Electronic product personalized marketing method based on big data accurate portrait

Abstract

The invention discloses an electronic product personalized marketing method based on a big data accurate portrait, and relates to the technical field of big data processing, and the method comprises the following steps: multi-source data collection and cross-domain fusion processing: obtaining a multi-source data set related to an electronic product from each platform of the Internet by using a web crawler and a streaming processing engine, the multi-source data set is subjected to cleaning classification and alignment fusion processing to generate multi-modal data, and then the multi-modal data is updated in real time to dynamically reflect data changes, so that a foundation is laid for subsequent data analysis; according to the method, multi-source data is collected by using a web crawler and a streaming processing engine, quantitative analysis is performed on the data by using a Hofscode culture dimension model and a Maslow demand level theory, a user portrait is constructed by using a distributed data flow engine Apache Flink to dynamically update a label, and a GAN (GAN Adversarial Generative Network) anomaly detection and evaluation analysis method is applied, so that the user portrait is obtained. And comprehensive and accurate data acquisition and processing are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of big data processing, and particularly to a personalized marketing method for electronic products based on accurate big data profiling. Background Art

[0002] In today's digital age, the competition in the electronic product market is extremely fierce. In order to increase market share and sales performance, enterprises pay more and more attention to personalized marketing. The development of big data technology provides strong support for the personalized marketing of electronic products. By collecting and analyzing a large amount of user data, enterprises can build accurate profiles for users, so as to achieve personalized product recommendations and marketing activities.

[0003] However, the existing personalized marketing methods for electronic products based on accurate big data profiling still have some technical defects:

[0004] For example, on the one hand, different cultural backgrounds shape people's unique values, aesthetic concepts and consumption habits, and the economic development level, infrastructure construction and customs in different regions may all affect users' needs and usage habits of electronic products. However, the existing accurate big data profiling methods often ignore the influence of cultural factors and regional differences on users' purchase decisions, and it is difficult to accurately classify and profile users under different cultural backgrounds and regional needs, which affects the marketing effect.

[0005] On the other hand, the motivation for users to purchase electronic products is often complex and diverse. In addition to the functions and performance of the products, it is also affected by various factors such as psychology, society and emotion, which makes the purchase motivation of users change dynamically with the changes of time, environment and personal experience. The existing accurate big data profiling methods are usually based on static data for analysis and modeling, and cannot track the dynamic changes of users' purchase motivation in real time. If enterprises cannot timely understand the changes in users' purchase motivation and still carry out marketing according to the original profile, there will be a lack of real-time performance and the marketing information will be out of touch with the actual needs of users due to the changes in purchase motivation.

[0006] The above information disclosed in the background art section is only used to enhance the understanding of the background of the present disclosure, and therefore it may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0007] The purpose of the present invention is to provide a personalized marketing method for electronic products based on accurate big data profiling. The present invention uses the Hofstede cultural dimension model and Maslow's hierarchy of needs theory to quantitatively analyze data, constructs user profiles by dynamically updating tags in Apache Flink, and performs anomaly detection and evaluation analysis using a GAN network to solve the problems in the above background art.

[0008] To achieve the above object, the present invention provides the following technical solutions: A personalized marketing method for electronic products based on big data precise profiling, comprising the following steps:

[0009] S1. Multi-source data collection and cross-domain fusion processing: Use web crawlers and streaming processing engines to obtain multi-source data sets related to electronic products from various platforms on the Internet, and perform cleaning classification, alignment and fusion processing on the multi-source data sets to generate multi-modal data. Then, update the multi-modal data in real time to dynamically reflect data changes, laying a foundation for subsequent data analysis;

[0010] S2. Construct a multi-dimensional user profile: Combine the Hofstede cultural dimension model and Maslow's hierarchy of needs theory to perform quantitative analysis of cultural regions and interest motivations on the multi-modal data, and generate dynamic and static labels. And dynamically update the labels through the distributed data stream engine Apache Flink to construct a user profile model;

[0011] S3. Detect demand anomalies and correct the profile: Use the generative adversarial network GAN to detect abnormal features of time decay, interest drift and cycle stages in the preliminary user profile, and establish a dynamic interest decay model based on the abnormal features. Train and correct the user profile model through the generative adversarial network GAN;

[0012] S4. Generation and execution of personalized marketing strategies: Develop and execute personalized differential marketing strategies according to the corrected user profile model. The personalized differential marketing strategies include hierarchical marketing strategies and multi-channel reach methods;

[0013] S5. Marketing effect evaluation and quantitative feedback: Construct a multi-dimensional evaluation model, input the behavioral feedback data of users who purchase electronic products into the GAN evaluation network, and output the evaluation results of the marketing health coefficient. Optimize the marketing strategy by quantifying the behavioral feedback data according to the evaluation results;

[0014] S6. Model iteration to optimize the marketing strategy: Use online learning technology to feedback the evaluation results to the user profile model for iterative learning, and continuously optimize the personalized differential marketing strategy.

[0015] Optionally, the steps for obtaining the multi-source data set are as follows:

[0016] Search for e-commerce platforms where electronic products are listed on the Internet, track the social platforms of users registered on the e-commerce platforms to obtain multi-source data, define the content according to the types of multi-source data, and configure the collection tools of web crawlers and streaming processing engines. Among them, the multi-source data set includes user basic information, behavioral data, regional characteristics and social data;

[0017] According to the web page structures and data formats of e-commerce platforms and social platforms, use the Scrapy framework to write a web crawler program, configure the parameters of the web crawler for request headers, crawling interval time, and crawling depth, and comply with the Robots protocol;

[0018] After crawling multi-source data through the web crawler, use the downsampling algorithm to obtain the number of multi-source data samples for dynamic sampling, and transmit the number of multi-source data samples to the streaming processing engine in real time. After performing preliminary filtering and conversion on the number of multi-source data samples, removing invalid data and duplicate data, a multi-source data set is obtained, denoted as Msd.

[0019] Optionally, the steps for generating the multi-modal data are as follows:

[0020] Clean and classify the obtained multi-source data set according to user basic information, behavior data, regional characteristics, and social data, and obtain the multi-source data set as and Bui represents the set of user basic information, Bhd represents the set of behavior data, Rcs represents the set of regional characteristics, Scd represents the set of social data, A x 、B y 、C i 、D j respectively represent subsets of the sets of user basic information, behavior data, regional characteristics, and social data;

[0021] Based on the graph embedding model and setting a determination threshold, use the identity association algorithm to perform cross-domain alignment of user identities from the user basic information and behavior data in the multi-source data set. Among them, the expression of the identity association algorithm is In the formula, u1(A x ,B y ) represents the user basic information and behavior data extracted from the e-commerce platform, and u2(A x ,B y ) represents the user basic information and behavior data extracted from the social platform, represents the vector representation of the e-commerce platform user u1 in the graph embedding space, represents the vector representation of the social platform user u2 in the graph embedding space, Similarity represents the similarity of cross-platform user identity association, cos(,) represents the cosine similarity function, and the value range is [-1,1]. When the absolute value of Similarity is greater than the determination threshold, it is determined as the same user, otherwise it is not;

[0022] Use the TF-IDF method for text feature extraction to extract the social features of users. Among them, the expression for the extraction of social features is Wherein, TF-IDF(D3, Scd) represents the extraction of the user's social features by the TF-IDF method, TF(D3, Scd) represents the number of times the user's D3 emotional comment on the electronic product appears in the current set of social data Scd, DF(D3) represents the total number of D3 emotional comments included, and Scd' represents the total number of the set of social data Scd;

[0023] Extract the corresponding regional features according to the user identity, and generate a regional sensitivity factor based on the regional features. Among them, the calculation formula of the regional sensitivity factor is δ Rcs =∑ω i ·C i and ∑ω i =1, i = 1, 2, 3. In the formula, δ Rcs represents the regional sensitivity factor, ∑ represents the summation symbol, ω i represents the weight coefficient of the corresponding subset in the regional feature set, and C i represents the subset of the regional feature set;

[0024] Construct the initial multimodal data tensor τ0 for each user according to the user identity, feature data and time dimension, and then use the variable streaming tensor fusion to update the multimodal data. The update expression of the multimodal data is τ1 = μτ0 + (1 - μ)Δτ, and μ ∈ (0, 1). In the formula, τ1 represents the updated multimodal data tensor, Δτ represents the real-time new data increment, and μ represents the decay factor.

