Top-k frequent pattern mining method and system based on subjective and objective evaluation

Abstract

The invention relates to a top-k frequent pattern mining method and system based on subjective and objective evaluation, and belongs to the field of data mining. The problem that a top-k mode excavated by a traditional technology is often difficult to meet user requirements due to lack of an effective mode evaluation mode is solved. According to the technical scheme, the method comprises the following steps: firstly, coding modes by using a coding technology of a minimum DFS sequence; then, a pattern evaluation model based on a twin neural network (Graph Patterns Evaluation Model, GPEM) is built, a partial order relationship between pattern pairs is learned, and subjective preference prediction of the pattern is realized under the condition that subjective scoring of the pattern is difficult; and finally, designing and fusing subjective and objective mode interestingness evaluation functions to guide top-k mode mining. The invention provides a top-k frequent pattern mining method and system based on subjective and objective evaluation. A top-k frequent pattern mining task fusing user subjective preference and pattern objective information is realized.

Description

Technical Field

[0001] The present invention relates to the field of data mining, and in particular to a top- k Frequent pattern mining method and system. Background Art

[0002] Frequent pattern mining is one of the core issues in the field of data mining. Its purpose is to discover patterns whose support is not less than a specified threshold from large-scale graph data. In recent years, frequent pattern mining based on a single large graph has been widely used in social network analysis, bioinformatics, cheminformatics, etc., and has attracted great attention from the academic community. Existing frequent pattern mining methods for a single large graph aim to mine a complete result set. However, the computational cost of such methods increases exponentially with the expansion of the input data scale, resulting in severe challenges in practical applications on large-scale graph data. In addition, the result set on a large graph is usually large in scale, which greatly increases the difficulty for users to understand and use it. In view of this, the researchers designed the top- k Frequent pattern mining algorithm to find the top ranked k Frequent patterns of names. Such algorithms generally use "objective" interest indicators such as support and pattern size to quantitatively evaluate and rank patterns. In practice, some researchers use support as a criterion for pattern evaluation, which will result in the high-frequency patterns mined being too simple in structure, thus limiting the diversity of the results. On the other hand, although many researchers use pattern size as a metric and design corresponding top- k The pattern mining algorithm takes into account the structural information of the pattern, but there is still a situation where the mining results do not match the actual interests of the user. Based on the above situation, it is particularly important to design a pattern evaluation method that can take into account both the user's subjective preferences and the objective information of the pattern. Summary of the invention

[0003] In view of the problems existing in the existing objective mining algorithms, this paper designs a top- k Frequent pattern mining method and system.

[0004] A top- k Frequent pattern mining methods, including: Preprocess the pattern; Build a subjective evaluation prediction model (Graph Patterns Evaluation Model, GPEM for short). The model uses a twin neural network to learn the partial order relationship between pattern pairs and uses a single tower structure as the backbone network to achieve subjective evaluation of the pattern. Using the partial order relationship between patterns, we construct a pattern topological sort to achieve global sorting of patterns, compare the ranking of the predicted scores with the actual ranking, and evaluate the similarity of the rankings; Integrate user subjective preferences and objective information of patterns to achieve top- k Frequent pattern mining.

[0005] Preferably, the preprocessing of the pattern specifically includes: First, for a single large graph, the present invention uses the common neighbor-aware random walk sampling algorithm to perform graph sampling; then, pattern mining is performed on the sampled graph to obtain a set of patterns with different structures and sizes, and a neural network model is constructed to score the patterns. This is a challenging task because the data structure graph contains different numbers of nodes and edges, usually without a common node order, and the lack of vector representation of the pattern makes feature learning of the subgraph complicated, and it is difficult to accurately score the pattern set one by one; according to the structure of the graph and the node and edge labels, the pattern is standardized into the minimum DFS encoding form; then, one-hot encoding is used to convert the classification features into input suitable for the machine learning model.

