[0004] To address the above problems, the present invention provides a multi-truth discovery method based on attribute fusion, which considers and combines the influence of auxiliary attributes on multi-truth discovery, and can improve the accuracy of experimental results of the multi-truth discovery method in the field of data integration.

[0005] To achieve the above objectives, this multi-truth discovery method based on attribute fusion specifically includes the following steps:

[0006] Step 1: Calculate the expertise of the data source using discrete auxiliary attributes;

[0007] Step 1-1: Count the data sources i Provided belongs to the collection The number of objects, in, Indicates data source s i Regarding the value set provided by object o on the discrete auxiliary attribute D, N(d, s i ) indicates the data source s i The number of objects of type d provided;

[0008] Step 1-2: Count the data that belong to the set The number of objects, Among them, N(d) represents the set of all data sources in the dataset The number of objects of type d provided;

[0009] Step 1-3: Calculate the data source s i Contribution rate of the category to which object o belongs on the discrete auxiliary attribute D

[0010]

[0011] Step 1-4: Calculate the data source s i The expertise of o on the discrete auxiliary attribute D

[0012]

[0013] Step 2: Calculate the data source consensus using continuous auxiliary attributes;

[0014] Step 2-1: Calculate the adaptive bucket distance of object o based on the continuous value provided by the data source about the continuous auxiliary attribute C of object o

[0015] Step 2-2: Get the value based on the adaptive bucket distance common equal intervals, where Represents a collection of data sources Regarding the set of values provided by an object o on a continuous auxiliary attribute C, Represents a value set The maximum value in Represents a value set The minimum value in Indicates data source s i Regarding the length of the value set provided by object o on the continuous auxiliary attribute C, count the continuous values of the auxiliary attribute The number of values in the interval will be The number of values in the interval from low to high is recorded as

[0016] Step 2-3: Calculate the data source s i Consensus on object o

[0017]

[0018] in, Indicates data source s i Regarding the continuous value provided by the object o on the continuous auxiliary attribute C;

[0019] Step 3: Use multi-truth attribute to calculate the activity of multi-truth attribute value;

[0020] Step 3-1: For multiple truth attribute values Calculate its activity γ v ,

[0021]

[0022] in, Indicates data source s i Regarding the multiple truth-valued attribute value sets provided by object o, Represents the set of all data sources that provide attribute value v for object o. Represents the sum of the frequencies of all attribute values;

[0023] Step 4: Use an iterative method to obtain the true value pseudo label and the initial reliability of the data source;

[0024] Step 4-1: Set the data sourcei The set of values provided by object o on its multiple truth-valued attributes The value in s is used as the data source i A provided union value;

[0025] Step 4-2: According to the process of step 4-1, call the iterative single truth value discovery method TruthFinder to calculate the probability that the joint value provided by the data source becomes the true value and the initial reliability τ of the data source s ;

[0026] Step 4-3: Split the joint value with the highest probability of being the true value for object o into single values, and use these single values as the true values to obtain the pseudo labels of the attribute values;

[0027] Step 5: Combine the expertise and consensus of the data source with the activity of the multi-truth attribute value to obtain the support of the data source for the attribute value and calculate the data source s i Support for multiple true-valued attribute values v

[0028] Step 6: Construct a source-value vector for each attribute value;

[0029] Each attribute value in the true attribute value set corresponds to a source-value vector, and the length of the vector corresponds to the number of all data sources. The value of each position in the vector is the support of the attribute value corresponding to the data source at that position, and the source-value vector is constructed for the multi-truth attribute value v in, Represents the collection of all data sources;

[0030] Step 7: Use a neural network to capture the relationship between the data source and the attribute value, and predict the probability of the attribute value becoming the true value;

[0031] Step 7-1: Use a three-layer neural network model to learn the relationship between the data source and the multi-truth attribute values. The first layer is a fully connected layer with 125 nodes, and the second layer is a fully connected layer with 65 nodes. The first two layers use the ReLU function as the activation function. A dropout layer is used after the first and second layers. The third layer, the output layer, contains one node and uses the sigmoid function as the activation function.