[0025] Optionally, the steps of quantitative analysis of the cultural region by the Hofstede cultural dimension model are as follows:

[0026] According to the C1 geographical location in the user's regional feature Rcs, obtain the six cultural dimension scores of the corresponding region from the database of the Hofstede cultural dimension model, including power distance, individualism, masculinity, uncertainty avoidance, long-term orientation, indulgence and restraint, which are respectively labeled as PDI, IDV, MAS, UAI, LTO, IVR;

[0027] After normalizing the six cultural dimension scores of the corresponding user's geographical location, construct a cultural feature vector. Among them, the expression of the cultural feature vector is In the formula, V hcdt represents the normalized user cultural preference vector, and the dimension is 6×1. PDI, IDV, MAS, UAI, LTO, and IVR respectively represent the original scores of the six Hofstede cultural dimensions, and the range is 0 - 100 points;

[0028] Generate cultural sensitivity factors through linear combination. Then, the cultural sensitivity factors serve as dynamic labels for cultural feature vectors to recommend marketing strategies for users to purchase electronic products. Among them, the calculation formula for the cultural sensitivity factor is ζ hcdt = ω hcdt T ·V hcdt , and ω hcdt ∈R 6 . In the formula, ζ hcdt represents the cultural sensitivity factor, ω hcdt represents the weight vector corresponding to the user's cultural preference vector V hcdt , and ω hcdt is determined by optimizing through A / B testing. T represents the transpose, and V hcdt represents the user's cultural preference vector.

[0029] Optionally, the steps for quantitative analysis of the interest motivation by the Maslow's hierarchy of needs theory are as follows:

[0030] Map the user's behavior data set Bhd and social data set Scd to the five levels of Maslow's needs, including physiological needs, safety needs, social needs, esteem needs, and self-actualization needs;

[0031] Use the BERT model to analyze the emotional intensity corresponding to the user's behavior in the five levels of Maslow's needs. Then, the emotional intensity serves as a dynamic label for the user's motivation preference vector. Among them, the calculation formula for the emotional intensity is In the formula, represents the emotional intensity, and the value range is mapped to the interval [0,1]. l represents the five levels of Maslow's needs. BERT(Bhd l ,Scd l ) represents the analysis of the emotional intensity of the behavior data set Bhd and social data set Scd on the five levels of Maslow's needs using the BERT model. σ represents the Sigmoid activation function, ω l represents the weight matrix corresponding to the five levels of Maslow's needs. T represents the transpose, represents the hidden vector of the behavior data set Bhd and social data set Scd on the five levels of Maslow's needs output by the BERT model, and b l represents the bias term corresponding to the five levels of Maslow's needs. Calculate the weights of each need level in the five levels of Maslow's needs based on the emotional intensity combined with the user's behavior frequency. Then, the calculation formula for the weights of the five levels of Maslow's needs is In the formula, ω mhn，l represents the weight of the l-th level of Maslow's needs, represents the number of times the behavior of the user's behavior data set Bhd and social data set Scd appears in the l'-th level of the five levels of Maslow's needs, Denote the total number of behaviors that the behavioral data set Bhd and the social data set Scd of the user appear in Maslow's five - level needs;

[0032] Use the weights ω of each need level mhn，l Construct a user motivation preference vector, where the expression of the user motivation preference vector is In the formula, V mhn Denote the user motivation preference vector, and the dimension is 5×1, ω mhn，1 、ω mhn，2 、ω mhn，3 、ω mhn，4 、ω mhn，5 Respectively denote the weights of Maslow's five - level needs.

[0033] Optionally, the steps for the distributed data stream engine to dynamically update tags are as follows:

[0034] Access the real - time event stream of the user's behavioral data set Bhd through the distributed data stream engine ApacheFlink;

[0035] Set a sliding window, and calculate the short - term interest of the user according to the sliding window and calculate the interest score. The calculation formula of the interest score is In the formula, is t Denote the interest score of the user at the current time point t, Denote the behavioral weight of the sliding - window user at the time point t - t L The time point, t represents the current time point, t L Denote the window length, and t L = 600 seconds, γ represents the decay coefficient;

[0036] Extract the initial tags of the geographical sensitivity factor δ Rcs 、the cultural sensitivity factor ζ hcdt and the emotional intensity And continuously update the user tags using the state backend of the distributed data stream engine. The update calculation formula of the user tags is In the formula, Denote the new tags updated from the initial tags corresponding to the geographical sensitivity factor δ Rcs 、the cultural sensitivity factor ζ hcdt and the emotional intensity , α represents the smoothing factor, and 0 < α < 1, is t Denote the interest score of the user;

[0037] The steps for constructing the user portrait model are as follows:

[0038] First, generate a static feature vector for the set Bui of user basic information and label it as V Bui , and the dimension is 4×1;

[0039] Combine the cultural feature vector V hcdt , the user motivation preference vector V hcdt and the static label V Bui to synthesize and splice them into a user portrait feature vector. The expression of the user portrait feature vector is In the formula, V user represents the user portrait feature vector, and the dimension is 15×1, represents the vector splicing operation;

[0040] Use the Wide&Deep model to jointly train the user portrait. The Wide part captures the explicit interaction relationships of culture, motivation, and static features, and the Deep part mines the implicit feature relationships through a multi-layer neural network;

[0041] And according to the user portrait model, output the user label probability and the recommendation priority. The calculation formula for the output user label probability is P(K|V user ) = σ(ω Wide&Deep T ·V suer +b Wide&Dee ). In the formula, P(K|V user ) represents the output user label probability, P(|) represents the conditional probability, K represents the user target label for purchasing electronic products, σ represents the Sigmoid activation function, ω Wide&Deep represents the weight vector of the Wide&Deep model, T represents the transpose, and b Wide&Dee represents the bias term of the Wide&Deep model.