[0006] Preferably, the subjective evaluation prediction model GPEM is constructed, the model adopts a twin neural network to learn the partial order relationship between pattern pairs, and uses a single tower structure as a backbone network to achieve subjective evaluation of the pattern, specifically including: The subjective score prediction model built is a twin neural network model based on a combination of a convolutional neural network (CNN) and a multi-layer perceptron (MLP). Some samples are selected from the pattern, and the user sets preferences. A partial order relationship is formed through pairwise combination, which serves as the input of the twin neural network. The first part is the CNN module, which specifically includes: for two parameter-sharing CNN modules, each module has three convolutional layers, and each convolutional layer is connected to a ReLU activation function to introduce nonlinear transformation to improve the expressiveness of the model; followed by a maximum pooling layer (MaxPooling) to downsample the feature map to reduce the size of the feature map and retain key features; the final output of the network is converted into a one-dimensional vector through a Flatten layer as the input of the subsequent MLP module; The second part is the MLP module, which specifically includes: after CNN, two parameter-sharing MLP modules are built, each of which integrates activation function, batch normalization (Batch Normalization) and Dropout regularization mechanisms are designed to further process low-dimensional features and predict preference scores. Specifically, the first fully connected layer of the network maps the input vector to a 1024-dimensional high-dimensional feature space to enhance the expressiveness of the original features. The ReLU activation function is used after each fully connected layer. The BatchNorm1d layer is introduced after the activation function to standardize the data of each batch and keep the input of each layer within a stable distribution range, thereby accelerating network training and improving model stability. Subsequently, the Dropout layer is added to reduce the network's dependence on specific features and reduce the risk of overfitting by randomly disabling a part of neurons and their connections during training, thereby improving the adaptability of the model. On this basis, the network gradually reduces the dimension of the features (from 1024 to 512, then to 256, 128, 64, 32, and finally to 1). The design of each fully connected layer is designed to further compress the feature space while retaining the most critical feature information. The terminal layer of the network reduces the feature dimension to 1 and outputs a preference score for the prediction task of the model. In the output layer of the model, the softmax function is first used to convert the output of the network into a probability distribution. Subsequently, the model parameters are optimized by minimizing the cross entropy loss to guide the gradient descent process and make the predicted probability distribution as close to the true label distribution as possible. Next, after constructing a parameter-sharing twin neural network, the model further adopts a single-tower structure as its backbone network. It inputs a single vector and outputs a preference score, thereby constructing a model for subjective evaluation.

[0007] Preferably, the partial order relationship between patterns is utilized to construct pattern topological sorting to achieve global sorting of patterns, and the ranking of predicted scores is compared with the actual ranking to evaluate the similarity of the rankings, which specifically includes: Since the input pattern is presented in the form of a partially ordered pair relationship and there is a lack of global sorting results, it is impossible to directly evaluate the model performance by comparing the model prediction results with the actual sorting results. To solve this problem, the present invention proposes a global sorting strategy based on preference relations, namely: according to the partially ordered relationship between the patterns, a pattern topological sorting is constructed to achieve pattern global sorting. Specifically, for each pair of pattern vectors in the data, a corresponding directed edge is constructed in a directed graph according to the preference relationship, and the in-degree and out-degree information of all nodes are updated. Before performing the topological sorting, it is first necessary to perform a ring detection on the graph through a depth-first search. If a ring is detected in the graph, it indicates that there is a mutually dependent preference relationship and an effective topological sorting cannot be generated. In this case, the ring is reduced to a superpoint, and all nodes in the ring are regarded as equivalent, thereby converting the original graph into a new directed acyclic graph. On this basis, the new directed acyclic graph is topologically sorted. If there is a ring in the graph and it is reduced to a superpoint, the nodes in the superpoint have the same topological sorting. This method effectively constructs a preference-based global sorting sequence. The ranking of the predicted score is compared with the topological ranking of the original partially ordered pair to evaluate the similarity of the ranking.

