User clustering method fusing user portrait features
By combining LSTM encoder and EM algorithm, the adaptability and interpretability issues of traditional static clustering algorithms in dynamic user data are solved, realizing dynamic adaptive clustering of user profile features and improving the stability and interpretability of clustering.
Patent Information
- Application Number
- CN202511074482.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-01
- Publication Date
- 2025-11-07
AI Technical Summary
Traditional static clustering algorithms are poorly adapted to dynamic user data and have weak interpretability, making it difficult to effectively perform dynamic adaptive clustering of multimodal user data.
We employ an LSTM encoder for semantic association learning, combine multi-head self-attention and Attention-Mask mechanisms for latent variable reconstruction, utilize a pre-built EM algorithm for clustering, and fuse user profile features to achieve dynamic adaptive clustering.
It improves the ability to model high-dimensional sparse data, realizes automatic feature extraction of nonlinear relationships and stability and interpretability of user clustering, and supports clustering analysis of dynamic user needs.
Smart Images
Figure CN120910588A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of user behavior analysis, in particular to a user clustering method fusing user portrait features. BACKGROUND
[0002] Under the development trend of big data and data mining, how to effectively and dynamically adaptively cluster data under multi-modal user data has become a challenging technical problem.
[0003] Traditional static clustering methods such as K-Means and DBSCAN algorithms perform very well on static data sets, but have poor adaptability in dynamic user data. SUMMARY
[0004] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a user clustering method fusing user portrait features, which solves the problems of poor adaptability and weak interpretability of traditional static clustering algorithms.
[0005] To achieve the above purpose, the present application provides the following scheme:
[0006] A user clustering method fusing user portrait features comprises:
[0007] Collecting user data and performing normalization processing on the user data to obtain normalized data;
[0008] Using an LSTM encoder to perform semantic association learning on the normalized data to obtain a latent variable mean and a latent variable standard deviation;
[0009] Using a reparameterization method to perform back propagation on the latent variable of the LSTM encoder according to the latent variable mean and the latent variable standard deviation;
[0010] When the LSTM encoder performs time series modeling, using a multi-head self-attention mechanism to assign weights to the latent variable of the LSTM encoder;
[0011] Using a decoder to sequentially perform latent variable reconstruction and latent variable mapping on the output of the LSTM encoder according to an Attention-Mask mechanism to obtain a user latent demand representation;
[0012] Using a pre-constructed EM algorithm to perform clustering processing on the user latent demand representation to obtain a user clustering result.
[0013] Preferably, the normalized data includes user information embedding, sentiment analysis embedding, social behavior embedding, and consumption behavior embedding.
[0014] Preferably, semantic correlation learning is performed on the normalized data by using an LSTM encoder to obtain a latent variable mean and a latent variable standard deviation, including:
[0015] The LSTM encoder is constructed, and the LSTM encoder includes a plurality of LSTM encoding units, and a first full connection layer and a second full connection layer connected with the LSTM encoding units.
[0016] A mean feature extraction formula is constructed, and an expression of the mean feature extraction formula is: ; wherein, is the latent variable mean; represents the first full connection layer; represents the LSTM encoding unit; represents an input latent variable;
[0017] A standard deviation feature extraction formula is constructed, and an expression of the standard deviation feature extraction formula is: F ; wherein, is the latent variable standard deviation; F represents the second full connection layer;
[0018] The normalized data is calculated by using the mean feature extraction formula and the standard deviation feature extraction formula to obtain the latent variable mean and the latent variable standard deviation.
[0019] Preferably, the latent variable of the LSTM encoder is back propagated by using a reparameterization method according to the latent variable mean and the latent variable standard deviation, including:
[0020] A reparameterization formula is constructed, and an expression of the reparameterization formula is: ; wherein, is the potential demand at t period; is a sampling noise;
[0021] The latent variable of the LSTM encoder is back propagated by using the reparameterization formula.
[0022] Preferably, when the LSTM encoder performs time series modeling, the latent variable of the LSTM encoder is weighted by using a multi-head self-attention mechanism, including:
[0023] The multi-head self-attention mechanism is constructed, and an expression of the multi-head self-attention mechanism is: ; wherein, is a latent variable attention result at t period, is an attention network layer, is a latent variable at i period;
[0024] The multi-head self-attention mechanism is used to weight the latent variables of the LSTM encoder.