[0032] Step 7-2: Train the weight matrix of each layer, obtain the output of the model through forward propagation, and obtain the probability of each multi-valued attribute value being true through the sigmoid function. Use the binary cross entropy loss function to measure the error between the predicted value and the true value:

[0033]

[0034] Among them, y iis the label of the i-th sample in the training set, p i is the output of the i-th sample of the above neural network model;

[0035] The weight matrix and bias matrix of each layer descend along the gradient until the model training is completed and the parameters converge;

[0036] Step 7-3: Divide the data into a training set:test set ratio of 1:9, obtain true value pseudo labels for the training set through step 4, and then train the model until the parameters converge;

[0037] Step 8: Get the true label of the attribute value by comparing it with the set threshold and obtain multiple true values of the object;

[0038] Step 8-1: Input the test set vector into the trained model and output the probability of each attribute value in the test set being the true value;

[0039] Step 8-2: When the probability of the attribute value is greater than the threshold K, the value is determined to be true, otherwise the value is determined to be false, and multiple true values of the object are obtained.

[0040] Adaptive bucket distance of object o in step 2-1 The calculation formula is as follows:

[0041]

[0042] in, Represents a collection of data sources Regarding the set of values provided by an object o on a continuous auxiliary attribute C, Represents a value set The maximum value in Represents a value set The minimum value in Indicates data source s i Provides the length of the set of values in the continuous auxiliary attribute C for object o.

[0043] The single truth value discovery method called in step 4-2 is an iteration-based single truth value discovery method TruthFinder.

[0044] Calculate the data source s in step 5 i Support for multiple true-valued attribute values v The calculation formula is as follows:

[0045]

[0046] When the data source i No value v is provided for object o, that is, value v does not belong to At this time, the data source s i The support for value v is 0.

[0047] In step 7-2, the RMSProp algorithm is used to complete the gradient descent.

[0048] Compared with the existing technology, this multi-truth discovery method based on attribute fusion takes into account the impact of auxiliary attributes on multi-truth discovery, regards the multi-truth discovery task as a classification problem, and uses neural networks to automatically learn the underlying data features between data sources and attribute values, and finally infers all the true values. In the face of multi-source noisy and conflicting data environments, it can automatically identify, judge, and infer complete true values or data close to the true values, thereby assisting users in decision-making and improving decision-making efficiency and quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is the flowchart of the multi-truth discovery method based on attribute fusion;

[0050] Figure 2 This is the model structure diagram of the multi-truth discovery method based on attribute fusion. DETAILED DESCRIPTION

[0051] This multi-truth discovery method based on attribute fusion is an unsupervised multi-truth discovery method. Considering the influence of auxiliary attributes on multi-truth discovery, the data source expertise and consensus obtained by the auxiliary attributes and the activity of multi-truth attribute values are fused to obtain the support degree of the data source for the attribute value. The multi-truth discovery problem is transformed into a classification problem. The neural network is used to automatically learn the underlying data features between the data source and the attribute value, thereby predicting multiple truth values, which can improve the accuracy of multi-truth discovery.

[0052] The present invention will be further described below with reference to the accompanying drawings and a list of all correct actors in the movie Harry Potter as an example.

[0053] An object is an objective entity in the real world. For example, the movie "Harry Potter" in Table 1 is an object. A data source is an information source that collects, stores, and provides data related to the object. For example, the IMDB website, Filmcrave website, Goodfilms website, and Movieindider website in Table 1 are all data sources. Attributes with multiple truth values are called multi-truth attributes. For example, the list of actors in Table 1 is a multi-truth attribute. Attributes other than multi-truth attributes are called auxiliary attributes. Depending on the type of auxiliary attribute value, attributes that provide discrete values are called discrete auxiliary attributes, and attributes that provide continuous values are called continuous auxiliary attributes. For example, the movie genre in Table 1 is a discrete auxiliary attribute, and the movie duration is a continuous auxiliary attribute. The multi-truth discovery problem is to find all correct values from the conflicting data provided by multiple data sources in the multi-truth attributes, that is, to find the correct list of all actors in the movie "Harry Potter" in Table 1.

[0054] Table 1 Information about the movie Harry Potter provided by different websites

[0055]

[0056] The multi-truth discovery method based on attribute fusion specifically includes the following steps:

[0057] Step 1: Calculate the expertise of the data source using discrete auxiliary attributes:

[0058] Step 1-1: Collect statistics from the data source i Provided belongs to the collection The number of objects, that is in, Indicates data source s i Regarding the value set provided by object o on the discrete auxiliary attribute D, N(d, s i ) indicates the data source s i The number of objects of type d provided.

[0059] Step 1-2: Statistical data belongs to the set The number of objects, that is Among them, N(d) represents the set of all data sources in the dataset The number of objects of type d provided.

