Dynamic truth value discovery model method for non-uniformly distributed data

Abstract

The invention relates to a dynamic truth value discovery model method for non-uniformly distributed data. The method comprises the following steps: acquiring a data source and a data object; defining a truth value discovery problem as an optimization problem at each current timestamp before the timestamp; based on a coordinate descent method, obtaining a solution formula of an estimated truth value; measuring the confidence level of a truth value by adopting an interval estimation method, converting an optimization problem into likelihood estimation by using a maximum posteriori estimation method, and obtaining a confidence source weight by using the variance of a source error; and after the timestamps, for each current timestamp, calculating a data standard deviation, and adopting a truth value evaluation mode of a stability mode and a volatility mode to obtain a truth value of the current timestamp. The problem that an existing dynamic truth value discovery method is prone to being limited by a single model, and huge overhead is generated is solved. The method considers the distribution and occurrence frequency of data points, can more accurately estimate the authenticity of each data, and is superior to a current truth value discovery method in precision and calculation efficiency.

Description

Technical Field

[0001] The present invention relates to a truth value discovery technology, and more particularly to a dynamic truth value discovery model method for unevenly distributed data. The dynamic truth value discovery model method can be applied to multiple fields, such as finance, medical care, social media analysis, and industrial manufacturing. It can help identify abnormal behaviors, discover hidden patterns, predict trends, and support decision making. Background Art

[0002] As an effective method to resolve data conflicts, truth value discovery has been widely used in many fields such as big data processing and knowledge base construction [7-10]. Domestic and foreign scholars have conducted a lot of research on the truth value discovery problem [1]. The original methods are usually rule-based, such as majority voting and averaging. However, in fact, data sources are not equally reliable, so these methods perform poorly when a large number of data sources provide low-quality data.

[0003] The truth discovery problem was first formally defined by Yin et al. [2], who proposed the TruthFinder model, which uses Bayesian distribution to iteratively estimate the reliability and truth of data sources. Since then, a large number of advanced methods have been proposed, all of which follow the following principle: statements provided by reliable data sources are often more likely to be credible, and data sources that always provide credible statements are often more reliable [1]. Pasternack et al. [3] proposed the Investment model, in which data sources uniformly "invest" their reliability in their declared values ​​and determine the truth value through weighted voting. Li et al. [4] proposed the CRH model, which normalizes the truth discovery task into an optimization framework to minimize the overall weighted distance between the declared value and the aggregated truth value. Meir et al. [5] proposed the P-TD model, which estimates the ability of workers by using the average distance to other workers. Allouche et al. [6] gave more weight to votes with less noise scale based on the Condorcet noise model when estimating the truth value. Wang et al. [7] proposed the TVD-RA framework, which uses an incentive mechanism based on reverse auctions to exclude malicious workers to improve data quality, thereby improving the accuracy of truth discovery. Wu et al. [8] proposed the TIRA model, which uses a hierarchical graph autoencoder based on object-source reliability aggregation to capture multiple crowdsourcing relationships and map these relationships into object representations for truth discovery. Tian et al. [9] considered inconsistent attribute credibility, inconsistent object reliability, and inconsistent object group reliability for truth discovery. The CATD model

[10] introduced confidence intervals and combined them with source weight estimation, which to some extent solved the impact of the long tail phenomenon. The ETCIBoot model

[11] used confidence intervals to consider the long tail phenomenon of objects based on the previous work and applied the model to distributed scenarios to process large-scale data. All of the above works are designed based on static data, but there are still a lot of dynamically changing time series data in reality. In dynamic scenarios, the truth value and source weight will change dynamically over time. If the static method is directly used on dynamic data, the truth value discovery is performed once at each timestamp, which will incur a huge computational overhead.

[0004] At present, some dynamic truth value discovery methods have been proposed. Li et al.

[12] calculated the truth value based on the optimization model, performed only one calculation at each timestamp, and calculated the source weight by the cumulative error of the current timestamp, which effectively reduced the computational cost. Pang et al.

[13] proposed a personalized privacy-preserving truth value discovery framework on crowdsourcing data streams, which can accurately infer the truth value in real time while meeting the user's personal privacy requirements. Tan Long et al.

[14] combined the Map-Reduce framework and the Bayesian truth value discovery model and implemented an efficient big data incremental truth value discovery algorithm using the Incoop incremental framework. Li Tianyi et al.

