User feature point classification method, device and electronic device based on pre-training

Abstract

In the embodiment of the present invention, a method, device and electronic device for classifying user feature points based on pre-training are provided, which belongs to the field of data processing technology. The method includes: after obtaining raw data from multiple social media, setting a pre-processing vector corresponding to the raw data; establishing data update rules for the first deep learning model, inputting the standard data into the first deep learning model, and forming query data by constructing a custom query; obtaining a second training parameter set by inputting the query data into the second deep learning model for pre-training; using the first deep learning model M1 to process the data to be classified X, obtaining a first set of feature point sets F1, and integrating all the second feature point sets into a second feature point set F2* that finally represents the user feature points through a fusion mechanism. The algorithm in the present invention uses the method of pre-training and fine-tuning to improve the accuracy and efficiency of user interest classification.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method, device and electronic device for classifying user feature points based on pre-training. Background Art

[0002] User interest point prediction is an important research direction in the fields of machine learning, data mining and recommendation systems. It aims to predict which content, products or services users may be interested in in the future by analyzing multiple data sources such as user historical behavior, preferences, personal information, contextual environment, etc. This prediction technology is widely used in many scenarios such as e-commerce recommendations, social media content recommendations, targeted advertising, search engine personalized result sorting, etc., and user feature point classification is an important means of predicting user interest points.

[0003] The prior art has the following problems in the process of classifying user feature points:

[0004] 1. Low accuracy: Existing algorithms rely on a large amount of user behavior data and do not handle new users or sparse data well, resulting in low accuracy of model classification.

[0005] 2. Insufficient generalization ability: Existing algorithms perform well on specific data sets, but perform poorly on new or unknown data sets and lack generalization ability.

[0006] 3. High consumption of computing resources: For large-scale user data, traditional classification algorithms require long-term training and consume a lot of computing resources.

[0007] 4. Adaptability and timeliness: User interests may change rapidly over time, and traditional algorithms are difficult to update in real time.

[0008] The above problems become technical problems that need to be solved. Summary of the invention

[0009] In view of this, the embodiments of the present invention provide a method, device and electronic device for classifying user feature points based on pre-training, which at least partially solve the problems existing in the prior art.

[0010] In a first aspect, an embodiment of the present invention provides a method for classifying user feature points based on pre-training, comprising:

[0011] After obtaining the original data from multiple social media, a preprocessing vector Y=[SS, IS, TS] corresponding to the original data is set, where SS represents the offset evaluation value of the data, IS represents the influence evaluation value of the data, and TS represents the timeliness evaluation value of the data. The comprehensive score CS of each piece of data in the original data is calculated by the preprocessing function G. By comparing the comprehensive score CS with the preset value CS0, the data whose comprehensive score CS is greater than the preset value CS0 is retained to obtain standard data;

[0012] Assume w t are all the parameters of the first deep learning model M1 at time t, η is the learning rate, γ is the correction term, ▽L(w t-1 ) is the loss function L at time t-1 with respect to parameter w t-1 The gradient of v establishes the data update rule for the first deep learning model: t =γv t-1 +η▽L(w t-1 ), w t =w t-1 -v t , v t and v t-1 denote the cumulative correction terms at time t and time t-1, γv t-1 Indicates the accumulated correction value at the last moment, η▽L(w t-1 ) is the instant update value of the current gradient multiplied by the learning rate;

[0013] Input the standard data into the first deep learning model, calculate the loss function L and the gradient ▽L(w) of the loss function with respect to the parameters of the first deep learning model according to the predicted output and the true label, and when the loss change of the first deep learning model is less than a preset value, use all parameters w' obtained by training the first deep learning model as the first training parameter set;

[0014] By building a custom query engine, the data including the user's geographic location, user's comment content, user's interactive messages and user's access records are queried from the standard data to form query data;

[0015] The parameters in the first training parameter set are used as the initial parameters m of the second deep learning model M2 at time y y , τ is the global learning rate of the second deep learning model, β1 and β2 are the first decay rate and the second decay rate of the second deep learning model respectively, and the data update rule is constructed for the second deep learning model: m y =β1m y-1 +(1-β1)▽L(w y-1 ), v y =β2v y-1 +(1-β2)(▽L(wy-1 )) 2 , m y is the gradient ▽L(w y-1 ), the exponentially weighted average of y is an exponentially weighted average of the square of the gradient, and a second training parameter set is obtained by inputting the query data into the second deep learning model for pre-training;

[0016] Use the first deep learning model M1 to process the data to be classified X to obtain the first set of feature points F1. Apply p-1 different data transformations to the data to be classified X to obtain p-1 view data of data F1. Process the p view data through M1 respectively to obtain p-1 feature point sets F2, F3, ...F p , all feature sets F1, F2, ...F p Perform fusion to obtain the first fusion feature F*;

[0017] The second deep learning model M2 is randomly mutated q times to obtain an integrated model EMB={M21, M22, ..., M2q} containing q mutated second deep learning models, the first fusion feature F* is input into each model in the integrated model EMB to obtain multiple second feature point sets F21, F22, ..., F2q, and all second feature point sets F2i are integrated into the final second feature point set F2* that represents the user feature points through a fusion mechanism.