[0060] Steps 1-3: Calculate data source s i Contribution rate of the category to which object o belongs on the discrete auxiliary attribute D The formula is as follows:

[0061]

[0062] Steps 1-4: Calculate data source s i The expertise of o on the discrete auxiliary attribute D The formula is as follows:

[0063]

[0064] Data Sources i Multiple discrete values can be provided for the discrete auxiliary attribute D. In Table 1, the IMDB website provides two values of "fantasy, adventure" for the movie type attribute of the movie "Harry Potter". Accordingly,

[0065]

[0066]

[0067] Step 2: Calculate the data source consensus using continuous auxiliary attributes:

[0068] Step 2-1: Calculate the adaptive bucket distance of object o on the continuous auxiliary attribute C The formula is as follows:

[0069]

[0070] in, Represents a collection of data sources Regarding the set of values provided by an object o on a continuous auxiliary attribute C, Represents a value set The maximum value in Represents a value set The minimum value in Indicates data source s i Provides the length of the set of values in the continuous auxiliary attribute C for object o.

[0071] Step 2-2: Get the value based on the adaptive bucket distance common equal intervals, statistical auxiliary attribute continuous values The number of values in the interval will be The number of values in the interval from low to high is recorded as

[0072] Step 2-3, calculate data source s i Consensus on object o The calculation formula is as follows:

[0073]

[0074] in, Indicates data source s i The continuous value provided by the continuous auxiliary attribute C of object o.

[0075] Step 3: Use multi-truth attribute to calculate the activity of multi-truth attribute value:

[0076] Step 3-1: For multiple true value attributes Calculate its activity γ v , the calculation formula is as follows:

[0077]

[0078] in, Indicates data source s i Regarding the multiple truth-valued attribute value sets provided by object o, represents the set of all data sources that provide attribute values v for object θ, Represents the sum of the frequencies of all attribute values.

[0079] Step 4: Use an iterative method to obtain the true value pseudo label and the initial reliability of the data source:

[0080] Step 4-1, data source s i The set of values provided by object θ on its multi-valued attributes The value in is considered as the data source s i A joint value is provided. Accordingly, the value set provided by the data source IMDB website for the actor list of the movie "Harry Potter" is {Dainel Radcliffe; Emma Waston; Rupert Grint}. All values in the multi-truth attribute value set {DainelRadcliffe; Emma Waston; Rupert Grint} are regarded as a joint value "Dainel Radcliffe; Emma Waston; Rupert Grint" provided by the IMDB website.

[0081] Step 4-2: Call the iterative single truth value discovery method TruthFinder according to the process of step 4-1 to calculate the probability that the joint value provided by the data source becomes the true value and the initial reliability τ of the data source. s .

[0082] Step 4-3: Split the joint value with the highest probability of being the true value for object o into single values, and use these single values as true values to obtain pseudo labels for the attribute values.

[0083] Step 5: Combine the expertise and consensus of the data source with the activity of the multi-truth attribute value to obtain the support degree of the data source for the attribute value:

[0084] Calculation data sources i Support for multiple true-valued attribute values v

[0085]

[0086] Among them, if the data source s i If no value v is provided, then the value v does not belong to the data source s i About object collections In the value set provided by the multi-valued attribute A, the data source s i The support for value v is 0.

[0087] Step 6. Construct a source-value vector for each attribute value:

[0088] Each attribute value in the true attribute value set corresponds to a source-value vector, and the length of the vector corresponds to the number of all data sources. The value of each position in the vector is the support of the attribute value corresponding to the data source at that position, and the source-value vector is constructed for the multi-truth attribute value v in, represents the set of all data sources, and accordingly, the source-value vector of the multi-truth attribute value “Dainel Radcliffe” in Table 1

[0089]

[0090] Step 7: Use a neural network to capture the complex relationship between data sources and attribute values, and predict the probability of the attribute value becoming the true value:

[0091] Step 7-1. The multi-truth discovery method based on attribute fusion uses a three-layer neural network model to learn the complex relationship between data sources and multi-truth attribute values. The first layer is a fully connected layer containing 125 nodes, and the second layer is a fully connected layer containing 65 nodes. The first two layers use the ReLU function as the activation function. In order to prevent the model from overfitting, a dropout layer is used after the first and second layers. The third output layer contains one node and uses sigmoid as the activation function.