[15] transformed the problem into an optimization problem by maximizing the credibility evaluation cycle of the data source under the premise that the cumulative error of the prediction does not exceed a given threshold, thereby improving efficiency. Yao et al.

[16] proposed an OTD model based on an online algorithm to capture the temporal patterns in time series data, including smooth evolution patterns and seasonal trends. While modeling the time series, it helps the truth value discovery component improve the accuracy of the truth value estimation. Zhi et al.

[17] proposed the EvolvT model, which uses a hidden Markov model to capture the dynamics of truth values ​​with potential temporal correlations and considers source dependencies to process numerical dynamic data. Zhang et al.

[18] used a dynamic truth discovery framework in social media perception scenarios to identify the true values ​​of measured variables that have not been reported by exploring the latent and lagged correlations between variables. Li et al.

[19] proposed the ASRA model, which uses an iteration-based method to calculate source weights only at specific timestamps to reduce the iteration frequency. The model uses previously calculated source weights to analyze the error of inferring the true value of the object. If the prediction error is small, the previously calculated source weights are used to infer the true value of the object at the current timestamp. Yang et al.

[20] proposed the DSWC model to capture the relationship between unit error and source weight evolution. At each timestamp, if the unit error is less than a threshold, the previous weight is used to approximate the true value, otherwise an iterative process is performed at the current timestamp to calculate the source weight and the object true value. Yan et al.

[21] proposed a source reputation-based source long-term quality modeling method, which effectively resolves the conflict between heterogeneous data and enables the integration of source weights in different truth discovery tasks. Yu et al.

[22] considered object similarity and combined the TES model to mine temporal patterns in historical truth values.

[0005] However, existing dynamic methods ignore the dynamic changes of the long-tail phenomenon of data sources and use point estimation methods to evaluate the true value of objects. Although the literature [16,22] introduced time series prediction technology, which can use historical information to compensate for the impact of data missing to a certain extent, these methods directly predict the true value, ignoring the dynamic nature of the source and are easily limited by a single model.

[0006] In dynamic scenarios, the number of data source declarations and the coverage of data object declarations will change over time, and this change usually occurs irregularly, showing a non-smooth evolution trend. Directly using the existing truth value discovery method for unevenly distributed data at each timestamp to solve the problem of uneven data distribution will cause huge computational overhead.

[0007] References:

[0008] [1] Yaliang Li, Jing Gao, Chuishi Meng, Qi Li, Lu Su, Bo Zhao, Wei Fan, and Jiawei Han: A survey on truth discovery. ACM SIGKDD Explorations Newsletter, 2015, 17(2): 1-16.

[0009] [2]Xiaoxin Yin,Jiawei Han,and Philip S.Yu:Truth discovery with multiple conflicting information providers on the web.In Proceedings of the13th ACM SIGKDDInternational Conference on Knowledge Discovery and DataMining.2007:1048-1052.

[0010] [3]Jeff Pasternack and Dan Roth:Making better informed trustdecisions with generalized fact-finding.IJCAI.Vol.11.2011.

[0011] [4]Qi Li,Yaliang Li,Jing Gao,Bo Zhao,Wei Fan,Jiawei Han:Resolvingconflicts in heterogeneous data by truth discovery and source reliabilityestimation.Proceedings of the 2014ACM SIGMOD international conference onManagement of data.2014:1187-1198.

[0012] [5]Reshef Meir,Ofra Amir,Omer Ben-Porat,Tsviel Ben Shabat,GalCohensius,Lirong Xia:Frustratingly easy truth discovery.Proceedings of theAAAI Conference on Artificial Intelligence.2023,37(5):6074-6083.

[0013] [6]Tahar Allouche, Lang,Florian Yger:Truth-tracking viaapprovalvoting:Size matters.Proceedings of the AAAI Conference onArtificialIntelligence.2022,36(5):4768-4775.

[0014] [7]Han Wang,Anfeng Liu,Neal N.Xiong,Shaobo Zhang,Tian Wang:TVD-RA:ATruthful Data Value Discovery Based Reverse Auction IncentiveSystem forMobile Crowd Sensing.IEEE Internet of Things Journal,2023.[8]Gongqing Wu,Xingrui Zhuo,Liangzhu Zhou,Xianyu Bao,Richang Hong,Xindong Wu:TIRA:TruthInference Via Reliability Aggregation onObject-Source Graph.IEEE Transactionson Knowledge and DataEngineering,2022.

