An association rule learning method with privacy protection

Abstract

The application relates to an association rule learning method with privacy protection, and belongs to the field of data security.The method comprises the following steps: user grouping: dividing users into three groups, and the users in the three groups have similar contribution degrees in frequent pattern mining; candidate frequent 1-item set construction: collecting first group user data to construct a candidate item set which may be a frequent 1-item set; frequent 1-item set checking: collecting second group user data to screen the candidate item set obtained by the first group of users, and checking the frequent item set; frequent pattern mining: using a speculation mechanism to construct a candidate frequent pattern based on the frequent item set checked by the first group of users, checking the candidate frequent pattern by using third group users, and finally realizing the mining of the frequent pattern.The application can safely use data, and can quickly mine the frequent pattern appearing in the data, so that the safe mining of the association rule between data is realized.

Description

Technical Field

[0001] This invention belongs to the field of data security and relates to a method for learning association rules with privacy protection. Background Technology

[0002] With the rapid development of network technology, a massive number of vehicle terminal devices have been connected to the network, generating a large amount of data, such as marketing data, trajectory data, and medical data. Association rule learning is a data mining technique used to discover relationships between items in large-scale datasets. Frequent pattern mining is one method within this technique, its main goal being to identify frequently occurring itemsets in a dataset—that is, a group of items that frequently appear together. Frequent pattern mining can be used to discover frequently occurring itemsets in large-scale datasets, thereby identifying potential relationships and generating association rules. With the advent of the big data era, the scale of data is becoming increasingly massive, and the requirements for data mining are also becoming increasingly demanding. Frequent pattern mining can effectively discover important patterns and information, which is of great significance for business optimization and improving decision-making effectiveness. It has wide applications, including marketing, medical diagnosis, image recognition, and cybersecurity. The results of frequent pattern mining are highly interpretable, providing clear support and suggestions for decision-makers, and have the ability to update patterns, enabling the timely discovery of new frequent patterns to adapt to business adjustments and changes.

[0003] Frequent pattern mining can discover frequently occurring patterns or association rules in large-scale datasets, providing support and suggestions for business decisions, promoting business optimization, and improving decision-making effectiveness. However, frequent pattern mining also presents privacy concerns because the process involves statistical analysis of raw data, potentially revealing sensitive information such as user personal details and trade secrets. To address this issue, researchers have proposed several privacy protection methods, the most commonly used being differential privacy technology. Local differential privacy technology can provide high-quality data analysis results while protecting data privacy, enabling the secure and effective application of data mining. In conclusion, frequent pattern mining can provide important support and suggestions for business decisions, but it is crucial to pay attention to data privacy and employ appropriate privacy protection technologies to safeguard sensitive data.

[0004] Currently, there are some research works on privacy-preserving frequent pattern mining. Li S, Mu N, and Le J, in "Privacy preserving frequent itemset mining: Maximizing data utility based on database reconstruction" [in Computers & Security, vol. 84, pp. 17-34, 2019], studied a highly privacy-preserving frequent pattern mining algorithm based on database reconstruction. This algorithm maximizes data utility by identifying relevant frequent itemsets and utilizing the remaining frequent itemsets for database reconstruction, and employs a hiding strategy to further reduce the risk of confidential knowledge leakage. Li J, Gan W, and Gui Y, in "Frequent itemset mining with local differential privacy" [in ACM International Conference on Information & Knowledge Management, pp. 1146-1155, 2022], proposed an interactive query response method based on filling and sampling algorithms combined with local differential privacy for efficient and accurate identification of frequent itemsets. A data collector generates frequent candidate sets, and users randomize their responses within a reduced domain to achieve local differential privacy. Wu N, Zou Y, and Shan C, in "A Frequent Itemset Mining Method Based on Local Differential Privacy" [in Web Information Systems and Applications, pp. 225-236, 2021], propose an improved data perturbation method that introduces a Hidden Markov Model to avoid a large privacy budget and improve the efficiency of the algorithm. This method protects data privacy while enhancing the usability of the mining results and rapidly generating candidate sets.