[0008] Preferably, the fusion of subjective evaluation and objective evaluation realizes top- k Frequent pattern mining, including: Top- k The frequent pattern mining problem is defined as: given a graph G and an integer k, the top k pattern set S is mined from the graph G; this paper proposes a new interest index Itr to evaluate the pattern, aiming to mine the top- k Frequent patterns; The interestingness of pattern Q is defined as follows: In the formula, Itr(Q) is the interest index, , α represents the subjective score weight; model score is the subjective score, which is obtained by the model, and the objective score is given by the formula Decision, among which is the support of pattern Q, is the size of the pattern; by integrating the subjective score and the objective score, the top- k A frequent pattern.

[0009] A top- k Frequent pattern mining system, including: A data preprocessing module, used to preprocess the pattern; Design the model module. First, build a parameter-sharing twin neural network GPEM to learn the partial order relationship between pattern pairs. Then extract a single tower model from the trained twin neural network to predict the preference score of a single input vector. The global sorting module uses the partial order relationship between patterns to construct a pattern topological sorting to achieve global sorting of patterns, compares the ranking of the predicted scores with the actual ranking, and evaluates the similarity of the rankings; The subjective and objective evaluation combination module is used to achieve top- k Frequent pattern mining.

[0010] Preferably, the data preprocessing module includes: First, for a single large graph, the present invention uses the common neighbor-aware random walk sampling algorithm to perform graph sampling; then, pattern mining is performed on the sampled graph to obtain a set of patterns with different structures and sizes, and a neural network model is constructed to score the patterns. This is a challenging task because the data structure graph contains different numbers of nodes and edges, usually without a common node order, and the lack of vector representation of the pattern makes feature learning of the subgraph complicated, and it is difficult to accurately score the pattern set one by one; according to the structure of the graph and the node and edge labels, the pattern is standardized into the minimum DFS encoding form; then, one-hot encoding is used to convert the classification features into input suitable for the machine learning model.

[0011] Preferably, the design model module includes: The subjective score prediction model built is a twin neural network model based on a combination of a convolutional neural network (CNN) and a multi-layer perceptron (MLP). Some samples are selected from the pattern, and the user sets preferences. A partial order relationship is formed through pairwise combination, which serves as the input of the twin neural network. The first part is the CNN module, which specifically includes: for two parameter-sharing CNN modules, each module has three convolutional layers, and each convolutional layer is connected to a ReLU activation function to introduce nonlinear transformation to improve the expressiveness of the model; followed by a maximum pooling layer (MaxPooling) to downsample the feature map to reduce the size of the feature map and retain key features; the final output of the network is converted into a one-dimensional vector through a Flatten layer as the input of the subsequent MLP module; The second part is the MLP module, which specifically includes: after CNN, two parameter-sharing MLP modules are built, each of which integrates activation function, batch normalization (Batch Normalization) and Dropout regularization mechanisms are designed to further process low-dimensional features and predict preference scores. Specifically, the first fully connected layer of the network maps the input vector to a 1024-dimensional high-dimensional feature space to enhance the expressiveness of the original features. The ReLU activation function is used after each fully connected layer. The BatchNorm1d layer is introduced after the activation function to standardize the data of each batch and keep the input of each layer within a stable distribution range, thereby accelerating network training and improving model stability. Subsequently, the Dropout layer is added to reduce the network's dependence on specific features and reduce the risk of overfitting by randomly disabling a part of neurons and their connections during training, thereby improving the adaptability of the model. On this basis, the network gradually reduces the dimension of the features (from 1024 to 512, then to 256, 128, 64, 32, and finally to 1). The design of each fully connected layer is designed to further compress the feature space while retaining the most critical feature information. The terminal layer of the network reduces the feature dimension to 1 and outputs a preference score, which is used for the prediction task of the model. In the output layer of the model, the softmax function is first used to convert the output of the network into a probability distribution. Subsequently, the model parameters are optimized by minimizing the cross entropy loss to guide the gradient descent process and make the predicted probability distribution as close to the true label distribution as possible. Next, after constructing a parameter-sharing twin neural network, the model further adopts a single-tower structure as its backbone network. It inputs a single vector and outputs a preference score, thereby constructing a model for subjective evaluation.