[0042] Optionally, the steps for the GAN to detect abnormal features are as follows:

[0043] Use the generative adversarial network GAN, and use the generator G and the discriminator D to perform anomaly detection on the user portrait feature vector V user , that is, input the user portrait feature vector V user into the generator G to generate a reconstructed portrait G(V user ), use the discriminator D to distinguish between the user portrait feature vector V user and the reconstructed portrait G(V user ), and detect the anomaly score;

[0044] Calculate the reconstruction error as the anomaly index. Among them, the calculation formula for the reconstruction error is In the formula, e(V user ) represents the user portrait feature vector Vuser and the reconstruction error between the reconstructed image G(V user ), expressed as the square of the L2 norm; is expressed as the square of the L2 norm;

[0045] Set a dynamic threshold and compare it with the anomaly index for analysis to determine it as an abnormal feature. Among them, the dynamic threshold is set to θ = 90%, and when e(V user ) > θ, it is judged as an abnormal feature of the user portrait, otherwise vice versa;

[0046] Train the GAN according to anomaly detection. By defining the loss functions of the generator G and the discriminator D, among them, the expression of the loss function of the generator G is and D(G(V user )) → 1, which is used to minimize the reconstruction error and deceive the discriminator D at the same time. In the formula, represents the loss function of the generator G, and E(V user , p) represents the expectation of the normal probability of the data distribution of the normal user portrait feature vector V user , μ represents the trade-off coefficient for balancing the reconstruction error and the adversarial loss, and D(G(V user )) represents the discriminant probability of minimizing the reconstructed image G(V user ), and p represents the normal probability of the data distribution of the normal user portrait feature vector V user ;

[0047] The expression of the loss function of the discriminator D is In the formula, represents the loss function of the discriminator D, and D(V user ) represents maximizing the discriminant probability of the normal user portrait feature vector V user ;

[0048] The steps for establishing the dynamic interest decay model are as follows:

[0049] Adjust the historical behavior weights of users by using an exponential decay method for time. Among them, the expression for adjusting the weights by time decay is and Δt = t - t0. In the formula, ω(t) represents the weight dynamically adjusted with time decay, t represents the current time point, t0 represents the historical time point when the user first had an action on the electronic product, and Δt represents the time difference, represents the decay rate for controlling the user's interest;

[0050] Aggregate the behavior influence after time decay through the weighted time difference and dynamically update the interest score. Among them, the update calculation formula for the interest score is UI t = ω(t) · is t, where UI t represents the interest score updated at time point t;

[0051] Then, use Fourier transform to extract the periodic interest pattern to detect and identify the periodic changes in the user's interest fluctuations. Among them, the expression of the periodic interest pattern is In the formula, F(f) represents the periodic intensity corresponding to frequency f, and T0 represents the total length of the time window, represents the imaginary unit, and f represents the frequency index, and the range is f = 0, 1, 2,..., T0 - 1, and the period is

[0052] Optionally, the construction steps of the multi-dimensional evaluation model are as follows:

[0053] According to the user portrait model and the personalized differential marketing strategy, define the evaluation index of the multi-dimensional evaluation model as the marketing health coefficient, including the matching degree between the user portrait and the recommended marketing strategy, the timeliness of the overlap between the marketing strategy touch time and the user demand window, and the cultural compatibility between the cultural preferences in the user portrait and the recommended marketing strategy;

[0054] Input the user's behavior feedback data and the user portrait feature V into the GAN evaluation network user , the generator G of the GAN generates a feature vector simulating the healthy marketing result, and outputs the health coefficient through the defined discriminator D loss function. The behavior feedback data includes click-through rate, conversion rate, and average order value, which are respectively calibrated as ctr, cvr, and aov. Among them, the expression of the discriminator D loss function is and (X') = (ctr, cvr, aov, V user ), in the formula, represents the discriminator D loss function in the GAN, E(X') represents the expectation of the distribution of the real marketing feedback data X', D(X') represents the output probability of the discriminator for the real marketing feedback data X', and X′ represents the real marketing feedback data of the electronic product to the user, including click-through rate ctr, conversion rate cvr, average order value aov, and user portrait V user ;

[0055] Construct the feature vector of the electronic product as V ecp , and calculate the matching degree based on the similarity between the user portrait feature and the feature of the recommended electronic product. Then, the calculation formula of the matching degree is Relevance = cos(V user , V ecp ), in the formula, Relevance represents the matching degree, and cos(,) represents the cosine similarity;

[0056] The timeliness is quantified by a time decay function to calculate the effectiveness of the reach timing, and the formula for timeliness is where Timeliness represents timeliness, χ represents the decay coefficient that controls the decay rate of timeliness due to time deviation, |t 触达 -t 需求 | represents the absolute value of the difference between the timestamp of the actual reach to the user and the midpoint timestamp of the user demand window t 触达 ; 需求

[0057] To calculate the cultural compatibility by evaluating the conflict degree between the marketing strategy and the user cultural vector V hcdt , the formula for cultural compatibility is CulturalFit = 1 - ||V 策略文化 -V hcdt ||1, where CulturalFit represents cultural compatibility, V 策略文化 represents the marketing strategy cultural label vector, ||||1 represents the L1 norm, which measures the sum of the absolute values of the differences between the marketing strategy cultural label vector V 策略文化 and the user cultural vector V hcdt .

[0058] A computer device includes: a memory and a processor; the memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned personalized marketing method for electronic products based on big data accurate profiling are implemented.

[0059] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned personalized marketing method for electronic products based on big data accurate profiling are implemented.

[0060] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0061] ​Through the use of web crawlers and streaming processing engines to collect multi-source data, quantitative analysis of the data using Hofstede's cultural dimension model and Maslow's hierarchy of needs theory, the use of the distributed data stream engine Apache Flink to dynamically update tags to build user portraits, and the use of the adversarial generative network GAN for anomaly detection and evaluation analysis, the present invention realizes comprehensive and accurate data collection and processing, provides solid data support for precision marketing, accurately constructs user portraits, enables marketing to better fit the cultural regions and interest motives of users; formulates personalized differential marketing strategies according to user portraits and executes them, improving the pertinence and accuracy of marketing; a scientific marketing effect evaluation and quantitative feedback mechanism can timely discover the deficiencies in marketing strategies; the iterative optimization of the model ensures that marketing strategies can dynamically adapt to market changes and user needs, thereby improving marketing efficiency and input-output ratio, enhancing the competitiveness of enterprises in the electronic product market, and better meeting the diverse needs of users. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings.

[0063] Figure 1 It is a flowchart of the personalized marketing method for electronic products based on big data precise portraits of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0064] Example embodiments will now be described more fully with reference to the accompanying drawings. However, the example embodiments can be implemented in various forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that this disclosure will be more complete and comprehensive, and will fully convey the concept of the example embodiments to those skilled in the art.

[0065] The present invention provides a personalized marketing method for electronic products based on big data precise portraits as Figure 1 shown, including the following steps:

[0066] S1. Multi-source data collection and cross-domain fusion processing: Use web crawlers and streaming processing engines to obtain multi-source data sets related to electronic products from various platforms on the Internet, and perform cleaning classification, alignment and fusion processing on the multi-source data sets to generate multi-modal data, and then update the multi-modal data in real time to dynamically reflect data changes, laying a foundation for subsequent data analysis;

[0067] Specifically, the steps for obtaining the multi-source data set are as follows:

[0068] An e-commerce platform that searches for electronic products to be listed on the Internet tracks the social platforms of registered users on the e-commerce platform to obtain multi-source data, defines content according to the types of multi-source data, and configures collection tools for web crawlers and streaming processing engines. Among them, the multi-source data set includes user basic information, behavior data, geographical features, and social data. User basic information covers the age, gender, occupation, and consumption ability data registered by users on the e-commerce platform; behavior data includes the page click stream of users on electronic products, shopping cart operations, and search keywords; geographical features are the geographical locations parsed from IP addresses, and the economic level and climate conditions of the corresponding geography are extracted from the geographical database; social data obtains users' social dynamics, friend relationships, and emotional comment data on the social platform from the social media API interface.

[0069] According to the web structures and data formats of the e-commerce platform and the social platform, use the Scrapy framework to write a web crawler program, configure parameters such as the request header, crawling interval time, and crawling depth for the web crawler, and abide by the Robots protocol to use the written appropriate web crawler to crawl the public multi-source data from the e-commerce platform and the social platform, and avoid causing excessive pressure on the e-commerce platform and the social platform.

[0070] After using the web crawler to crawl the multi-source data, use the downsampling algorithm to obtain the number of multi-source data samples for dynamic sampling to prevent data overload. Among them, the expression of the downsampling algorithm is In the formula, P represents the proportion of the downsampled data samples in the total data, n represents the downsampled data samples, N represents the total number of sample data crawled by the web crawler, and the number of multi-source data samples is transmitted to the streaming processing engine in real time. After preliminary filtering and conversion of the number of multi-source data samples, invalid data and duplicate data are removed to obtain a multi-source data set, designated as Msd.