[0025] Preferably, the output of the LSTM encoder is sequentially subjected to latent variable reconstruction and latent variable mapping by the decoder according to an Attention-Mask mechanism to obtain a user latent demand representation, including:
[0026] The decoder is constructed and includes two layers of the LSTM encoding unit and a third fully connected layer of the two layers of the LSTM encoding unit.
[0027] The output of the LSTM encoder is subjected to variable reconstruction by the two layers of the LSTM encoding unit of the decoder according to an Attention-Mask mechanism to obtain reconstructed data.
[0028] The third fully connected layer of the decoder is used to map the latent variables of the reconstructed data to obtain the user latent demand representation.
[0029] Preferably, the user latent demand representation is subjected to clustering processing by a pre-constructed EM algorithm to obtain a user clustering result, including:
[0030] A responsibility degree formula is constructed, and the expression of the responsibility degree formula is: wherein, is the output of the responsibility degree formula. represents the probability density function of the kth Gaussian distribution. is the mixing weight of the kth Gaussian distribution. is the number of classification clusters. represents the covariance corresponding to the kth Gaussian distribution.
[0031] The user latent demand representation is calculated by the responsibility degree formula to obtain responsibility degree data.
[0032] Preferably, the user latent demand representation is subjected to clustering processing by a pre-constructed EM algorithm to obtain a user clustering result, including:
[0033] A mixing weight update formula is constructed, and the expression of the mixing weight update formula is: wherein, is the number of categories of the classification cluster.
[0034] A mean update formula is constructed, and the expression of the mean update formula is:
[0035] A covariance matrix update formula is constructed, and the expression of the covariance matrix update formula is: ;
[0036] The responsibility degree data is iteratively clustered by using the mixed weight update formula, the mean update formula and the covariance matrix update formula to obtain the user clustering result.
[0037] Preferably, the user latent demand representation is clustered by using a pre-constructed EM algorithm to obtain a user clustering result, comprising:
[0038] The silhouette coefficient and DBI index of the user clustering result are calculated to obtain a silhouette value and a DBI value.
[0039] When the silhouette value and the DBI value meet a preset clustering threshold, the user clustering result is output.
[0040] The present application discloses the following technical effects:
[0041] The present application provides a user clustering method fusing user portrait features, which integrates the latent variable modeling capability of a variational autoencoder and the parameter optimization mechanism of an expectation maximization algorithm, solves the problems of poor adaptability and weak interpretability of traditional static clustering algorithms, and realizes optimization of high-dimensional sparse data, modeling of nonlinear relationships and automatic feature extraction. BRIEF DESCRIPTION OF DRAWINGS
[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.
[0043] Figure 1 A user clustering process schematic diagram fusing user portrait features is provided for the embodiments of the present application.
[0044] Figure 2 A model architecture diagram is provided for the embodiments of the present application.
[0045] Figure 3 An EM clustering schematic diagram is provided for the embodiments of the present application. DETAILED DESCRIPTION
[0046] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.
[0047] The application aims to provide a user clustering method fusing user portrait features, to solve the problems of poor adaptability and weak interpretability of traditional static clustering algorithms.
[0048] In order to make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the application will be further described in detail below in combination with the drawings and specific embodiments.
[0049] Figure 1 The user clustering process diagram fusing user portrait features provided by the embodiments of the application is shown in Figure 1 The application provides a user clustering method fusing user portrait features, which comprises the following steps:
[0050] Step 100: Collecting user data and performing normalization processing on the user data to obtain normalized data;
[0051] Step 200: Performing semantic association learning on the normalized data by using an LSTM encoder to obtain mean of latent variables and standard deviation of latent variables;
[0052] Step 300: Performing back propagation on the latent variables of the LSTM encoder by using a reparameterization method according to the mean of latent variables and the standard deviation of latent variables;
[0053] Step 400: Assigning weights to the latent variables of the LSTM encoder by using a multi-head self-attention mechanism when the LSTM encoder performs time series modeling;
[0054] Step 500: Performing latent variable reconstruction and latent variable mapping on the output of the LSTM encoder by using a decoder according to an Attention-Mask mechanism, to obtain user latent demand representation;
[0055] Step 600: Performing clustering processing on the user latent demand representation by using a pre-constructed EM algorithm to obtain user clustering results.
[0056] Preferably, the normalized data comprises user information embedding, sentiment analysis embedding, social behavior embedding and consumption behavior embedding.