[0012] Preferably, the global sorting module comprises: Since the input pattern is presented in the form of a partially ordered pair relationship and there is a lack of global sorting results, it is impossible to directly evaluate the model performance by comparing the model prediction results with the actual sorting results. To solve this problem, the present invention proposes a global sorting strategy based on preference relations, namely: according to the partially ordered relationship between the patterns, a pattern topological sorting is constructed to achieve pattern global sorting. Specifically, for each pair of pattern vectors in the data, a corresponding directed edge is constructed in a directed graph according to the preference relationship, and the in-degree and out-degree information of all nodes are updated. Before performing the topological sorting, it is first necessary to perform a ring detection on the graph through a depth-first search. If a ring is detected in the graph, it indicates that there is a mutually dependent preference relationship and an effective topological sorting cannot be generated. In this case, the ring is reduced to a superpoint, and all nodes in the ring are regarded as equivalent, thereby converting the original graph into a new directed acyclic graph. On this basis, the new directed acyclic graph is topologically sorted. If there is a ring in the graph and it is reduced to a superpoint, the nodes in the superpoint have the same topological sorting. This method effectively constructs a preference-based global sorting sequence. The ranking of the predicted score is compared with the topological ranking of the original partially ordered pair to evaluate the similarity of the ranking.

[0013] Preferably, the subjective and objective evaluation combining module comprises: Top- k The frequent pattern mining problem is defined as: given a graph G and an integer k, the top k ranked pattern set S is mined from the graph G; this paper proposes a new interest index Itr to evaluate the pattern, aiming to mine the top- k Frequent patterns; The interestingness of pattern Q is defined as follows: In the formula, Itr(Q) is the interest index, , α represents the subjective score weight; model score is the subjective score, which is obtained by the model, and the objective score is given by the formula Decision, among which is the support of pattern Q, is the size of the pattern; by integrating the subjective score and the objective score, the top- k A frequent pattern. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. It is obvious that the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0015] Figure 1 This invention is a top- k A method flow chart of an embodiment of a frequent pattern mining method.

[0016] Figure 2 This invention is a top- k Schematic diagram of the model design module of the frequent pattern mining system.

[0017] Figure 3 This invention is a top- k Schematic diagram of the global ranking module of the frequent pattern mining system.

[0018] Figure 4 This invention is a top- k Comparison chart of model results of frequent pattern mining system.

[0019] Figure 5 This invention is a top- k Model stability test diagram under different mining algorithms of frequent pattern mining system. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] The purpose of this invention is to provide a top- k Frequent pattern mining method and system aims to mine the top- k A frequent pattern.

[0022] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] refer to Figure 1 , which is a top- k A method flow chart of an embodiment of a frequent pattern mining method, the method comprising the following steps: Preprocess the pattern. The specific steps include: Perform graph sampling on a single large graph to generate a sampling graph; Perform pattern mining on the sampling graph to obtain a set of patterns; Normalize the pattern into a minimal DFS encoding form based on the graph structure and node and edge labels; Use one-hot encoding to convert categorical features into input suitable for machine learning models.

