A targeted marketing method for private domain traffic platform based on user purchase frequency analysis

Abstract

The present invention discloses a method for targeted marketing of a private domain traffic platform based on user purchase frequency analysis, including S1, a formed user behavior data set; S2, a formed user comprehensive data set; S3, a dynamic behavior feature data set; S4, a GNN-Transformer hybrid model is used to train a user behavior prediction model based on a dynamic behavior feature data set; S5, a multi-level user portrait is constructed in combination with the output result of the user behavior prediction model; S6, a marketing strategy is dynamically optimized based on the prediction results of the user portrait and the user behavior prediction model; S7, a marketing content that meets user needs is automatically generated in combination with the user portrait and the current context; S8, a closed-loop feedback system is established. The present invention realizes the refined management and dynamic optimization of marketing strategies through in-depth analysis and application of user purchase frequency and behavior data, and significantly improves the accuracy and effect of marketing activities.

Description

Technical Field

[0001] The present invention relates to the technical field of private domain traffic platforms, and in particular to a private domain traffic platform targeted marketing method based on user purchase frequency analysis. Background Art

[0002] With the rapid development of e-commerce and social media, the behavioral data of users on various platforms has become increasingly rich and complex. These data include not only the user's purchase history, but also multi-dimensional information such as the user's browsing history, interactive behavior, social sharing, etc. Traditional marketing strategies mainly rely on users' static attributes and historical behavior data. This method is difficult to capture the dynamic changes in user behavior, resulting in difficulty in improving marketing accuracy and user satisfaction. Therefore, how to effectively analyze and utilize these rich user behavior data to build a more refined and personalized marketing strategy has become an urgent problem to be solved by the current private domain traffic platform.

[0003] The existing technology has made some progress in user behavior analysis and targeted marketing, but there are still many shortcomings. First, traditional targeted marketing methods mainly rely on static user attributes and historical behavior data, and lack real-time capture and analysis of dynamic changes in user behavior. This method is difficult to provide accurate marketing suggestions when user behavior changes rapidly, resulting in poor marketing results. Secondly, most of the existing user behavior prediction models are based on simple statistical analysis or shallow machine learning algorithms, which cannot fully utilize the complex relationships and potential patterns in multi-dimensional user data, resulting in low prediction accuracy and reliability. In addition, existing marketing strategy optimization methods usually adopt preset rules or simple feedback mechanisms, lack intelligence and dynamic adjustment capabilities, and cannot optimize marketing strategies in a timely manner according to user behavior changes and feedback data. Therefore, how to provide a targeted marketing method for a private domain traffic platform based on user purchase frequency analysis is a problem that technicians in this field urgently need to solve. Summary of the invention

[0004] One object of the present invention is to propose a targeted marketing method for a private domain traffic platform based on user purchase frequency analysis. The present invention realizes the refined management and dynamic optimization of marketing strategies by in-depth analysis and application of user purchase frequency and behavior data, which not only significantly improves the accuracy and effectiveness of marketing activities, but also improves the user's shopping experience and satisfaction, bringing substantial competitive advantages and commercial value to the private domain traffic platform.

[0005] A private domain traffic platform targeted marketing method based on user purchase frequency analysis according to an embodiment of the present invention comprises the following steps:

[0006] S1. On the private domain traffic platform, through embedded user behavior tracking technology, collect multi-dimensional data of users in the private domain traffic platform to form a user behavior data set;

[0007] S2, integrating the collected user behavior data set with third-party data sources to form a comprehensive user data set using data enhancement technology;

[0008] S3. Based on the improved time series pattern mining algorithm, the short-term behavior fluctuation characteristics, medium-term behavior trend characteristics and long-term behavior pattern characteristics of the user's purchase frequency are dynamically extracted from the user's comprehensive data set to form a dynamic behavior feature data set;

[0009] S4. Build a user behavior prediction model and use the GNN-Transformer hybrid model to train the user behavior prediction model based on the dynamic behavior feature dataset;

