A heterogeneous information fusion method and system based on a graph neural network

Abstract

The application discloses a heterogeneous information fusion method and system based on a graph neural network, and relates to the technical field of heterogeneous information fusion. The method comprises the following steps: collecting information data, preprocessing the information data, and taking the preprocessed data as basic data; setting conversion standards of basic data of different categories, converting the basic data according to the conversion standards to obtain elementary data; extracting and identifying corresponding entities according to the elementary data and taking the entities as standard entity objects; acquiring the collection information of the elementary data, judging the data quality of the elementary data according to the collection information and the standard entity objects, and matching the data quality to obtain medium data; forming a heterogeneous information graph according to the medium data, selecting a graph neural network model to process the heterogeneous information graph, and obtaining a heterogeneous information fusion result. The application improves the accuracy of the heterogeneous information fusion based on the graph neural network.

Description

Technical Field

[0001] This application relates to the technical field of heterogeneous information fusion, and in particular to a heterogeneous information fusion method and system based on graph neural networks. Background Technology

[0002] Heterogeneous information fusion refers to the process of effectively integrating information data from different data sources, formats, and types. Heterogeneous information fusion makes information data more comprehensive and accurate. This comprehensiveness and accuracy help to uncover deeper connections and patterns between information data, thereby unlocking greater data value. Furthermore, heterogeneous information fusion can break down data barriers and promote data interconnection between different organizations, departments, or systems. This facilitates data sharing and exchange, improving the utilization rate of data resources. However, during the heterogeneous information fusion process, information from different sources has quality differences. These differences in information data quality will affect the quality of the fused information data. This impact may lead to the fused information data failing to accurately reflect the actual situation, thus reducing the effectiveness of information data fusion. Currently, data cleaning is used to reduce data quality differences, such as removing missing values, duplicate data, and erroneous data. However, it is clear that in addition to these fundamental issues, the inherent differences in data quality due to data accuracy significantly reduce the effectiveness of heterogeneous information fusion and the accuracy of the resulting fused information data. Summary of the Invention

[0003] The purpose of this invention is to provide a heterogeneous information fusion method and system based on graph neural networks to solve the problems mentioned in the background art.

[0004] Firstly, this application provides a heterogeneous information fusion method based on graph neural networks, which adopts the following technical solution:

[0005] Collect information data, preprocess the information data, and use the preprocessed data as the base data;

[0006] Set conversion standards for different categories of basic data, and convert the basic data according to the conversion standards to obtain elementary data;

[0007] Extract and identify the corresponding entities from elementary data and use them as standard entity objects;

[0008] Obtain information on primary data collection, determine the data quality of primary data based on the collected information and standard entity objects, and obtain intermediate data based on the data quality.

[0009] Heterogeneous information graphs are generated based on medium-sized data. A graph neural network model is selected to process the heterogeneous information graphs, and the heterogeneous information fusion results are obtained.

[0010] Preferably, the steps of acquiring the collection information of primary data, judging the data quality of primary data based on the collection information and standard entity objects, and matching the data quality to obtain intermediate data are as follows:

[0011] Acquire the collection information of elementary data, the collection information including data generation time and collection association information;

[0012] The collected association information includes the source of collection, the occasion of collection, and the environment of collection. The accuracy of the primary data is analyzed based on the collected association information and standard entity objects.

[0013] The validity period of elementary data is determined by the data generation time and the validity period.

[0014] Determine the data type of elementary data and, in conjunction with the data generation time and acquisition environment, determine the completeness of the elementary data.

[0015] Data quality is calculated by setting data weight ratios for accuracy, validity, and completeness, respectively, based on accuracy, validity, completeness, and their corresponding data weight ratios.

[0016] Set data quality standards, and match data that meets the standards from the primary data and record them as intermediate data.

[0017] Preferably, the step of analyzing the accuracy of elementary data based on collected association information and standard entity objects specifically includes:

[0018] Based on the source of the primary data, determine whether the original outputter has subjective consciousness;

[0019] If the original outputter has no subjective consciousness, then obtain the original outputter's historical output records, extract the data accuracy rate from the historical output records, and use the data accuracy rate as the accuracy of the primary data.

[0020] If the original outputter has subjective consciousness, then obtain the original outputter's data output habits, and determine the data subjectivity of the primary data based on the data output habits and the collection context;

[0021] Based on the data output habits, the emotional state of the original outputter is obtained, and the emotional level of the primary data is obtained based on the emotional state.

[0022] The information fusion requirements of standard entity objects are obtained, and the accuracy of primary data is obtained by combining data subjectivity and data sentiment.

[0023] Preferably, the step of obtaining the data output habits of the original outputter if the original outputter has subjective consciousness, and determining the data subjectivity of the primary data based on the data output habits and the collection context, specifically includes:

[0024] Collect words that describe subjective feelings and record them as subjective words;

[0025] Determine whether the elementary data describes objective facts. If the elementary data describes objective facts, then calculate the actual frequency (AP) of subjective terms used in the elementary data.

[0026] Extract the average frequency (AB) of subjective vocabulary used by the original outputter in the data output habits;

[0027] Extract the occurrence time AF of the objective facts described by the primary data, obtain the contact time AJ of the last contact between the original outputter and the objective fact-related information, and the contact frequency AL between the original outputter and the objective fact-related information.

[0028] Collect historical data from various collection scenarios and calculate the data accuracy AZ in the historical data.

[0029] The subjectivity AG of the data is calculated using the subjectivity association function. The subjectivity association function is as follows:

[0030]

[0031] Where a1, a2, and a3 are scaling factors and are greater than 0, and AS is the data generation time;

[0032] If elementary vocabulary does not describe objective facts, then the subjectivity of the data is at its maximum.