[0057] Further, the semantic association learning on the normalized data by using the LSTM encoder to obtain the mean of latent variables and the standard deviation of latent variables comprises the following steps:
[0058] Constructing the LSTM encoder; the LSTM encoder comprises a plurality of LSTM encoding units, and a first full connection layer and a second full connection layer connected with the LSTM encoding units;
[0059] Constructing a mean feature extraction formula; the expression of the mean feature extraction formula is: ; wherein, is the latent variable mean; denotes the first fully connected layer; denotes the LSTM encoding unit; denotes the input latent variable;
[0060] a standard deviation feature extraction formula is constructed; the expression of the standard deviation feature extraction formula is: F ; wherein, is the latent variable standard deviation; F denotes the second fully connected layer;
[0061] The mean feature extraction formula and the standard deviation feature extraction formula are used to calculate the normalized data, and the latent variable mean and the latent variable standard deviation are obtained.
[0062] Specifically, according to the latent variable mean and the latent variable standard deviation, the latent variable of the LSTM encoder is back propagated by using the reparameterization method, including:
[0063] A reparameterization formula is constructed; the expression of the reparameterization formula is: ; wherein, is the potential demand at t period; is the sampling noise;
[0064] The reparameterization formula is used to back propagate the latent variable of the LSTM encoder.
[0065] Further, when the LSTM encoder performs time series modeling, the latent variable of the LSTM encoder is weighted by using a multi-head self-attention mechanism, including:
[0066] The multi-head self-attention mechanism is constructed; the expression of the multi-head self-attention mechanism is: ; wherein, is the latent variable attention result at t period, is the attention network layer, is the latent variable at i period;
[0067] The multi-head self-attention mechanism is used to weight the latent variable of the LSTM encoder.
[0068] Specifically, according to the Attention-Mask mechanism, the output of the LSTM encoder is sequentially reconstructed and mapped by using the decoder, and the user potential demand representation is obtained, including:
[0069] The decoder is constructed; the decoder comprises two layers of the LSTM encoding unit and a third fully connected layer of the two layers of the LSTM encoding unit;
[0070] According to the Attention-Mask mechanism, the output of the LSTM encoder is reconstructed by using the two layers of the LSTM encoding unit of the decoder, to obtain reconstruction data;
[0071] The latent variable mapping is performed on the reconstruction data by using the third fully connected layer of the decoder, to obtain the user latent demand representation.
[0072] Further, the user latent demand representation is clustered by using a pre-constructed EM algorithm to obtain a user clustering result, comprising:
[0073] A responsibility degree formula is constructed; the expression of the responsibility degree formula is: ; wherein, is the output of the responsibility degree formula; represents the probability density function of the kth Gaussian distribution; is the mixing weight of the kth Gaussian distribution; is the number of classification clusters; represents the covariance corresponding to the kth Gaussian distribution;
[0074] The user latent demand representation is calculated by using the responsibility degree formula to obtain responsibility degree data.
[0075] Specifically, the user latent demand representation is clustered by using a pre-constructed EM algorithm to obtain a user clustering result, comprising:
[0076] A mixing weight update formula is constructed; the expression of the mixing weight update formula is: ; wherein, is the number of categories of classification clusters;
[0077] A mean update formula is constructed; the expression of the mean update formula is: ;
[0078] A covariance matrix update formula is constructed; the covariance matrix update formula is: ;
[0079] The responsibility degree data is iteratively clustered by using the mixing weight update formula, the mean update formula, and the covariance matrix update formula to obtain the user clustering result.
[0080] Further, the user latent demand representation is clustered by using a pre-constructed EM algorithm to obtain a user clustering result, comprising:
[0081] The silhouette coefficient and DBI index of the user clustering result are calculated to obtain a silhouette value and a DBI value;
[0082] When the silhouette value and the DBI value meet a preset clustering threshold, the user clustering result is output.
[0083] Specifically, the EM-VAE clustering algorithm is designed, and under the large model COT inference, the multi-dimensional user portrait of each description can be explained and manually intervened. In order to realize the clustering of the user's potential demand, the potential demand is regarded as a latent variable, and then the dynamic self-adaptive clustering algorithm of EM-VAE is designed. The design and implementation of VAE in the image field: the background of VAE in the image field, the variational autoencoder (VAE) is based on the AutoEncoder architecture, and is mathematically transformed, mainly reflected in the modeling of latent variables and the assumption of Gaussian distribution. In the image generation field, the self-encoding and high-quality generation effect of the image are realized. The core design is to realize the self-learning of the image in a latent space, and to generate an image close to the original data. The VAE model is composed of the following two parts:
[0084] 1) Encoder: Encoder is the encoder of VAE. The role of the encoder is to reduce the dimension of the image number and represent the image embedding. Through the modeling of latent variables, the mean and variance of the data are generated. The structure of the encoder is generally a multi-layer convolution and pooling structure, which realizes dimension reduction through a multi-layer neural network.