[0018] According to a specific implementation of the embodiment of the present disclosure, after obtaining the original data from multiple social media, setting a preprocessing vector Y corresponding to the original data includes:

[0019] By constructing the first calculation formula, calculate the value of SS:

[0020]

[0021] N is the total number of words recognized in the text Ti, It is vocabulary In text The frequency in It is vocabulary The sentiment polarity score of

[0022] By constructing the second calculation formula, the value of IS is calculated:

[0023]

[0024] Fl represents the number of followers, Lk represents the number of likes, Sr represents the number of shares or reposts, Cm represents the number of comments, Tc represents the time difference between the content release time and the present time, Ms is the time decay factor, and Qc represents the content quality evaluation value;

[0025] By constructing the third calculation formula, the value of TS is calculated:

[0026]

[0027] is the current time, is the data release time, is the attenuation coefficient.

[0028] According to a specific implementation of the embodiment of the present disclosure, the calculating of the comprehensive score CS of each piece of data in the original data by the preprocessing function G includes:

[0029] By weighting the calculation results of the first calculation formula, the second calculation formula and the third calculation formula, a comprehensive score CS of each data in the original data is obtained.

[0030] According to a specific implementation of the embodiment of the present disclosure, establishing a data update rule for the first deep learning model includes:

[0031] Using a federated learning algorithm, only the parameters of the first deep learning model are updated without updating the entire model structure;

[0032] Perform a decay operation on the parameters generated by the old data in the standard data.

[0033] According to a specific implementation of the embodiment of the present disclosure, after obtaining the original data from the multiple social media, the method further includes:

[0034] Clean, standardize and transform the collected data, including text denoising, grammar standardization, and removal of irrelevant content;

[0035] Extract keywords from text and visual features from avatars to prepare matching input formats for model training.

[0036] According to a specific implementation of the embodiment of the present disclosure, after constructing the data update rule for the second deep learning model, the method further includes:

[0037] right and The value of is corrected:

[0038]

[0039]

[0040] According to the revised and To update the parameters:

[0041] .

[0042] According to a specific implementation of the embodiment of the present disclosure, the method further includes:

[0043] After the first deep learning model and the second deep learning model are deployed, the models are iteratively updated based on the feedback, including adjusting algorithm parameters, retraining, or further fine-tuning.

[0044] According to a specific implementation of the embodiment of the present disclosure, the method further includes:

[0045] The user's interests are predicted using a pre-trained and fine-tuned first deep learning model and a second deep learning model, so as to predict and output the user's interest points based on the input user data.

[0046] In a second aspect, an embodiment of the present invention provides a user feature point classification device based on pre-training, comprising:

[0047] A setting module, after obtaining raw data from multiple social media, sets a preprocessing vector Y=[SS, IS, TS] corresponding to the raw data, where SS represents the offset evaluation value of the data, IS represents the influence evaluation value of the data, and TS represents the timeliness evaluation value of the data. The comprehensive score CS of each piece of data in the raw data is calculated through the preprocessing function G, and the data with a comprehensive score CS greater than the preset value CS0 is retained by comparing the comprehensive score CS with the preset value CS0 to obtain standard data;

[0048] Update module, set w t are all the parameters of the first deep learning model M1 at time t, η is the learning rate, γ is the correction term, ▽L(w t-1 ) is the loss function L at time t-1 with respect to parameter w t-1 The gradient of v establishes the data update rule for the first deep learning model: t =γv t-1 +η▽L(w t-1 ), w t =w t-1 -v t , v t and v t-1 denote the cumulative correction terms at time t and time t-1, γv t-1 Indicates the accumulated correction value at the last moment, η▽L(w t-1 ) is the instant update value of the current gradient multiplied by the learning rate;

[0049] An input module inputs the standard data into the first deep learning model, calculates the loss function L and the gradient ▽L(w) of the loss function with respect to the parameters of the first deep learning model according to the predicted output and the true label, and when the loss change of the first deep learning model is less than a preset value, uses all parameters w' obtained by training the first deep learning model as the first training parameter set;

[0050] A formation module is formed, by constructing a custom query device, querying and obtaining data including user geographic location, user comment content, user interactive messages and user access records from the standard data to form query data;

[0051] The construction module uses the parameters in the first training parameter set as the initial parameters m of the second deep learning model M2 at time y y , τ is the global learning rate of the second deep learning model, β1 and β2 are the first decay rate and the second decay rate of the second deep learning model respectively, and the data update rule is constructed for the second deep learning model: m y =β1m y-1 +(1-β1)▽L(w y-1 ), v y =β2v y-1 +(1-β2)(▽L(w y-1 )) 2 , m y is the gradient ▽L(w y-1 ), v y is an exponentially weighted average of the square of the gradient, and a second training parameter set is obtained by inputting the query data into the second deep learning model for pre-training;

[0052] The transformation module uses the first deep learning model M1 to process the data to be classified X to obtain the first set of feature points F1, applies p-1 different data transformations to the data to be classified X to obtain p-1 view data of the data F1, and processes the p view data through M1 respectively to obtain p-1 feature point sets F2, F3, ...F p , all feature sets F1, F2,...F p Perform fusion to obtain the first fusion feature F*;

[0053] The integration module performs q random mutations on the second deep learning model M2 to obtain an integrated model EMB={M21, M22, ..., M2q} containing q mutated second deep learning models, inputs the first fusion feature F* into each model in the integrated model EMB to obtain multiple second feature point sets F21, F22, ..., F2q, and integrates all the second feature point sets F2i into the final second feature point set F2* that represents the user feature points through the fusion mechanism.