[0033] Preferably, the step of analyzing the original outputter's emotional state based on data output habits and obtaining the data sentiment level of the primary data based on the emotional state specifically includes:

[0034] Obtain the degree of emotional fluctuation of the original outputter, calculate the average degree of emotional fluctuation based on the degree of emotional fluctuation, and record it as the original emotional degree BY;

[0035] Vocabulary, symbols, and facial expressions used to express emotions are recorded as emotion terms, and the actual frequency of use of emotion terms in elementary data is statistically analyzed.

[0036] Extract the average frequency of emotional words used by the original outputter in the data output habits, calculate the difference between the actual frequency of emotional words and the average frequency of emotional words, and record it as the actual emotional score BS;

[0037] Extract the objective facts of the original output from the elementary data, and find the fact type of the objective facts;

[0038] Find the degree of emotional fluctuation in the output of information providers of similar factual information, calculate the average value of the emotional fluctuation and record it as the information emotionality BX;

[0039] The sentiment score (BQ) of the data is calculated using the sentiment correlation function. The sentiment correlation function is as follows:

[0040]

[0041] Among them, b1, b2, and b3 are scaling factors and are greater than 0.

[0042] Preferably, the step of obtaining the information fusion requirements of standard entity objects and combining data subjectivity and data sentiment to obtain the accuracy of primary data specifically includes:

[0043] Obtain the information fusion requirements of standard entity objects and determine whether the information fusion requirements only require objective facts;

[0044] If the information fusion requirement only requires objective facts, then a data subjectivity threshold is set, and the primary data that does not reach the data subjectivity threshold is selected and recorded as the first data. The data sentiment of the first data is used as the accuracy of the primary data, and the accuracy of primary data other than the first data is 0.

[0045] If the information fusion requirement is not limited to objective facts, then the first weight ratios for data subjectivity and data sentiment are set separately, and the accuracy of the primary data is calculated based on the data subjectivity, data sentiment and the corresponding first weight ratios.

[0046] Preferably, the step of obtaining the effective duration of elementary data and determining the validity of the elementary data based on the data generation time and the effective duration specifically includes:

[0047] Obtain the effective duration of elementary data, obtain the real-time time, and calculate the time difference between the data generation time and the real-time time.

[0048] Determine whether the time difference exceeds the valid duration. If the time difference exceeds the valid duration, the validity of the elementary data is 0.

[0049] If the time difference does not exceed the valid duration, then elementary data of the same fact type will be grouped into a type data set;

[0050] Extract the actual meaning of data from the type dataset, and extract the most frequent actual meaning as the standard meaning;

[0051] Extract the actual meaning of elementary data, compare the similarity between the actual meaning and the standard meaning, and record it as the validity of elementary data.

[0052] Preferably, the step of obtaining the data type of elementary data and determining the completeness of the elementary data by combining the data generation time and the acquisition environment specifically includes:

[0053] Obtain the necessary attributes of elementary data, and determine whether the elementary data is complete based on the necessary attributes. If the elementary data is incomplete, the completeness of the elementary data is 0.

[0054] If the elementary data is complete, then obtain the data type of the elementary data, extract the basic completeness standard of the data type, and obtain the percentage of elementary data that meets the basic completeness standard and record it as the completeness percentage.

[0055] Determine the degree of interference caused by the data acquisition environment to the elementary data based on the data acquisition environment;

[0056] Set completeness weight ratios for completeness percentage and interference level respectively, and calculate the completeness of elementary data based on completeness percentage, interference level and corresponding completeness weight ratio.

[0057] Preferably, the step of determining the degree of interference caused by the acquisition environment to the primary data based on the acquisition environment specifically includes:

[0058] Obtain the interference factors affecting elementary data, and count the interference factors contained in the collection environment and record them as interference factors;

[0059] Obtain the interference range of the interfering factor, obtain the actual interference value of the interfering factor, and determine the degree of interference based on the actual interference value and interference range;

[0060] The basic level of interference caused by the data acquisition environment is obtained by summing all interference factors and their corresponding levels.

[0061] The transmission duration and number of transmissions of elementary data are statistically analyzed, and the degree of transmission interference is obtained based on the transmission duration and number of transmissions.

[0062] The actual interference level of elementary data is obtained by superimposing the basic interference level and the transmission interference level.

[0063] Secondly, this application provides a heterogeneous information fusion system based on graph neural networks, which adopts the following technical solution:

[0064] A heterogeneous information fusion system based on graph neural networks includes:

[0065] The data processing module collects information data, preprocesses the information data, and uses the preprocessed data as the base data.

[0066] The conversion standard module sets conversion standards for different categories of basic data, and converts the basic data according to the conversion standards to obtain elementary data;

[0067] The entity recognition module extracts and identifies the corresponding entities based on elementary data and uses them as standard entity objects.

[0068] The data quality module acquires the collection information of primary data, judges the data quality of primary data based on the collection information and standard entity objects, and obtains intermediate data based on the data quality.

[0069] The fusion results module generates a heterogeneous information graph based on medium-sized data, selects a graph neural network model to process the heterogeneous information graph, and obtains the heterogeneous information fusion results.

[0070] In summary, this application includes at least one of the following beneficial technical effects:

[0071] 1. The accuracy of primary data is analyzed based on collected correlation information and standard entity objects. Collected correlation information includes the data source, collection context, and collection environment. The validity of primary data is determined by the data generation time and validity period, while the completeness of primary data is determined by combining the data generation time and collection environment. This comprehensive approach yields the overall data quality, which is then matched according to data quality standards. By using the data's collection correlation information to determine accuracy, validity, and completeness, and comprehensively evaluating data quality, the comprehensiveness of heterogeneous information fusion based on graph neural networks is improved.