[0085] 2) Decoder: Decoder is the decoder of VAE. Decoder is based on Encoder to realize the recovery and reconstruction of image. Its structure is generally opposite to Encoder. Through the decoding of Decoder in the latent space, the reconstruction of the original image is recovered, and this process is self-supervised learning.
[0086] The training target of VAE is to maximize the evidence lower bound (ELBO), which includes:
[0087] 1) Reconstruction error: measure the difference between the original image and the reconstructed image.
[0088] 2) KL divergence: measure the difference between two distributions. In VAE, the KL divergence is the separation degree between the distribution of latent variables and the prior Gaussian distribution. The greater the value is, the greater the difference between them is, and vice versa.
[0089] Preferably, the VAE is migrated to the field of user demand clustering. The VAE models the latent variable, which can also be applied to the field of AI marketing, and can be used to mine potential demand and user behavior patterns from multi-dimensional user data, and to realize potential demand modeling and representation. In the scenario of user demand clustering, the VAE can be used to:
[0090] 1) Learn the potential demand of the user: from the user's consumption behavior and social behavior, learn the user's behavior pattern while realizing the embedding dimension reduction, which is beneficial to reasoning acceleration.
[0091] 2) Learn the deep structure relationship: through the self-supervised learning of VAE, the potential correlation between potential demand, user behavior and user portrait, and the correlation between such relationship and user recommendation is mined.
[0092] 3) User clustering analysis: the purpose of modeling the potential demand of the user is to realize the clustering of the user's customer group, which is convenient for coordinating the filtering recommendation.
[0093] Optionally, before modeling, it is assumed that the latent variable is subject to Gaussian distribution, which can be regarded as a simplification of the real complex world. On the basis of demand modeling, through the modeling of the latent space of VAE, a mixed Gaussian distribution (GMM) is introduced, so that the user realizes multi-class clustering in multiple dimensions, and through the probability type soft clustering, the time sequence variability of the demand can be met, and adaptive dynamic clustering is realized.
[0094] Reference Figure 2 , EM-VAE algorithm modeling, model design, EM-VAE algorithm design, which includes the following parts.
[0095] Further, the input data and the embedding layer, since the user data is diverse, mainly contains the following four parts of information:
[0096] 1) User information embedding (UserEmbedding): representing the basic information of the user.
[0097] 2) Emotion analysis embedding (EmotionEmbedding): embedding based on emotion analysis, reflecting the emotional response of the user.
[0098] 3) Social behavior embedding (SocialBehaviorEmbedding): embedding generated through social behavior data (such as likes, comments, etc.).
[0099] 4) Consumption behavior embedding (ConsumptionEmbedding): user consumption behavior history, such as purchase records.
[0100] In terms of data processing, due to the large differences between multi-dimensional data, it is necessary to do normalization processing, including the length of the text and the dimension of the embedding.
[0101] Specifically, LSTM encoder: time modeling and latent space inference. There is a time dependence between the user's behavior, which can be learned through the long short-term memory network to learn the semantic relationship of the text and the time dependence of the user's behavior. The encoder is part of the encoder, and due to the existence of time dependence, the structure of the encoder is designed as a multi-time step t LSTM on the embedding layer, which is used to learn the semantic association, and the uppermost layer of the encoder is twice linear fully connected layer, output the mean of the latent variable and standard deviation , the formula is as follows:
[0102]
[0103]
[0104] where, is the mean prediction value of the neural network layer to the latent variable, is the variance prediction value of the neural network layer to the latent variable, preferably, the reparameterization technique, VAE uses the reparameterization technique, so that the latent variable can be optimized by gradient through back propagation, that is:
[0105]
[0106] where, ϵ is the sampling noise, ϵ~N(0, I), where N(0, I) is from the standard normal distribution, and I is the unit matrix.