[0054] In a third aspect, an embodiment of the present invention further provides an electronic device, the electronic device comprising:

[0055] at least one processor; and,

[0056] a memory communicatively connected to the at least one processor; wherein,

[0057] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the user feature point classification method based on pre-training in any of the first aspects or any implementation of the first aspect.

[0058] In a fourth aspect, an embodiment of the present invention further provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable the computer to execute the pre-trained user feature point classification method in the aforementioned first aspect or any implementation of the first aspect.

[0059] In the fifth aspect, an embodiment of the present invention further provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer executes the user feature point classification method based on pre-training in the aforementioned first aspect or any implementation of the first aspect.

[0060] The user feature point classification scheme based on pre-training in the embodiment of the present invention includes: after obtaining raw data from multiple social media, setting a pre-processing vector Y=[SS, IS, TS] corresponding to the raw data, SS represents the offset evaluation value of the data, IS represents the influence evaluation value of the data, and TS represents the timeliness evaluation value of the data, calculating the comprehensive score CS of each data in the raw data through the pre-processing function G, and retaining the data whose comprehensive score CS is greater than the preset value CS0 by comparing the comprehensive score CS with the preset value CS0 to obtain standard data; setting w t are all the parameters of the first deep learning model M1 at time t, η is the learning rate, γ is the correction term, ▽L(w t-1) is the loss function L at time t-1 with respect to parameter w t-1 The gradient of v establishes the data update rule for the first deep learning model: t =γv t-1 +η▽L(w t-1 ), w t =w t-1 -v t , v t and v t-1 denote the cumulative correction terms at time t and time t-1, γv t-1 Indicates the accumulated correction value at the last moment, η▽L(w t-1 ) is the instant update value of the current gradient multiplied by the learning rate; the standard data is input into the first deep learning model, and the loss function L and the gradient ▽L(w) of the loss function with respect to the parameters of the first deep learning model are calculated according to the predicted output and the true label. When the loss change of the first deep learning model is less than the preset value, all parameters w' obtained by training the first deep learning model are used as the first training parameter set; by constructing a custom query, data including user geographic location, user comment content, user interactive messages and user access records are queried from the standard data to form query data; the parameters in the first training parameter set are used as the initial parameters m of the second deep learning model M2 at time y y , τ is the global learning rate of the second deep learning model, β1 and β2 are the first decay rate and the second decay rate of the second deep learning model respectively, and the data update rule is constructed for the second deep learning model: m y =β1m y-1 +(1-β1)▽L(w y-1 ), v y =β2v y-1 +(1-β2)(▽L(w y-1 )) 2 , m y is the gradient ▽L(w y-1 ), the exponentially weighted average of y is the exponential weighted average of the square of the gradient, the query data is input into the second deep learning model for pre-training to obtain a second training parameter set; the first deep learning model M1 is used to process the data to be classified X to obtain a first set of feature points F1, p-1 different data transformations are applied to the data to be classified X to obtain p-1 view data of the data F1, and the p view data are processed by M1 respectively to obtain p-1 feature point sets F2, F3, ...F p , all feature sets F1, F2, ...F pThe first fusion feature F* is obtained by fusion; the second deep learning model M2 is randomly mutated q times to obtain an integrated model EMB={M21, M22, ..., M2q} containing q mutated second deep learning models, the first fusion feature F* is input into each model in the integrated model EMB, and multiple second feature point sets F21, F22, ..., F2q are obtained, and all second feature point sets F2i are integrated into the second feature point set F2* that finally represents the user feature points through the fusion mechanism. This scheme predicts the interests of users on social media through public data on users' social media (such as personal profiles, avatars, posts and comment data, etc.). The algorithm in the present invention uses pre-training and fine-tuning methods to improve the accuracy and efficiency of user interest classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, 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.

[0062] Figure 1 A schematic diagram of a flow chart of a method for classifying user feature points based on pre-training provided in an embodiment of the present invention;

[0063] Figure 2 A schematic diagram of another method for classifying user feature points based on pre-training provided in an embodiment of the present invention;

[0064] Figure 3 A schematic diagram of the structure of a user feature point classification device based on pre-training provided in an embodiment of the present invention;

[0065] Figure 4 A schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0066] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0067] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present disclosure.

[0068] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein may be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present disclosure, it should be understood by those skilled in the art that an aspect described herein may be implemented independently of any other aspect, and two or more of these aspects may be combined in various ways. For example, any number of aspects described herein may be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein may be used to implement this device and / or practice this method.

[0069] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The drawings only show components related to the present disclosure rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0070] Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, it will be understood by those skilled in the art that the aspects described may be practiced without these specific details.

[0071] The embodiment of the present disclosure provides a user feature point classification method based on pre-training. The user feature point classification method based on pre-training provided in this embodiment can be executed by a computing device, which can be implemented as software, or as a combination of software and hardware, and the computing device can be integrated in a server, terminal device, etc.