[0072] 2. Determine whether the original outputter has subjective intent. Different data accuracy levels are obtained based on the presence or absence of subjective intent. If the original outputter has subjective intent, their data output habits are obtained. The subjectivity of the primary data is determined based on the data output habits and the collection context. Furthermore, the sentiment level of the primary data is obtained by analyzing the data output habits. Finally, the accuracy of the primary data is obtained by combining this with the information fusion requirements of standard entity objects. By evaluating the accuracy of the data based on subjectivity and sentiment, the influence of the original outputter on data accuracy is fully considered, thus improving the accuracy of heterogeneous information fusion based on graph neural networks.

[0073] 3. The original emotional score is calculated based on the degree of emotional fluctuation of the original outputter, and the actual emotional score is obtained based on the use of emotional terms, including words, symbols, and facial expressions that express emotions. Simultaneously, the output emotional scores of information outputters with similar objective facts to those of the original outputter are compared to obtain the information emotional score. Furthermore, the information emotional score is evaluated from multiple perspectives, taking into account not only the emotional influence of the original outputter but also the inverse influence of the objective facts in the information data on the original outputter's emotions. This ultimately confirms the data emotional score, improving the practicality of heterogeneous information fusion based on graph neural networks. Attached Figure Description

[0074] Figure 1This is a schematic diagram illustrating the specific steps of an embodiment of a heterogeneous information fusion method based on graph neural networks according to the present invention.

[0075] Figure 2 This is a schematic diagram of the module connections of an embodiment of a heterogeneous information fusion system based on graph neural networks according to the present invention. Detailed Implementation

[0076] The following examples and... Figures 1-2 The present invention will be described in further detail, but the embodiments of the present invention are not limited thereto.

[0077] This invention discloses a heterogeneous information fusion method based on graph neural networks, specifically including the following steps:

[0078] Step S1: Collect information data, preprocess the information data, and use the preprocessed data as the basic data.

[0079] Heterogeneous data typically comes from different data sources and may contain issues such as missing data, duplication, and noise. Preprocessing the data, or data cleaning, involves identifying fundamental problems like duplication and missing information.

[0080] Step S2: Set conversion standards for different categories of basic data, and convert the basic data according to the conversion standards to obtain elementary data.

[0081] Heterogeneous data may have different formats, structures, and units. Data transformation requires unifying these different formats, structures, and units to facilitate subsequent analysis and processing. For example, converting location information into a unified geographic coordinate system.

[0082] Step S3: Extract and identify the corresponding entities based on the elementary data and use them as standard entity objects.

[0083] When integrating heterogeneous data, it is necessary to identify and match the same entities or attributes in different data sources, such as user IDs and mobile phone numbers.

[0084] Step S4: Obtain the collection information of primary data, determine the data quality of primary data based on the collection information and standard entity objects, and obtain intermediate data based on the data quality.

[0085] Step S5: Generate a heterogeneous information graph based on the medium data, select a graph neural network model to process the heterogeneous information graph, and obtain the heterogeneous information fusion result.

[0086] Graph neural networks (GNNs) use neural networks to learn graph-structured data, extracting and discovering features and patterns within it. GNNs leverage these heterogeneous relationships, integrating node information from different relationships through specific fusion mechanisms (such as attention mechanisms and message passing mechanisms) to generate richer node representations. The fused information is then further integrated using a selected GNN to obtain more accurate and subjective results in the fusion of heterogeneous information.

[0087] In practical applications, after heterogeneous information is fused, further analysis and mining of the fused data are usually required to obtain valuable information. However, if there are differences in data quality, the accuracy of the analysis and mining will also be affected. For example, if there is noise or bias in the data, the analysis results may lead to misleading conclusions. Therefore, in the process of heterogeneous information fusion, data of poor quality can be filtered and removed based on data quality matching information, thereby improving the effectiveness and accuracy of heterogeneous information fusion.

[0088] The steps for obtaining primary data collection information, determining the data quality of primary data based on the collection information and standard entity objects, and matching the data quality to obtain intermediate data are as follows:

[0089] Step S41: Obtain the collection information of primary data, including data generation time and collection association information.

[0090] The data generation time is the point in time when the data is output. For example, if the data collected is user voice, then the point in time when the user speaks is the data generation time of the user voice.

[0091] Step S42: Collect related information, including the source of collection, the occasion of collection, and the environment of collection. Analyze the accuracy of the primary data based on the collected related information and standard entity objects.

[0092] Step S43: Obtain the effective duration of the elementary data, and determine the validity of the elementary data based on the data generation time and the effective duration.

[0093] Step S44: Obtain the data type of the primary data, and determine the completeness of the primary data by combining the data generation time and the acquisition environment.

[0094] Step S45: Set the data weight ratios for accuracy, validity, and completeness respectively, and calculate the data quality based on accuracy, validity, completeness, and the corresponding data weight ratios.

[0095] Step S46: Set data quality standards, and match data that meets the data quality standards from the primary data and record them as intermediate data.

[0096] For example, by setting a data quality threshold, it can be assumed that primary data that meets the data quality threshold can be matched with standard entity objects, and these data are recorded as medium data.

[0097] In practical applications, the accuracy, validity, and completeness of data all affect data quality. Inaccurate data can directly lead to flawed decision-making and even serious consequences. For example, if the standard entity is a user, inaccurate location data will severely impact subsequent user recommendations and business processing. During data entry, if a field requires an email address, but the user enters a string of gibberish or an incorrect email address, that field is invalid. Such invalid data will affect subsequent data processing and analysis. Similarly, incomplete data will also hinder subsequent data analysis and information fusion.

[0098] The steps for analyzing the accuracy of elementary data based on collected related information and standard entity objects are as follows:

[0099] Step S421: Based on the source of the primary data, determine whether the original outputter has subjective consciousness.