[0015] [9]Hui Tian,Wenwen Sheng,Hong Shen,Can Wang:Truth findingbyreliability estimation on inconsistent entities for heterogeneousdatasets.Knowledge-Based Systems,2020,187:104828.

[0016]

[10] Qi Li,Yaliang Li,Jing Gao,Lu Su,Bo Zhao,Murat Demirbas,WeiFan:Aconfidence-aware approach for truth discovery on long-tail data.Proceedingsof the VLDB Endowment,2014,8(4):425-436.

[0017]

[11] Houping Xiao,Jing Gao,Qi Li,Fenglong Ma,Lu Su,Yunlong Feng,AidongZhang:Towards confidence interval estimation in truthdiscovery.IEEETransactions on Knowledge and Data Engineering,2018,31(3):575-588.

[0018]

[12] Yaliang Li,Qi Li,Jing Gao,Lu Su,Bo Zhao,Wei Fan:On thediscoveryof evolving truth.Proceedings of the 21th acm sigkddinternational conferenceon knowledge discovery and data mining.2015:675-684.

[0019]

[13] Xiaoyi Pang, Zhibo Wang, Defang Liu, John CSLui, Qian Wang, JuRen: Towards Personalized Privacy-Preserving Truth Discovery Over CrowdsourcedData Streams. IEEE / ACM Transactions on Networking, 2022, 30(1): 327-340.

[0020]

[14] Tan Long, Zhang Xiaoqi, Jia Li, et al. An efficient incremental truth value discovery algorithm for big data[J]. Journal of Harbin Engineering University, 2019, 40(4): 805-812.

[0021]

[15] Li Tianyi, Gu Yu, Ma Qian, et al. A continuous truth discovery technique for multi-source sensor data streams[J]. Journal of Software, 2016, 27(7): 1655-1670.

[0022]

[16] Liuyi Yao, Lu Su, Qi Li, Yaliang Li, Fenglong Ma, JingGao, AidongZhang: Online truth discovery on time series data. Proceedings of the 2018siaminternational conference on data mining. Society for Industrial and Applied Mathematics, 2018: 162-170.

[0023]

[17] Shi Zhi, Fan Yang, Zheyi Zhu, Qi Li, Zhaoran Wang, Jiawen Han: Dynamic truth discovery on numerical data. 2018 IEEE International Conference on Data Mining (ICDM). IEEE, 2018: 817-826.

[0024]

[18] Daniel Yue Zhang,Jose Badilla,Yang Zhang,Dong Wang:Towardsreliable missing truth discovery in online social media sensingapplications.2018IEEE / ACM International Conference on Advances in SocialNetworks Analysis and Mining(ASONAM).IEEE,2018:143-150.

[0025]

[19] Tianyi Li,Yu Gu,Xiangmin Zhou,Qian Ma,Ge Yu:An effective andefficient truth discovery framework over data streams.Proceedings of the 20thInternational Conference on Extending Database Technology.Springer,2017:180-191.

[0026]

[20] Yi Yang,Quan Bai,Qing Liu:Dynamic source weight computation fortruth inference over data streams.Proceedings of the 18th internationalconference on autonomous agents and multiagent systems.2019:277-285.

[0027]

[21] Lei Yan,Kan Yang,Shouyi Yang:Reputation-based truth discoverywith long-term quality of source in Internet of Things.IEEE Internet ofThings Journal,2021,9(7):5410-5421.

[0028]

[22] Xiaohan Yu, Zhongxin Shao, Keming Dong, Chao Chen: A Novel TruthDiscovery Approach for Health Recommendation Systems. IEEE Transactions on Consumer Electronics, 2023. Summary of the invention

[0029] In view of the fact that the existing dynamic truth value discovery methods mentioned in the background technology are easily limited by a single model and generate huge overhead, a dynamic truth value discovery model method for unevenly distributed data is proposed.