[0072] See also Figure 1 and Figure 2 The present disclosure provides a method for classifying user feature points based on pre-training, including:

[0073] S101, after obtaining the original data from multiple social media, set the preprocessing vector Y=[SS,IS,TS] corresponding to the original data, where SS represents the offset evaluation value of the data, IS represents the influence evaluation value of the data, and TS represents the timeliness evaluation value of the data. The comprehensive score CS of each data in the original data is calculated through the preprocessing function G, and the data with a comprehensive score CS greater than the preset value CS0 is retained by comparing the comprehensive score CS with the preset value CS0 to obtain standard data.

[0074] Data can be collected in real time from multiple social media platforms through APIs or data scraping tools. The collected data can include:

[0075] Text content: posts and comments posted by users.

[0076] Metadata: publishing time, location, user ID, number of likes, number of shares, number of comments, etc.

[0077] It can remove invalid data: such as empty values, duplicate data, and data in non-target languages, convert date, time, and number fields into a unified format, and remove HTML tags, special characters, stop words, etc.

[0078] If the API does not directly provide sentiment scores, you can use natural language processing (NLP) technology (such as sentiment dictionary-based methods, machine learning models, etc.) to perform sentiment analysis on the text content, calculate the sentiment score, and then obtain the offset evaluation value SS.

[0079] The comprehensive influence score of each post is calculated by combining the number of likes, shares, comments and user influence to obtain the influence evaluation value IS of the data.

[0080] In order to obtain standard data, a comprehensive scoring formula can be designed, which combines information from multiple dimensions:

[0081] Comprehensive score (CS) = α × offset rating (SS) + β × influence score (IS) + γ × timeliness score (TS).

[0082] S102, let w t are all the parameters of the first deep learning model M1 at time t, η is the learning rate, γ is the correction term, ▽L(w t-1 ) is the loss function L at time t-1 with respect to parameter w t-1 The gradient of v establishes the data update rule for the first deep learning model: t =γv t-1 +η▽L(w t-1 ), w t =w t-1 -v t , vt and v t-1 denote the cumulative correction terms at time t and time t-1, γv t-1 Indicates the accumulated correction value at the last moment, η▽L(w t-1 ) is the instant update value of the current gradient multiplied by the learning rate.

[0083] A first deep learning model M1 can be designed, and a cumulative correction term can be added to the first deep learning model M1 to improve the accuracy of data training of the first deep learning model M1 through the cumulative correction term.

[0084] Specifically, use w t represents all the parameters of the model M1 at time t, η represents the learning rate, which controls the step size of parameter update, and represents the correction term coefficient, which usually takes a value between [0,1) and is used to control the contribution of historical gradients. ∇L(w t−1 ) represents the loss function L with respect to parameter w at time t−1 t−1 The gradient of , which indicates how the parameters should be adjusted to minimize the loss. t and v t−1 denote the cumulative correction terms at time t and time t−1, respectively, which combine the historical gradient and the current gradient to update the parameters.

[0085] Data update rules can be established based on the given rules:

[0086] Calculate the cumulative correction term vt: Combined with the cumulative correction term v at the previous moment t−1 and the current gradient ∇L(w t−1 ), adjust the contribution of historical gradients by the momentum coefficient γ, and add the instant update value of the current gradient multiplied by the learning rate η. The specific formula is:

[0087] v t =γv t-1 +η▽L(w t-1 )

[0088] Cumulative Modifier v t is the historical cumulative correction term γv t−1 and the current gradient update η∇L(w t−1 ) is the weighted sum of .

[0089] The cumulative correction term vt can be used to update the model parameter wt. The specific formula is:

[0090] w t =w t-1 -v t

[0091] The new parameter w t is the old parameter w t−1Subtract the cumulative correction term vt, and update the parameters along the negative gradient direction (that is, the direction of loss reduction), but the step size and direction of the update are affected by the cumulative correction term v t of adjustment.

[0092] S103, input the standard data into the first deep learning model, calculate the loss function L and the gradient ▽L(w) of the loss function with respect to the parameters of the first deep learning model according to the predicted output and the true label, and when the loss change of the first deep learning model is less than a preset value, use all parameters w' obtained by training the first deep learning model as the first training parameter set.

[0093] You can choose a suitable loss function for the first deep learning model based on actual needs. Use the loss function to calculate the difference between the predicted output and the true label, that is, the loss value. According to the chain rule, calculate the gradient of the loss function with respect to the model parameters (▽L(w)). Use an optimization algorithm to update the model parameters according to the gradient to reduce the loss value.

[0094] After each iteration, check whether the stopping condition is met. Specifically, the loss value at the end of each epoch can be recorded, and the change in the loss value of the adjacent epoch can be calculated. If this change is less than the preset threshold, the model is considered to have converged and the training process can be stopped.

[0095] When the stopping condition is met, all parameters of the current model (w') are saved as the first training parameter set. These parameters will be used for subsequent model evaluation and prediction tasks.

[0096] S104, by constructing a custom query, query and obtain data including user geographic location, user comment content, user interactive messages and user access records from the standard data to form query data.