[0100] Subjective consciousness refers to the uniqueness and individuality exhibited by an individual when thinking, perceiving, and experiencing the world. It involves personal experience, emotions, and cognition, and is contrasted with objective facts and logical analysis. Therefore, if the original output is a person, it possesses subjective consciousness; however, if the original output is a machine such as a sensor or detector, it does not possess subjective consciousness. For example, if the source of the 25-degree Celsius temperature reading is a sensor, then the original output is the sensor, and it does not possess subjective consciousness.

[0101] Step S422: If the original outputter has no subjective consciousness, then obtain the original outputter's historical output records, extract the data accuracy rate from the historical output records, and use the data accuracy rate as the accuracy of the primary data.

[0102] If the original outputter has no subjective intent, the output data is not subject to subjective influence. In this case, the accuracy of the primary data is obtained based on the actual data accuracy rate. For example, a sensor detects temperature and outputs temperature data. Data errors arise from reasons such as the sensor's accuracy or malfunction, but the sensor itself will not subjectively tamper with the temperature data.

[0103] Step S423: If the original outputter has subjective consciousness, then obtain the original outputter's data output habits, and determine the data subjectivity of the primary data based on the data output habits and the collection context.

[0104] Step S424: Analyze the original outputter's emotional state based on data output habits, and obtain the data emotion level of the primary data based on the emotional state.

[0105] Step S425: Obtain the information fusion requirements of standard entity objects, and combine data subjectivity and data sentiment to obtain the accuracy of primary data.

[0106] In practical applications, data accuracy is influenced by the outputter. If the outputter is free from subjective bias, the accuracy of the primary data is determined by the actual accuracy of their output. However, if the outputter is subjective, the accuracy will be affected by their subjective feelings and emotions. In the example, let's consider a user as the original outputter. For instance, if user A says the temperature today is 26 degrees Celsius, due to their subjective feelings, they might exaggerate the temperature, making it less than 26 degrees Celsius. Therefore, when judging data accuracy, it's necessary to consider the degree of subjectivity and emotional influence on the data, and to combine this with standard entity objects.

[0107] If the original data outputter has subjective intent, the steps to determine the subjectivity of the primary data based on their data output habits and the data collection context are as follows:

[0108] Step S4231: Collect words that describe subjective feelings and record them as subjective words.

[0109] Subjective consciousness refers to the uniqueness and individuality exhibited by an individual when thinking, perceiving, and experiencing the world. It involves personal experience, emotions, and cognition, and is contrasted with objective facts and logical analysis. Examples include words like "like," "feel," and "believe."

[0110] Step S4232: Determine whether the elementary data is describing objective facts. If the elementary data is describing objective facts, then calculate the actual frequency (AP) of subjective words used in the elementary data.

[0111] Objective facts can be verified by scientific methods or logical reasoning and are not influenced by personal subjective will; they exist independently in time and space. For example, "Today is October 25, 2024" is an objective fact because it is not affected by personal feelings or opinions, and models can be trained to determine whether it is an objective fact.

[0112] Step S4233: Extract the average frequency AB of subjective words used by the original outputter in the data output habits.

[0113] Step S4234: Extract the occurrence time AF of the objective facts described by the primary data, obtain the contact time AJ of the last contact between the original outputter and the objective fact-related information, and the contact frequency AL between the original outputter and the objective fact-related information.

[0114] Step S4235: Collect historical data from the collection scenarios and calculate the data accuracy AZ in the historical data.

[0115] The accuracy of information collected varies depending on the context. For example, information collected in everyday situations may contain errors due to users joking around, resulting in lower accuracy. Higher accuracy indicates fewer subjective errors, which we consider to be lower subjectivity. In more serious settings, such as academic research, the accuracy of the collected data will be even higher.

[0116] Step S4236: Calculate the data subjectivity AG using the subjectivity correlation function. Subjectivity correlation function:

[0117]

[0118] Where a1, a2, and a3 are scaling factors and are greater than 0, and AS is the data generation time.

[0119] Step S4237: If elementary vocabulary is not describing objective facts, then the subjectivity of the data is at its maximum value.

[0120] If the original output is not describing objective facts, it is considered that the original output is describing personal feelings. For example, "I think this machine is very useful" is not describing objective facts, but only expressing the user's feelings. Therefore, the information is entirely derived from the user's subjectivity, and the subjectivity is at its maximum.

[0121] In practical applications, the greater the subjectivity of information, the less objective it is, and the greater the influence of the original contributor's subjectivity, thus reducing the accuracy of the information. When subjective words like "think" or "feel" appear in information, the information is mixed with the describer's subjective judgment, reducing its credibility. However, due to different users' language habits, some users habitually add subjective words to their descriptions, but in reality, there isn't much subjective content. Therefore, it's necessary to confirm this based on the average frequency of users' daily use of subjective words. When the actual frequency is greater than the average frequency, it indicates that the user has added more subjective content, resulting in greater subjectivity. Simultaneously, when users describe objective facts, due to the varying times when these facts occurred, and the users' memories being less vivid, they automatically fill in the gaps with subjective content, thus including subjective elements in the information. The greater the time gap between the objective facts and the user's description, the more blurred the memory becomes. Furthermore, if the user's last contact was a long time ago, the memory is also more blurred, resulting in greater subjectivity. Conversely, the higher the frequency of contact, the deeper the memory, and the lower the subjectivity. For example, if User A is the original outputter and their output describes the process of Experiment A, which takes a long time to complete, and User A has not been exposed to any related information about Experiment A, then the subjective nature of User A's data is high. On the other hand, User B frequently watches videos of Experiment A's procedures, therefore User B's data has a lower degree of subjectivity.