[0030] The technical solution of the present invention is:

[0031] A dynamic truth value discovery model method for unevenly distributed data includes the following steps:

[0032] Data acquisition phase: First, obtain the data sources at different times under the required scenarios, and select unevenly distributed data objects according to the requirements of different scenarios;

[0033] Initialization phase: at each current timestamp t≤T before timestamp T, define the truth discovery problem as an optimization problem to solve; based on the coordinate descent method, fix the source weight, find the partial derivative of the objective function with respect to the estimated true value, and obtain the solution formula for the estimated true value; use the interval estimation method to measure the confidence level of the true value, use the maximum a posteriori estimation method to transform the optimization problem into likelihood estimation, and build an incremental model; in order to capture the long-tail phenomenon of the data source, use the variance of the source error to reflect the reliability of the source, minimize the source weight estimation error variance by considering the worst case of the source error variance, and use the upper bound of the confidence interval as an estimate of the source declaration error to obtain the confident source weight;

[0034] Dynamic estimation stage: After timestamp T, for each current timestamp t>T, the data standard deviation is calculated, and a threshold is set to determine the degree of data fluctuation; according to the degree of data fluctuation, the truth value evaluation method of stability method and volatility method is adopted to obtain the true value of the current timestamp.

[0035] Further, the method specifically includes the following steps:

[0036] Data acquisition phase: First, obtain the data sources at different times under the required scenarios, and select unevenly distributed data objects according to the requirements of different scenarios;

[0037] In the initialization phase, for each current timestamp t≤T before timestamp T, the truth value discovery problem is defined as an optimization problem to be solved, that is, to minimize the difference between each timestamp and the actual truth value. The distance between t To minimize the true value of all timestamps The global loss L T ,as follows:

[0038]

[0039] The first calculation statement value and the estimated true value The second term introduces a smoothing factor η to smooth the changing trend of the estimated true value at adjacent timestamps; the larger η is, the more the minimum value of the objective function depends on the time correlation constraint; the third term is introduced to prevent the source weight The value is 0, It means that source s provides a set of declared data objects at timestamp t, and is subject to the constraint that the sum of the weights W of all sources S is 1, that is,

[0040] Based on coordinate descent, fixed source weights Find the objective function with respect to the estimated true value The partial derivative of , we get the solution formula for estimating the true value:

[0041]

[0042] The interval estimation method is further used to measure the confidence level of the true value. The less the source statement coverage of the object, the wider the confidence interval of the estimated true value, and the credibility of the interval estimation evaluation will also be reduced to a certain extent; for each data object o, use Indicates the data source set that provides the declared value for the object at timestamp t; first use the Bootstrap self-service sampling method to select Based on N sets of sample data, we use formula (2) to get the true value estimation set And take the average of the set to get Next, assuming that the declared error of each data source follows a normal distribution, we use express and Based on the Bootstrap results, a Estimators of related statistics Thus we get The two-sided confidence interval estimate with a confidence level of α is:

[0043]

[0044] in, and The left and right boundaries of the confidence interval for estimating the true value of object o at timestamp t;

[0045] Based on the interval estimation of the estimated true value, the relative width is obtained IW o,t Reflects the confidence of the confidence interval and the declared coverage of the data object; the object popularity is calculated using the following formula

[0046]

[0047] Where o′ is all declared objects at the current timestamp. This formula captures the declaration coverage of object o at timestamp t, that is, The higher it is, the greater the claim coverage of object o, indicating that it is given higher attention;

[0048] In formula (1), the source weight is calculated It is necessary to revisit all previous data. To avoid repeated calculation of previous data when new data arrives, the maximum a posteriori estimation method is used to transform the optimization problem into likelihood estimation, build an incremental model, and source weights The MAP is estimated as:

[0049]

[0050] in α, β are the prior parameters of the Gamma distribution, and the declaration errors of different objects are based on their object popularity Perform a weighted sum to incorporate object popularity into the source weight estimate;

[0051] In order to capture the long tail phenomenon of the data source, the source error is used Variance To reflect the reliability of the source, that is, the smaller the variance, the greater the source weight; calculation The (1-α) confidence interval is as follows:

[0052]

[0053] in, and The left and right boundaries of the confidence interval for the declared value of source s at timestamp t; due to source error Normal distribution The sum of squares of source errors follows a chi-square distribution

[0054] When optimizing the objective function of formula (1), by considering The worst case of is used to minimize the source weight estimation error variance, using the upper bound of the confidence interval As an estimate of the source declaration error, the confident source weight is obtained Based on formula (5), it evolves as follows:

[0055]

[0056] in,

[0057] In the dynamic estimation stage, after timestamp T, for each current timestamp t>T, the data standard deviation std is calculated t , and set a threshold Φ to judge the degree of data fluctuation; different true value evaluation methods are used according to the degree of data fluctuation:

[0058] Stability method (std t ≤φ), first use formula (2) to calculate the estimated true value Since the data fluctuation is relatively small at this time, Bootstrap sampling is omitted compared to the initialization stage; it is assumed that the estimated true value follows a normal distribution, that is, Get the estimated true value The two-sided confidence interval of is estimated as follows, where Z α / 2 The corresponding Z value when the corresponding confidence level is α:

[0059]

[0060] In terms of source weight estimation, compared with the initialization stage, the confidence-based source weight estimation is omitted, and the time series prediction method is used alone to estimate the source weight, using the stability of time series data to compensate for the impact of data sparsity; specifically, based on historical data, the ARIMA-LSTM hybrid model is used to estimate the time series evolution trend of the source weight, and the source weight is predicted online; the autoregressive integrated moving average ARIMA model is used to capture the linear trend in the time series data, and the long short-term memory network LSTM is used to capture the nonlinear fluctuation characteristics in the time series data; the predicted values ​​of the ARIMA model and the LSTM model are then aggregated based on the errors of their respective models to obtain the time series prediction source weight:

[0061]

[0062] in, The source weights for the current time stamp t predicted based on all previous time stamps 1:t-1 using the ARIMA model, is the source weight of the current timestamp t predicted by the LSTM model based on all previous timestamps 1:t-1; ∈ ARIMA ,∈ LSTM are the prediction errors of different models respectively;

[0063] In the current situation, Substitute into formula (2) to get the true value of the current timestamp;

[0064] Volatility method (std t >Φ), since the data fluctuates greatly at this time, a more accurate evaluation of the true value and source weight is required; first, based on formula (2), the estimated true value is calculated Calculate the trusted source weight based on formula (7) Calculate the time series prediction source weight based on formula (9) Then calculate the relative width of source s based on the confidence interval of the source error variance And order As a weight pair and Perform weighted aggregation to finally obtain the comprehensive source weight of source s

[0065]

[0066] by As a weight, the model considers the time series prediction source weight more when the data source declaration value is sparse, so as to improve the accuracy of reliability assessment of small sample data sources; finally, Substitute into formula (2) to get the true value of the current timestamp.

[0067] Furthermore, in the initialization phase, for the solution formula to obtain the estimated true value, the declaration coverage of different data objects is different, that is, due to objective factors and the different popularity of objects at different times, a certain object is subject to different numbers of source declarations at different times; correspondingly, due to equipment loss, fault repair, and energy threshold, the number of declarations provided by the source at different times is also different, so the credibility of the truth value evaluation is also different; in the case of insufficient data, it is impossible to make an accurate estimate of the object's true value at the current timestamp, and the credibility of the true value estimate is low.

[0068] The beneficial effects of the present invention are:

[0069] The present invention combines concepts and techniques from data mining, machine learning, and statistics. This approach aims to automatically discover true, accurate, or meaningful information in data. When faced with unevenly distributed data, the values ​​of different features or attributes may have different frequencies or probabilities of occurrence in the data set. This may lead to biases or inaccuracies in the model when analyzing and predicting the data. Confidence-based methods can more accurately estimate the authenticity of each data point by considering the distribution and frequency of occurrence of the data points.

[0070] The present invention takes into account the true value estimation errors caused by insufficient data volume and the change of stream data distribution over time, and implements an online true value discovery method in large-scale stream data scenarios. This technology is superior to the current true value discovery method in terms of accuracy and computational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 The present invention is a flow chart of the dynamic truth value discovery model method for unevenly distributed data. DETAILED DESCRIPTION

[0072] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0073] This paper will focus on studying the dynamic evolution of the uneven distribution of data objects and the long-tail phenomenon of data sources, and improve the accuracy of truth value discovery by distinguishing the number of declarations of data sources and the declared coverage of data objects at different times.

[0074] First, we obtain data sources at different times in different scenarios through the Internet and experiments, and select unevenly distributed data objects according to the needs of different scenarios, such as commodity sales in e-commerce, word frequency distribution in natural language processing, and user behavior in recommendation systems. Since the number of objects considered in actual research is often large, different sources such as official websites and forums often have interests and field preferences, so it is almost impossible to provide statements for all objects, resulting in uneven data distribution and even long tail phenomena. Under this data distribution, the following experiments are conducted:

[0075] In the initialization phase, for each current timestamp t≤T before timestamp T, the truth value discovery problem is defined as an optimization problem to be solved, that is, to minimize the difference between each timestamp and the actual truth value. The distance between t To minimize the true value of all timestamps The global loss L T ,as follows:

[0076]

[0077] The first calculation statement value and the estimated true value The second term introduces a smoothing factor η to smooth the changing trend of the estimated true value at adjacent timestamps. The larger η is, the more the minimum value of the objective function depends on the time correlation constraint. The third term is introduced to prevent the source weight The value is 0, It means that source s provides a set of declared data objects at timestamp t, and is subject to the constraint that the sum of the weights W of all sources S is 1, that is,

[0078] Based on coordinate descent, fixed source weights Find the objective function with respect to the estimated true value The partial derivative of , we get the solution formula for estimating the true value:

[0079]

[0080] Different data objects have different declaration coverage, that is, due to objective factors, such as the different popularity of objects at different times, a certain object receives different numbers of source declarations at different timestamps; correspondingly, due to equipment loss, fault repair, energy threshold and other reasons, the number of declarations provided by the source at different times is also different, so the credibility of the true value evaluation is also different. In the case of insufficient data, it is impossible to make an accurate estimate of the true value of the object at the current timestamp, and the credibility of the true value estimate is low. We further use the interval estimation method to measure the confidence level of the true value. The less the source declaration coverage of the object, the wider the confidence interval of the estimated true value, and the credibility of the evaluation based on the interval estimation will also be reduced to a certain extent. For each data object o, we use Indicates the data source set that provides the declared value for the object at timestamp t. First, use the Bootstrap self-service sampling method to select Based on N sets of sample data, we use formula (2) to get the true value estimation set And take the average of the set to get Next, assuming that the declared error of each data source follows a normal distribution, we use express and Based on the Bootstrap results, a Estimators of related statistics Thus we get The two-sided confidence interval estimate with a confidence level of α is:

[0081]

[0082] in, and The left and right bounds of the confidence interval for the estimated true value of object o at timestamp t.

[0083] Based on the interval estimate of the estimated true value, we get the relative width IW o,t It reflects the confidence of the confidence interval and the declared coverage of the data object. The object popularity is calculated using the following formula:

[0084]

[0085] Where o′ is all declared objects at the current timestamp. This formula captures the declaration coverage of object o at timestamp t, that is, The higher it is, the greater the declaration coverage of object o, indicating that it is more highly valued.

[0086] In formula (1), the source weight is calculated It is necessary to revisit all previous data. To avoid repeated calculation of previous data when new data arrives, the maximum a posteriori estimation method (MAP) is used to transform the optimization problem into likelihood estimation, build an incremental model, and use the source weights. The MAP is estimated as:

[0087]

[0088] in α, β are the prior parameters of the Gamma distribution, and the declaration errors of different objects are based on their object popularity A weighted sum is performed to incorporate object popularity into the source weight estimate.

[0089] In order to capture the long tail phenomenon of the data source, the source error is used Variance To reflect the reliability of the source, that is, the smaller the variance, the greater the source weight. The (1-α) confidence interval is as follows:

[0090]

[0091] in, and The left and right bounds of the confidence interval for the declared value of source s at timestamp t. Normal distribution The sum of squares of source errors follows a chi-square distribution

[0092] When optimizing the objective function of formula (1), by considering The worst case of is used to minimize the source weight estimation error variance, using the upper bound of the confidence interval As an estimate of the source declaration error, the confident source weight is obtained Based on formula (5), it evolves as follows:

[0093]

[0094] in,

[0095] In the dynamic estimation stage, after timestamp T, for each current timestamp t>T, the data standard deviation std is calculated t , and set a threshold Φ to determine the degree of data fluctuation. Different true value evaluation methods are used according to the degree of data fluctuation:

[0096] (1) Stability method (std t ≤Φ), first use formula (2) to calculate the estimated true value Since the data fluctuation is relatively small at this time, we omit Bootstrap sampling compared to the initialization stage. Assume that the estimated true value follows a normal distribution, that is, Get the estimated true value The two-sided confidence interval of is estimated as follows, where Z α / 2 The corresponding Z value when the corresponding confidence level is α:

[0097]

[0098] In terms of source weight estimation, compared with the initialization stage, we omit the confidence-based source weight estimation and use the time series prediction method alone to estimate the source weight, using the stability of time series data to compensate for the impact of data sparsity. Specifically, based on historical data, the ARIMA-LSTM hybrid model is used to estimate the time series evolution trend of the source weight and predict the source weight online. The autoregressive integrated moving average (ARIMA) model is used to capture the linear trend in the time series data, and the long short-term memory network (LSTM) is used to capture the nonlinear fluctuation characteristics in the time series data. The predicted values ​​of the ARIMA model and the LSTM model are then aggregated based on the errors of their respective models to obtain the time series prediction source weight:

[0099]

[0100] in, The source weights for the current time stamp t predicted based on all previous time stamps 1:t-1 using the ARIMA model, is the source weight of the current timestamp t predicted by the LSTM model based on all previous timestamps 1:t-1. ∈ ARIMA ,∈ LSTM are the prediction errors of different models.