[0097] You can use a database management tool or a database library of a programming language to connect to a database that stores standard data. Construct an SQL query, connect the user information table, comment table, interactive message table, and access record table through a JOIN operation, and filter out data containing geographic location, comment content, interactive messages, and access records.

[0098] You can check whether the returned data is complete and accurate, adjust the data format as needed, export the query results to a preset storage file, or display them in a visualization tool.

[0099] S105: Using the parameters in the first training parameter set as the initial parameters m of the second deep learning model M2 at time y y, τ is the global learning rate of the second deep learning model, β1 and β2 are the first decay rate and the second decay rate of the second deep learning model respectively, and the data update rule is constructed for the second deep learning model: m y =β1m y-1 +(1-β1)▽L(w y-1 ), v y =β2v y-1 +(1-β2)(▽L(w y-1 )) 2 , m y is the gradient ▽L(w y-1 ), v y It is an exponentially weighted average of the square of the gradient. The second training parameter set is obtained by inputting the query data into the second deep learning model for pre-training.

[0100] After the query data is prepared, the architecture of the second deep learning model is designed according to the task requirements, and the parameters in the first training parameter set are used as the initial parameters w0.

[0101] The query data can be fed into the second deep learning model M2 in batches, and the model parameters are iteratively updated according to the preset data update rules. During the update process, the loss value during the training process is monitored to ensure that the model is learning.

[0102] The model performance can be evaluated using the validation set to ensure that the model is not overfitting or underfitting. When the model performance meets the requirements, the current model parameters are saved as the second training parameter set.

[0103] S106, using the first deep learning model M1 to process the data to be classified X, obtaining a first set of feature points F1, applying p-1 different data transformations to the data to be classified X, obtaining p-1 view data of the data F1, respectively processing the p view data through M1, obtaining p-1 feature point sets F2, F3, ...F p , all feature sets F1, F2, ...F p The fusion is performed to obtain the first fusion feature F*.

[0104] Use the deep learning model M1 to process the original data to be classified X, extract features that are helpful for classification or prediction tasks, and generate the first set of feature points F1.

[0105] By applying p-1 different data transformations to the original data X, these transformations may be rotation, scaling, translation, feature subset selection, PCA transformation, etc., revealing different characteristics of the data from different angles or dimensions, and each transformation generates a new data representation, namely view data.

[0106] For each transformed view data, the same deep learning model M1 is used again to process it, and p-1 sets of new feature point sets F2, F3, ..., Fp are obtained respectively, thereby obtaining the feature diversity that the model can capture under different data representations.

[0107] The feature set F1 obtained directly from the original data is merged with all the feature sets F2, F3, ..., Fp obtained through transformation. The fusion method can be simple concatenation, weighted summation, or a more complex feature fusion strategy (such as kernel fusion). Through fusion, the advantages of each view can be combined to form a more comprehensive and discriminative feature representation, that is, the first fused feature F*.

[0108] S107, perform q random mutations on the second deep learning model M2 to obtain an integrated model EMB={M21, M22, ..., M2q} containing q mutated second deep learning models, input the first fusion feature F* into each model in the integrated model EMB to obtain multiple second feature point sets F21, F22, ..., F2q, and integrate all the second feature point sets F2i into the final second feature point set F2* that represents the user feature points through a fusion mechanism.

[0109] The second deep learning model M2 is randomly mutated q times. Mutations may include, but are not limited to, fine-tuning of weights, minor modifications to the network structure (such as adding or removing layers, changing activation functions), changes in training parameters, etc. Each mutation generates a new model variant, denoted as M21, M22, ..., M2q. These variant models constitute a model ensemble EMB, each of which is unique to a certain extent and can understand data from different perspectives.

[0110] The first fusion feature F* obtained previously is input into each variant model M2i (i = 1, 2, ..., q) in the integrated model EMB. Each model extracts features from F* based on its unique perspective and learning ability to generate q second feature point sets F21, F22, ..., F2q. This approach further enhances the diversity and robustness of the features.

[0111] Similar to the feature fusion in the first stage, all the second feature point sets F21, F22, ..., F2q are integrated through a fusion mechanism to form the final second feature point set F2*. The fusion method can also be concatenation, weighted average or other fusion strategies, the purpose of which is to integrate the information provided by all variant models in order to achieve a more powerful and stable feature representation than a single model.

[0112] The final F2* is a highly abstract and rich user feature representation refined through two-stage deep learning and ensemble learning. It can be used as the core input for application scenarios such as advanced analysis, recommendation systems, and personalized services, further improving the accuracy and personalization of the system.

[0113] This two-stage method of combining deep learning with ensemble learning not only takes into account the multi-view characteristics of the data itself, but also introduces diversity at the model level through model variation and integration, thereby improving the comprehensiveness and stability of feature extraction and representation.

[0114] According to a specific implementation of the embodiment of the present disclosure, after obtaining the original data from multiple social media, setting a preprocessing vector Y corresponding to the original data includes:

[0115] By constructing the first calculation formula, calculate the value of SS:

[0116]

[0117] N is the total number of words recognized in the text Ti, It is vocabulary In text The frequency in It is vocabulary The sentiment polarity score of

[0118] By constructing the second calculation formula, the value of IS is calculated:

[0119]

[0120] Fl represents the number of followers, Lk represents the number of likes, Sr represents the number of shares or reposts, Cm represents the number of comments, Tc represents the time difference between the content release time and the present time, Ms is the time decay factor, and Qc represents the content quality evaluation value;

[0121] By constructing the third calculation formula, the value of TS is calculated:

[0122]

[0123] is the current time, is the data release time, is the attenuation coefficient.