[0122] The steps for analyzing the emotional state of the original contributor based on data output habits, and then obtaining the data sentiment level of the primary data based on that emotional state, are as follows:

[0123] Step S4241: Obtain the emotional fluctuation level of the original outputter, calculate the average emotional fluctuation level based on the emotional fluctuation level and record it as the original emotional level BY.

[0124] The degree of emotional fluctuation of the original sender can be determined by physiological information such as heart rate, or by authorized information such as facial expressions. For example, the average difference between the original sender's heart rate and the normal average heart rate can be calculated to obtain the average degree of emotional fluctuation.

[0125] Step S4242: Record the words, symbols and expressions that express emotions as emotion terms, and count the actual usage frequency of emotion terms in the elementary data.

[0126] Some words are used to express the emotions of the original sender, such as "awkward," "bored," "anxious," etc.

[0127] Step S4243: Extract the average frequency of emotional words used by the original outputter in the data output habits, calculate the difference between the actual frequency of emotional words and the average frequency of emotional words, and record it as the actual emotional score BS.

[0128] Some users frequently use emotional words in their daily lives, so using emotional words doesn't necessarily mean the user has emotions. For example, on the internet, users use smiley faces to express friendliness, but they don't actually have any emotion; they're just adding smiley faces habitually. Similarly, typing "hahahahaha" is simply a message response, while typing "hahahahahahahahahahaha" expresses happiness. Therefore, the frequency of emotional word usage can be used to determine a user's actual emotional level.

[0129] Step S4244: Extract the objective facts of the original output from the primary data and find the fact type of the objective facts.

[0130] Extract the objective facts given by the original outputter and find similar objective facts. For example, if the objective fact described by the original outputter is that today's weather is sunny, then the fact type of the objective fact is weather.

[0131] Step S4245: Find the output emotional fluctuation level of information outputters of the same type of factual information, calculate the average value of the output emotional fluctuation level and record it as the information emotional level BX.

[0132] The degree of emotional fluctuation in the output of information providers is obtained by using the actual emotional intensity method, and then the average value of the emotional fluctuation in the output of all information providers is calculated to obtain the information emotional intensity.

[0133] The emotions conveyed in information data are influenced not only by the individual but also by the objective facts being described. For example, when describing an objective fact such as a date, most contributors will exhibit relatively neutral emotions. However, when describing an objective fact such as the loss of a valuable item, most contributors will show anxiety and greater emotional fluctuations. If describing this information generally causes emotional fluctuations among contributors, then the information carries a greater degree of emotion.

[0134] Step S4246: Calculate the data sentiment score BQ using the sentiment correlation function. Sentiment correlation function:

[0135]

[0136] Among them, b1, b2, and b3 are scaling factors and are greater than 0.

[0137] In practical applications, besides the subjectivity of the original contributor affecting the accuracy of information, the contributor's emotions also influence data accuracy. Emotional state affects an individual's cognitive processes and information processing methods; heightened emotions can lead to information bias, thus affecting accuracy. Conversely, the further back in time the objective facts occurred, the calmer the user's emotions tend to be, resulting in less emotional involvement in their descriptions. A higher level of initial emotional intensity indicates greater susceptibility to emotional influence, making it easier for the user to transmit personal emotions during information output.

[0138] The steps for obtaining information fusion requirements for standard entity objects and combining data subjectivity and data sentiment to determine the accuracy of primary data are as follows:

[0139] Step S4251: Obtain the information fusion requirements of the standard entity object and determine whether the information fusion requirements only require objective facts.

[0140] Different industries have different needs for heterogeneous information fusion. For example, in the field of intelligent manufacturing, heterogeneous system integration technology can break down barriers between different devices and systems, enabling seamless data connection and flow. In this case, heterogeneous information fusion only requires objective facts, i.e., accurate data. However, in the field of product recommendation, heterogeneous information fusion needs to obtain the user's actual experience. Therefore, the user's subjective feelings and emotions play a crucial role in product recommendations, thus requiring subjective data and emotional data.

[0141] Step S4252: If the information fusion requirement only requires objective facts, then set a data subjectivity threshold, filter out the primary data that does not reach the data subjectivity threshold and record it as the first data, use the data sentiment of the first data as the accuracy of the primary data, and the accuracy of primary data other than the first data is 0.

[0142] Due to subjective factors, the supplementary content may deviate from or contain errors in relation to the actual events. If only objective facts are required, and subjective content does not represent objective facts, then data with a high degree of subjectivity should be filtered out to reduce non-objective information. Emotions can also affect the accuracy of objective descriptions; therefore, after filtering data based on subjectivity, the emotional level of the data is used as a primary measure of accuracy.

[0143] Step S4253: If the information fusion requirement is not only objective facts, then set the first weight ratio for data subjectivity and data sentiment respectively, and calculate the accuracy of the primary data based on the data subjectivity, data sentiment and the corresponding first weight ratio.

[0144] In practical applications, the accuracy of data needs to be assessed based on the specific circumstances. For information fusion that only requires objective facts, the subjectivity and emotions of the original contributor become interfering factors, thus reducing the accuracy of information mixed with subjectivity and emotion. However, for information fusion that requires the original contributor's subjectivity and emotion, information mixed with subjectivity and emotion can provide more valuable information, thus increasing its accuracy.

[0145] The steps for obtaining the validity period of elementary data and determining the validity of elementary data based on the data generation time and validity period are as follows:

[0146] Step S431: Obtain the effective duration of the primary data, obtain the real-time time, and calculate the time difference between the data generation time and the real-time time.

[0147] Step S432: Determine whether the time difference exceeds the valid duration. If the time difference exceeds the valid duration, the validity of the elementary data is 0.

[0148] If the data has exceeded its validity period, it means the data is invalid and no longer useful, therefore its validity is 0.