[0071] Specifically, the generation steps of multi-modal data are as follows:

[0072] Clean and classify the obtained multi-source data set according to user basic information, behavior data, geographical features, and social data to obtain the multi-source data set as And Bui represents the set of user basic information, Bhd represents the set of behavior data, Rcs represents the set of geographical features, Scd represents the set of social data, A x 、B y 、C i 、D j respectively represent subsets of the sets of user basic information, behavior data, geographical features, and social data.

[0073] Based on the graph embedding model and setting a determination threshold, use the identity association algorithm to cross-domain align user identities from the user basic information and behavior data in the multi-source dataset. Among them, the expression of the identity association algorithm is Similarity(u1(A x ,B y ), wherein, u1(A x ,B y ) represents the user basic information and behavior data extracted from the e-commerce platform, u2(A x ,B y ) represents the user basic information and behavior data extracted from the social platform, represents the vector representation of the e-commerce platform user u1 in the graph embedding space, represents the vector representation of the social platform user u2 in the graph embedding space, Similarity represents the similarity of cross-platform user identity association, cos(,) represents the cosine similarity function, and the value range is [-1,1]. When the absolute value of Similarity is greater than the determination threshold, it is determined as the same user, otherwise vice versa;

[0074] Use the TF-IDF method for text feature extraction to extract the social features of users. Among them, the extraction expression of the social features is wherein, TF-IDF(D3,Scd) represents the extraction of the social features of users by the TF-IDF method, TF(D3,Scd) represents the number of times the user's D3 emotional comment on electronic products appears in the current social data set Scd, DF(D3) represents the total number of D3 emotional comments included, and Scd' represents the total number of the social data set Scd;

[0075] Extract the corresponding regional features according to the user identity, and generate a regional sensitivity factor according to the regional features. Among them, the calculation formula of the regional sensitivity factor is δ Rcs =∑ω i ·C i , and Σω i =1, i = 1, 2, 3. In the formula, δ Rcs represents the regional sensitivity factor, Σ represents the summation symbol, ω i represents the weight coefficient of the corresponding subset in the regional feature set, C i represents the subset of the regional feature set;

[0076] Construct the initial multi-modal data tensor for each user, calibrated as τ0, based on user identity, feature data, and time dimension, and then use the changing streaming tensor fusion to update the multi-modal data. The update expression for the multi-modal data is τ1 = μτ0 + (1 - μ)Δτ, where μ ∈ (0, 1). In the formula, τ1 represents the updated multi-modal data tensor, Δτ represents the real-time new data increment, and μ represents the attenuation factor.

[0077] Furthermore, a multi-source dataset related to electronic products is widely crawled from various platforms on the Internet through web crawlers, covering user basic information, behavior data, geographical features, social data, etc. The streaming processing engine is used to achieve real-time acquisition and preliminary processing of data. That is, the method of combining web crawlers and streaming processing engines is an important basis for integrating multi-source data, improving data quality, and ensuring data real-time. While cleaning, classifying, aligning, and fusing the acquired multi-source dataset, a real-time update mechanism is established to enable the multi-modal data to dynamically reflect data changes, thereby providing comprehensive, accurate, and timely basic data for subsequent data analysis, and then being able to construct a more accurate portrait for users and achieve personalized marketing of electronic products.

[0078] S2. Construct a multi-dimensional user portrait: Combine the Hofstede cultural dimension model and Maslow's hierarchy of needs theory to quantitatively analyze the multi-modal data in terms of cultural region and interest motivation, generate dynamic and static tags, and dynamically update the tags through the distributed data stream engine Apache Flink to construct a user portrait model;

[0079] Specifically, the steps for quantitatively analyzing the cultural region by the Hofstede cultural dimension model are as follows:

[0080] According to the C1 geographical location in the regional characteristics Rcs of the user, obtain the six cultural dimension scores of the corresponding region from the database of the Hofstede cultural dimension model, including power distance, individualism, masculinity, uncertainty avoidance, long-term orientation, indulgence and restraint, which are respectively calibrated as PDI, IDV, MAS, UAI, LTO, IVR;

[0081] After normalizing the six cultural dimension scores of the corresponding user's geographical location, construct a cultural feature vector. Among them, the expression of the cultural feature vector is In the formula, V hcdt represents the normalized user cultural preference vector, and the dimension is 6×1. PDI, IDV, MAS, UAI, LTO, and IVR respectively represent the original scores of the six Hofstede cultural dimensions, and the range is 0 - 100 points;

[0082] Generate cultural sensitivity factors through linear combination. Then, the cultural sensitivity factors serve as dynamic labels for the cultural feature vectors to recommend marketing strategies for users to purchase electronic products. Among them, the calculation formula for the cultural sensitivity factor is ζ hcdt = ω hcdt T ·V hcdt , and ω hcdt ∈R 6 . In the formula, ζ hcdt represents the cultural sensitivity factor, ω hcdt represents the weight vector corresponding to the user's cultural preference vector V hcdt , and ω hcdt is determined by A / B test optimization. T represents the transpose, and V hcdt represents the user's cultural preference vector.

[0083] Specifically, the quantitative analysis steps of Maslow's hierarchy of needs for interest motivation are as follows:

[0084] Map the user's behavior data set Bhd and social data set Scd to Maslow's five levels of needs, including physiological needs, safety needs, social needs, respect needs, and self-actualization needs;

[0085] Use the BERT model to analyze the emotional intensity corresponding to the user's behavior in Maslow's five levels of needs. Then, the emotional intensity serves as a dynamic label for the user's motivation preference vector. Among them, the calculation formula for the emotional intensity is In the formula, represents the emotional intensity, and the value range is mapped to the [0,1] interval. l represents the five levels of Maslow's needs. BERT(Bhd l ,Scd l ) represents the analysis of the emotional intensity of the behavior data set Bhd and social data set Scd in Maslow's five levels of needs using the BERT model. σ represents the Sigmoid activation function, ω l represents the weight matrix corresponding to Maslow's five levels of needs. T represents the transpose, represents the hidden vector of the behavior data set Bhd and social data set Scd output by the BERT model in Maslow's five levels of needs, and b l represents the bias term corresponding to Maslow's five levels of needs. Calculate the weights of each need level in Maslow's five levels of needs based on the emotional intensity combined with the user's behavior frequency. Then, the calculation formula for the weights of Maslow's five levels of needs is In the formula, ω mhn，l represents the weight of the l-th level of Maslow's needs, represents the number of times the behavior of the user's behavior data set Bhd and social data set Scd appears in the l'-th level of needs in Maslow's five levels of needs, Denote the total number of behaviors that the user's behavioral data set Bhd and social data set Scd appear in Maslow's five - layer needs;

[0086] Use the weights ω of each need level mhn，l Construct a user motivation preference vector, where the expression of the user motivation preference vector is In the formula, V mhn Denote the user motivation preference vector, and the dimension is 5×1, ω mhn，1 、ω mhn，2 、ω mhn，3 、ω mhn，4 、ω mhn，5 Respectively denote the weights of Maslow's five - layer needs.

