Dynamic temporal network clustering method and system based on multi-layer feature cascade

Abstract

The present invention discloses a dynamic temporal network clustering method and system with multi-layer feature cascade. The method comprises: obtaining a dynamic temporal network from social data, and representing the dynamic temporal network as an adjacency matrix through preset rules; inputting the adjacency matrix into a pre-trained variational autoencoder to obtain multiple feature outputs that match the number of levels of the pre-trained variational autoencoder; performing cluster analysis based on the multiple feature outputs and the K-means algorithm and outputting social statistical clustering results; wherein the pre-trained variational autoencoder is obtained by training with an improved loss function; the improved loss function is a function containing time features and fusion features. In the present invention, by adding time features and fusion features to the improved loss function, the influence of historical moments can be added to the data analysis, and at the same time, the information complementarity between the multiple layers of features is utilized to make the performance of the data features smoother, thereby improving the accuracy of the data clustering analysis.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and in particular relates to a multi-layer feature cascade dynamic temporal network clustering method and system. Background Art

[0002] Networks are ubiquitous in nature and human society, especially in today's rapidly evolving technology. Networks can be analogized to graphs, where individuals and their relationships can be compared to vertices and edges. Over the past decade, with the rise of big data and cloud computing, the field of time series clustering has undergone significant changes and developments. Improving the efficiency, quality, and sophistication of time series clustering methods has become a major trend.

[0003] In social networks, nodes typically represent users, and friendships or interactions between users form edges. Existing social network clustering methods can be divided into two categories: coupled graph-based methods and two-stage methods. Coupled graph-based methods convert the dynamic network in the social process into a static network and then apply a graph clustering algorithm, but this method ignores cluster drift. To address this issue, the two-stage method first independently obtains community division results from the dynamic network data at each moment, and then matches the community division results to track the evolution of the community. However, because this method analyzes clustering accuracy and cluster drift separately, the final features can only contain clustering accuracy or cluster drift, which in turn makes the accuracy of social network clustering results low. Summary of the Invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a multi-layer feature cascade dynamic temporal network clustering method and system.

[0005] The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0006] In a first aspect, the present invention provides a multi-layer feature cascade dynamic temporal network clustering method, comprising:

[0007] Obtain a dynamic temporal network from social data and represent it as an adjacency matrix using preset rules;

[0008] Input the adjacency matrix into the pre-trained variational autoencoder to obtain multiple feature outputs that match the number of layers of the pre-trained variational autoencoder;

[0009] Perform cluster analysis based on multiple feature outputs and the K-means algorithm and output social statistical clustering results;

[0010] Among them, the pre-trained variational autoencoder is trained through an improved loss function; the improved loss function is a function containing time features and fusion features.

[0011] Optionally, cluster analysis is performed based on multiple feature outputs and a K-means algorithm, and social statistical clustering results are output, including:

[0012] Multiple feature outputs are weighted according to their importance to obtain feature cascade results;

[0013] The feature concatenation results are clustered using the K-means algorithm to output the social statistical clustering results.

[0014] Optionally, multiple feature outputs are weighted according to their importance to obtain a feature concatenation result, including:

[0015] Setting a plurality of first weight coefficients that are negatively correlated with the levels of the pre-trained variational autoencoder to obtain an importance degree; wherein the number of the first weight coefficients is equal to the number of levels of the pre-trained variational autoencoder, and the sum of the plurality of first weight coefficients is 1;

[0016] Multiply multiple feature outputs by their matching importance, and superimpose the multiplication results to obtain the feature cascade result.

[0017] Optionally, the improved loss function includes three parts: reconstruction error, cluster drift, and encoder probability distribution;

[0018] The reconstruction error, clustering drift, and encoder probability distribution are added together to obtain an improved loss function.

[0019] Among them, the reconstruction error is used to calculate the reconstruction error between the training high-order information matrix and the reconstructed training high-order information matrix; the clustering drift is used to calculate the feature difference of the training high-order information matrix at adjacent moments; the encoder probability distribution is used to measure the difference between the pre-trained variational autoencoder and the predefined probability distribution.

