Time-space truth data collection method based on trust and attention

Abstract

The invention discloses a spatio-temporal data truth collection method based on trust and attention. Key challenges of service quality reduction caused by false data submitted by malicious workers, sparse data space-time faults and the like exist in crowd sensing network data collection. Therefore, the invention provides a data truth discovery method based on trust and a data recovery method based on a time-space relationship. After the platform obtains worker data, a part of basic truth phases are aggregated through completely trusted workers and highly trusted workers, and a sparse truth phase data matrix is filled through the part of truth phases; then, the platform identifies the space-time relation of the basic real data through an attention-based space-time true phase recovery method, and reconstructs unperceived or unreliable residual data; therefore, after the complete true phase data is obtained, the platform can evaluate the credibility of suspicious and unknown workers, high-quality workers are continuously identified, meanwhile, the high-credibility complete spatio-temporal data true phase is obtained, and a high-quality crowd sensing service network is constructed.

Description

Technical Field

[0001] The present invention belongs to the field of trusted data collection in mobile crowd intelligence networks, and in particular to real data collection for untrusted workers and sparse spatiotemporal data in crowd intelligence networks. Background Art

[0002] Mobile crowd sensing has become a popular data collection service paradigm, mainly by utilizing ubiquitous intelligent sensing devices carried by mobile workers to complete various sensing tasks. The mobile crowd sensing network consists of three main parts: a service platform that publishes data collection tasks and builds service applications, workers who perform data collection tasks, and consumers who use service applications. Nowadays, applications based on mobile crowd sensing have covered a wide range of fields, such as air quality monitoring, traffic flow estimation, and geographic information exploration. The mobile crowd sensing network uses the mobility of workers and the ubiquity of sensing devices to achieve large-scale, fine-grained urban spatiotemporal data collection.

[0003] In actual scenarios, the distribution of workers in urban areas often presents the characteristics of "globally sparse and locally dense". Specifically, for the same task, there may be multiple different workers who can upload perception data, and due to their different perception devices, perception capabilities, and perception perspectives, the multi-source data reported by different workers may conflict with each other. On the other hand, mobile workers and sensors cannot cover the entire area of ​​the city, which will inevitably lead to data missing. These data conflicts and data missing may result in low quality of constructed service applications, which in turn may lead to adverse effects and may even endanger the property and life safety of consumers. Therefore, truth estimation becomes a key issue in crowd sensing, aiming to estimate accurate and complete truth data from the conflicting and sparse data collected by workers.

[0004] To address this problem, researchers have proposed a class of methods called truth discovery, which mainly attempts to alleviate data conflicts based on multi-source observations provided by recruiting multiple workers to perform the same task. Existing truth discovery methods all assume that most workers in the crowd-sensing network are trustworthy, and thus obtain reliable estimated truth values ​​through dynamic weighted averaging. However, even if the vast majority of workers are trustworthy, the accuracy of the truth estimate will be affected by a small number of malicious workers due to weighted averaging. What's more serious is that the accuracy of these methods is directly related to the proportion of malicious workers. When the number of malicious workers in the worker group exceeds half, the results of classic truth discovery will be manipulated at will by malicious workers.

[0005] Data recovery is another method to solve the problem of true value estimation. It attempts to solve the problem of data sparsity by using the perceived data to infer the unperceived data through the spatiotemporal correlation of the perception matrix. Studies have shown that as long as each row and column in the spatiotemporal matrix contains a certain amount of data, the current data recovery methods can achieve extremely high recovery accuracy under high missing rates. Current data recovery methods actually face similar trust challenges as truth discovery methods, because they often assume that the data of the perceived task is completely reliable, and thus infer the unperceived data based on this. But in fact, if the service platform mistakenly uses toxic data as the benchmark data for data recovery, it will lead to further propagation and aggravation of the error, making the final estimated true value completely too error-prone to be used.