[0005] Local differential privacy-preserving frequent pattern mining is an effective means of protecting data privacy, but it still faces two main problems. First, as the data scale and dimensionality increase, the space of frequent candidate patterns grows exponentially, leading to a significant decrease in the execution efficiency of mining algorithms and the usability of mining results. Second, due to the introduction of noise in local differential privacy, errors occur in data frequency estimation, resulting in inaccurate candidate patterns and thus reducing the accuracy of frequent pattern mining. To address the first problem, researchers have proposed various improved algorithms, such as projection-based algorithms, pruning-based algorithms, and sampling-based algorithms. These algorithms can effectively reduce the number of frequent candidate patterns and improve the execution efficiency and usability of mining results. For the second problem, researchers have proposed some effective solutions, such as random response techniques, Laplace mechanisms, and exponential mechanisms. These techniques protect data privacy by introducing noise, and while protecting data privacy, they also strive to ensure the accuracy of data mining as much as possible. In summary, local differential privacy-preserving frequent pattern mining is an important research area. By improving algorithms and introducing appropriate privacy-preserving techniques, problems such as large data scale and noise introduction can be effectively solved, thereby improving the efficiency and accuracy of frequent pattern mining. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a privacy-preserving association rule learning method that can securely use data and quickly mine frequent patterns in the data, thereby achieving secure mining of association rules between data.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A privacy-preserving association rule learning method is proposed. First, to improve the accuracy and efficiency of data mining, a uniform user grouping mechanism is proposed. This mechanism groups users together to complete association rule mining, effectively reducing the risk of a single user answering the task multiple times and accelerating the completion of the mining task. Simultaneously, this mechanism can also reduce the impact of data differences between different groups on the accuracy of the mined patterns. Second, to reduce the noise in frequent pattern frequency estimation under local differential privacy protection, a privacy-preserving method is proposed. This method segments the data into area codes and interval codes, and uses hash functions and local differential privacy mechanisms respectively to protect data privacy. It also enables support counting of the data, effectively reducing the noise level during the protection process in trajectory frequent pattern mining. The method specifically includes the following steps:

[0009] S1: User Grouping: The edge server divides users into three groups based on their characteristics and assigns the three sub-tasks of frequent pattern mining to different groups.

[0010] S2: Construction of candidate frequent 1-itemsets: Based on the data from the first group, the edge server constructs frequent 1-itemsets. To reduce the input domain and facilitate focusing on smaller candidate sets, an input domain reduction method is adopted. The trajectory of each user is randomly sampled, and one item p is randomly reported to reduce the size of the input domain. Then, a privacy protection method is used to protect the items reported by users, and a corresponding frequency estimation algorithm is used to calculate the frequency of a single item. Finally, the Top-2k set is obtained by descending the frequency of item occurrence as candidate frequent 1-itemsets.

[0011] S3: Verification of Frequent 1-Itemsets: In the previous task, the overall frequency of candidate frequent 1-itemsets was underestimated, so the items in the frequent 1-itemsets and their frequencies need to be verified. In this task, each user needs to answer whether the item contains the i-th item in the Top-2k candidate frequent 1-itemsets. If it does, the i-th position is encoded as 1; otherwise, it is 0, resulting in a binary string V of length 2k. User data privacy is involved in the answering process, so a privacy protection mechanism is used to protect the user's answer vector V. Afterward, all user responses are aggregated, and the server uses the frequency estimation corresponding to the privacy protection method to calculate the frequency of the candidate frequent 1-itemsets. These are then sorted in descending order of frequency to obtain the top k frequent items and their frequencies, thus verifying the frequent 1-itemsets.

[0012] S4: Frequent Pattern Mining: Based on the top-k frequent 1-itemsets, a frequency inference mechanism is used to generate top-1.5k candidate frequent patterns. The third group of users checks the candidate frequent patterns to answer whether each pattern appears in their own trajectory transactions. Similarly, a privacy protection mechanism is used to protect the privacy of each user and the corresponding frequency estimation method is used to determine the final frequent patterns.