[0149] Step S433: If the time difference does not exceed the valid duration, then the elementary data of the same fact type are formed into a type data set.

[0150] Among the collected data, many data points express the same fact. For example, if a user enters address information on platform 1 and a user also enters address information on platform 2, then both pieces of information describe the user's address, and therefore the fact type is the same.

[0151] Step S434: Extract the actual meaning of the data in the type data set, and extract the most frequent actual meaning as the standard meaning.

[0152] Data of the same type describes an entity or fact, but they may also differ. For example, user A is called "Zhang San" on platform 1, "Zhang San" on platform 2, and "Mr. Zhang" on platform 3. Therefore, the most common meaning is that the user's name is "Zhang San", which is the standard meaning of "Zhang San" as a user's name.

[0153] Step S435: Extract the actual meaning of the elementary data, compare the similarity between the actual meaning and the standard meaning, and record it as the validity of the elementary data.

[0154] In practical applications, the greater the deviation of the data, the lower the validity of the data. For example, if user A's name on platform number four is "Sunflower is Singing", the similarity to "Zhang San" is too low. Obviously, the deviation from the user's name is too large, and it cannot provide effective information about the user's name. Therefore, the validity is low.

[0155] The steps for determining the data type of elementary data and assessing its completeness based on the data generation time and acquisition environment are as follows:

[0156] Step S441: Obtain the essential attributes of the elementary data, and determine whether the elementary data is complete based on the essential attributes. If the elementary data is incomplete, the completeness of the elementary data is 0.

[0157] Different types of data have their essential attributes; only data with these essential attributes is considered complete. For example, for text data, the presence of periods, exclamation marks, and question marks indicates a complete sentence. For temperature acquisition devices, the data is considered complete only if it ends in degrees Celsius.

[0158] Step S442: If the elementary data is complete, obtain the data type of the elementary data, extract the basic completeness standard of the data type, and obtain the percentage of elementary data that meets the basic completeness standard and record it as the completeness percentage.

[0159] Essential attributes only indicate the basic completeness of the data; they don't necessarily represent completeness for information fusion. For example, address information can be divided into different smallest units, such as district, building, and house number. Based on the data type of the primary data, the basic completeness standard generally required for information fusion is determined. The completeness percentage is obtained based on the proportion of the primary data containing the basic completeness standard data. For example, if the basic completeness standard is province, city, district, town, and building, but the information only goes up to the district, then the completeness percentage is 60%.

[0160] Step S443: Determine the degree of interference caused by the acquisition environment to the primary data based on the acquisition environment.

[0161] Step S444: Set the completeness weight ratio for completeness percentage and interference level respectively, and calculate the completeness of elementary data based on completeness percentage, interference level and corresponding completeness weight ratio.

[0162] In practical applications, incomplete data can lead to incomplete information, thus affecting the effectiveness of information fusion and further impacting its usability. For example, in predictive models, if complete driver information (such as age and gender) is missing, the model's prediction results may be severely affected.

[0163] The steps for determining the degree of interference caused by the acquisition environment to elementary data based on the acquisition environment are as follows:

[0164] Step S4431: Obtain the interference factors affecting elementary data, and count the interference factors contained in the collection environment and record them as interference factors.

[0165] Interference factors are those that interfere with primary data. For example, when collecting user voice data, noise will interfere with the user's voice, so noise is an interference factor. If there is no noise in the collection environment, then noise is not an interference factor. If there is noise, then noise is an interference factor.

[0166] Step S4432: Obtain the interference range of the interfering factor, obtain the actual interference value of the interfering factor, and obtain the degree of the interfering factor based on the actual interference value and the interference range.

[0167] Interference range refers to the range of interference caused by interfering factors. For example, the interference range of noise is 50 decibels to 100 decibels. If the actual interference value is 80 decibels, then (80-50) / (100-50) = 60%, so the interference factor level is 60%.

[0168] Step S4433: Summing up all interference factors and their corresponding degrees of interference to obtain the basic interference level caused by the data acquisition environment.

[0169] The basic interference level is obtained by superimposing the values ​​of the levels of all interference factors.

[0170] Step S4434: Calculate the transmission duration and number of transmissions of elementary data, and obtain the degree of transmission interference based on the transmission duration and number of transmissions.

[0171] Step S4435: Superimpose the basic interference level and the transmission interference level to obtain the actual interference level of the elementary data.

[0172] In practical applications, when data needs to be transmitted multiple times, each transmission carries a certain risk of data loss. For example, in UDP transmission, without a proper retransmission mechanism, data packets may be lost due to network instability or other reasons. The cumulative risk of data loss increases with the number of transmissions. Longer transmission times may also increase the risk of data loss. For instance, in unstable network conditions, prolonged transmissions are more susceptible to network interruptions or data packet loss. Data loss severely impacts data integrity; data preprocessing can only remove simple missing data. Some data lost during transmission may not be detected, therefore, the longer the transmission duration and the more transmissions, the lower the data integrity.

[0173] A heterogeneous information fusion system based on graph neural networks, which applies the above-described heterogeneous information fusion method based on graph neural networks, includes:

[0174] The data processing module collects information data, preprocesses the information data, and uses the preprocessed data as the base data.

[0175] The conversion standard module sets conversion standards for different categories of basic data, and converts the basic data according to the conversion standards to obtain elementary data.

[0176] The entity recognition module extracts and identifies the corresponding entities based on elementary data and uses them as standard entity objects.

[0177] The data quality module acquires the collection information of primary data, judges the data quality of primary data based on the collection information and standard entity objects, and obtains intermediate data based on the data quality.

[0178] The fusion results module generates a heterogeneous information graph based on medium-sized data, selects a graph neural network model to process the heterogeneous information graph, and obtains the heterogeneous information fusion results.