[0006] In addition, current research on data recovery does not fully consider the problem of spatiotemporal breaks, where spatiotemporal breaks include the lack of observation data in all time slices of a specific sensing sub-area (temporal breaks) and the lack of observation data in all sensing sub-areas at a specific moment (spatial breaks). In crowd sensing, especially in large-scale and high-resolution data collection, spatiotemporal breaks are actually very common. Especially in untrusted environments, the available worker data is more sparse, which may lead to more obvious spatiotemporal breaks. The inevitable spatiotemporal breaks undoubtedly increase the challenge of extracting complex spatiotemporal data associations from limited spatiotemporal information and insufficient context. Therefore, the present invention proposes a trust- and attention-based spatiotemporal data truth collection method to estimate accurate true values ​​in the presence of malicious workers and spatiotemporal breaks. Summary of the invention

[0007] The present invention discloses a method for collecting spatiotemporal data truth based on trust and attention. Different from the previous methods that mainly focus on single-stage truth discovery or data recovery, the present invention introduces a trust mechanism and makes full use of the synergy of these two processes, thereby improving the accuracy of the estimated truth value. In the truth discovery stage, we adopt a conservative strategy to obtain the benchmark truth value, and minimize the error by pre-excluding the data reported by suspicious workers. Then, we expand the scope of data recovery from only recovering the task data that was not perceived due to the absence of workers to also recovering the perceived tasks of suspicious workers, thereby providing a data benchmark for identifying suspicious workers. After obtaining accurate and complete estimated truth values ​​through the above operations, we continuously update the worker trust based on the comparison between the worker data and the estimated truth value. In this way, while continuously identifying high-quality workers, we can collect highly reliable complete spatiotemporal data truth and build a high-quality crowd intelligence perception service network.

[0008] The technical solution of the invention is as follows:

[0009] (1) In the crowd-sensing network, there are M workers who can perform the sensing task, denoted as S = {s1, s2, ..., s M}; Initially, the service platform initially knows a small proportion of fully trusted workers, denoted as S F ∈S, since its number is often very small in real scenarios, |S F |<<|S|; the service platform will set an indication To represent workers i Whether to be recruited to perform task T in round t j ;if It means that worker s i Recruited in round t to perform task T j At the same time, the platform uses Indicates the execution of task T in round t j All the data submitted by the workers at that time; using the truth matrix Represents all real data values, consisting of N tasks and T rounds of data; the service platform will update the trust of the corresponding worker in each round according to the quality of the worker's reported data. The trust set of all workers in each round t is recorded as in, Representing workers i Trust in round t;

[0010] (2) Obtaining truth data based on worker trust: In this stage, the service platform divides workers into fully trusted workers, highly trusted workers, and low trusted workers based on trust and thresholds ξ1 and ξ2. greater than ξ1, worker s i is considered a fully trusted worker. If its trust is greater than ξ2 and less than ξ1, it is considered a high-trust worker, otherwise it is a low-trust worker; ξ1 and ξ2 can be set to 0.9 and 0.5; through the above division, different processing measures are taken for the data reported by workers of different levels to obtain the truth data; the truth data includes the following two categories: gold benchmark truth value and silver benchmark truth value. The following is a method for obtaining these two truth values

[0011] a. Method for obtaining the true value of the gold benchmark:

[0012] Since fully trusted workers are considered to be completely trustworthy a priori, the platform obtains the gold benchmark truth value by directly taking the average of all task data reported by fully trusted workers.

[0013]

[0014] in, For task T jAll fully trusted workers recruited in round t, binary variable Indicates whether there are fully trusted workers executing task T in round t j However, considering the rarity of fully trusted workers a priori, in order to effectively apply the data submitted by highly trusted workers identified a posteriori, the platform can further perform truth discovery from the data submitted by highly trusted workers to construct the silver benchmark truth

[0015] b. Method for obtaining the true value of the silver benchmark:

[0016] The silver benchmark true value is calculated by iterating the following two equations:

[0017] Weight update:

[0018]

[0019] Truth value update:

[0020]

[0021] in, represents the weight of the worker, represents the estimated silver benchmark truth value, set Indicates that in round t, task T j All highly trusted workers recruited, and d(·,·) represents the distance function; the above two processes will be iterated repeatedly until the termination condition is reached; the final calculated at the end of the iteration is Represents the silver benchmark truth value finally estimated by the platform;

[0022] (3) Acquisition of spatiotemporal data truth based on attention data recovery: After obtaining the truth data through the above method, it is merged and integrated into a true value data matrix The data in this data matrix are considered highly reliable;

[0023]

[0024] in, Reliable benchmark data collected, Indicates whether there is a high-trust worker performing task T in round t j ;Since this matrix is ​​often sparsely missing, the platform will further use the perception matrix Y′ to perform the data recovery process;

[0025] a. Prefill non-fracture missing data:

[0026] Considering the sensing matrix Y′ in period k (k) Including both break missing data and non-break missing data, if Y′(k) There is no data submitted in the n0th row and t0th column, and the platform starts from Y′ (k) Remove the n0th row and t0th column from the matrix to obtain a low-dimensional matrix without fracture data. And use matrix decomposition methods to explore the associations between rows and columns and use these associations to fill Missing data in M Aux To mark the location of the pre-filled data; at this time, the perception matrix obtained by the above method All missing data are broken data; then the platform will conduct the next step of spatiotemporal information mining based on the spatiotemporal breaks;

[0027] b. Spatial information extraction:

[0028] The platform will first convert the perception matrix in period k As a graph, divide it into in Represents task T j The observation vector in all rounds; use N i In addition to task T i Other tasks besides (can be called adjacent nodes in the graph), and can calculate and any adjacent node observation vector The relationship between i，j as follows

[0029]

[0030] in, is a dimensional transformation matrix, represents a learnable weight matrix, T represents the transposition operation, and LeakyReLU(·) is an activation function calculated as follows

[0031]

[0032] After obtaining the spatial correlation between tasks through the above method, the platform can i Estimation Node data

[0033]

[0034] Here, σ(·) represents a nonlinear activation function, which is used to enhance the expressiveness of the network and enable the network to learn more spatial representations; after obtaining the initial estimate of each node The platform then decodes D into a matrix with the original dimensions

[0035]

[0036] here, represents the weight matrix, b is the bias term; in the data matrix In the process, due to the complex nonlinear transformation, the original benchmark truth data may have changed. In order to protect the benchmark truth obtained in the truth discovery phase from being modified by the network, the platform will replace the benchmark truth into the matrix through the following formula middle

[0037]

[0038] c. Time information extraction:

[0039] After processing the spatial fractures, the platform further uses the multi-layer self-attention mechanism of Transformer to capture the time series features corresponding to the perception data between different rounds to estimate the time fracture data; the platform uses a multi-head self-attention mechanism based on time mask and an embedding mechanism with position encoding to encode the data, denoted as E, and uses it as the input of Transformer:

[0040]

[0041] in, represents the weight matrix, b E is the bias, d m is the dimension of the model, and represents the result of position encoding; to enhance the attention of adjacent vectors, the platform uses a thresholded Gaussian kernel and a i，j Represents the correlation between vectors i and j;

[0042]

[0043] [TM(x)][i,j]=x i，j +a i，j .

[0044] Here, ρ is the threshold and β is the shape parameter; in order to improve the stability of the true value estimation, a multi-head self-attention mechanism is used

[0045]

[0046] TMMHSA(x)=Concat(h1,h2,…,h i )W O .

[0047] Here, h i represents the index of the header, and The parameter matrix The platform maps each row of the embedding to a query vector W Q , one dimension is d k The key vector W K , and a dimension d v The value vector W V Then, TMMHSA(X) is input into the feedforward network layer to obtain the predicted spatiotemporal representation Z st and the complete spatiotemporal prediction matrix

[0048] Z st =FFN(TMMHSA(E)) L .

[0049]

[0050] Here, represents the weight matrix, and is a bias term; next, in order to protect the benchmark truth value obtained in the truth value discovery phase from being modified, the platform will restore the truth value in the truth value discovery phase;

[0051]

[0052] When constructing the loss function of the neural network, given the unknowability of missing data in actual scenarios, the platform uses the error between the network's reconstruction of the observed data and the benchmark true value data as the indicator of the loss function, and calls the self-supervised task corresponding to this loss function the observed data reconstruction task, which is expressed as In addition, considering the difference in quality between the gold-medal benchmark data and the silver-medal benchmark data in the perception matrix, λ1 and λ2 are used to represent the weights of the gold-medal benchmark data and the silver-medal benchmark data in the loss function respectively;

[0053]

[0054] Considering that the pre-filled restored data can also be used as a supervisory signal outside the benchmark truth, we further introduce the pre-filled data reconstruction task, denoted as To enhance the self-supervision effect,

[0055]

[0056] Among them, l(·,·) represents the element-level error function; through the above method, we can obtain the complete spatiotemporal truth data of the reconstruction period k and update the evaluation and trustworthiness of auxiliary worker data;

[0057] (4) Worker trust update based on recovered truth value: After obtaining the estimated truth value through the above method, the service platform will evaluate the data quality by comparing the data reported by the workers with the estimated truth value, and then update the trust information of each recruited worker; the trust of each worker is modeled as a Beta distribution, using Indicates that workers i In round t, task T j Reporting data Data quality