[0024] Build a parameter-sharing twin neural network GPEM to learn the partial order relationship between pattern pairs, and extract a single tower model from the trained twin neural network to predict the preference score of a single input vector; the specific steps include: The subjective score prediction model built is a twin neural network model based on a combination of a convolutional neural network (CNN) and a multi-layer perceptron (MLP). Some samples are selected from the pattern, and the user sets preferences. A partial order relationship is formed through pairwise combination, which serves as the input of the twin neural network. The first part is the CNN module, which specifically includes: for two parameter-sharing CNN modules, each module has three convolutional layers, and each convolutional layer is connected to a ReLU activation function to introduce nonlinear transformation to improve the expressiveness of the model; followed by a maximum pooling layer (MaxPooling) to downsample the feature map to reduce the size of the feature map and retain key features; the final output of the network is converted into a one-dimensional vector through a Flatten layer as the input of the subsequent MLP module; The second part is the MLP module, which specifically includes: after CNN, two parameter-sharing MLP modules are built, each of which integrates activation function, batch normalization (Batch Normalization) and Dropout regularization mechanisms are designed to further process low-dimensional features and predict preference scores. Specifically, the first fully connected layer of the network maps the input vector to a 1024-dimensional high-dimensional feature space to enhance the expressiveness of the original features. The ReLU activation function is used after each fully connected layer. The BatchNorm1d layer is introduced after the activation function to standardize the data of each batch and keep the input of each layer within a stable distribution range, thereby accelerating network training and improving model stability. Subsequently, the Dropout layer is added to reduce the network's dependence on specific features and reduce the risk of overfitting by randomly disabling a part of neurons and their connections during training, thereby improving the adaptability of the model. On this basis, the network gradually reduces the dimension of the features (from 1024 to 512, then to 256, 128, 64, 32, and finally to 1). The design of each fully connected layer is designed to further compress the feature space while retaining the most critical feature information. The terminal layer of the network reduces the feature dimension to 1 and outputs a preference score for the prediction task of the model. In the output layer of the model, the softmax function is first used to convert the output of the network into a probability distribution. Subsequently, the model parameters are optimized by minimizing the cross entropy loss to guide the gradient descent process and make the predicted probability distribution as close to the true label distribution as possible. Next, after constructing a parameter-sharing twin neural network, the model further adopts a single-tower structure as its backbone network. It inputs a single vector and outputs a preference score, thereby constructing a model for subjective evaluation.

[0025] Use topological sorting to construct the global preference order of the pattern and obtain the real sorting result; compare the ranking of the predicted score with the topological ranking of the original partial order pair and evaluate the similarity of the ranking; the specific steps include: According to the partial order relationship between patterns, the present invention adopts a method combining topological sorting with depth-first search to construct a global pattern sorting sequence; before executing the topological sorting, it is first necessary to perform a ring detection on the graph through a depth-first search. If a ring is detected in the graph, it indicates that there is a mutually dependent preference relationship and an effective topological sorting cannot be generated; in this case, the ring is reduced to a superpoint, and all nodes in the ring are regarded as equivalent, thereby converting the original graph into a new directed acyclic graph; on this basis, the new directed acyclic graph is topologically sorted; if there is a ring in the graph and it is reduced to a superpoint, the nodes in the superpoint have the same topological sorting; this method effectively constructs a preference-based global sorting sequence; compare the ranking of the predicted score with the topological ranking of the original partial order pair, and evaluate the similarity of the ranking.

[0026] Integrate subjective and objective evaluation to achieve top- k Frequent pattern mining, the specific steps include: Top- k The frequent pattern mining problem is defined as: given a graph G and an integer k, the top k ranked pattern set S is mined from the graph G; this paper proposes a new interest index Itr to evaluate the pattern, aiming to mine the top- k Frequent patterns; The interestingness of pattern Q is defined as follows: In the formula, Itr(Q) is the interest index, , α represents the subjective score weight; model score is the subjective score, which is obtained by the model, and the objective score is given by the formula Decision, among which is the support of pattern Q, is the size of the pattern; by integrating the subjective score and the objective score, the top- k A frequent pattern.