[0124] According to a specific implementation of the embodiment of the present disclosure, the calculating of the comprehensive score CS of each piece of data in the original data by the preprocessing function G includes:

[0125] By weighting the calculation results of the first calculation formula, the second calculation formula and the third calculation formula, a comprehensive score CS of each data in the original data is obtained.

[0126] According to a specific implementation of the embodiment of the present disclosure, establishing a data update rule for the first deep learning model includes:

[0127] Using a federated learning algorithm, only the parameters of the first deep learning model are updated without updating the entire model structure;

[0128] Perform a decay operation on the parameters generated by the old data in the standard data.

[0129] According to a specific implementation of the embodiment of the present disclosure, after obtaining the original data from the multiple social media, the method further includes:

[0130] Clean, standardize and transform the collected data, including text denoising, grammar standardization, and removal of irrelevant content;

[0131] Extract keywords from text and visual features from avatars to prepare matching input formats for model training.

[0132] According to a specific implementation of the embodiment of the present disclosure, after constructing the data update rule for the second deep learning model, the method further includes:

[0133] right and The value of is corrected:

[0134]

[0135]

[0136] According to the revised and To update the parameters:

[0137] .

[0138] According to a specific implementation of the embodiment of the present disclosure, the method further includes:

[0139] After the first deep learning model and the second deep learning model are deployed, the models are iteratively updated based on the feedback, including adjusting algorithm parameters, retraining, or further fine-tuning.

[0140] According to a specific implementation of the embodiment of the present disclosure, the method further includes:

[0141] The user's interests are predicted using a pre-trained and fine-tuned first deep learning model and a second deep learning model, so as to predict and output the user's interest points based on the input user data.

[0142] Corresponding to the above method embodiment, see Figure 3 The embodiment of the present invention further discloses a user feature point classification device 30 based on pre-training, comprising:

[0143] Setting module 301, after obtaining raw data from multiple social media, sets a preprocessing vector Y=[SS, IS, TS] corresponding to the raw data, where SS represents the offset evaluation value of the data, IS represents the influence evaluation value of the data, and TS represents the timeliness evaluation value of the data, and calculates the comprehensive score CS of each data in the raw data through the preprocessing function G, and compares the comprehensive score CS with a preset value CS0, retaining data with a comprehensive score CS greater than the preset value CS0, to obtain standard data;

[0144] Update module 302, set w t are all the parameters of the first deep learning model M1 at time t, η is the learning rate, γ is the correction term, ▽L(w t-1 ) is the loss function L at time t-1 with respect to parameter w t-1 The gradient of v establishes the data update rule for the first deep learning model: t =γv t-1 +η▽L(w t-1 ), w t =w t-1 -v t , v t and v t-1 denote the cumulative correction terms at time t and time t-1, γv t-1 Indicates the accumulated correction value at the last moment, η▽L(w t-1 ) is the instant update value of the current gradient multiplied by the learning rate;

[0145] The input module 303 inputs the standard data into the first deep learning model, calculates the loss function L and the gradient ▽L(w) of the loss function with respect to the parameters of the first deep learning model according to the predicted output and the true label, and when the loss change of the first deep learning model is less than a preset value, uses all parameters w' obtained by training the first deep learning model as the first training parameter set;

[0146] Forming module 304, by building a custom query, querying and obtaining data including user geographic location, user comment content, user interactive messages and user access records from the standard data to form query data;

[0147] Construction module 305, using the parameters in the first training parameter set as the initial parameters m of the second deep learning model M2 at time y y , τ is the global learning rate of the second deep learning model, β1 and β2 are the first decay rate and the second decay rate of the second deep learning model respectively, and the data update rule is constructed for the second deep learning model: m y =β1m y-1 +(1-β1)▽L(w y-1 ), v y =β2v y-1 +(1-β2)(▽L(w y-1 )) 2 , m y is the gradient ▽L(w y-1 ), the exponentially weighted average of y is an exponentially weighted average of the square of the gradient, and a second training parameter set is obtained by inputting the query data into the second deep learning model for pre-training;

[0148] The transformation module 306 uses the first deep learning model M1 to process the data to be classified X to obtain a first set of feature points F1, applies p-1 different data transformations to the data to be classified X to obtain p-1 view data of the data F1, and processes the p view data through M1 respectively to obtain p-1 feature point sets F2, F3, ...F p , all feature sets F1, F2,...F p Perform fusion to obtain the first fusion feature F*;

[0149] The integration module 307 performs q random mutations on the second deep learning model M2 to obtain an integrated model EMB={M21, M22, ..., M2q} including q mutated second deep learning models, inputs the first fusion feature F* into each model in the integrated model EMB to obtain multiple second feature point sets F21, F22, ..., F2q, and integrates all the second feature point sets F2i into the final second feature point set F2* that represents the user feature points through a fusion mechanism.