[0087] Specifically, the steps for the distributed data stream engine to dynamically update tags are as follows:

[0088] Access the real - time event stream of the user's behavioral data set Bhd through the distributed data stream engine ApacheFlink;

[0089] Set a sliding window, set it to a ten - minute window with a two - minute slide, and calculate the short - term interest of the user according to the sliding window and calculate the interest score. The formula for calculating the interest score is In the formula, is t Denote the interest score of the user at the current time point t, Denote the behavioral weight of the user in the sliding window at the time point t - t L The time point, t represents the current time point, t L Denote the window length, and t L =600 seconds, γ represents the decay coefficient, which is used to control the decay speed of the historical behavior weight over time;

[0090] Extract the initial tags of the geographical sensitivity factor δ Rcs 、the cultural sensitivity factor ζ hcdt And the emotional intensity And continuously update the user tags using the state backend of the distributed data stream engine. The formula for updating the user tags is In the formula, Denote the new tags after updating the initial tags corresponding to the geographical sensitivity factor δ Rcs 、the cultural sensitivity factor ζ hcdt And the emotional intensity α represents the smoothing factor, and 0 < α < 1, is t Denote the interest score of the user.

[0091] Specifically, the steps for constructing the user portrait model are as follows:

[0092] First, generate a static feature vector for the set Bui of user basic information and mark it as V Bui , and the dimension is 4×1;

[0093] Combine the cultural feature vector V hcdt , the user motivation preference vector V hcdt and the static label V Bui to comprehensively splice them into a user portrait feature vector. The expression of the user portrait feature vector is In the formula, V user represents the user portrait feature vector, and the dimension is 15×1, represents the vector splicing operation;

[0094] Use the Wide&Deep model to jointly train the user portrait. The Wide part captures the explicit interaction relationships of culture, motivation, and static features, and the Deep part mines the implicit feature relationships through a multi-layer neural network;

[0095] And according to the user portrait model, output the user label probability and the recommendation priority. The calculation formula for the output user label probability is P(K|V user ) = σ(ω Wide&Deep T ·V user +b Wide&Dee ). In the formula, P(K|V user ) represents the output user label probability, P(|) represents the conditional probability, K represents the user target label for purchasing electronic products, σ represents the Sigmoid activation function, ω Wide&Deep represents the weight vector of the Wide&Deep model, T represents the transpose, and b Wide&Dee represents the bias term of the Wide&Deep model.

[0096] Furthermore, it should be noted that by adopting the method of combining the Hofstede cultural dimension model and the Maslow's hierarchy of needs theory, quantitative analysis is carried out on multi-modal data. For example, with the help of the Hofstede cultural dimension model, the cultural regions are quantified to deeply analyze the influence of different regional cultures on user behavior; the Maslow's hierarchy of needs theory is used to quantify the interest motivation to accurately grasp the user's need level. Based on this quantitative analysis, dynamic and static labels are generated to comprehensively and meticulously depict user characteristics; and the distributed data stream engine Apache Flink is used to dynamically update the labels to ensure that the user portrait keeps up with user behavior and market changes, thereby achieving the effect of more accurately depicting user characteristics and reflecting user demand changes in real time, thus constructing a more comprehensive, accurate, and dynamic multi-dimensional user portrait, providing strong support for the personalized marketing of electronic products, and improving the pertinence and effectiveness of marketing.

[0097] S3. Detect and correct the abnormal features in the portrait by demand: Use the Generative Adversarial Network (GAN) to detect the abnormal features of time decay, interest drift, and periodic stage in the user's preliminary portrait, establish a dynamic interest decay model based on the abnormal features, and train and correct the user portrait model through the Generative Adversarial Network (GAN);

[0098] Specifically, the steps for GAN to detect abnormal features are as follows:

[0099] Use the Generative Adversarial Network (GAN), and use the generator G and the discriminator D to perform abnormal detection on the user portrait feature vector V user That is, input the user portrait feature vector V user into the generator G to generate a reconstructed portrait G(V user ), use the discriminator D to distinguish between the user portrait feature vector V user and the reconstructed portrait G(V user ), and detect the abnormal score;

[0100] Calculate the reconstruction error as the abnormal index. Among them, the calculation formula of the reconstruction error is In the formula, e(V user ) represents the reconstruction error between the user portrait feature vector V user and the reconstructed portrait G(V user ), represents the square of the L2 norm;

[0101] Set a dynamic threshold, and compare and analyze it with the abnormal index to determine it as an abnormal feature. Among them, the dynamic threshold is set to θ = 90%, when e(V user ) > θ, it is judged as an abnormal feature of the user portrait, otherwise it is not;

[0102] Train the GAN according to the abnormal detection. By defining the loss functions of the generator G and the discriminator D. Among them, the expression of the loss function of the generator G is and D(G(V user )) → 1, which is used to minimize the reconstruction error and deceive the discriminator D at the same time. In the formula, represents the loss function of the generator G, E(V user , p) represents the expectation of the normal probability of the data distribution of the normal user portrait feature vector V user , μ represents the trade-off coefficient for balancing the reconstruction error and the adversarial loss, D(G(V user )) represents the minimum discriminant probability of the reconstructed portrait G(V user ), and p represents the normal probability of the data distribution of the normal user portrait feature vector V user ;

[0103] The loss function expression of discriminator D is In the formula, denotes the loss function of discriminator D, and D(V user ) represents maximizing the discrimination probability for the feature vector V of the normal user portrait user .

[0104] Specifically, the steps to establish the dynamic interest decay model are as follows:

[0105] Adjust the historical behavior weight of the user by adopting an exponential decay method for time. Among them, the expression for the time decay adjustment weight is and Δt = t - t0. In the formula, ω(t) represents the weight dynamically adjusted with time decay, t represents the current time point, t0 represents the historical time point when the user first had an action on the electronic product, and Δt represents the time difference represents the decay rate of controlling the user's interest;

[0106] Aggregate the behavior influence after time decay through the weighted time difference, and dynamically update the interest score. Among them, the update calculation formula for the interest score is UI t = ω(t)·is t , in the formula, UI t represents the interest score updated at time t;

[0107] Then use the Fourier transform to extract the periodic interest pattern to detect and identify the periodic change of the user's interest fluctuation. Among them, the expression for the periodic interest pattern is In the formula, F(f) represents the periodic intensity corresponding to the frequency f, and T0 represents the total length of the time window represents the imaginary unit, and f represents the frequency index, and the range is f = 0, 1, 2, …, T0 - 1, and the period is

[0108] Specifically, the steps to correct the user portrait model are as follows:

[0109] Train the GAN according to the user portrait feature vector V user ;

[0110] Define the classification loss and adversarial loss of the user portrait;

[0111] Synchronously optimize the generator G through gradient descent;

[0112] Use the generator G to generate the corrected normal portrait to obtain the corrected user portrait model V′ user .

[0113] Furthermore, through the powerful feature detection ability of GAN, a comprehensive analysis is carried out on the user's preliminary portrait, and abnormal features such as time decay, interest drift, and cycle stages are accurately identified. After identifying the abnormal features, a dynamic interest decay model is established, fully considering the changing law of the user's interest over time. Then, the adversarial generative network GAN is used again to train and correct the user portrait model, so that the portrait can fit the user's changing interests and needs in real time, and the deviation of the user portrait can be discovered and corrected in time, improving the accuracy and timeliness of the user portrait, and further enabling the personalized marketing of electronic products to better meet the real needs of users and improve the conversion rate and effect of marketing.

[0114] S4. Generation and implementation of personalized marketing strategies: Develop and implement personalized differential marketing strategies according to the corrected user portrait model. The personalized differential marketing strategies include hierarchical marketing strategies and multi-channel reach methods. Among them, the hierarchical marketing strategies include cultural region adaptation strategies, interest motivation-driven hierarchical preferential promotion strategies, and dynamic pricing strategies.