[0027] refer to Figure 2, which is a top- k Schematic diagram of the model design module of the frequent pattern mining system, the system includes the following modules: First, a parameter-sharing twin neural network GPEM is built to learn the partial order relationship between pattern pairs. Then, a single tower model is extracted from the trained twin neural network to predict the preference score of a single input vector. Specifically, the following are included: The subjective score prediction model built is a twin neural network model based on a combination of a convolutional neural network (CNN) and a multi-layer perceptron (MLP). Some samples are selected from the pattern, and the user sets preferences. A partial order relationship is formed through pairwise combination, which serves as the input of the twin neural network. The first part is the CNN module, which specifically includes: for two parameter-sharing CNN modules, each module has three convolutional layers, each convolutional layer is connected to a ReLU activation function and a maximum pooling layer (MaxPooling); the final output of the network is converted into a one-dimensional vector through the Flatten layer as the input of the subsequent MLP module; The second part is the MLP module, which specifically includes: after CNN, two parameter-sharing MLP modules are built, each of which integrates activation function, batch normalization and Dropout regularization mechanism; specifically, the first fully connected layer of the network maps the input vector to a high-dimensional feature space of 1024 dimensions, and the ReLU activation function is used after each fully connected layer. The BatchNorm1d layer is introduced after the activation function, and then the Dropout layer is added; on this basis, the network gradually reduces the dimension of the feature (from 1024 to 512, then to 256, 128, 64, 32, and finally to 1), and the design of each fully connected layer is to further compress the feature space while retaining the most critical feature information; the terminal layer of the network reduces the feature dimension to 1 and outputs a preference score for the prediction task of the model; In the output layer of the model, the softmax function is first used to convert the output of the network into a probability distribution. Subsequently, the model parameters are optimized by minimizing the cross entropy loss to guide the gradient descent process and make the predicted probability distribution as close to the true label distribution as possible. Next, after constructing a parameter-sharing twin neural network, the model further adopts a single-tower structure as its backbone network. It inputs a single vector and outputs a preference score, thereby constructing a model for subjective evaluation.

[0028] refer to Figure 3 , which is a top- k Schematic diagram of a global ranking module of a frequent pattern mining system, the system comprising the following steps: Use topological sorting to construct the global preference order of the pattern to obtain the actual sorting result; compare the ranking of the predicted score with the topological ranking of the original partial order pair to evaluate the similarity of the ranking; specifically include: Since the input patterns are presented in the form of partially ordered pair relationships and lack global sorting results, it is impossible to directly evaluate the model performance by comparing the model prediction results with the actual sorting results. To solve this problem, the present invention proposes a global sorting strategy based on preference relations, namely: according to the partial order relationship between patterns, a pattern topological sorting is constructed to achieve global sorting of patterns. Specifically, for each pair of pattern vectors in the data, a corresponding directed edge is constructed in a directed graph according to the preference relationship, and the in-degree and out-degree information of all nodes is updated. Before performing the topological sorting, a depth-first traversal search is first performed to obtain the global sorting results. The graph is tested for loops. If a loop is detected in the graph, it indicates that there is a mutually dependent preference relationship and an effective topological sorting cannot be generated. In this case, the loop is reduced to a superpoint and all nodes in the loop are considered equivalent, thereby converting the original graph into a new directed acyclic graph. On this basis, the new directed acyclic graph is topologically sorted. If there is a loop in the graph and it is reduced to a superpoint, the nodes in the superpoint have the same topological sorting. This method effectively constructs a global sorting sequence based on preference. The ranking of the predicted score is compared with the topological ranking of the original partial order pair to evaluate the similarity of the rankings.