[0150] See also Figure 4 The embodiment of the present invention further provides an electronic device 60, the electronic device comprising:

[0151] at least one processor; and,

[0152] a memory communicatively connected to the at least one processor; wherein,

[0153] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the user feature point classification method based on pre-training in the aforementioned method embodiment.

[0154] An embodiment of the present invention further provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable the computer to execute the aforementioned method embodiment.

[0155] An embodiment of the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer executes the user feature point classification method based on pre-training in the aforementioned method embodiment.

[0156] Reference below Figure 4 , which shows a schematic diagram of the structure of an electronic device 60 suitable for implementing the embodiment of the present disclosure. The electronic device in the embodiment of the present disclosure may include, but is not limited to, mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), vehicle terminals (such as vehicle navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 4 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0157] like Figure 4 As shown, the electronic device 60 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. In the RAM 603, various programs and data required for the operation of the electronic device 60 are also stored. The processing device 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0158] Typically, the following devices may be connected to the I / O interface 605: an input device 606 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, a gage, a gyroscope, etc.; an output device 607 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 608 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 609. The communication device 609 may allow the electronic device 60 to communicate with other devices wirelessly or by wire to exchange data. Although Figure 4The electronic device 60 is shown with various devices, but it should be understood that it is not required to implement or have all the devices shown. More or fewer devices may be implemented or have instead.

[0159] In particular, according to an embodiment of the present disclosure, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device 609, or installed from a storage device 608, or installed from a ROM 602. When the computer program is executed by the processing device 601, the above-mentioned functions defined in the method of the embodiment of the present disclosure are executed.

[0160] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A user feature point classification method based on pre-training, It is characterized in that include: After obtaining raw data from multiple social media, a preprocessing vector Y=[SS, IS, TS] corresponding to the raw data is set, where SS represents the offset evaluation value of the data, IS represents the influence evaluation value of the data, and TS represents the timeliness evaluation value of the data. The comprehensive score CS of each piece of data in the raw data is calculated by the preprocessing function G, and the data whose comprehensive score CS is greater than the preset value CS0 is retained to obtain standard data, wherein data is collected from multiple social media platforms in real time through an API or a data crawler tool, and the collected data includes text content and metadata; w t are all the parameters of the first deep learning model M1 at time t, η is the learning rate, γ is the correction term, ▽L(w t-1 ) is the loss function L at time t-1 with respect to parameter w t-1 The gradient of v establishes the data update rule for the first deep learning model: t =γv t-1 +η▽L(w t-1 ), w t =w t-1 -v t , v t and v t-1 denote the cumulative correction terms at time t and time t-1, γv t-1 Indicates the accumulated correction value at the last moment, η▽L(w t-1 ) is the instant update value of the current gradient multiplied by the learning rate; Input the standard data into the first deep learning model, calculate the loss function L and the gradient ▽L(w) of the loss function with respect to the parameters of the first deep learning model according to the predicted output and the true label, and when the loss change of the first deep learning model is less than a preset value, use all parameters w' obtained by training the first deep learning model as the first training parameter set; By building a custom query engine, the data including the user's geographic location, user's comment content, user's interactive messages and user's access records are queried from the standard data to form query data; The parameters in the first training parameter set are used as the initial parameters m of the second deep learning model M2 at time y y , τ is the global learning rate of the second deep learning model, β 1 and β 2 are the first decay rate and the second decay rate of the second deep learning model, respectively. The data update rule is constructed for the second deep learning model: y =β 1 m y-1 +(1-β 1 )▽L(w y-1 ), v y =β 2 v y-1 +(1-β 2 )(▽L(w y-1 )) 2 , m y is the gradient ▽L(w y-1 ), v y is an exponentially weighted average of the square of the gradient, and a second training parameter set is obtained by inputting the query data into the second deep learning model for pre-training; Use the first deep learning model M1 to process the data to be classified X to obtain the first set of feature points F 1 , apply p-1 different data transformations to the data to be classified X, and obtain data F 1 The p-1 view data of the image are processed by M1 respectively to obtain the p-1 feature point set F 2 , F 3 , ...F p , all feature sets F 1 , F 2 , ...F p Perform fusion to obtain the first fusion feature F*; The second deep learning model M2 is randomly mutated q times to obtain an integrated model EMB={M21, M22, ..., M2q} including q mutated second deep learning models, the first fusion feature F* is input into each model in the integrated model EMB to obtain multiple second feature point sets F21, F22, ..., F2q, and all the second feature point sets F2i are integrated into the second feature point set F2* that finally represents the user feature points through the fusion mechanism; After obtaining the original data from the plurality of social media, setting a preprocessing vector Y corresponding to the original data includes: By constructing the first calculation formula, calculate the value of SS: N is the total number of words recognized in the text Ti, It is vocabulary In text The frequency in It is vocabulary The sentiment polarity score of By constructing the second calculation formula, the value of IS is calculated: Fl represents the number of followers, Lk represents the number of likes, Sr represents the number of shares or reposts, Cm represents the number of comments, Tc represents the time difference between the content release time and the present time, Ms is the time decay factor, and Qc represents the content quality evaluation value; By constructing the third calculation formula, the value of TS is calculated: is the current time, is the data release time, is the attenuation coefficient.