[0020] Optionally, the calculation formula of the reconstruction error is expressed as:

[0021]

[0022] Among them, L 1’ represents the reconstruction error; M t Indicates that the high-order information matrix is ​​trained at time t, which is the adjacency matrix W trained at time t t After PMI matrix transformation, we get: Represents the reconstructed training high-order information matrix at time t; mse represents the mean square error; α represents the second weight coefficient.

[0023] Optionally, the calculation formula for cluster drift is expressed as:

[0024]

[0025] Among them, L 2’ represents cluster drift, Indicates M t The output features after the i-th layer of the pre-trained variational autoencoder at time t, Indicates M t The output features after the jth layer of the pre-trained variational autoencoder at time t-1, i and j both represent the number of layers of the pre-trained variational autoencoder, and l is the total number of layers of the pre-trained variational autoencoder.

[0026] Optionally, the calculation formula of the improved loss function is expressed as:

[0027]

[0028] Among them, L all’ represents the improved loss function, represents the output probability distribution of the pre-trained variational autoencoder The difference between it and the predefined probability distribution p(z); is the input data M t The probability distribution of the output latent variable z; φ represents the network parameters of the pre-trained variational autoencoder; D KL represents the KL divergence.

[0029] Optionally, the training process of the pre-trained variational autoencoder includes:

[0030] Obtain a training dynamic temporal network from social data, and represent the training dynamic temporal network as a training adjacency matrix using preset rules;

[0031] The training adjacency matrix is ​​input into the variational autoencoder, and the variational autoencoder is trained in the direction of decreasing the improved loss function to obtain a pre-trained variational autoencoder.

[0032] Optionally, after performing cluster analysis based on multiple feature outputs and a K-means algorithm and outputting social statistical clustering results, a multi-layer feature cascade dynamic time series network clustering method further includes:

[0033] Use social statistical clustering results for image analysis, information retrieval, or data compression.

[0034] In the second aspect, the present invention provides a multi-layer feature cascade dynamic temporal network clustering system, comprising: a processor, a storage medium and a bus, the storage medium storing machine-readable instructions executable by the processor, when the multi-layer feature cascade dynamic temporal network clustering system is running, the processor and the storage medium communicate through the bus, and the processor executes the machine-readable instructions to perform the steps of the method of the first aspect mentioned above.

[0035] The present invention provides a method and system for clustering dynamic time series networks with multi-layer feature cascades. Among them, a method for clustering dynamic time series networks with multi-layer feature cascades includes: obtaining a dynamic time series network from social data, and expressing the dynamic time series network as an adjacency matrix through preset rules; inputting the adjacency matrix into a pre-trained variational autoencoder to obtain multiple feature outputs that match the number of levels of the pre-trained variational autoencoder; performing cluster analysis based on the multiple feature outputs and the K-means algorithm and outputting social statistical clustering results; wherein, the pre-trained variational autoencoder is obtained by training with an improved loss function; the improved loss function is a function containing time features and fusion features. In the present invention, by adding time features to the improved loss function, when performing cluster analysis of dynamic time series networks, the impact of historical moments on current moments can be added to the data analysis, thereby improving the accuracy of social data cluster analysis; in addition, by adding fusion features and utilizing the information complementarity between multiple layers of features, the performance of data features is made smoother, further improving the accuracy of social data cluster analysis.

[0036] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A schematic diagram of a process flow of a multi-layer feature cascade dynamic time series network clustering method provided by an embodiment of the present invention;

[0038] Figure 2 A multi-layer feature extraction model constructed using a variational autoencoder provided in an embodiment of the present invention;

[0039] Figure 3 A schematic diagram of a multi-layer feature cascade dynamic temporal network clustering system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The present invention provides a dynamic time series network clustering method with multi-layer feature cascade, which uses a variational autoencoder to extract features from the adjacency matrix and takes into account the factors of historical moments. Compared with ordinary autoencoders, variational autoencoders have better generation ability, data compression, dimensionality reduction ability, good probabilistic interpretability and stronger noise interference resistance. By gradually refining the coarse features of each layer, the main features of the adjacency matrix are extracted more effectively, which is better than traditional methods. In addition, time smoothness is introduced to take into account the influence of historical information. In addition, by cascading the features extracted from multiple levels, a complete feature representation is obtained.