[0029] refer to Figure 4 , which is a top- kModel result comparison chart of pattern mining system; shows the accuracy experimental results of each model on 6 real data sets, p is the ratio of the number of predicted patterns, starting from 0.05, increasing by 0.05 each time until 0.20; the information of the 6 sets of graph data sets is: Amazon is a product joint procurement network, when the frequency of two products being purchased by customers at the same time reaches a certain number, an edge will be formed; DBLP is a paper publishing network diagram, each node represents a paper, and each edge represents the cooperation relationship between two authors; Mico is a network built by Microsoft co-authors, each node represents an author, and each edge represents the cooperation relationship between two authors; Patent is a patent Citation network, each node represents a patent, each edge is represented as a citation relationship; Twitch is a social network graph, each node represents a Twitch user, each edge represents the attention of two users; Twitter is a social network graph, each node represents a Twitter user, each edge represents the attention of two users; As can be seen from the figure, except in the DBLP dataset, the accuracy of GPEM is slightly lower than that of the CNN+MLP model when the sampling rate is 0.20, and its performance in other cases is better than the two comparison models; For the Mico dataset and Twitch dataset, the accuracy of GPEM has a trend of first decreasing and then increasing. The reason is that when the proportion of the dataset selection is small, the number of samples is not enough to fully train the model, resulting in the weak generalization ability of the model in the small sample stage, resulting in a short-term decrease in accuracy; However, as the sample size increases, the model can better learn the data features, and the accuracy increases accordingly.

[0030] refer to Figure 5 , which is a top- k Model stability test diagram of the frequent pattern mining system under different mining algorithms; using three different mining algorithms, three pattern sets of the same size are obtained, and the pattern scores are obtained according to the interest function Itr. The ranking is obtained from the scores to evaluate the use of the model for top- kStability of frequent pattern mining; k is the number of patterns. On 6 graph data sets, the number of patterns increases from 100 to 500 in increments of 100 to test the changes in the accuracy of the three methods; the information of the 6 graph data sets is as follows: Amazon is a product joint procurement network. When the frequency of two products being purchased by customers at the same time reaches a certain number, an edge will be formed; DBLP is a paper publishing network graph, each node represents a paper, and each edge represents the cooperative relationship between two authors; Mico is a network built by Microsoft co-authors, each node represents an author, and each edge represents the cooperative relationship between two authors; Patent is a patent citation network, each node represents a patent, and each edge is represented as a citation relationship; Twitch is a social network graph, each node represents a Twitch user, and each edge represents the attention of two users; Twitter is a social network graph, each node represents a Twitter user, and each edge represents the attention of two users; It can be seen from the figure that as the scale of the graph increases, the accuracy of all algorithms also increases, and the model shows good stability.

[0031] In summary, compared with the prior art, the present invention discloses a top- k The frequent pattern mining method and system have the following beneficial effects: (1) The present invention proposes a pattern representation method combining minimum DFS coding and One-hot coding, which solves the representation problem of pattern vectors; (2) The present invention designs a pattern subjective evaluation prediction model, which adopts a twin neural network to learn the information of user partial order relationship pairs, and then extracts a single tower structure as the backbone network to realize the subjective evaluation of the pattern; (3) The present invention proposes an interest measurement index based on user subjective preference and objective pattern information, aiming to mine the top- k frequent patterns; (4) A large number of experiments were conducted on 6 real graph datasets to verify the performance of GPEM; the experimental results show that compared with other models, GPEM performs well in various indicators, and its accuracy can reach up to 93% when the test set sampling ratio is only 20%.

[0032] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A top- k The frequent pattern mining method is characterized by: The following steps are involved: Preprocess the pattern; Build a subjective evaluation prediction model GPEM, use a twin neural network to learn the partial order relationship between pattern pairs, and use a single tower structure as the backbone network to achieve subjective evaluation of the pattern; Using the partial order relationship between pattern pairs, we construct the pattern topological sorting to achieve global sorting of patterns. Integrate user subjective preferences and objective information of patterns to achieve top- k Frequent pattern mining.

2. A top-down method based on subjective and objective evaluation according to claim 1 k The frequent pattern mining method is characterized by: The preprocessing of the pattern specifically includes: Firstly, a single large graph is sampled using the common neighbor-aware random walk sampling algorithm to generate a sampling graph. Secondly, pattern mining is performed on the sampling graph to obtain the pattern. Subsequently, the pattern is standardized into the minimum DFS encoding form according to the structure of the graph and the node and edge labels. Finally, One-hot encoding is used to convert the pattern into the input of the machine learning model.