[0179] The implementation principle of this system is as follows: First, the data processing module collects information data, preprocesses it, and uses the preprocessed data as basic data. The conversion standard module sets conversion standards for different categories of basic data and converts the basic data according to these standards to obtain primary data. The entity recognition module extracts and identifies corresponding entities from the primary data and uses them as standard entity objects. The data quality module obtains the collection information of the primary data, including data generation time and collection association information, including the collection source, collection occasion, and collection environment. Based on the collection association information and standard entity objects, the accuracy of the primary data is analyzed. The validity of the primary data is determined based on the data generation time and effective duration, and the completeness of the primary data is determined by combining the data generation time and collection environment. The data quality is calculated by comprehensively considering accuracy, validity, and completeness, and intermediate data is obtained based on the data quality matching. Finally, the fusion result module forms a heterogeneous information graph based on the intermediate data, selects a graph neural network model to process the heterogeneous information graph, and obtains the heterogeneous information fusion result.

[0180] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A heterogeneous information fusion method based on graph neural networks, characterized in that, Includes the following steps: Collect information data, preprocess the information data, and use the preprocessed data as the base data; Set conversion standards for different categories of basic data, and convert the basic data according to the conversion standards to obtain elementary data; Extract and identify the corresponding entities from elementary data and use them as standard entity objects; Obtain information on primary data collection, determine the data quality of primary data based on the collected information and standard entity objects, and obtain intermediate data based on the data quality. A heterogeneous information graph is formed based on medium-sized data. A graph neural network model is selected to process the heterogeneous information graph to obtain the heterogeneous information fusion result. The steps of acquiring information from primary data collection, determining the data quality of primary data based on the collected information and standard entity objects, and matching the data quality to obtain intermediate data are as follows: Acquire the collection information of elementary data, the collection information including data generation time and collection association information; The collected association information includes the source of collection, the occasion of collection, and the environment of collection. The accuracy of the primary data is analyzed based on the collected association information and standard entity objects. The validity period of elementary data is determined by the data generation time and the validity period. Determine the data type of elementary data and, in conjunction with the data generation time and acquisition environment, determine the completeness of the elementary data. Data quality is calculated by setting data weight ratios for accuracy, validity, and completeness, respectively, based on accuracy, validity, completeness, and their corresponding data weight ratios. Set data quality standards, and match data that meets the data quality standards from the primary data and record them as medium data; The step of analyzing the accuracy of elementary data based on collected related information and standard entity objects specifically includes: Based on the source of the primary data, determine whether the original outputter has subjective consciousness; If the original outputter has no subjective consciousness, then obtain the original outputter's historical output records, extract the data accuracy rate from the historical output records, and use the data accuracy rate as the accuracy of the primary data. If the original outputter has subjective consciousness, then obtain the original outputter's data output habits, and determine the data subjectivity of the primary data based on the data output habits and the collection context; Based on the data output habits, the emotional state of the original outputter is obtained, and the emotional level of the primary data is obtained based on the emotional state. The information fusion requirements of standard entity objects are obtained, and the accuracy of primary data is obtained by combining data subjectivity and data sentiment. The step of obtaining the data subjectivity of the primary data based on the data output habits of the original outputter and the data collection context, if the original outputter has subjective consciousness, is as follows: Collect words that describe subjective feelings and record them as subjective words; Determine whether the elementary data describes objective facts. If the elementary data describes objective facts, then calculate the actual frequency (AP) of subjective terms used in the elementary data. Extract the average frequency (AB) of subjective vocabulary used by the original outputter in the data output habits; Extract the occurrence time AF of the objective facts described by the primary data, obtain the contact time AJ of the last contact between the original outputter and the objective fact-related information, and the contact frequency AL between the original outputter and the objective fact-related information. Collect historical data from various collection scenarios and calculate the data accuracy AZ in the historical data. The subjectivity AG of the data is calculated using the subjectivity association function. The subjectivity association function is as follows: in, , , The scaling factor is greater than 0, and AS is the data generation time. If elementary vocabulary does not describe objective facts, then the subjectivity of the data is at its maximum.

2. The heterogeneous information fusion method based on graph neural networks according to claim 1, characterized in that, The steps of analyzing the original outputter's emotional state based on data output habits and obtaining the data sentiment level of the primary data based on the emotional state are as follows: Obtain the degree of emotional fluctuation of the original outputter, calculate the average degree of emotional fluctuation based on the degree of emotional fluctuation, and record it as the original emotional degree BY; Vocabulary, symbols, and facial expressions used to express emotions are recorded as emotion terms, and the actual frequency of use of emotion terms in elementary data is statistically analyzed. Extract the average frequency of emotional words used by the original outputter in the data output habits, calculate the difference between the actual frequency of emotional words and the average frequency of emotional words, and record it as the actual emotional score BS; Extract the objective facts of the original output from the elementary data, and find the fact type of the objective facts; Find the degree of emotional fluctuation in the output of information providers of similar factual information, calculate the average value of the emotional fluctuation and record it as the information emotionality BX; The sentiment score (BQ) of the data is calculated using the sentiment correlation function. The sentiment correlation function is as follows: in, , , It is a scaling factor and is greater than 0.

3. The heterogeneous information fusion method based on graph neural networks according to claim 2, characterized in that, The steps for obtaining the accuracy of primary data by combining the information fusion requirements of standard entity objects with data subjectivity and data sentiment are as follows: Obtain the information fusion requirements of standard entity objects and determine whether the information fusion requirements only require objective facts; If the information fusion requirement only requires objective facts, then a data subjectivity threshold is set, and the primary data that does not reach the data subjectivity threshold is selected and recorded as the first data. The data sentiment of the first data is used as the accuracy of the primary data, and the accuracy of primary data other than the first data is 0. If the information fusion requirement is not limited to objective facts, then the first weight ratios for data subjectivity and data sentiment are set separately, and the accuracy of the primary data is calculated based on the data subjectivity, data sentiment and the corresponding first weight ratios.