[0115] Specifically, the steps for formulating personalized differential marketing strategies are as follows:

[0116] A1. Hierarchical marketing strategies: Based on the corrected user portrait model, users are divided into a cultural adaptation layer, a motivation-driven layer, and a price-sensitive layer, and corresponding cultural region adaptation strategies, interest motivation-driven hierarchical preferential promotion strategies, and dynamic pricing strategies are generated according to the divided levels. The weighted score is used to determine the priority of the strategies.

[0117] B1. The cultural region adaptation strategy is to generate an adaptation marketing strategy based on the matching degree between the cultural feature vector V hcdt and the electronic product, and push the user sharing copywriting.

[0118] B2. The interest motivation-driven hierarchical preferential promotion strategy is to map the preferential promotion strategy based on the user motivation preference vector V mhn and adjust the discount range of the electronic product in combination with the weight of Maslow's hierarchy of needs, and then dynamically generate a hierarchical preferential marketing strategy and push it to the user.

[0119] B3. The dynamic pricing strategy is to establish a price elasticity model, calculate the user price sensitivity and demand elasticity, and then use reinforcement learning for dynamic price adjustment to obtain the maximum benefit.

[0120] A2. Multi-channel reach: Calculate the priority of the channels based on the user's historical reach feedback information on the electronic product, and use a time series model to optimize the reach time.

[0121] Furthermore, by analyzing the corrected user portraits, deeply understanding the characteristics of different users in terms of cultural regions, interest motives, etc. is an important basis for formulating a hierarchical marketing strategy. Among them, the cultural region adaptation strategy can adjust the marketing plan according to the cultural characteristics of different regions; the interest motive-driven hierarchical preferential promotion strategy gives different discounts and promotional activities based on users' interests and purchase motives; the dynamic pricing strategy can flexibly adjust the product price according to the market and user demands; at the same time, combined with multi-channel touch methods, communicate and interact with users in an all-round and multi-angle manner, not only improving the accuracy and effectiveness of marketing, but also being able to better attract target customers, enhance users' participation and purchase willingness, thereby increasing product sales and market share, and achieving more efficient marketing of electronic products.

[0122] S5. Marketing Effect Evaluation and Quantitative Feedback: Construct a multi-dimensional evaluation model, input the behavioral feedback data of users who purchase electronic products into the GAN evaluation network, and output the evaluation results of the marketing health coefficient. Optimize the marketing strategy by quantifying the behavioral feedback data according to the evaluation results. Among them, the behavioral feedback data includes click-through rate, conversion rate, and average order value; the marketing health coefficient includes matching degree, timeliness, and cultural compatibility.

[0123] Specifically, the construction steps of the multi-dimensional evaluation model are as follows:

[0124] According to the user portrait model and the personalized differential marketing strategy, define the evaluation index of the multi-dimensional evaluation model as the marketing health coefficient, including the matching degree between the user portrait and the recommended marketing strategy, the timeliness of the overlap between the marketing strategy touch time and the user demand window, and the cultural compatibility between the cultural preferences in the user portrait and the recommended marketing strategy.

[0125] Input the behavioral feedback data of users and the user portrait feature V into the GAN evaluation network user , the generator G of the GAN generates a feature vector simulating the healthy marketing result, and outputs the health coefficient through defining the loss function of the discriminator D. And the behavioral feedback data includes click-through rate, conversion rate, and average order value, which are respectively calibrated as ctr, cvr, and aov. Among them, the expression of the loss function of the discriminator D is and (X') = (ctr, cvr, aov, V user ), in the formula, represents the loss function of the discriminator D in the GAN, E(X') represents the expectation of the distribution of the real marketing feedback data X', D(X') represents the output probability of the discriminator for the real marketing feedback data X', and X′ represents the real marketing feedback data of the electronic product to the user, including click-through rate ctr, conversion rate cvr, average order value aov and user portrait V user ;

[0126] Construct the feature vector of the electronic product as Vecp , calculate the matching degree based on the similarity between the user portrait features and the features of the recommended electronic products. The calculation formula for the matching degree is Relevance = cos(V user , V ecp ). In the formula, Relevance represents the matching degree, and cos(,) represents the cosine similarity;

[0127] Quantify the effectiveness of the touch opportunity through the time decay function to calculate the timeliness. The calculation formula for the timeliness is In the formula, Timeliness represents the timeliness, χ represents the decay coefficient that controls the decay speed of the timeliness due to time deviation, and |t 触达 -t 需求 | represents the absolute value of the difference between the timestamp of the actual touch of the user and the midpoint timestamp of the user demand window t 触达 ; 需求 The absolute value of the difference between the timestamp of the actual touch of the user and the midpoint timestamp of the user demand window t

[0128] Evaluate the conflict degree between the marketing strategy and the user culture vector V hcdt to calculate the cultural compatibility. The calculation formula for the cultural compatibility is CulturalFit = 1 - ||V 策略文化 -V hcdt ||1. In the formula, CulturalFit represents the cultural compatibility, V 策略文化 represents the marketing strategy cultural label vector, and || ||1 represents the L1 norm, which measures the sum of the absolute values of the differences between the marketing strategy cultural label vector V 策略文化 and the user culture vector V hcdt .

[0129] Specifically, the steps to obtain the evaluation result are as follows:

[0130] Collect the real-time behavioral feedback data of the user, including click-through rate ctr, conversion rate cvr, and average order value aov;

[0131] Concatenate the behavioral feedback data with the user portrait feature vector to form an input vector, then (X') = (ctr, cvr, aov, V user );

[0132] Use the discriminator D in GAN to output a three-dimensional health coefficient. Among them, the expression for the discriminator D to output the health coefficient is D = (Relevance, Timeliness, CulturalFit);

[0133] And perform weighted aggregation on the matching degree Relevance, timeliness Timeliness, and cultural compatibility CulturalFit to calculate the comprehensive health score. Among them, the calculation formula for the comprehensive health score is score = ωR ·Relevance + ω T ·Timeliness + ω C ·Cultural Fit, and ω R +ω T +ω C = 1, where score represents the comprehensive health score, ω R 、ω T 、ω C respectively represent the weights corresponding to the matching degree Relevance, timeliness Timeliness, and cultural compatibility Cultural Fit;

[0134] Dynamically modify the priority of the hierarchical marketing strategy according to the health coefficient.

[0135] Furthermore, by constructing a multi-dimensional evaluation model, it covers behavioral feedback data such as click-through rate, conversion rate, and average order value, as well as marketing health coefficients such as matching degree, timeliness, and cultural compatibility. Input the behavioral feedback data of users who purchase electronic products into the GAN evaluation network, and use the powerful data analysis and evaluation capabilities of GAN to output the evaluation results of the marketing health coefficients. Then, based on these evaluation results, conduct quantitative analysis on the behavioral feedback data to clarify the problems and deficiencies in the marketing strategy, and then comprehensively and accurately evaluate the marketing effect, be able to timely discover the shortcomings in the marketing strategy, provide a strong basis for the precise optimization of the marketing strategy, make the marketing activities more in line with the market and user needs, improve the marketing efficiency and input-output ratio, and enhance the competitiveness of the enterprise in the electronic product market.

[0136] S6. Iteratively optimize the marketing strategy with the model: Use online learning technology to feedback the evaluation results to the user portrait model for iterative learning, and continuously optimize the personalized differential marketing strategy.

[0137] Furthermore, online learning technology can timely feedback the marketing effect evaluation results to the user portrait model, allowing the model to perform iterative learning. With the continuous input of new data, the user portrait model can be continuously updated and optimized, more accurately reflecting the latest characteristics and needs of users. And based on the updated user portrait model, continuously adjust and optimize the personalized differential marketing strategy to ensure that the marketing strategy always matches the actual situation of users, making the marketing strategy able to dynamically adapt to market changes and user needs, enhancing the flexibility and pertinence of the marketing activities, improving the accuracy and effect of marketing, thereby helping the enterprise maintain a competitive advantage in the electronic product market and achieve more efficient marketing goals.