[0010] S5. Combine the output results of the user behavior prediction model to build a multi-level user portrait, including static portrait, dynamic portrait and situational portrait;

[0011] S6. Dynamically optimize marketing strategies based on the prediction results of user portraits and user behavior prediction models;

[0012] S7. Combine user portraits and current situations to automatically generate marketing content that meets user needs, and push it to target users through the multi-channel reach mechanism of the private domain traffic platform;

[0013] S8. Establish a closed-loop feedback system to monitor and evaluate the implementation effect of marketing strategies in real time, dynamically adjust and optimize marketing strategies based on user interaction data and feedback information, and optimize marketing effects and user satisfaction.

[0014] Optionally, the S1 includes the following steps:

[0015] S11. Embed data collection modules in the front-end and back-end of the private domain traffic platform, collect various types of user behavior data in the private domain traffic platform in real time through API interface and script technology, and build a user behavior data collection matrix M u,t :

[0016]

[0017] Among them, M u,t represents the behavior data collection matrix for user u at time t, m ij Represents user u i The j-th behavior data at time t;

[0018] S12. Collect the user's purchase frequency data by monitoring the user's shopping cart, order history and payment record, and record the timestamp T of each purchase event.i and product IDP i ;

[0019] S13. Collect the user's browsing history data by tracking the user's browsing path, page dwell time, and click behavior in the private domain traffic platform, and record the timestamp T of each page browsing event. j and page IDPG j ;

[0020] S14. Interaction behavior data collection: Collect user interaction behavior data by monitoring user interaction behaviors such as likes, comments, shares, and favorites on the private domain traffic platform, and record the timestamp T of each interaction event. k and interaction type I k ;

[0021] S15. Collect the user's social sharing data by monitoring the user's social sharing activities in the private domain traffic platform, including sharing content, sharing channels and sharing times, and record the timestamp T of each sharing event. l and sharing channels l ;

[0022] S16. Integrate the purchase frequency data, browsing record data, interactive behavior data and social sharing data to form a complete user behavior data set D:

[0023] D={(T i , P i ), (T j , P.G. j ), (T k , I k ), (T l , S l )}.

[0024] Optionally, S2 includes the following steps:

[0025] S21, preliminarily clean and filter the collected user behavior data set D, remove noise data and duplicate data, and perform missing value filling and data standardization processing;

[0026]

[0027] Where D represents the initial data set, D noise represents the noisy data set, D clean Represents the cleaned user behavior dataset;

[0028]

[0029] Among them, μ is the mean, σ is the standard deviation, and D normIt is a standardized user behavior dataset;

[0030] S22. Obtain relevant user behavior data from social media platforms and e-commerce platforms through API interfaces or data exchange protocols, including social interaction data, purchase history data, and user evaluation data:

[0031] D 3rd ={d|d∈API1∪API2∪…∪API n};

[0032] Among them, D 3rd Represents third-party datasets, API i represents the i-th third-party data source;

[0033] S23, the preprocessed user behavior dataset D norm Align and match data with third-party data sources using a unique user identifier U i To associate and match users from different data sources:

[0034] M aligned ={(u i , t j , d k )|u i ∈U,t j ∈T,d k ∈D 3rd ∪D norm};

[0035] Among them, M aligned represents the aligned dataset, U represents the user set, and T represents the time set;

[0036] S24, using the improved Bayesian inference data fusion technology, the preprocessed user behavior dataset D norm Integrate data from third-party data sources to form a comprehensive user data set D u , the improved Bayesian inference data fusion technology introduces the temporal dependency and purchase frequency characteristics of user behavior to enhance the accuracy and reliability of data fusion;

[0037] D u ={f(M aligned )|f:M aligned →D u};

[0038] Among them, the improved Bayesian inference data fusion function f is expressed as:

[0039] f(M aligned )=argmax θ P(M aligned|θ)P(θ);

[0040] Among them, θ represents the model parameters, P(M aligned |θ) represents the aligned dataset M under the given model parameter θ aligned The likelihood of , P(θ) is the prior probability of the model parameter θ, argmax θ It means to find the posterior probability P(M aligned |θ) the largest parameter θ;

[0041] Calculate the fused user data set by maximizing the product of conditional probability and prior probability;

[0042] S25, the user comprehensive data set D u Stored in the user data platform, establish a complete user data archive.