[0013] Further, in step S1, user grouping specifically includes: grouping users within the edge server management area using parallel K-means++ to divide the edge-managed terminal users into three groups with similar contributions to frequent pattern mining; specifically including the following steps:

[0014] S11: Calculate feature vectors: The edge server calculates the feature vectors of each user. For example, mining frequent trajectory patterns can calculate peak travel rate, average speed and hot spot travel rate during the journey and form feature vectors. Mining frequent patterns of consumer users can calculate the number of purchases, peak purchase rate and popular product purchase rate to collect user features.

[0015] S12: Parallel initialization. Each edge server uses the K-means++ initialization method to select cluster centers c1 = x[i], i = random(1,n) on the local subset, where x[i] represents the feature vector of the i-th data point; n represents the total number of feature vectors in the local subset; and initializes the local clustering model.

[0016] S13: Find local cluster centers. Perform K-means++ clustering independently on each participant, using locally initialized cluster centers, and iteratively update the local clustering model based on the data in the local subset. For the t-th cluster center, t∈[1,3], select according to the following formula:

[0017] c t = x[i], i = argmax(D(x[i]) 2 x[i]! = c1,c2,c3

[0018] Where D(x[i]) represents the minimum distance from the feature vector x[i] to the currently selected cluster center, that is:

[0019] D(x[i])=min(||x[i]-c1|| 2 ,||x[i]-c2|| 2 ,||x[i]-c3|| 2 )

[0020] S14: Update local cluster centers. Based on the selected new cluster centers, recalculate the distance from each sample point to the new cluster center and assign the sample point to the nearest cluster center. Iterate through S13 until three distinct cluster centers satisfy the convergence condition are obtained, resulting in the initialized cluster center set C. loc ={c1,c2,c3};

[0021] S15: Global Cluster Center Merging: After each iteration, the cluster centers of each participant are merged into a global cluster center to obtain the global clustering result. First, each local cluster center C = {C[1]} loc C[2] loc ,...,C[k] loc}, where C[k] loc The local cluster centers of the k-th edge server are calculated, and the three global center sets P of each local center are obtained. glo ={p1,p2,p3};

[0022] S16: User Grouping: Three groups were obtained based on the global cluster centers, and a certain number of users were randomly selected from these groups and evenly distributed among them to answer three tasks related to frequent pattern mining.

[0023] Furthermore, in step S2, the construction of candidate frequent 1-itemsets specifically includes the following steps:

[0024] S21: Random sampling: For each user's transaction A random sample of an item p is shown below:

[0025] p = Random_Sample(d)

[0026] in, Represents the total set of transactions for all users;

[0027] S22: Privacy Protection Method: Use a privacy protection method to protect the user-reported item p to ensure data privacy;

[0028] S23: Estimating the support count f (frequency) of item p: Using the corresponding estimation mechanism of the privacy protection method, the support count is estimated on the noisy data;

[0029] S24: Filtering candidate frequent one-itemsets P: Sort the items in descending order of frequency of occurrence, and select the top-2k most frequent items as candidate frequent one-itemsets P, that is:

[0030] Sorted_Items = Sort(p,f)

[0031] P = Sorted_Items[0:Top-2k]

[0032] Sort(p,f) represents sorting items p in descending order of frequency f to obtain the set Sorted_Items; the top 2k items from this set constitute a candidate frequent one-itemset P. In constructing the candidate frequent one-itemset based on the data transaction set, random sampling and privacy protection methods reduce the size of the input domain and protect data privacy, thereby improving the efficiency of frequency estimation. Finally, the top-2k items sorted in descending order of frequency are used as the candidate frequent one-itemset. Although the actual frequency of items is underestimated, this underestimation is global, meaning that the frequency is underestimated for each item to a certain extent, and therefore does not have an additional impact on its ranking.