[0058]

[0059] Here, represents the task T in round t j The estimated true value of If the data submitted by the worker exceeds the threshold ε, high-quality data, thereby increasing the worker's trust; conversely, if the value of data quality is If it is lower than the threshold ε, the trust of the corresponding worker is reduced; the service platform can obtain its trust by calculating the expectation of the Beta distribution, that is,

[0060]

[0061] Here, and Represents workers s i Execute task T in round t j The beta distribution parameter of When we control and At the same information decay rate, according to data quality Additional enlargement To increase trust The value of is regarded as a positive trust incentive for high-quality data submission.

[0062]

[0063] Here, represents the exponential decay parameter, and μ is an incentive parameter factor, which is used to regulate the platform's positive incentive for reliable data. The larger μ is, the stronger the incentive effect of reporting high-quality data on trust is.

[0064] Likewise, when When, according to the data quality Additional enlargement To reduce trust The value of is regarded as a negative trust penalty for low-quality data submission.

[0065]

[0066] Here, μ represents a penalty parameter factor, which is used to regulate the platform's negative trust penalty for unreliable data. The larger μ is, the stronger the penalty effect on trust for each report of low-quality data.

[0067] In this way, by continuously identifying high-quality workers to perform tasks, it is possible to ultimately obtain credible truth data.

[0068] Beneficial Effects

[0069] The present invention discloses a spatiotemporal data truth collection method based on trust and attention, which makes full use of the synergy of these two processes, thereby improving the accuracy of the estimated true value. In the truth discovery stage, the platform adopts a conservative strategy to obtain the benchmark truth value, and minimizes the error by pre-excluding the data reported by suspicious workers. Then, the platform uses an attention-based data recovery scheme, which recovers the unperceived and untrusted data by integrating the spatial block based on the graph neural network and the temporal block based on the Transformer. In particular, the present invention expands the scope of data recovery from only recovering the task data that was not perceived due to the absence of workers in the past to also recovering the perceived tasks of suspicious workers, thereby providing a data benchmark for suspicious identification of suspicious workers, thereby solving the problem of worker data quality identification. After obtaining an accurate and complete estimated true value through the above operations, we further use the Beta distribution to model the trust of each worker, and continuously update the worker trust based on the comparison with the estimated true value. In this way, the crowd intelligence perception network can identify credible workers and obtain high-quality truth data in the environment of malicious workers and data spatiotemporal fractures. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 The mechanism diagram of the method of the present invention is

[0071] Figure 2 Comparison chart of the mean absolute percentage error of truth discovery as the malicious rate changes

[0072] Figure 3 Comparison chart of the mean absolute percentage error of data recovery as the malicious rate changes

[0073] Figure 4 The trust changes with the number of rounds DETAILED DESCRIPTION

[0074] In order to facilitate the understanding of the present invention, the present invention will be described more comprehensively and meticulously below in conjunction with the accompanying drawings and preferred embodiments of the specification, but the protection scope of the present invention is not limited to the following specific embodiments.

[0075] Unless otherwise defined, all professional terms used below have the same meanings as those generally understood by those skilled in the art. The professional terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the scope of protection of the present invention.

[0076] Unless otherwise specified, various raw materials, reagents, instruments and equipment used in the present invention can be purchased from the market or prepared by existing methods.

[0077] Example:

[0078] In a smart city, it is necessary to monitor the environment in the city in real time, such as temperature, noise, traffic flow, etc., for data perception. At this time, there are a large number of workers in the city who hold sensing devices (such as mobile collection, smart watches, etc.) distributed in various areas of the city, so the platform can request workers to participate in data perception by giving certain rewards for data perception. However, there are some malicious workers who fabricate false data, and the distribution of workers in urban areas often presents the characteristics of "globally sparse and locally dense", resulting in spatiotemporal breaks in the data. Therefore, the present invention discloses a spatiotemporal data truth collection method based on trust and attention, which estimates the spatiotemporal data truth while identifying trusted workers. The experimental method is as follows: suppose there are 10,000 workers in the city. In order to fully simulate the diverse trust of workers in real scenarios, all workers are divided into three attribute groups: fully trusted, highly trusted, and untrustworthy. Specifically, for fully trusted workers, their reliability has been verified by the service platform, and the reported data is directly equal to the ground truth of the task. For highly trusted workers, we take into account error factors such as equipment accuracy, and set the data reported by them to add a small amount of observation error on the basis of the ground truth (the maximum deviation value is set to 10% of the true value). For untrusted workers, we set a 90% probability that they will maliciously report false data when reporting, and the remaining 10% probability that they will pretend to be highly trusted workers and report data truthfully. The proportion of untrusted workers in the total worker population is called the malicious rate, represented by r. During this experiment, we recorded the algorithm performance corresponding to different malicious rates r∈[0.1,0.9] to analyze the sensitivity of different algorithms to the malicious rate r. At the same time, the data sparse rate was set to 30%.