[0138] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data for software simulation to get a formula that is closest to the actual situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0139] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wired or wireless (such as infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer, or a data storage device such as a server or data center that includes one or more collections of available media. The available medium can be a magnetic medium (such as a floppy disk, hard disk, magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0140] It should be understood that in various embodiments of the present application, the magnitudes of the serial numbers of the above processes do not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0141] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professionals can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0142] The above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present application, and all should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An electronic product personalized marketing method based on big data precise profiling, characterized in that, The steps are as follows: S1. Multi-source data collection and cross-domain fusion processing: Use web crawlers and streaming processing engines to obtain multi-source data sets related to electronic products from various platforms on the Internet, and perform cleaning, classification, alignment, and fusion processing on the multi-source data sets to generate multi-modal data. Then, update the multi-modal data in real time to dynamically reflect data changes, laying a foundation for subsequent data analysis; S2. Constructing a multi-dimensional user portrait: Combining the Hofstede cultural dimension model and Maslow's hierarchy of needs theory, conduct quantitative analysis of cultural regions and interest motives on the multi-modal data, generate dynamic and static tags, and dynamically update the tags through the distributed data stream engine Apache Flink to construct a user portrait model; ​ ​ ​ ​ 2. The personalized marketing method for electronic products based on big data precise profiling according to claim 1, wherein ​ ​ ​ ​ 3. The personalized marketing method for electronic products based on big data precise profiling according to claim 2, wherein ​ The obtained multi-source dataset is cleaned and classified according to user basic information, behavior data, regional characteristics, and social data, and the multi-source dataset obtained is and Bui represents the set of user basic information, Bhd represents the set of behavior data, Rcs represents the set of regional characteristics, Scd represents the set of social data, A x 、B y 、C i 、D j respectively represent subsets of the sets of user basic information, behavior data, regional characteristics, and social data; Based on the graph embedding model and setting a determination threshold, use the identity association algorithm to cross-domain align user identities from the user basic information and behavior data in the multi-source dataset. Among them, the expression of the identity association algorithm is In the formula, u1(A x ,B y ) represents the user basic information and behavior data extracted from the e-commerce platform, and u2(A x ,B y ) represents the user basic information and behavior data extracted from the social platform, represents the vector representation of the e-commerce platform user u1 in the graph embedding space, represents the vector representation of the social platform user u2 in the graph embedding space. Similarity represents the similarity of cross-platform user identity association, cos(,) represents the cosine similarity function, and its value range is [-1,1]. When the absolute value of Similarity is greater than the determination threshold, it is determined to be the same user, otherwise it is not; Use the TF-IDF method for text feature extraction to extract the social features of users. Among them, the extraction expression of social features is In the formula, TF-IDF(D3, Scd) represents the extraction of the social features of users by the TF-IDF method. TF(D3, Scd) represents the number of times the user's D3 emotional comment on electronic products appears in the current set of social data Scd. DF(D3) represents the total number of social data sets containing the D3 emotional comment. Scd' represents the total number of the social data set Scd; Extract the corresponding regional characteristics according to the user identity, and generate a regional sensitivity factor based on the regional characteristics. The calculation formula of the regional sensitivity factor is δ Rcs =∑ω i ·C i , and ∑ω i =1, i = 1, 2, 3. In the formula, δ Rcs represents the regional sensitivity factor, ∑ represents the summation symbol, ω i represents the weight coefficient of the corresponding subset in the regional characteristics set, and C i represents the subset of the regional characteristics set; Construct the initial multimodal data tensor for each user, labeled as τ0, based on user identity, feature data, and time dimension, and then use the variable streaming tensor fusion to update the multimodal data. The update expression for the multimodal data is τ1 = μτ0 + (1 - μ)Δτ, where μ ∈ (0, 1). In the formula, τ1 represents the updated multimodal data tensor, Δτ represents the real-time new data increment, and μ represents the attenuation factor.

4. The personalized marketing method for electronic products based on big data precise profiling according to claim 3, wherein The steps for the quantitative analysis of cultural regions by the Hofstede cultural dimension model are as follows: According to the C1 geographical location in the regional characteristics Rcs of the user, obtain the six cultural dimension scores of the corresponding region from the database of the Hofstede cultural dimension model, including power distance, individualism, masculinity, uncertainty avoidance, long-term orientation, indulgence and restraint, labeled as PDI, IDV, MAS, UAI, LTO, and IVR respectively; After normalizing the scores of the six cultural dimensions corresponding to the geographical location of the user, a cultural feature vector is constructed. The expression of the cultural feature vector is In the formula, V hcdt represents the normalized user cultural preference vector, and the dimension is 6×1. PDI, IDV, MAS, UAI, LTO, and IVR respectively represent the original scores of the six Hofstede cultural dimensions, and the range is 0-100 points; Generate cultural sensitivity factors through linear combination. Then, the cultural sensitivity factors serve as dynamic labels for cultural feature vectors to recommend marketing strategies for users to purchase electronic products. Among them, the calculation formula for the cultural sensitivity factor is ζ hcdt = ω hcdt T ·V hcdt , and ω hcdt ∈R 6 . In the formula, ζ hcdt represents the cultural sensitivity factor, ω hcdt represents the weight vector corresponding to the user's cultural preference vector V hcdt , and ω hcdt is determined by optimizing through A / B testing. T represents the transpose, and V hcdt represents the user's cultural preference vector.

5. The personalized marketing method for electronic products based on big data precise profiling according to claim 4, characterized in that, The steps for the quantitative analysis of interest motivation by the Maslow's hierarchy of needs theory are as follows: Map the user's behavior data set Bhd and social data set Scd to the five levels of needs in Maslow's hierarchy, including physiological needs, safety needs, social needs, esteem needs, and self-actualization needs; Analyze the emotional intensity corresponding to the five levels of Maslow's needs from the user behavior using the BERT model. Then the emotional intensity is used as the dynamic label of the user motivation preference vector. Among them, the calculation formula of the emotional intensity is In the formula,[[]]END]] represents the emotional intensity, and the value range is mapped to the interval [0, 1]. l represents the five levels of Maslow's needs. BERT(Bhd l , Scd l ) represents the analysis of the emotional intensity of the behavior data set Bhd and the social data set Scd on the five levels of Maslow's needs using the BERT model. σ represents the Sigmoid activation function. ω l represents the weight matrix corresponding to the five levels of Maslow's needs. T represents the transpose.[[]]END]] represents the hidden vector of the behavior data set Bhd and the social data set Scd output by the BERT model on the five levels of Maslow's needs. b l represents the bias term corresponding to the five levels of Maslow's needs. Combine the emotional intensity with the user behavior frequency to calculate the weight of each need level in the five levels of Maslow's needs. Then the calculation formula of the weight of the five levels of Maslow's needs is In the formula, ω mhn，l represents the weight of the l-th level of Maslow's needs.[[]]END]] represents the number of times the behavior of the user's behavior data set Bhd and social data set Scd appears in the l'-th level of the five levels of Maslow's needs.[[]]END]] represents the total number of times the behavior of the user's behavior data set Bhd and social data set Scd appears in the five levels of Maslow's needs.[[]]END]] Using the weights ω of each hierarchy of needs mhn，l Construct a user motivation preference vector, where the expression of the user motivation preference vector is In the formula, V mhn represents the user motivation preference vector, and its dimension is 5×1, ω mhn，1 , ω mhn，2 , ω mhn，3 , ω mhn，4 , ω mhn，5 respectively represent the weights of Maslow's five-level needs.