[0101] In the current situation, Substitute into formula (2) to obtain the true value of the current timestamp.

[0102] (2) Volatility method (std t >Φ), since the data fluctuates greatly at this time, we need a more accurate evaluation of the true value and source weight. First, based on formula (2), the estimated true value is calculated Calculate the trusted source weight based on formula (7) Calculate the time series prediction source weight based on formula (9) Then calculate the relative width of source s based on the confidence interval of the source error variance And order As a weight pair and Perform weighted aggregation to finally obtain the comprehensive source weight of source s

[0103]

[0104] by As a weight, the model considers the time series prediction source weight more when the data source declaration value is sparse, so as to improve the accuracy of the reliability assessment of small sample data sources. Substitute into formula (2) to get the true value of the current timestamp.

[0105] The above-mentioned embodiment only expresses one implementation mode of the present invention, and its description is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. A dynamic truth value discovery model method for unevenly distributed data, characterized in that: The following steps are involved: Data acquisition phase: First, obtain the data sources at different times under the required scenarios, and select unevenly distributed data objects according to the requirements of different scenarios; Initialization phase: At each current timestamp t≤T before timestamp T, the truth-finding problem is defined as an optimization problem to be solved; Based on the coordinate descent method, the source weight is fixed, the partial derivative of the objective function with respect to the estimated true value is calculated, and the solution formula for the estimated true value is obtained; The interval estimation method is used to measure the confidence level of the true value, and the maximum a posteriori estimation method is used to transform the optimization problem into likelihood estimation to build an incremental model. In order to capture the long-tail phenomenon of the data source, the variance of the source error is used to reflect the reliability of the source. The variance of the source weight estimation error is minimized by considering the worst case of the source error variance. The upper bound of the confidence interval is used as an estimate of the source declaration error, thereby obtaining the confident source weight. Dynamic estimation stage: After timestamp T, for each current timestamp t>T, the data standard deviation is calculated, and a threshold is set to determine the degree of data fluctuation; According to the degree of data fluctuation, the stability method and volatility method are used to evaluate the truth value and obtain the true value of the current timestamp.