[0041] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0042] In order to improve the accuracy of cluster analysis results in social data, an embodiment of the present invention provides a dynamic temporal network clustering method with multi-layer feature cascade. Figure 1 The present invention provides a multi-layer feature cascade dynamic time series network clustering method. Figure 1 As shown in Figure 2, the dynamic temporal network clustering method of multi-layer feature cascade includes:

[0043] S101. Obtain a dynamic temporal network from social data, and represent the dynamic temporal network as an adjacency matrix according to preset rules.

[0044] It should be noted that the social data in the embodiments of the present invention can be interactive information data between people. For example: information data of calls between people within a certain range or information data of interactions through instant messaging software. In a dynamic time series network, users represent nodes, and friendships or interactions between users form edges. Dynamic time series networks can also be divided into time windows, such as one day or one week. Each time window corresponds to a snapshot of the network, and each snapshot is an independent network graph. An adjacency matrix can be constructed from the snapshots of each time window, where the elements of the adjacency matrix indicate whether there is an edge between nodes.

[0045] For example, there is a dynamic time series network consisting of N nodes, then the adjacency matrix at each moment is an N*N matrix, where N vj Indicates whether there is a connection between the vth and jth nodes. The prediction rule settings mainly include: if the edge is unweighted, then 0 and 1 are used to indicate whether there is a connection between the two nodes. If the edge is weighted, then the weight is used to represent the relationship between the two nodes. At this point, the dynamic time series network can be represented as an adjacency matrix.

[0046] Mathematical definition of adjacency matrix:

[0047] A single dynamic sequential network can be represented by a graph G = V, E, where V = {v1, ..., v n} is a set of nodes, v1,…,v n represents the nodes in the network, n represents the number of nodes, E={(v i , v j )} is a set of edges used to represent the interaction between vertices, v i , v j Similarly, it represents the nodes in the network, i and j represent the i-th and j-th nodes respectively. A whole dynamic time series network can be aggregated by multiple independent dynamic time series networks at different moments, that is, Among them G t ={V t , E t} represents the network at time t.

[0048] S102: Input the adjacency matrix into the pre-trained variational autoencoder to obtain multiple feature outputs that match the number of layers of the pre-trained variational autoencoder.

[0049] It should be noted that in the embodiments of the present invention, the number of layers of the pre-trained variational autoencoder can specifically be the number of encoder layers in the pre-trained variational autoencoder. Preferably, it can be set to three layers. In addition, the number of encoder layers can be flexibly adjusted according to actual needs and is not limited in the embodiments of the present invention.

[0050] S103: Perform cluster analysis based on multiple feature outputs and the K-means algorithm and output social statistical clustering results.

[0051] Among them, the pre-trained variational autoencoder is trained through an improved loss function; the improved loss function is a function containing time features and fusion features.

[0052] The embodiment of the present invention provides a multi-layer feature cascade dynamic time series network clustering method, including: obtaining a dynamic time series network from social data, and expressing the dynamic time series network as an adjacency matrix through preset rules; inputting the adjacency matrix into a pre-trained variational autoencoder to obtain multiple feature outputs that match the number of levels of the pre-trained variational autoencoder; performing cluster analysis based on the multiple feature outputs and the K-means algorithm and outputting social statistical clustering results; wherein the pre-trained variational autoencoder is obtained by training with an improved loss function; the improved loss function is a function containing time features and fusion features. In the present invention, by adding time features to the improved loss function, when performing cluster analysis of the dynamic time series network, the impact of historical moments on the current moment can be added to the data analysis, thereby improving the accuracy of the social data cluster analysis; in addition, by adding fusion features and utilizing the information complementarity between multiple layers of features, the performance of the data features is smoother, further improving the accuracy of the social data cluster analysis.

[0053] In order to clearly illustrate a multi-layer feature cascade dynamic temporal network clustering method provided by an embodiment of the present invention, Figure 2 The multi-layer feature extraction model constructed using the variational autoencoder provided in the embodiment of the present invention. Figure 2 As shown, the deep feature decomposition based on the variational autoencoder represents the data flow during model training. The cluster analysis of multi-layer cascade features represents the process of feature concatenation and clustering using kmeans on the output feature data after processing by the pre-trained variational autoencoder during model application. It should be noted that the variational autoencoder is illustrated as having three layers in this embodiment.