2. The method according to claim 1, It is characterized in that The method of calculating the comprehensive score CS of each piece of data in the original data by using the preprocessing function G includes: By weighting the calculation results of the first calculation formula, the second calculation formula and the third calculation formula, a comprehensive score CS of each data in the original data is obtained.

3. The method according to claim 2, It is characterized in that The data update rules for the first deep learning model are established, including: Using a federated learning algorithm, only the parameters of the first deep learning model are updated without updating the entire model structure; Perform a decay operation on the parameters generated by the old data in the standard data.

4. The method according to claim 3, It is characterized in that After obtaining the raw data from the plurality of social media, the method further comprises: Clean, standardize and transform the collected data, including text denoising, grammar standardization, and removal of irrelevant content; Extract keywords from text and visual features from avatars to prepare matching input formats for model training.

5. The method according to claim 1, It is characterized in that The method further comprises: After the first deep learning model and the second deep learning model are deployed, the models are iteratively updated based on the feedback, including adjusting algorithm parameters, retraining, or further fine-tuning.

6. The method according to claim 5, It is characterized in that The method further comprises: The user's interests are predicted using a pre-trained and fine-tuned first deep learning model and a second deep learning model, so as to predict and output the user's interest points based on the input user data.

7. A user feature point classification device based on pre-training, It is characterized in that include: A setting module, after obtaining raw data from multiple social media, sets a preprocessing vector Y=[SS, IS, TS] corresponding to the raw data, where SS represents the offset evaluation value of the data, IS represents the influence evaluation value of the data, and TS represents the timeliness evaluation value of the data, and calculates the comprehensive score CS of each data in the raw data through the preprocessing function G, and retains the data whose comprehensive score CS is greater than the preset value CS0 by comparing the comprehensive score CS with the preset value CS0 to obtain standard data, wherein data is collected from multiple social media platforms in real time through APIs or data crawling tools, and the collected data includes text content and metadata; Update module, w t are all the parameters of the first deep learning model M1 at time t, η is the learning rate, γ is the correction term, ▽L(w t-1 ) is the loss function L at time t-1 with respect to parameter w t-1 The gradient of v establishes the data update rule for the first deep learning model: t =γv t-1 +η▽L(w t-1 ), w t =w t-1 -v t , v t and v t-1 denote the cumulative correction terms at time t and time t-1, γv t-1 Indicates the accumulated correction value at the last moment, η▽L(w t-1 ) is the instant update value of the current gradient multiplied by the learning rate; An input module inputs the standard data into the first deep learning model, calculates the loss function L and the gradient ▽L(w) of the loss function with respect to the parameters of the first deep learning model according to the predicted output and the true label, and when the loss change of the first deep learning model is less than a preset value, uses all parameters w' obtained by training the first deep learning model as the first training parameter set; A formation module is formed, by constructing a custom query device, querying and obtaining data including user geographic location, user comment content, user interactive messages and user access records from the standard data to form query data; The construction module uses the parameters in the first training parameter set as the initial parameters m of the second deep learning model M2 at time y y , τ is the global learning rate of the second deep learning model, β 1 and β 2 are the first decay rate and the second decay rate of the second deep learning model, respectively. The data update rule is constructed for the second deep learning model: y =β 1 m y-1 +(1-β 1 )▽L(w y-1 ), v y =β 2 v y-1 +(1-β 2 )(▽L(w y-1 )) 2 , m y is the gradient ▽L(w y-1 ), v y is an exponentially weighted average of the square of the gradient, and a second training parameter set is obtained by inputting the query data into the second deep learning model for pre-training; The transformation module uses the first deep learning model M1 to process the data to be classified X to obtain the first set of feature points F 1 , apply p-1 different data transformations to the data to be classified X, and obtain data F 1 The p-1 view data are processed by M1 respectively to obtain the p-1 feature point set F 2 , F 3 , ...F p , all feature sets F 1 , F 2 , ...F p Perform fusion to obtain the first fusion feature F*; The integration module performs q random mutations on the second deep learning model M2 to obtain an integrated model EMB={M21, M22, ..., M2q} containing q mutated second deep learning models, inputs the first fusion feature F* into each model in the integrated model EMB to obtain multiple second feature point sets F21, F22, ..., F2q, and integrates all the second feature point sets F2i into the final second feature point set F2* that represents the user feature points through a fusion mechanism; After obtaining the original data from the plurality of social media, setting a preprocessing vector Y corresponding to the original data includes: By constructing the first calculation formula, calculate the value of SS: N is the total number of words recognized in the text Ti, It is vocabulary In text The frequency in It is vocabulary The sentiment polarity score of By constructing the second calculation formula, the value of IS is calculated: Fl represents the number of followers, Lk represents the number of likes, Sr represents the number of shares or reposts, Cm represents the number of comments, Tc represents the time difference between the content release time and the present time, Ms is the time decay factor, and Qc represents the content quality evaluation value; By constructing the third calculation formula, the value of TS is calculated: is the current time, is the data release time, is the attenuation coefficient.

8. An electronic device, It is characterized in that The electronic device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the user feature point classification method based on pre-training as described in any one of claims 1 to 6.

Images