[0079] Two types of benchmark algorithms, truth value discovery and data recovery, were used for comparison. Specifically, in the truth value discovery method, the following four types of classic benchmark algorithms were compared.

[0080] Average truth method: This method takes the average of the data submitted by all workers in a specific task and uses the average of the data as the aggregate truth value.

[0081] Median Truth Method: This method takes the median of the data submitted by all workers in a specific task and uses the median of the data as the aggregate truth value.

[0082] Iterative truth calculation method: This method is the most widely used truth discovery method. It gives each data a dynamic weight and minimizes the weighted deviation between the worker's reported data and the estimated truth value through continuous iteration, and finally converges to the final result of the estimated truth value.

[0083] Weighted Average Truth Method: This method is similar to the Average Truth Method, but adds a confidence estimate for each worker, so that each time the truth estimate is made, a weighted average is taken based on the confidence.

[0084] In addition to the truth value discovery method, 7 currently mainstream data recovery methods are also compared to evaluate the advantages of the present invention in data recovery methods.

[0085] KNN: This method calculates the average value of the k nearest sensed data to the data to be recovered as the inference result of data recovery and is widely used in spatiotemporal linear interpolation.

[0086] MF: This method recovers missing data by exploiting low-rank matrix factorization.

[0087] GAIN: This method mainly combines the joint distribution model and the class probability model to fill in the missing data based on the perceived benchmark data in the data matrix.

[0088] BiGRU: This method converts the problem into a multivariate time series prediction problem, using a bidirectional recurrent unit model and a gating scheme to capture the temporal correlation between sensory data and thus enable inference of missing data.

[0089] GAT: This method models the perception task as a node of a graph, and sets the pairwise geographical distances between nodes as the edge weights of the graph structure, thereby using an attention-based graph neural network to capture the temporal correlation between perception data and achieve inference of missing data.

[0090] Transformer (abbreviated as TRSFM when drawing): This method learns the contextual information of the data through positional encoding and multi-layer self-attention mechanism to achieve inference of missing data.

[0091] GAMC: This method makes full use of the graph attention mechanism and the spatiotemporal attention mechanism to analyze the spatiotemporal correlation within the perception data, and proposes to use MF to pre-fill the fractured data, which has achieved outstanding results in the current data recovery with spatiotemporal fractures.

[0092] During the experimental stage, we mainly evaluated the above indicators through the mean absolute percentage error. The smaller the percentage error of an algorithm, the better its performance.

[0093] The experimental results of the inventive method are given below.

[0094] Figure 2 This is a comparison chart of the mean absolute percentage error of truth discovery as the malicious rate changes. The accuracy of the method of the present invention and all benchmark algorithms will decrease as the malicious rate increases. This is a normal phenomenon, because when the number of untrusted workers increases, the noise of the data submitted by the workers will become larger and larger, resulting in a larger deviation in the final truth discovery. It is worth noting that the sensitivity of each method to the change of the malicious rate is different. It can be clearly seen that the error increase of the method of the present invention proposed by us is very small under different malicious rates, and basically maintains stable performance, and is always better than all benchmark algorithms, which shows that the trust mechanism in the method of the present invention plays a key role.

[0095] Figure 3 This is a comparison chart of the mean absolute percentage error of data recovery as the malicious rate changes. Overall, the method of the present invention shows the best performance under each malicious rate, and the data recovery effect changes less with the change of the malicious rate. It can be seen that the method of the present invention also shows extremely strong robustness to the worker malicious rate during the data recovery process.