[0107] Optionally, time-based attention mechanism, in time modeling, due to the different weights of user behavior, there are also different weights in time sequence, so the latent variable should adopt multi-head self-attention mechanism, give different weights to different time steps t and different behaviors, or say different attention. The mathematical representation of self-attention mechanism is as follows:
[0108]
[0109] Specifically, decoder: capture demand latent variable, the task of the decoder is to generate the user's demand representation from the latent space. The output of the decoder represents the prediction of the user's behavior, in addition, through the reconstruction of the latent variable, the demand latent variable can be generated, and these latent variables can better reflect the user's demand pattern. The decoder uses fully connected layer (FC) to combine the latent variables The reconstruction of the user behavior corresponds to the transformation of the user behavior as a representation of the user's needs:
[0110] Through the design of the Encoder structure described above, the mean and variance of the user's potential needs are generated, and then on this basis, the Decoder is designed as a neural network structure of two-layer LSTM and FC full connection layer, which realizes the reconstruction of the user behavior. Similarly, the Attention-Mask mechanism will be introduced to realize the mask of the subsequent user behavior, and better ensure that the time step t cannot observe the user information of the time step t+1. Through the reconstruction of the user behavior by the Decoder, the potential demand representation behind the user behavior is learned.
[0111] Reference Figure 3 , the optimization process of EM combined with VAE, E step: inferring the responsibility of latent variables. The EM algorithm is a multi-step solution to the probability of the user's potential demand classification, and then maximizes its probability, constantly updating the probability. Specifically, in the E step, first infer the responsibility of each user belonging to different clusters of customer groups. In mathematical modeling, for each latent variable , it can be regarded as subject to a high-dimensional mixed Gaussian distribution, and then the probability of the latent variable belonging to different Gaussian distributions is calculated by modeling, which is set as the responsibility , the mathematical formula is as follows:
[0112]
[0113] Among them, represents the probability density function of the kth Gaussian distribution of the latent variable, is the mixing weight of the kth Gaussian distribution.
[0114] M step: update model parameters, through the E step, the responsibility of the latent variable belonging to each category is calculated , in the M step, the probability needs to be maximized, and the mean , covariance and mixing weight are updated to improve the clustering effect:
[0115] 1) Update the mixing weight:
[0116]
[0117] 2) Update the mean:
[0118]
[0119] 3) Update the covariance matrix:
[0120]
[0121] Specifically, the evaluation and summary of clustering. Evaluation index of clustering. After EM-VAE clustering, the clustering results need to be analyzed, and the evaluation index has two kinds, namely the silhouette coefficient and the DBI index:
[0122] 1) Silhouette coefficient: This index represents the tightness and similarity of the latent variable in its clustered category with other sample points, and the difference with other categories. The larger the value, the better the clustering effect, and vice versa.
[0123] 2) DBI index: This index reflects the difference between the latent variable in its clustered category and other sample points. Just the opposite of the silhouette coefficient, the larger the DBI index value, the worse the clustering effect.
[0124] Through clustering evaluation, the threshold of clustering can be set, when the threshold is reached, the clustering of EM reaches the preset clustering target, otherwise, the E step and M step of EM will be repeated until the clustering target is reached. If the clustering effect is poor or cannot converge, the self-supervised training and learning of the VAE model will be repeated.
[0125] Preferably, the generation and use of the final clustering result, in the EM-VAE algorithm, if the convergence of the model is achieved, the high-quality latent demand embedding is obtained, and the dynamic adaptive clustering of the user's latent demand is also achieved. These information can realize the following analysis:
[0126] 1) User clustering: Through the clustering of the user's latent demand, the user-based coordinated filtering recommendation can be realized, and the accurate marketing of the recommendation is ensured.
[0127] 2) User behavior prediction: Through the behavior modeling of VAE, the user's behavior prediction is realized in the Decoder, and the reconstruction of the latent variable of the model is ensured.
[0128] In the design of the EM-VAE algorithm, the VAE algorithm reconstructs the user's behavior, realizes the modeling of the latent variable, combines the EM algorithm and the mixture Gaussian distribution, and uses to realize the adaptive dynamic clustering of the user. On the other hand, due to the diversity, real-time and dynamic nature of the demand, the dimensionality reduction of the user's features is also realized, and a dynamic soft clustering method is realized, which can capture the dynamic change characteristics of the latent variable.
[0129] The beneficial effects of the present application are as follows:
[0130] The present application integrates the latent variable modeling capability of the variational autoencoder and the parameter optimization mechanism of the expectation maximization algorithm, improves the processing effect of high-dimensional sparse data, modeling of nonlinear relationship and automatic feature extraction, and enhances the stability and interpretability of clustering.