[0092] Step 7-2: The core of the model is to train the weight matrix of each layer, obtain the output of the model through forward propagation, and finally obtain the probability of each multi-valued attribute value being true through the sigmoid function. The error between the predicted value and the true value is measured using the binary cross entropy loss function:

[0093]

[0094] Among them, y i is the label of the i-th sample in the training set, p i is the output of the i-th sample of the above neural network model.

[0095] Through training, the weight matrix and bias matrix of each layer are continuously adjusted along the direction of gradient descent until the model training is completed and the parameters converge. In order to speed up the training of the model, the RMSProp algorithm is used to complete the gradient descent.

[0096] Step 7-3: Divide the data into a ratio of training set: test set = 1:9, repeat step 4 to obtain true value pseudo labels for the training set, and then start training the model until the parameters converge.

[0097] Step 8: Get the true label of the attribute value by comparing it with the set threshold, thereby finding multiple true values of the object:

[0098] Step 8-1: Input the test set vector into the trained model and output the probability of each attribute value in the test set being the true value.

[0099] Step 8-2: When the probability of the attribute value is greater than the threshold K, the value is determined to be true; otherwise, the value is determined to be false, thereby obtaining multiple true values of the object.

[0100] To verify the accuracy and completeness of the truth value of this multi-truth discovery method based on attribute fusion, we took the book dataset and movie dataset as examples and compared this multi-truth discovery method with the Majority Voting multi-truth discovery method, the TruthFinder multi-truth discovery method, the LTM multi-truth discovery method, and the DART multi-truth discovery method. The details are as follows:

[0101] (1) Book Dataset: This is the dataset used in the experiment by Xueling LIN et al. at the Hong Kong University of Science and Technology. The dataset contains 54,591 different registered booksellers, providing 2,338,559 pieces of information for 210,206 books, each identified by an ISBN-13. The attributes of the books include "author", "price", "publication year", and "category". The original data contains noise, and the multi-truth attribute "author" is preprocessed to generate a unified format. 40 books were randomly selected and the author information was manually annotated. In each round of the experiment, 280 books were randomly selected as the training set and 120 books were randomly selected as the test set, and the experiment was repeated 10 times.

[0102] (2) Movie Dataset: This is the dataset used in the experiment conducted by Xueling Lin et al. at the Hong Kong University of Science and Technology. A movie is identified by the combination of its release year and title. The attributes of a movie include "director," "release year," and "movie genre." The original dataset contains noise, so the multi-truth attribute "director" is preprocessed to generate a unified format. 320 movies were randomly selected and their director information was manually annotated. In each round of the experiment, 220 movies were randomly selected as the training set and 10 movies were randomly selected as the test set. This was repeated 10 times.

[0103] Metrics used to evaluate the accuracy and completeness of multiple truth-finding methods include recall (the average percentage of predicted ground truth values among all object ground truth sets), precision (the average percentage of predicted ground truth values among all object ground truth sets predicted by the algorithm), and the harmonic mean F-measure (the harmonic mean of precision and recall). The experimental results are shown in Table 2.

[0104] Table 2 Comparison of different multi-truth discovery methods

[0105]

[0106] As shown in the table above, the recall, precision, and harmonic mean of this attribute fusion-based multi-truth discovery method are all at relatively high levels. It can be applied to many real-world scenarios, such as big data fusion systems, the construction of high-quality knowledge bases, crowd sensing, privacy protection, crowdsourcing aggregation, and the medical field.