[0096] Figure 4 The figure shows the change of trust with the number of rounds. The recognition accuracy of the worker category of the method of the present invention is significantly better than other benchmark methods. In the method of the present invention, the trust of the trustworthy workers increases with the increase of rounds, and the trust of the malicious workers decreases with the increase of rounds, which shows that the method of the present invention can effectively distinguish between trustworthy workers and malicious workers. In the benchmark method, the trust of the trustworthy workers decreases with the increase of rounds, and the trust of the malicious workers increases with the increase of rounds, which shows that the benchmark algorithm cannot effectively estimate the true value, which will cause worker misidentification in the trust assessment.

Claims

1. A trust-and-attention-based spatiotemporal truth data collection method, characterized in that The following steps are involved: (1) In the crowd-sensing network, there are M workers who can perform the sensing task, denoted as S = {s1, s2, ..., s M }; Initially, the service platform initially knows a small proportion of fully trusted workers, denoted as S F ∈S, since its number is often very small in real scenarios, |S F |<<|S|; the service platform will set an indication To represent workers i Whether to be recruited to perform task T in round t j ;if It means that worker s i Recruited in round t to perform task T j At the same time, the platform uses Indicates the execution of task T in round t j All the data submitted by the workers at that time; using the truth matrix Represents all real data values, consisting of N tasks and T rounds of data; the service platform will update the trust of the corresponding worker in each round according to the quality of the worker's reported data. The trust set of all workers in each round t is recorded as in, Representing workers i Trust in round t; (2) Obtaining truth data based on worker trust: In this stage, the service platform divides workers into fully trusted workers, highly trusted workers, and low trusted workers based on trust and thresholds ξ1 and ξ2. greater than ξ1, worker s i is considered a fully trusted worker. If its trust is greater than ξ2 and less than ξ1, it is considered a high-trust worker, otherwise it is a low-trust worker; ξ1 and ξ2 can be set to 0.9 and 0.5; through the above division, different processing measures are taken for the data reported by workers of different levels to obtain the truth data; the truth data includes the following two categories: gold benchmark truth value and silver benchmark truth value. The following is a method for obtaining these two truth values a. Method for obtaining the true value of the gold benchmark: Since fully trusted workers are considered to be completely trustworthy a priori, the platform obtains the gold benchmark truth value by directly taking the average of all task data reported by fully trusted workers. in, For task T j All fully trusted workers recruited in round t, binary variable Indicates whether there are fully trusted workers executing task T in round t j However, considering the rarity of fully trusted workers a priori, in order to effectively apply the data submitted by highly trusted workers identified a posteriori, the platform can further perform truth discovery from the data submitted by highly trusted workers to construct the silver benchmark truth b. Method for obtaining the true value of the silver benchmark: The silver benchmark true value is calculated by iterating the following two equations: Weight update: Truth value update: in, represents the weight of the worker, represents the estimated silver benchmark truth value, set Indicates that in round t, task T j All highly trusted workers recruited, and d(·,·) represents the distance function; the above two processes will be iterated repeatedly until the termination condition is reached; the final calculated at the end of the iteration is Represents the silver benchmark truth value finally estimated by the platform; (3) Acquisition of spatiotemporal data truth based on attention data recovery: After obtaining the truth data through the above method, it is merged and integrated into a true value data matrix The data in this data matrix are considered highly reliable; in, Reliable benchmark data collected, Indicates whether there is a high-trust worker performing task T in round t j ;Since this matrix is ​​often sparsely missing, the platform will further use the perception matrix Y′ to perform the data recovery process; a. Prefill non-fracture missing data: Considering the sensing matrix Y′ in period k (k) Including both break missing data and non-break missing data, if Y′ (k) There is no data submitted in the n0th row and t0th column, and the platform starts from Y′ (k) Remove the n0th row and t0th column from the matrix to obtain a low-dimensional matrix without fracture data. And use matrix decomposition methods to explore the associations between rows and columns and use these associations to fill Missing data in M Aux To mark the location of the pre-filled data; at this time, the perception matrix obtained by the above method All missing data are broken data; then the platform will conduct the next step of spatiotemporal information mining based on the spatiotemporal breaks; b. Spatial information extraction: The platform will first convert the perception matrix in period k As a graph, divide it into in Represents task T j The observation vector in all rounds; use N i In addition to task T i Other tasks besides (can be called adjacent nodes in the graph), and can calculate and any adjacent node observation vector The relationship between i,j as follows in, is a dimensional