[0054] Optionally, cluster analysis is performed based on multiple feature outputs and a K-means algorithm, and social statistical clustering results are output, including:

[0055] Multiple feature outputs are weighted according to their importance to obtain feature cascade results; the feature cascade results are clustered using the K-means algorithm to output social statistical clustering results.

[0056] Optionally, multiple feature outputs are weighted according to their importance to obtain a feature concatenation result, including:

[0057] A plurality of first weight coefficients negatively correlated with the levels of the pre-trained variational autoencoder are set to obtain importance levels; the number of the first weight coefficients is equal to the number of levels of the pre-trained variational autoencoder, and the sum of the plurality of first weight coefficients is 1; a plurality of feature outputs are multiplied by their matching importance levels, and the multiplication results are superimposed to obtain a feature cascade result.

[0058] Optionally, the improved loss function includes three parts: reconstruction error, cluster drift and encoder probability distribution; the reconstruction error, cluster drift and encoder probability distribution are added together to obtain the improved loss function.

[0059] Among them, the reconstruction error is used to calculate the reconstruction error between the training high-order information matrix and the reconstructed training high-order information matrix; the clustering drift is used to calculate the feature difference of the training high-order information matrix at adjacent moments; the encoder probability distribution is used to measure the difference between the pre-trained variational autoencoder and the predefined probability distribution.

[0060] Optionally, the calculation formula of the reconstruction error is expressed as:

[0061]

[0062] Among them, L 1’ represents the reconstruction error; M t Indicates that the high-order information matrix is ​​trained at time t, which is the adjacency matrix W trained at time t t After PMI matrix transformation, we get: Represents the reconstructed training high-order information matrix at time t; mse represents the mean square error; α represents the second weight coefficient.

[0063] Optionally, the calculation formula for cluster drift is expressed as:

[0064]

[0065] Among them, L 2’ represents cluster drift, Indicates M t The output features after the i-th layer of the pre-trained variational autoencoder at time t, Indicates M t The output features after the jth layer of the pre-trained variational autoencoder at time t-1, i and j both represent the number of layers of the pre-trained variational autoencoder, and l is the total number of layers of the pre-trained variational autoencoder.

[0066] In this embodiment, To M t Input the output of the pre-trained variational autoencoder.

[0067] Optionally, the calculation formula of the improved loss function is expressed as:

[0068]

[0069]

[0070] Among them, L all′ represents the improved loss function, represents the output probability distribution of the pre-trained variational autoencoder The difference between it and the predefined probability distribution p(z); is the input data M t The probability distribution of the output latent variable z; φ represents the network parameters of the pre-trained variational autoencoder; D KL represents the KL divergence.

[0071] The embodiment of the present invention also describes part of the derivation process of the improved loss function.

[0072] It should be noted that, in the embodiment of the present invention, the PMI matrix M is used instead of the adjacency matrix W to overcome the deficiency of low-order topological connectivity, where the element m in the M matrix is vj is defined as:

[0073]

[0074] Where v represents the row number in the M matrix, j represents the column number in the M matrix, K represents the number of negative sampling; k represents the kth node; d v d j and d k represent the degrees of the vth, jth, and kth nodes respectively.

[0075] The initial loss function L is obtained by applying the M matrix all ,as follows:

[0076]

[0077] The problem with the above formula is that it only uses one layer of extracted features, which will lose a lot of important information. Therefore, in the embodiment of the present invention, by using the extracted multi-layer features to simultaneously integrate time series smoothing and cascade multi-layer features to perform cluster analysis, the improved loss function L is obtained. all’ :

[0078]

[0079] Among them, L all’ represents the improved loss function, represents the output probability distribution of the pre-trained variational autoencoder The difference between the p(z) and the predefined probability distribution p(z); is the input data M t The probability distribution of the output latent variable z; φ represents the network parameters of the pre-trained variational autoencoder; D KL represents the KL divergence.

[0080] Optionally, the training process of the pre-trained variational autoencoder includes:

[0081] Obtain a training dynamic temporal network from social data, and represent the training dynamic temporal network as a training adjacency matrix using preset rules;

[0082] The training adjacency matrix is ​​input into the variational autoencoder, and the variational autoencoder is trained in the direction of decreasing the improved loss function to obtain a pre-trained variational autoencoder.

[0083] In an embodiment of the present invention, the Adam optimizer is used to update the parameters of the pre-trained variational autoencoder, and the L2 norm is used as the regularization term.