[0033] Furthermore, step S22 specifically includes: dividing project p into two parts: area code and interval code, i.e.<region,sub_region> For example, geographic location information can be divided into regional information and precise location information; a hash function can be applied to the area code portion to map the area code to a fixed-length hash value to protect the privacy of the area code, i.e., H = h(region); a local differential privacy mechanism can be applied to the interval number portion to protect the privacy of the interval number. The local differential privacy mechanism can be a randomized algorithm used to add noise to the interval number to hide its true value. Assuming the true value of sub_region is x, the random response mechanism obtains the perturbed value y by perturbing x. That is:

[0034]

[0035] Where Pr[y|x] represents the probability of obtaining the perturbation value y given the true value x. When y equals x, that is, when y = x, the probability of the perturbation is... Where e represents the base of the natural logarithm, ε is the privacy budget, and k is the range of values ​​for the interval number sub_region. In this case, if the perturbation value y is equal to the true value x, it means that the sub_region has not been perturbed; when y is not equal to x, i.e., y≠x, the probability of perturbation is... In this case, the perturbation value y is not equal to the true value x, indicating that the sub-region is perturbed. To finely select the degree of interval division r, this invention considers minimizing the expected error, i.e., minimizing the difference between the estimated frequency and the true frequency. The implementation process is as follows:

[0036]

[0037] Among them, 2 r The number of area codes; the number of users assigned to each area code is n1, n2, ..., n. 2r ; f is the estimated frequency of the j-th interval number in the i-th interval number; ij Then it is its true frequency.

[0038] Furthermore, step S23 specifically includes: querying using a hash counter to obtain the number of region codes, which is n. region The frequency of interval numbers under a known area code can be obtained by using the local differential privacy frequency estimation method, as expressed by:

[0039]

[0040] Where C(x) is the number of times each user reports the value x; n regionq1 represents the total number of users, which is an unbiased estimate of the actual user count; q2 represents the probability of a disturbance when the disturbance value equals the actual value; and q3 represents the probability of a disturbance when the disturbance value is not equal to the actual value.

[0041] Furthermore, step S3 specifically includes the following steps:

[0042] S31: Encoding the user's response to the candidate frequent 1-itemset: Let T i Let P be the user's current transaction, P be the set of candidate frequent 1-itemsets, and V[T] be the user's current transaction. i [j] represents the current transaction T i The inclusion status of item j in set P;

[0043]

[0044] That is, if transaction T i If item j in V[T] appears in set P, then V[T] will be... i If [j] is set to 1, otherwise set to 0; in this way, the elements in the V vector will be automatically set to 1 or 0 according to the inclusion of items in the P set in the transaction, which can realize the support statistics of candidate frequent 1-itemsets.

[0045] S32: Privacy Protection of Uploaded Response V: The privacy protection method in step S22 is also used to protect the user-uploaded V to ensure data privacy;

[0046] S33: Estimate the occurrence of items in V: Similarly, use the estimation mechanism in S23 to estimate the noisy data and count the support of the Top-2k items;

[0047] S34: Check frequent 1-itemsets: Sort the items in descending order of frequency of occurrence, and select the top-k most frequent items as the final frequent 1-itemsets.

[0048] Furthermore, in step S4, the mining of frequent patterns involves using the frequent one-itemsets from the previous task to construct candidate frequent patterns. The mining of frequent patterns can be achieved by verifying these candidate patterns. Specifically, this includes the following steps:

[0049] S41: Constructing Candidate Frequent Patterns: Based on the top-k frequent 1-itemsets, a frequency inference mechanism is used to generate top-1.5k candidate frequent patterns. The formula for the inference mechanism is: Let pattern X be a set of top-2k candidate frequent patterns I = {x1, x2, ..., x...} 1.5k A subset of}, i.e. The frequency of any item p is known and is f(p), 1≤i≤m; then the expression for the predicted frequency A(X) of the itemset is:

[0050]

[0051] Among them, f max This represents the maximum frequency of all items in pattern X; γ is an adjustment parameter, typically set to 0.9; by multiplying the frequency of each item appearing in pattern X by 0.9, the influence of the maximum frequency item on the predicted frequency formula is supplemented; finally, a candidate set of possible frequent patterns is constructed using values ​​between 0 and 1. Candidate frequent patterns can be quickly constructed by predicting frequencies.