3. A top-down method based on subjective and objective evaluation according to claim 1 k The frequent pattern mining method is characterized by: The subjective evaluation prediction model GPEM adopts a twin neural network to learn the partial order relationship between pattern pairs, and uses a single tower structure as a backbone network to achieve subjective evaluation of the pattern, specifically including: The subjective evaluation prediction model GPEM is a twin neural network model based on a combination of a convolutional neural network and a multi-layer perceptron. Some samples are selected from the pattern, and the user sets preferences. The partial order relationship is formed through pairwise combination as the input of the twin neural network. The first part is the convolutional neural network module, which specifically includes: designing two parameter-sharing convolutional neural network modules, each of which has three convolutional layers, and each convolutional layer is connected to a ReLU activation function and a maximum pooling layer; the final output of the convolutional neural network module is converted into a one-dimensional vector through a Flatten layer as the input of the subsequent multi-layer perceptron; The second part is the multi-layer perceptron module, which specifically includes: after the convolutional neural network, two multi-layer perceptron modules are connected, each of which integrates activation function, batch normalization and Dropout regularization mechanism; specifically, the first fully connected layer of the module maps the input vector to a 1024-dimensional high-dimensional feature space, and the ReLU activation function is used after each fully connected layer, and then the batch normalization layer and Dropout layer are introduced; on this basis, the network gradually reduces the dimension of the feature, and the design of each fully connected layer aims to further compress the feature space while retaining the most critical feature information; the terminal layer of the network reduces the feature dimension to 1 and outputs a preference score for the prediction task of the model; In the output layer of the model, the softmax function is first used to convert the output of the network into a probability distribution. Subsequently, the model parameters are optimized by minimizing the cross entropy loss to guide the gradient descent process so that the predicted probability distribution is as close to the true label distribution as possible. Next, after constructing a parameter-sharing twin neural network, the model further adopts a single-tower structure as its backbone network. The input of a single vector outputs a score, thereby constructing a model for subjective evaluation.

4. A top-down method based on subjective and objective evaluation according to claim 1 k The frequent pattern mining method is characterized by: The method of utilizing the partial order relationship between pattern pairs to construct pattern topological sorting and realize pattern global sorting specifically includes: According to the partial order relationship between patterns, a pattern topological sorting is constructed to achieve global sorting of patterns. Before executing the topological sorting, it is first necessary to perform a loop detection on the graph through depth-first traversal to find the strongly connected components in the graph and shrink them into super-points to form a directed acyclic graph. The nodes in the super-points have the same topological sorting. Then, a topological sorting is performed on the topological graph, and finally a preference order for all patterns is constructed.

5. A top-down method based on subjective and objective evaluation according to claim 1 k The frequent pattern mining method is characterized by: The integration of user subjective preferences and mode objective information achieves top- k Frequent pattern mining, including: Top- k The frequent pattern mining problem is defined as: given a graph G and an integer k, the top k pattern set S is mined from the graph G; a new interest index Itr is proposed to evaluate the pattern, aiming to mine the top- k Frequent patterns; The interestingness of pattern Q is defined as follows: In the formula, Itr(Q) is the interest index, , α represents the subjective score weight; model score is the subjective score, which is obtained by the model, and the objective score is given by the formula Decision, among which is the support of pattern Q, is the size of the pattern; by integrating the subjective score and the objective score, the top- k A frequent pattern.

6. A top- k The frequent pattern mining system is characterized by: The following steps are involved: A data preprocessing module, used to preprocess the pattern; Design a model module to build a subjective evaluation prediction model GPEM, use a twin neural network to learn the partial order relationship between pattern pairs, and use a single tower structure as the backbone network to achieve subjective evaluation of the pattern; The global sorting module is used to use the partial order relationship between pattern pairs to construct pattern topological sorting to achieve global sorting of patterns; The subjective and objective evaluation combination module is used to integrate user subjective preferences and objective information of the model to achieve top- k Frequent pattern mining.

Images