transformation matrix, represents a learnable weight matrix, represents the transposition operation, LeakyReLU(·) is an activation function calculated as follows After obtaining the spatial correlation between tasks through the above method, the platform can i Estimation Node data Here, σ(·) represents a nonlinear activation function, which is used to enhance the expressiveness of the network and enable the network to learn more spatial representations; after obtaining the initial estimate of each node The platform then decodes D into a matrix with the original dimensions here, represents the weight matrix, b is the bias term; in the data matrix In the process, due to the complex nonlinear transformation, the original benchmark truth data may have changed. In order to protect the benchmark truth obtained in the truth discovery phase from being modified by the network, the platform will replace the benchmark truth into the matrix through the following formula middle c. Time information extraction: After processing the spatial fractures, the platform further uses the multi-layer self-attention mechanism of Transformer to capture the time series features corresponding to the perception data between different rounds to estimate the time fracture data; the platform uses a multi-head self-attention mechanism based on time mask and an embedding mechanism with position encoding to encode the data, denoted as E, and uses it as the input of Transformer: in, represents the weight matrix, b E is the bias, d m is the dimension of the model, and represents the result of position encoding; to enhance the attention of adjacent vectors, the platform uses a thresholded Gaussian kernel and a i,j Represents the correlation between vectors i and j; [TM(x)][i,j]=x i,j +a i,j . Here, ρ is the threshold and β is the shape parameter; in order to improve the stability of the true value estimation, a multi-head self-attention mechanism is used TMMHSA(X)=Concat(h1,h2,…,h i )W O . Here, h i represents the index of the header, and The parameter matrix The platform maps each row of the embedding to a query vector W Q , one dimension is d k The key vector W K , and a dimension d v The value vector W V Then, TMMHSA(X) is input into the feedforward network layer to obtain the predicted spatiotemporal representation Z st and the complete spatiotemporal prediction matrix FROM st =FFN(TMMHSA(E)) L . Here, represents the weight matrix, and is a bias term; next, in order to protect the benchmark truth value obtained in the truth value discovery phase from being modified, the platform will restore the truth value in the truth value discovery phase; When constructing the loss function of the neural network, given the unknowability of missing data in actual scenarios, the platform uses the error between the network's reconstruction of the observed data and the benchmark true value data as the indicator of the loss function, and calls the self-supervised task corresponding to this loss function the observed data reconstruction task, which is expressed as In addition, considering the difference in quality between the gold-medal benchmark data and the silver-medal benchmark data in the perception matrix, λ1 and λ2 are used to represent the weights of the gold-medal benchmark data and the silver-medal benchmark data in the loss function respectively; Considering that the pre-filled restored data can also be used as a supervisory signal outside the benchmark truth, we further introduce the pre-filled data reconstruction task, denoted as To enhance the self-supervision effect, in, Represents the element-level error function; through the above method, we can obtain the complete spatiotemporal truth data of the reconstruction period k and update the evaluation and trustworthiness of auxiliary worker data; (4) Worker trust update based on recovered truth value: After obtaining the estimated truth value through the above method, the service platform will evaluate the data quality by comparing the data reported by the workers with the estimated truth value, and then update the trust information of each recruited worker; the trust of each worker is modeled as a Beta distribution, using Indicates that workers i In round t, task T j Reporting data Data quality Here, represents the task T in round t j The estimated true value of If the data submitted by the worker exceeds the threshold ε, high-quality data, thereby increasing the worker's trust; conversely, if the value of data quality is If it is lower than the threshold ε, the trust of the corresponding worker is reduced; the service platform can obtain its trust by calculating the expectation of the Beta distribution, that is, Here, and Represents workers s i Execute task T in round t j The beta distribution parameter of When we control and At the same information decay rate, according to data quality Additional enlargement To increase trust The value of is regarded as a positive trust incentive for high-quality data submission. Here, represents the exponential decay parameter, and μ is an incentive parameter factor, which is used to regulate the platform's positive incentive for reliable data. The larger μ is, the stronger the incentive effect of reporting high-quality data on trust is. Likewise, when When, according to the data quality Additional enlargement To reduce trust The value of is regarded as a negative trust penalty for low-quality data submission. Here, μ represents a penalty parameter factor, which is used to regulate the platform's negative trust penalty for unreliable data. The larger μ is, the stronger the penalty effect on trust for each report of low-quality data. In this way, by continuously identifying high-quality workers to perform tasks, it is possible to ultimately obtain credible truth data.

Images