[0131] The various embodiments described in this specification are intended to be exemplary only. The various embodiments were chosen and described in order to best explain the principles of the application and its practical application, to thereby enable others skilled in the art to best utilize the application, and to best enable the present application to be performed with determination by those skilled in the art.
[0132] The principles and implementations of the present application have been described in the specification with reference to specific examples. The above description is only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation and application range will be changed. In view of the above, the content of the specification should not be understood as a limitation of the present application.
Claims
1. A user clustering method fusing user profile features, characterized in that, The application relates to a user demand prediction method and device. The application comprises the following steps: Collecting user data and normalizing the user data to obtain normalized data; Using an LSTM encoder to learn semantic association of the normalized data to obtain mean and standard deviation of latent variables; Using a reparameterization method to back-propagate the latent variables of the LSTM encoder according to the mean and standard deviation of the latent variables; Using a multi-head self-attention mechanism to assign weights to the latent variables of the LSTM encoder when the LSTM encoder performs time series modeling; Using a decoder to sequentially perform latent variable reconstruction and latent variable mapping on the output of the LSTM encoder according to an Attention-Mask mechanism to obtain a user latent demand representation; 2.The user clustering method of claim 1, wherein, Using a pre-constructed EM algorithm to cluster the user latent demand representation to obtain a user clustering result. 3.The user clustering method of claim 2, wherein, The normalized data comprises user information embedding, sentiment analysis embedding, social behavior embedding and consumption behavior embedding. The LSTM encoder is constructed, and the LSTM encoder comprises a plurality of LSTM encoding units and a first full connection layer and a second full connection layer connected with the LSTM encoding units; constructing a mean feature extraction formula; an expression of the mean feature extraction formula is: ; wherein, is the latent variable mean; denotes the first fully connected layer; denotes the LSTM encoding unit; denotes the input latent variable; A standard deviation feature extraction formula is constructed; an expression of the standard deviation feature extraction formula is: F ; wherein, is the latent variable standard deviation; F represents the second full connection layer; The mean and standard deviation of the latent variables are calculated by using the mean feature extraction formula and the standard deviation feature extraction formula on the normalized data. 4.The user clustering method of claim 3, wherein, The reparameterization formula is used to back-propagate the latent variables of the LSTM encoder. Constructing a reparameterization formula; the expression of the reparameterization formula is: ; wherein, is the potential demand at time t; is the sampling noise; The multi-head self-attention mechanism is used to assign weights to the latent variables of the LSTM encoder when the LSTM encoder performs time series modeling. 5.The user clustering method of claim 4, wherein, The decoder is constructed, and the decoder comprises two layers of the LSTM encoding units and a third full connection layer connected with the two layers of the LSTM encoding units. constructing the multi-head self-attention mechanism; an expression of the multi-head self-attention mechanism is: ; wherein, is a latent variable attention result at a t period, is an attention network layer, is a latent variable at an i period; The two layers of the LSTM encoding units of the decoder are used to reconstruct variables according to the Attention-Mask mechanism to obtain reconstruction data. 6.The user clustering method of claim 5, wherein, The third full connection layer of the decoder is used to map latent variables of the reconstruction data to obtain the user latent demand representation. The responsibility degree formula is used to calculate the user latent demand representation to obtain responsibility degree data. The pre-constructed EM algorithm is used to cluster the user latent demand representation to obtain the user clustering result. 7.The user clustering method of claim 6, wherein, constructing a responsibility formula; an expression of the responsibility formula is: ; wherein, is an output of the responsibility formula; denotes a probability density function of the kth Gaussian distribution; is a mixing weight of the kth Gaussian distribution; is a number of classification clusters; denotes a covariance corresponding to the kth Gaussian distribution; 8.The user clustering method of claim 7, wherein, A mixed weight update formula is constructed; an expression of the mixed weight update formula is: ; wherein, is a class number of the classification cluster. constructing a mean value updating formula; an expression of the mean value updating formula is: ; constructing a covariance matrix updating formula; the covariance matrix updating formula is: ; The responsibility degree data is iteratively clustered by using the mixed weight update formula, the mean update formula, and the covariance matrix update formula to obtain the user clustering result. 9.The user clustering method of claim 8, wherein, The user potential demand representation is clustered by using a pre-constructed EM algorithm to obtain a user clustering result, including: The silhouette coefficient and the DBI index of the user clustering result are calculated to obtain a silhouette value and a DBI value. When the silhouette value and the DBI value meet a preset clustering threshold, the user clustering result is output.