[0084] Optionally, after performing cluster analysis based on multiple feature outputs and a K-means algorithm and outputting social statistical clustering results, a multi-layer feature cascade dynamic time series network clustering method further includes:

[0085] Use social statistical clustering results for image analysis, information retrieval, or data compression.

[0086] It should be noted that in the embodiments of the present invention, after obtaining the social statistical clustering results, they can not only be used for image analysis, information retrieval or data compression tasks, but also for natural grouping, and then for friend recommendations or interest recommendations.

[0087] Furthermore, the multi-layer feature cascade dynamic temporal network clustering method provided by the embodiment of the present invention can be applied not only to social data analysis, but also to fields such as bioinformatics, transportation networks, and financial markets.

[0088] The dynamic time series network clustering method of multi-layer feature cascade provided by the embodiment of the present invention uses variational autoencoders to extract nonlinear features. Compared with the traditional matrix decomposition method, it has a good filtering effect on redundant information that causes deviations in clustering results. Secondly, a key issue in dynamic time series network clustering analysis is how to consider the impact of historical information on the current moment when performing clustering analysis. The present invention extracts multi-layer features and simultaneously integrates them into time series smoothing, which not only utilizes the information complementarity brought by multi-layer features, but also takes into account the impact of historical information on this basis; further, the present invention uses PMI matrix instead of adjacency matrix to overcome the shortcomings of low-order topological connectivity. The use of high-order information can make the extracted deep features more expressive, so as to depict the changes in dynamic time series networks, and further improve the accuracy of social data clustering analysis.

[0089] The method provided in the embodiments of the present invention can be applied to electronic devices. Specifically, the electronic devices can be desktop computers, portable computers, smart mobile terminals, servers, etc. This is not limited here; any electronic device that can implement the present invention falls within the scope of protection of the present invention.

[0090] Based on the same inventive concept, an embodiment of the present invention further provides a multi-layer feature cascade dynamic temporal network clustering system. Figure 3 A schematic diagram of a multi-layer feature cascade dynamic temporal network clustering system provided in an embodiment of the present invention includes: a processor 710, a storage medium 720, and a bus 730. The storage medium 720 stores machine-readable instructions executable by the processor 710. When the multi-layer feature cascade dynamic temporal network clustering system is running, the processor 710 and the storage medium 720 communicate via the bus 730, and the processor 710 executes the machine-readable instructions to perform the steps of the above-mentioned method embodiment. The specific implementation methods and technical effects are similar and will not be repeated here.

[0091] The storage medium may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage. Alternatively, the storage medium may be at least one storage device located away from the processor.

[0092] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.

[0093] It should be noted that the terms "first," "second," and the like are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure.

[0094] In the description of this specification, the reference terms "one embodiment," "some embodiments," "example," "specific example," or "some examples" mean that the specific features or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification.

[0095] Although the present invention is described herein in conjunction with various embodiments, in the process of implementing the claimed invention, those skilled in the art can understand and implement other variations of the disclosed embodiments by viewing the drawings and the disclosed content. In the description of the present invention, the word "comprising" does not exclude other components or steps, "one" or "a" does not exclude multiple situations, and "multiple" means two or more, unless otherwise clearly and specifically defined. In addition, certain measures are recorded in different embodiments, but this does not mean that these measures cannot be combined to produce good results.