4. The heterogeneous information fusion method based on graph neural networks according to claim 3, characterized in that, The step of obtaining the effective duration of elementary data and determining the validity of the elementary data based on the data generation time and the effective duration is as follows: Obtain the effective duration of elementary data, obtain the real-time time, and calculate the time difference between the data generation time and the real-time time. Determine whether the time difference exceeds the valid duration. If the time difference exceeds the valid duration, the validity of the elementary data is 0. If the time difference does not exceed the valid duration, then elementary data of the same fact type will be grouped into a type data set; Extract the actual meaning of data from the type dataset, and extract the most frequent actual meaning as the standard meaning; Extract the actual meaning of elementary data, compare the similarity between the actual meaning and the standard meaning, and record it as the validity of elementary data.

5. The heterogeneous information fusion method based on graph neural networks according to claim 4, characterized in that, The step of obtaining the data type of elementary data and determining the completeness of the elementary data by combining the data generation time and the acquisition environment is as follows: Obtain the necessary attributes of elementary data, and determine whether the elementary data is complete based on the necessary attributes. If the elementary data is incomplete, the completeness of the elementary data is 0. If the elementary data is complete, then obtain the data type of the elementary data, extract the basic completeness standard of the data type, and obtain the percentage of elementary data that meets the basic completeness standard and record it as the completeness percentage. Determine the degree of interference caused by the data acquisition environment to the elementary data based on the data acquisition environment; Set completeness weight ratios for completeness percentage and interference level respectively, and calculate the completeness of elementary data based on completeness percentage, interference level and corresponding completeness weight ratio.

6. The heterogeneous information fusion method based on graph neural networks according to claim 5, characterized in that, The step of determining the degree of interference caused by the acquisition environment to the primary data based on the acquisition environment is as follows: Obtain the interference factors affecting elementary data, and count the interference factors contained in the collection environment and record them as interference factors; Obtain the interference range of the interfering factor, obtain the actual interference value of the interfering factor, and determine the degree of interference based on the actual interference value and interference range; The basic level of interference caused by the data acquisition environment is obtained by summing all interference factors and their corresponding levels. The transmission duration and number of transmissions of elementary data are statistically analyzed, and the degree of transmission interference is obtained based on the transmission duration and number of transmissions. The actual interference level of elementary data is obtained by superimposing the basic interference level and the transmission interference level.

7. A heterogeneous information fusion system based on graph neural networks, characterized in that, By applying the heterogeneous information fusion method based on graph neural networks as described in any one of claims 1-6, the method includes: The data processing module collects information data, preprocesses the information data, and uses the preprocessed data as the base data. The conversion standard module sets conversion standards for different categories of basic data, and converts the basic data according to the conversion standards to obtain elementary data; The entity recognition module extracts and identifies the corresponding entities based on elementary data and uses them as standard entity objects. The data quality module acquires the collection information of primary data, judges the data quality of primary data based on the collection information and standard entity objects, and obtains intermediate data based on the data quality. The fusion results module generates a heterogeneous information graph based on medium-sized data, selects a graph neural network model to process the heterogeneous information graph, and obtains the heterogeneous information fusion results. The steps for obtaining primary data collection information, determining the data quality of primary data based on the collection information and standard entity objects, and matching the data quality to obtain intermediate data are as follows: Obtain information on the collection of elementary data, including data generation time and collection-related information. The collection of related information includes the source of collection, the occasion of collection, and the environment of collection. The accuracy of primary data is analyzed based on the collection of related information and standard entity objects. The validity period of elementary data is determined by the data generation time and the validity period. Determine the data type of elementary data and, in conjunction with the data generation time and acquisition environment, determine the completeness of the elementary data. Data quality is calculated by setting data weight ratios for accuracy, validity, and completeness, respectively, based on accuracy, validity, completeness, and their corresponding data weight ratios. Set data quality standards, and match data that meets the data quality standards from the primary data and record them as medium data; The steps for analyzing the accuracy of elementary data based on collected related information and standard entity objects are as follows: Based on the source of the primary data, determine whether the original outputter has subjective consciousness; If the original outputter has no subjective consciousness, then obtain the original outputter's historical output records, extract the data accuracy rate from the historical output records, and use the data accuracy rate as the accuracy of the primary data. If the original outputter has subjective consciousness, then obtain the original outputter's data output habits, and determine the data subjectivity of the primary data based on the data output habits and the collection context; Based on the data output habits, the emotional state of the original outputter is obtained, and the emotional level of the primary data is obtained based on the emotional state. The information fusion requirements of standard entity objects are obtained, and the accuracy of primary data is obtained by combining data subjectivity and data sentiment. If the original data outputter has subjective intent, the steps to determine the subjectivity of the primary data based on their data output habits and the data collection context are as follows: Collect words that describe subjective feelings and record them as subjective words; Determine whether the elementary data describes objective facts. If the elementary data describes objective facts, then calculate the actual frequency (AP) of subjective terms used in the elementary data. Extract the average frequency (AB) of subjective vocabulary used by the original outputter in the data output habits; Extract the occurrence time AF of the objective facts described by the primary data, obtain the contact time AJ of the last contact between the original outputter and the objective fact-related information, and the contact frequency AL between the original outputter and the objective fact-related information. Collect historical data from various collection scenarios and calculate the data accuracy AZ in the historical data. The subjectivity AG of the data is calculated using the subjectivity association function. The subjectivity association function is as follows: in, , , The scaling factor is greater than 0, and AS is the data generation time. If elementary vocabulary does not describe objective facts, then the subjectivity of the data is at its maximum.

Images