[0052] S42: Encode the user's response to the candidate frequent patterns. The implementation process is the same as step S41. The candidate frequent patterns contained by the user can be encoded as W for subsequent statistical counting.

[0053] S43: Response W for privacy protection upload: Use the same privacy protection method as in step S22 to protect the user-uploaded V to ensure data privacy;

[0054] S44: Estimate the frequency of candidate frequent patterns in W: Similarly, use the estimation mechanism in step S23 to estimate the noisy data and achieve the support count of the Top-1.5k candidate frequent patterns;

[0055] S45: Check Frequent Patterns: Sort the candidate frequent patterns in descending order of their frequency of occurrence, and select the top-k most frequent patterns to complete the final frequent pattern mining.

[0056] The beneficial effects of this invention are as follows:

[0057] (1) The user grouping method proposed in this invention divides the terminal users under edge server management area into three groups with similar contributions to frequent pattern mining by using parallel K-means++. This can effectively reduce the risk of a single user answering the task multiple times and speed up the completion of the mining task. At the same time, this method can also reduce the impact of data differences between different groups on the accuracy of frequent patterns.

[0058] (2) The privacy protection method proposed in this invention uses a combination of hash algorithm and local differential privacy mechanism to protect data privacy by using area code and interval code respectively, and collects support count of data through corresponding frequency estimation mechanism, which effectively reduces the impact of noise on frequency estimation in privacy protection during frequent pattern mining.

[0059] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0060] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0061] Figure 1 A model diagram of the frequent pattern mining system provided by this invention;

[0062] Figure 2 A diagram illustrating the implementation process of user grouping designed for this invention;

[0063] Figure 3 A diagram illustrating the privacy protection process designed for this invention. Detailed Implementation

[0064] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0065] Please see Figures 1-3 , Figure 1 The diagram shown is a model of a frequent pattern mining system, which includes the following components:

[0066] Terminal Layer: The terminal layer consists of vehicle devices that generate a large amount of driving trajectory data and divides these users into three groups. By collecting data from these groups, frequent trajectory patterns can be identified, ultimately providing car owners with richer and more personalized services.

[0067] Edge Layer: The edge layer consists of multiple roadside units equipped with edge servers. Geographically close to vehicles, it enables real-time data processing and response. This avoids transmitting data to the cloud, reducing the risk of data attacks and compromised links. Its primary task is to mine frequent trajectory patterns using data from vehicle responses.

[0068] Cloud layer: Cloud servers, as centers for large-scale data, can provide applications such as long-term data analysis, regular maintenance, and business decision-making. The collaboration between the cloud and the edge enables more comprehensive and efficient data processing and computing, improving business efficiency and data security.

[0069] To address the aforementioned system, this invention provides a method for mining frequent cloud-edge collaboration patterns, specifically comprising the following steps:

[0070] 1) User grouping: The edge server divides users into three groups based on their characteristics and assigns the three sub-tasks of frequent pattern mining to different groups.

[0071] 2) Construction of candidate frequent one-itemsets: The edge server constructs frequent one-itemsets based on the data of the first group. To reduce the input domain and focus on smaller candidate sets, an input domain reduction method is adopted. The trajectory of each user is randomly sampled and one item p is randomly reported to reduce the size of the input domain. Then, a privacy protection method is adopted to protect the items reported by the users, and a corresponding frequency estimation algorithm is used to calculate the frequency of a single item. Finally, the Top-2k set is obtained by descending the frequency of the items as candidate frequent one-itemsets.