2. The dynamic truth value discovery model method for unevenly distributed data according to claim 1, characterized in that: The specific steps include: Data acquisition phase: First, obtain the data sources at different times under the required scenarios, and select unevenly distributed data objects according to the requirements of different scenarios; In the initialization phase, for each current timestamp t≤T before timestamp T, the truth value discovery problem is defined as an optimization problem to be solved, that is, to minimize the difference between each timestamp and the actual truth value. The distance between t To minimize the true value of all timestamps The global loss L T ,as follows: The first calculation statement value and the estimated true value The second term introduces a smoothing factor η to smooth the changing trend of the estimated true value at adjacent timestamps; the larger η is, the more the minimum value of the objective function depends on the time correlation constraint; the third term is introduced to prevent the source weight The value is 0, It means that source s provides a set of declared data objects at timestamp t, and is subject to the constraint that the sum of the weights W of all sources S is 1, that is, δ(W) = 1, Based on coordinate descent, fixed source weights Find the objective function with respect to the estimated true value The partial derivative of , we get the solution formula for estimating the true value: The interval estimation method is further used to measure the confidence level of the true value. The less the source statement coverage of the object, the wider the confidence interval of the estimated true value, and the credibility of the interval estimation evaluation will also be reduced to a certain extent; for each data object o, use Indicates the data source set that provides the declared value for the object at timestamp t; first use the Bootstrap self-service sampling method to select Based on N sets of sample data, we use formula (2) to get the true value estimation set And take the average of the set to get Next, assuming that the declared error of each data source follows a normal distribution, we use express and Based on the Bootstrap results, a Estimators of related statistics Thus we get The two-sided confidence interval estimate with a confidence level of α is: in, and The left and right boundaries of the confidence interval for estimating the true value of object o at timestamp t; Based on the interval estimation of the estimated true value, the relative width is obtained IW o,t Reflects the confidence of the confidence interval and the declared coverage of the data object; the object popularity is calculated using the following formula Among them ′ For all declared objects at the current timestamp, this formula captures the declaration coverage of object o at timestamp t, that is, The higher it is, the greater the claim coverage of object o, indicating that it is given higher attention; In formula (1), the source weight is calculated It is necessary to revisit all previous data. To avoid repeated calculation of previous data when new data arrives, the maximum a posteriori estimation method is used to transform the optimization problem into likelihood estimation, build an incremental model, and source weights The MAP is estimated as: in α, β are the prior parameters of the Gamma distribution, and the declaration errors of different objects are based on their object popularity Perform a weighted sum to incorporate object popularity into the source weight estimate; In order to capture the long tail phenomenon of the data source, the source error is used Variance To reflect the reliability of the source, that is, the smaller the variance, the greater the source weight; calculation The (1-α) confidence interval is as follows: in, and The left and right boundaries of the confidence interval for the declared value of source s at timestamp t; due to source error Normal distribution The sum of the squares of the source errors follows a chi-square distribution When optimizing the objective function of formula (1), by considering The worst case of is used to minimize the source weight estimation error variance, using the upper bound of the confidence interval As an estimate of the source declaration error, the confident source weight is obtained Based on formula (5), it evolves as follows: in, In the dynamic estimation stage, after timestamp T, for each current timestamp t>T, the data standard deviation std is calculated t , and set a threshold Φ to judge the degree of data fluctuation; different true value evaluation methods are used according to the degree of data fluctuation: When std t ≤Φ, stability method: first use formula (2) to calculate the estimated true value Since the data fluctuation is relatively small at this time, Bootstrap sampling is omitted compared to the initialization stage; it is assumed that the estimated true value follows a normal distribution, that is, Get the estimated true value The two-sided confidence interval of is estimated as follows, where Z α / 2 The corresponding Z value when the corresponding confidence level is α: In terms of source weight estimation, compared with the initialization stage, the confidence-based source weight estimation is omitted, and the time series prediction method is used alone to estimate the source weight, using the stability of time series data to compensate for the impact of data sparsity; specifically, based on historical data, the ARIMA-LSTM hybrid model is used to estimate the time series evolution trend of the source weight, and the source weight is predicted online; the autoregressive integrated moving average ARIMA model is used to capture the linear trend in the time series data, and the long short-term memory network LSTM is used to capture the nonlinear fluctuation characteristics in the time series data; the predicted values ​​of the ARIMA model and the LSTM model are then aggregated based on the errors of their respective models to obtain the time series prediction source weight: in, The source weights for the current time stamp t predicted based on all previous time stamps 1:t-1 using the ARIMA model, is the source weight of the current timestamp t predicted by the LSTM model based on all previous timestamps 1:t-1; ∈ ARIMA ,∈ LSTM are the prediction errors of different models respectively; In the current situation, Substitute into formula (2) to get the true value of the current timestamp; When std t >Φ, volatility method: Since the data fluctuates greatly at this time, a more accurate evaluation of the true value and source weight is required; first, based on formula (2), the estimated true value is calculated Calculate the trusted source weight based on formula (7) Calculate the time series prediction source weight based on formula (9) Then calculate the relative width of source s based on the confidence interval of the source error variance And order As a weight pair and Perform weighted aggregation to finally obtain the comprehensive source weight of source s by As a weight, the model considers the time series prediction source weight more when the data source declaration value is sparse, so as to improve the accuracy of reliability assessment of small sample data sources; finally, Substitute into formula (2) to get the true value of the current timestamp.

3. The dynamic truth value discovery model method for unevenly distributed data according to claim 2, characterized in that: In the initialization phase, for the solution formula to obtain the estimated true value, the declaration coverage of different data objects is different, that is, due to objective factors and the different popularity of objects at different times, a certain object receives different numbers of source declarations at different timestamps; correspondingly, due to equipment loss, fault repair, and energy threshold, the number of declarations provided by the source at different times is also different, so the credibility of the truth value evaluation is also different; in the case of insufficient data, it is impossible to make an accurate estimate of the object's true value at the current timestamp, and the credibility of the true value estimate is low.

Images