6. The personalized marketing method for electronic products based on big data precise profiling according to claim 5, characterized in that The steps for the distributed data stream engine to dynamically update tags are as follows: Access the real-time event stream of the user's behavior data set Bhd through the distributed data stream engine ApacheFlink; Set a sliding window, and calculate the short-term interest of users based on the sliding window and calculate the interest score. The calculation formula of the interest score is In the formula, is t represents the interest score of the user at the current time point t, represents the behavior weight of the sliding window user at the time point t - t L , t represents the current time point, t L represents the window length, and t L = 600 seconds, and γ represents the decay coefficient; Extract the regional sensitivity factor δ Rcs , the cultural sensitivity factor ζ hcdt and the emotional intensity of the initial tags, and continuously update the user tags using the state backend of the distributed data flow engine. Among them, the update calculation formula of the user tags is In the formula represents the new tag updated from the initial tag corresponding to the regional sensitivity factor δ Rcs , the cultural sensitivity factor ζ hcdt and the emotional intensity , α represents the smoothing factor, and 0 < α < 1, is t represents the interest score of the user; The steps for constructing the user portrait model are as follows: First, generate a static feature vector for the set Bui of user basic information and mark it as V Bui , and the dimension is 4×1; The cultural feature vector V hcdt , the user motivation preference vector V hcdt and the static label V Bui are comprehensively spliced into the user portrait feature vector, and the expression of the user portrait feature vector is In the formula, V user represents the user portrait feature vector, and the dimension is 15×1, represents the vector splicing operation; Use the Wide&Deep model to jointly train the user portrait; And output the user label probability and recommendation priority according to the user portrait model. The calculation formula for the output user label probability is P(K|V user ) = σ(ω Wide&Deep T ·V user + b Wide&Dee ). In the formula, P(K|V user ) represents the output user label probability, P(|) represents the conditional probability, K represents the user target label for purchasing electronic products, σ represents the Sigmoid activation function, ω Wide&Deep represents the weight vector of the Wide&Deep model, T represents the transpose, and b Wide&Dee represents the bias term of the Wide&Deep model.

7. The personalized marketing method for electronic products based on big data precise profiling according to claim 6, wherein The steps for the GAN to detect abnormal features are as follows: Using the Generative Adversarial Network (GAN), the generator G and the discriminator D are used to perform anomaly detection on the user portrait feature vector V user That is, the user portrait feature vector V user is input into the generator G to generate a reconstructed portrait G(V user ), and the discriminator D is used to distinguish between the user portrait feature vector V user and the reconstructed portrait G(V user ), and the anomaly score is detected; By calculating the reconstruction error as an anomaly index, where the calculation formula of the reconstruction error is In the formula, e(V user ) represents the reconstruction error between the user portrait feature vector V user and the reconstructed portrait G(V user ), which is expressed as the square of the L2 norm; Set a dynamic threshold and conduct a comparative analysis with the abnormal indicators to determine abnormal features. Among them, the dynamic threshold is set to θ = 90%, and when e(V user ) > θ, it is judged as an abnormal feature of the user profile, otherwise it is not; Train the GAN according to anomaly detection by defining the loss functions of the generator G and the discriminator D. Among them, the expression of the loss function of the generator G is and D(G(V user )) → 1, which is used to minimize the reconstruction error and deceive the discriminator D at the same time. In the formula, is expressed as the loss function of the generator G, and E(V user , p) is expressed as the expectation of the normal probability of the data distribution of the normal user portrait feature vector V user , μ represents the trade-off coefficient for balancing the reconstruction error and the adversarial loss, and D(G(V user )) is expressed as minimizing the discrimination probability of the reconstructed portrait G(V user ), and p represents the normal probability of the data distribution of the normal user portrait feature vector V user ; The expression of the loss function of the discriminator D is In the formula,[[]] represents the loss function of the discriminator D, and D(V user ) represents maximizing the discrimination probability for the feature vector V of the normal user profile user ; The steps for establishing the dynamic interest attenuation model are as follows: Adjust the weight of the user's historical behavior by using an exponential decay method for time. Among them, the expression for adjusting the weight by time decay is and Δt = t - t0. In the formula, ω(t) represents the weight dynamically adjusted with time decay, t represents the current time point, t0 represents the historical time point when the user initially performed an action on the electronic product, and Δt represents the time difference represents the decay rate that controls the user's interest Dynamically update the interest score by aggregating the time-decayed behavior impact through the weighted time difference. Among them, the update calculation formula for the interest score is UI t = ω(t)·is t , where UI t represents the interest score updated at time point t; Then, the Fourier transform is used to extract the periodic interest patterns to detect and identify the periodic changes in the user interest fluctuations. Among them, the expression of the periodic interest pattern is In the formula, F(f) represents the periodic intensity corresponding to the frequency f, and T0 represents the total length of the time window, represents the imaginary unit, and f represents the frequency index, and the range is f = 0, 1, 2, …, T0 - 1, and the period is 8. The personalized marketing method for electronic products based on big data precise profiling according to claim 7, characterized in that, The steps for constructing the multi-dimensional evaluation model are as follows: According to the user portrait model and the personalized differential marketing strategy, define the evaluation index of the multi-dimensional evaluation model as the marketing health coefficient, including the matching degree between the user portrait and the recommended marketing strategy, the timeliness of the overlap between the marketing strategy reach time and the user demand window, and the cultural compatibility between the cultural preferences in the user portrait and the recommended marketing strategy; Input the user's behavioral feedback data and user portrait features V into the GAN evaluation network user , the generator G of the GAN generates a feature vector simulating the healthy marketing result, and outputs a health coefficient through the defined discriminator D loss function. The behavioral feedback data includes click-through rate, conversion rate, and average order value, which are respectively labeled as ctr, cvr, and aov. Among them, the expression of the discriminator D loss function is and (X') = (ctr, cvr, aov, V user ), where represents the discriminator D loss function in the GAN, E(X') represents the expectation of the distribution of the real marketing feedback data X', D(X') represents the output probability of the discriminator for the real marketing feedback data X', and X′ represents the real marketing feedback data of the electronic product to the user, including click-through rate ctr, conversion rate cvr, average order value aov, and user portrait V user ; Construct the feature vector of the electronic product as V ecp , calculate the matching degree based on the similarity between the user profile features and the features of the recommended electronic product. The calculation formula for the matching degree is Relevance = cos(V user , V ecp ). In the formula, Relevance represents the matching degree, and cos(,) represents the cosine similarity; Quantify the timeliness by calculating the effectiveness of the reach time through a time decay function. The formula for timeliness is In the formula, Timeliness represents timeliness, χ represents the decay coefficient that controls the decay rate of timeliness due to time deviation, |t 触达 -t 需求 | represents the absolute value of the difference between the t 触达 timestamp of the actual reached user and the midpoint timestamp of the user demand window t 需求 ; Evaluate the conflict degree between the marketing strategy and the user cultural vector V hcdt to calculate the cultural compatibility. The formula for cultural compatibility is CulturalFit = 1 - ||V 策略文化 - V hcdt ||1, where CulturalFit represents cultural compatibility, V 策略文化 represents the marketing strategy cultural label vector, and || ||1 represents the L1 norm, which measures the sum of the absolute values of the differences between the marketing strategy cultural label vector V 策略文化 and the user cultural vector V hcdt .

9. A computer device, comprising: A memory and a processor; The memory stores a computer program, characterized in that: when the processor executes the computer program, it implements the steps of the personalized marketing method for electronic products based on big data accurate portrait according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the personalized marketing method for electronic products based on big data accurate portrait according to any one of claims 1 to 8.

Images