[0096] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A multi-layer feature cascade dynamic temporal network clustering method, characterized by: include: Obtaining a dynamic temporal network from social data, and representing the dynamic temporal network as an adjacency matrix according to preset rules; The social data is the interactive information data between people; Inputting the adjacency matrix into a pre-trained variational autoencoder to obtain a plurality of feature outputs matching the number of layers of the pre-trained variational autoencoder; Perform cluster analysis based on the multiple feature outputs and the K-means algorithm and output social statistical clustering results; The pre-trained variational autoencoder is obtained by training an improved loss function; the improved loss function is a function containing time features and fusion features; The improved loss function includes three parts: reconstruction error, cluster drift and encoder probability distribution; Adding the reconstruction error, the cluster drift, and the encoder probability distribution to obtain the improved loss function; Among them, the reconstruction error is used to calculate the reconstruction error between the training high-order information matrix and the reconstructed training high-order information matrix; the cluster drift is used to calculate the feature difference of the training high-order information matrix at adjacent moments; the encoder probability distribution is used to measure the difference between the pre-trained variational autoencoder and the predefined probability distribution; The calculation formula of the improved loss function is expressed as: Among them, L all' represents the improved loss function, represents the output probability distribution of the pre-trained variational autoencoder The difference between it and the predefined probability distribution p(z); is the input data M t The output probability distribution of the output latent variable z; φ represents the network parameters of the pre-trained variational autoencoder; D KL represents KL divergence; Indicates M t The output features after passing through the i-th layer of the pre-trained variational autoencoder at time t, Indicates M t The output feature after passing through the jth layer of the pre-trained variational autoencoder at time t-1, i and j both represent the number of layers of the pre-trained variational autoencoder, l is the total number of layers of the pre-trained variational autoencoder, and α represents the second weight coefficient; M t Indicates that the high-order information matrix is ​​trained at time t, which is the adjacency matrix W trained at time t t After PMI matrix transformation, we get: Represents the reconstruction of the training high-order information matrix at time t, T represents the transpose processing, and mse represents the mean square error.

2. A multi-layer feature cascade dynamic time series network clustering method according to claim 1, characterized in that: Perform cluster analysis based on the multiple feature outputs and the K-means algorithm and output social statistical clustering results, including: Perform feature weighting on the plurality of feature outputs according to their importance to obtain a feature cascade result; The feature concatenation result is clustered using the K-means algorithm to output the social statistics clustering result.

3. The multi-layer feature cascade dynamic time series network clustering method according to claim 2, characterized in that: The step of weighting the plurality of feature outputs according to their importance to obtain a feature cascade result includes: Setting a plurality of first weight coefficients that are negatively correlated with the levels of the pre-trained variational autoencoder to obtain the importance level; wherein the number of the first weight coefficients is equal to the number of levels of the pre-trained variational autoencoder, and the sum of the plurality of first weight coefficients is 1; The plurality of feature outputs and the importance levels of their matches are multiplied, and the multiplication results are superimposed to obtain the feature cascade result.

4. The multi-layer feature cascade dynamic time series network clustering method according to claim 1, characterized in that: The calculation formula of the reconstruction error is expressed as: Among them, L 1' represents the reconstruction error; M t Indicates that the high-order information matrix is ​​trained at time t, which is the adjacency matrix W trained at time t t After PMI matrix transformation, we get: Represents the reconstructed training high-order information matrix at time t; mse represents the mean square error; α represents the second weight coefficient.

5. The multi-layer feature cascade dynamic temporal network clustering method according to claim 4, characterized in that: The calculation formula of the cluster drift is expressed as: Among them, L 2' represents the cluster drift, Indicates M t The output features after passing through the i-th layer of the pre-trained variational autoencoder at time t, Indicates M t The output features after passing through the jth layer of the pre-trained variational autoencoder at time t-1, where i and j both represent the number of layers of the pre-trained variational autoencoder, and l is the total number of layers of the pre-trained variational autoencoder.

6. The multi-layer feature cascade dynamic time series network clustering method according to claim 1, characterized in that: The training process of the pre-trained variational autoencoder includes: Obtaining a training dynamic temporal network from social data, and representing the training dynamic temporal network as a training adjacency matrix according to preset rules; The training adjacency matrix is ​​input into a variational autoencoder, and the variational autoencoder is trained in a direction in which the improved loss function decreases, to obtain the pre-trained variational autoencoder.

7. The multi-layer feature cascade dynamic time series network clustering method according to claim 1, characterized in that: After performing cluster analysis based on the plurality of feature outputs and the K-means algorithm and outputting the social statistical clustering results, the multi-layer feature cascade dynamic time series network clustering method further includes: The social statistical clustering results are used for image analysis, information retrieval or data compression.

8. A multi-layer feature cascade dynamic temporal network clustering system, characterized by: include: A processor, a storage medium, and a bus, wherein the storage medium stores machine-readable instructions executable by the processor. When the multi-layer feature cascade dynamic temporal network clustering system is running, the processor and the storage medium communicate via the bus, and the processor executes the machine-readable instructions to perform the steps of the method according to any one of claims 1 to 7.

Images