Claims

1. A multi-truth discovery method based on attribute fusion, characterized in that: The specific steps include: Step 1: Calculate the expertise of the data source using discrete auxiliary attributes; Step 1-1: Count the data sources i Provided belongs to the collection The number of objects, in, Indicates data source s i Regarding the value set provided by object o on the discrete auxiliary attribute D, N(d,s i ) indicates the data source s i The number of objects of type d provided; Step 1-2: Count the data that belong to the set The number of objects, Among them, N(d) represents the set of all data sources in the dataset The number of objects of type d provided; Step 1-3: Calculate the data source s i Contribution rate of the category to which object o belongs on the discrete auxiliary attribute D Step 1-4: Calculate the data source s i Expertise about object o on discrete auxiliary attributes D Step 2: Calculate the data source consensus using continuous auxiliary attributes; Step 2-1: Calculate the adaptive bucket distance T of object o based on the continuous value provided by the data source about the continuous auxiliary attribute C of object o o C ; Step 2-2: Get the value based on the adaptive bucket distance common equal intervals, where Represents a collection of data sources Regarding the set of values provided by an object o on a continuous auxiliary attribute C, Represents a value set The maximum value in Represents a value set The minimum value in Indicates data source s i Regarding the length of the value set provided by object o on the continuous auxiliary attribute C, count the continuous values of the auxiliary attribute The number of values in the interval will be The number of values in the interval from low to high is recorded as Step 2-3: Calculate the data source s i Consensus on object o in, Indicates data source s i Regarding the continuous value provided by the object o on the continuous auxiliary attribute C; Step 3: Use multi-truth attribute to calculate the activity of multi-truth attribute value; Step 3-1: For multiple truth attribute values Calculate its activity γ v , in, Indicates data source s i Regarding the multiple truth-valued attribute value sets provided by object o, Represents the set of all data sources that provide multiple truth-valued attribute values v for object o. Represents the sum of the frequencies of all attribute values; Step 4: Use an iterative method to obtain the true value pseudo label and the initial reliability of the data source; Step 4-1: Set the data source i The set of values provided by object o on its multiple truth-valued attributes The value in s is used as the data source i A provided union value; Step 4-2: Call the single truth value discovery method according to the process of step 4-1 to calculate the probability that the joint value provided by the data source becomes the true value and the initial reliability τ of the data source s ; Step 4-3: Split the joint value with the highest probability of being the true value for object o into single values, and use these single values as the true values to obtain the pseudo labels of the attribute values; Step 5: Combine the expertise and consensus of the data source with the activity of the multi-truth attribute value to obtain the support of the data source for the attribute value and calculate the data source s i Support for multiple true-valued attribute values v Step 6: Construct a source-value vector for each attribute value; Each attribute value in the true attribute value set corresponds to a source-value vector, and the length of the vector corresponds to the number of all data sources. The value of each position in the vector is the data source s corresponding to that position i Support for multiple true-valued attribute values v Construct source-value vector for multi-valued attribute value v in, Represents the collection of all data sources; Step 7: Use a neural network to capture the relationship between the data source and the attribute value, and predict the probability of the attribute value becoming the true value; Step 7-1: Use a three-layer neural network model to learn the relationship between the data source and the multi-truth attribute values. The first layer is a fully connected layer with 125 nodes, and the second layer is a fully connected layer with 65 nodes. The first two layers use the ReLU function as the activation function. A dropout layer is used after the first and second layers. The third layer, the output layer, contains one node and uses the sigmoid function as the activation function. Step 7-2: Train the weight matrix of each layer, obtain the output of the model through forward propagation, and obtain the probability of each multi-valued attribute value being true through the sigmoid function. Use the binary cross entropy loss function to measure the error between the predicted value and the true value: Where N is the number of samples, y i is the label of the i-th sample in the training set, p i is the output of the i-th sample of the above neural network model; The weight matrix and bias matrix of each layer descend along the gradient until the model training is completed and the parameters converge; Step 7-3: Divide the data into a training set:test set ratio of 1:9, obtain true value pseudo labels for the training set through step 4, and then train the model until the parameters converge; Step 8: Get the true label of the attribute value by comparing it with the set threshold and obtain multiple true values of the object; Step 8-1: Input the test set vector into the trained model and output the probability of each attribute value in the test set being the true value; Step 8-2: When the probability of the attribute value is greater than the threshold K, the value is determined to be true, otherwise the value is determined to be false, and multiple true values of the object are obtained.

2. The multi-truth discovery method based on attribute fusion according to claim 1 is characterized in that: Adaptive bucket distance of object o in step 2-1 The calculation formula is as follows: in, Represents a collection of data sources Regarding the set of values provided by an object o on a continuous auxiliary attribute C, Represents a value set The maximum value in Represents a value set The minimum value in Indicates data source s i Provides the length of the set of values in the continuous auxiliary attribute C for object o.

3. The multi-truth discovery method based on attribute fusion according to claim 1 is characterized in that: The single truth value discovery method called in step 4-2 is an iteration-based single truth value discovery method TruthFinder.

4. The multi-truth discovery method based on attribute fusion according to claim 1 is characterized in that: Calculate the data source s in step 5 i Support for multiple true-valued attribute values v The calculation formula is as follows: When the data source i There is no multiple truth attribute value v provided for object o, that is, the multiple truth attribute value v does not belong to At this time, the data source s i The support for value v is 0.

5. The multi-truth discovery method based on attribute fusion according to claim 1 is characterized in that: In step 7-2, the RMSProp algorithm is used to complete the gradient descent.

Images