[0072] 3) Verification of Frequent 1-Itemsets: Each user needs to answer whether their answer vector V contains the i-th item from the top-2k candidate frequent 1-itemsets. If it does, the i-th position is encoded as 1; otherwise, it is 0. This results in a binary string V of length 2k. User data privacy is involved during the answering process, so a privacy protection mechanism is used to protect the user's answer vector V. The server uses the corresponding frequency estimation to calculate the frequency of the candidate frequent 1-itemsets, sorts them in descending order of frequency, and obtains the top k frequent items and their frequencies, thus verifying the frequent 1-itemsets.

[0073] 4) Frequent Pattern Mining: Based on the top-k frequent one-itemsets, this section uses a frequency inference mechanism to generate top-1.5k candidate frequent patterns. The third group of users checks the candidate frequent patterns to answer whether each pattern appears in their own trajectory transactions. Similarly, privacy protection mechanisms are used to protect the privacy of each user, and corresponding frequency estimation methods are used to determine the final frequent patterns.

[0074] Please see Figure 2 , Figure 2 The diagram shows the user grouping method designed in this invention, which specifically includes the following steps:

[0075] 1) Calculate feature vectors: The edge server calculates the feature vectors of each user. For example, mining frequent trajectory patterns can calculate the peak travel rate, average speed and hot spot travel rate during the journey and form a feature vector. Mining frequent patterns of consumer users can calculate the number of purchases, peak purchase rate and popular product purchase rate to collect user features.

[0076] 2) Parallel initialization: Each edge server uses the K-means++ initialization method to select cluster centers c1 = x[i], i = random(1,n) on its local subset, where x[i] represents the feature vector of the i-th data point, and n represents the total number of data points in the local subset. It also initializes its local clustering model.

[0077] 3): Find local cluster centers. Perform K-means++ clustering independently on each participant, using locally initialized cluster centers, and iteratively update the local clustering model based on the data in the local subset. For the t-th cluster center (t∈[1,3]), select it according to the following formula:

[0078] c t = x[i], i = argmax(D(x[i]) 2 x[i]! = c1,c2,c3

[0079] Where D(x[i]) represents the minimum distance from the feature vector x[i] to the currently selected cluster center, that is:

[0080] D(x[i])=min(||x[i]-c1|| 2 ,||x[i]-c2|| 2 ,||x[i]-c3|| 2 )

[0081] 4): Update the local cluster centers. Based on the selected new cluster centers, recalculate the distance from each sample point to the new cluster center, and assign the sample points to the nearest cluster center. Iterate through S13 until three distinct cluster centers satisfy the convergence condition are obtained, thus obtaining the initialized cluster center set C. loc ={c1,c2,c3};

[0082] 5): Global cluster center merging: After each iteration, the cluster centers of each participant are merged into a global cluster center to obtain the global clustering result. First, each local cluster center C = {C[1]} loc C[2] loc ,...,C[k] loc}, where C[k] locThe local cluster centers of the k-th edge server are calculated, and the three global center sets P of each local center are obtained. glo ={p1,p2,p3};

[0083] 6): User grouping: Three groups were obtained based on the global cluster centers, and a certain number of users were randomly selected from these groups and evenly distributed among the three groups to answer three tasks related to frequent pattern mining.

[0084] Please see Figure 3 , Figure 3 The diagram shows the privacy protection process designed in this invention. The method specifically includes the following steps:

[0085] 1) Divide project p into two parts: area code and range code, i.e.<region,sub_region> For example, geographic location information can be divided into regional information and precise location information;

[0086] 2) Apply a hash function to the area code portion to map the area code to a fixed-length hash value to protect the privacy of the area code, i.e.:

[0087] H = h(region)

[0088] 3) Apply a local differential privacy mechanism to the interval number portion to protect its privacy. This mechanism can be a randomized algorithm used to add noise to the interval number to hide its true value. Assuming the true value of the sub_region is x, the random response mechanism perturbs x to obtain the perturbed value y. That is:

[0089]

[0090] Where Pr[y|x] represents the probability of obtaining the perturbation value y given the true value x. When y equals x, that is, when y = x, the probability of the perturbation is... Where e represents the base of the natural logarithm, ε is the privacy budget, and k is the range of values ​​for the interval number sub_region. In this case, if the perturbation value y is equal to the true value x, it means that the sub_region has not been perturbed; when y is not equal to x, i.e., y≠x, the probability of perturbation is... In this case, the perturbation value y is not equal to the true value x, indicating that the sub_region has been perturbed;

[0091] 5) To finely select the degree of interval division r, this section considers minimizing the expected error, that is, minimizing the difference between the estimated frequency and the true frequency. The implementation process is as follows:

[0092]

[0093] Among them, 2 r The number of area codes; the number of users assigned to each area code is n1, n2, ..., n. 2r ; f is the estimated frequency of the j-th interval number in the i-th interval number; ij Then it is its true frequency.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A privacy-preserving association rule learning method, characterized in that, The method specifically includes the following steps: S1: User Grouping: The edge server divides users into three groups based on their characteristics and assigns the three sub-tasks of frequent pattern mining to different groups. S2: Construction of candidate frequent one-itemsets: Based on the data from the first group, the edge server constructs frequent one-itemsets. Using a method of reducing the input domain, it randomly samples the trajectory of each user and randomly reports one item from it. Then, privacy-preserving methods are used to protect user-reported items, and corresponding frequency estimation algorithms are used to calculate the frequency of individual items. Finally, the items are sorted in descending order of frequency to obtain... Top -2 k The set is used as a candidate frequent 1-itemset; S3: Verification of Frequent 1-Itemsets: Each user in the second group needs to answer whether it contains... Top -2 k The first candidate in the frequent 1-item set i If the item is included, then the first item... i Each position is encoded as 1, otherwise as 0, resulting in a final value of length 2. k binary string V During the answering process, a privacy protection mechanism is employed to protect the user's answer vector. V The server uses a corresponding frequency estimation algorithm to calculate the frequency of candidate frequent 1-itemsets, sorts them in descending order of frequency, and obtains the top... k A set of frequent items and their frequencies, i.e., check the frequent 1-itemset; S4: Frequent Pattern Mining: Based on Top-k A frequent 1-itemset is generated using a frequency inference mechanism. Top -1.5 k A third group of users checks each candidate frequent pattern against the previous one, determining whether each pattern appears in their own trajectory transaction. Similarly, a privacy protection mechanism is used to protect the privacy of each user, and a corresponding frequency estimation method is used to determine the final frequent pattern.

2. The association rule learning method according to claim 1, characterized in that, In step S1, user grouping specifically includes: grouping users within the edge server management area using parallel K-means++ to divide the edge-managed terminal users into three groups with similar contributions to frequent pattern mining; specifically including the following steps: S11: Calculate feature vectors: Calculate the feature vectors of each user managed by the edge server; S12: Parallel initialization, each edge server selects cluster centers on its local subset using the K-means++ initialization method. ,in x [ i ] indicates the first i Feature vectors of data points; n This indicates the total number of feature vectors in the local subset; and initializes the local clustering model. S13: Find local cluster centers, perform K-means++ clustering independently on each participant, use locally initialized cluster centers, and iteratively update the local clustering model based on the data in the local subset. For the ... t Cluster centers, Choose according to the following formula: in, D ( x [ i ]) represents the eigenvector x [ i The minimum distance to the currently selected cluster centers, i.e.: S14: Update local cluster centers. Based on the selected new cluster centers, recalculate the distance from each sample point to the new cluster center and assign the sample point to the nearest cluster center. Iterate through S13 until three distinct cluster centers satisfy the convergence condition are obtained, thus obtaining the initialized set of cluster centers. ; S15: Global Cluster Center Merging: After each iteration, the cluster centers of each participant are merged into a single global cluster center to obtain the global clustering result. First, the local cluster centers... ,in It is the first k The local cluster centers of the edge server are calculated to obtain three global center sets for each local center. ; S16: User Grouping: Based on the global cluster centers, three groups are obtained, and a certain number of users are randomly selected from these groups and evenly distributed among them to answer three tasks related to frequent pattern mining.

3. The association rule learning method according to claim 2, characterized in that, In step S2, the construction of candidate frequent 1-itemsets specifically includes the following steps: S21: Random sampling: For each user's transaction Randomly sample an item p As shown below: in, Represents the total set of transactions for all users; S22: Privacy Protection Methods: Using privacy protection methods for user-reported items p Protect data to ensure privacy; S23: Estimated Items p Support count, i.e., frequency f The support score is calculated by using the corresponding estimation mechanism of privacy protection methods to estimate the noisy data. S24: Filter candidate frequent one-itemsets P Sort the items in descending order of frequency and select the most frequent ones. Top- 2 k Each item is a candidate frequent itemset P ,Right now: in, Sort ( p,f ) indicates according to frequency f Descending sort items p A set can be obtained ; and select the top 2 from them k The items constitute a candidate frequent itemset P .

4. The association rule learning method according to claim 3, characterized in that, Step S22 specifically includes: [The project is...] p It consists of two parts: the area code and the range code, i.e., < region , sub_region > Apply a hash function to the area code portion, mapping the area code to a fixed-length hash value to protect the privacy of the area code. Apply a local differential privacy mechanism to the interval number portion.

5. The association rule learning method according to claim 4, characterized in that, Step S23 specifically includes: obtaining the area code by querying the hash counter. region The number of The frequency of the interval number under the known area code is obtained by using the local differential privacy frequency estimation method, and the expression is: in, All reported values x Number of users; It represents the total number of users; This represents the probability that the perturbation value equals the true value. This represents the probability of a disturbance when the disturbance value is not equal to the true value.

6. The association rule learning method according to claim 3, characterized in that, Step S3 specifically includes the following steps: S31: Encoding the user's response to the candidate frequent 1-itemset: Let For the user's current transaction, For the set of candidate frequent 1-itemsets, Indicates the current transaction Items in exist The inclusion condition within a set; That is, if the transaction Items in exist If it appears in the set, then it will Set to 1 otherwise set to 0; S32: Answers uploaded with privacy protection V Similarly, the privacy method in step S22 is used to process user-uploaded data. V Protect data to ensure privacy; S33: Estimate V The occurrence of the project: The same estimation mechanism as S23 is used to estimate the noisy data, achieving... Top -2 k Support count for each project; S34: Check Frequent Itemsets: Sort items in descending order of frequency of occurrence and select the most frequent one. Top-k Each item is a final frequent 1-itemset.

7. The association rule learning method according to claim 6, characterized in that, Step S4 specifically includes the following steps: S41: Constructing candidate frequent patterns: based on Top-k A frequent 1-itemset is generated using a frequency inference mechanism. Top- 1.5 k Given a number of frequent candidate patterns, the inference mechanism can be implemented using the following formula: Let the pattern... X It is by Top -2 k A set of candidate frequent patterns A subset of the composition, namely any of the items The frequency of is known, and is , Then the predicted frequency of itemsets The expression is: in, Represented as a pattern X The highest frequency of occurrence among all items; It refers to adjusting parameters; S42: Encode the user's response to candidate frequent patterns. The implementation process is similar to step S41, encoding the candidate frequent patterns contained by the user into... W This is for subsequent statistical counting; S43: Answers uploaded with privacy protection W Similarly, the privacy protection method in step S22 is used to protect user-uploaded data. V Protect data to ensure privacy; S44: Estimate W The frequency of frequent candidate patterns is estimated using the same estimation mechanism as in step S23, which is applied to the noisy data. Top -1.5 k Support counts for each candidate frequent pattern; S45: Check Frequent Patterns: Sort candidate frequent patterns in descending order of frequency of occurrence, and select the one with the highest frequency. Top-k The goal is to identify the most frequent patterns, thus completing the final frequent pattern mining.

Images