[0043] Optionally, S3 includes the following steps:

[0044] S31, the user comprehensive data set D u The user purchase frequency data in is segmented by time window to form a segmented data set D seg , where each time window is of length Δt:

[0045] D seg ={d|d∈D u , time period t i ∈[t0+kΔt,t0+(k+1)Δt)};

[0046] Where k represents the index of the time window, t0 is the starting time, and t i is the time point after segmentation, and d represents the user’s purchase frequency data;

[0047] S32. Analyze the user purchase frequency data in each time window to extract the user's behavior fluctuation characteristics in the short term F short ;

[0048] S33, performing trend analysis on the user purchase frequency data in multiple continuous time windows to extract the user's mid-term behavior trend characteristics F mid ;

[0049] S34, perform pattern recognition on the user purchase frequency data in the entire time period to extract the user's long-term behavior pattern features F long ;

[0050] S35. When extracting short-term behavior trend features, medium-term behavior trend features, and long-term behavior trend features, an improved temporal pattern mining algorithm based on topic drive is used. The improved temporal pattern mining algorithm integrates the temporal dependency of user behavior and topic information to form topic behavior features F topic :

[0051] F topic ={f tp |f tp = Theme Relationship (F short , F mid , F long , T)};

[0052] Among them, f tp represents the behavioral characteristics after combining the topic information, and T represents the topic set;

[0053] S36, short-term behavior fluctuation characteristics F short 、Medium-term behavioral trend characteristics F mid 、Long-term behavior pattern characteristics F long and the subject behavior characteristics F topic Integrate to form a dynamic behavior feature dataset F dyn :

[0054] F dyn ={F short , F mid , F long , F topic}.

[0055] Optionally, the improved time series pattern mining algorithm introduces time-dependent modeling of user behavior and adopts a time series variational autoencoder method:

[0056]

[0057] Among them, d t is the purchase frequency data at time t, μ and σ are the mean and standard deviation respectively;

[0058] Combined with the topic information, the LDA topic model is used for association analysis:

[0059]

[0060] Among them, α i represents the weight of the i-th topic, t i is the i-th topic, the relevance (f tp , t i ) represents the relevance between the feature and the topic;

[0061] Based on the short-term behavior trend characteristics, medium-term behavior trend characteristics, and long-term behavior trend characteristics, the final topic behavior characteristics are obtained by maximizing the joint probability of the characteristics under all conditions:

[0062] f tp = argmax f P(f|F short , F mid , F long , T).

[0063] Optionally, S4 includes the following steps:

[0064] S41, dynamic behavior feature data set F dyn Converted into graph structure data G, where nodes represent users and products, and edges represent interaction behaviors;

[0065] S42. Build a user behavior prediction model, add a graph neural network layer, and use the adjacency matrix A and the node feature matrix H 0 Perform graph feature extraction and generate graph feature vector H (l+1) :

[0066]

[0067] Among them, H (l) represents the node feature matrix of the lth layer, W (l) is the weight matrix, σ is the activation function, A ij is an element in the adjacency matrix, c ij is the normalized coefficient between nodes i and j, N(i) represents the set of neighbor nodes of node i;

[0068] S43, embedding time information into the graph feature vector to form a time-enhanced feature matrix. The time information enables the user behavior prediction model to capture the time dependency of user behavior and the changing trend of purchase frequency:

[0069]

[0070] in, represents the feature after embedding time information, α k and β k is the time embedding parameter, t ij is the time interval between nodes i and j, and T is the time period;

[0071] S44. Use the graph attention network to apply the attention mechanism on the time-enhanced feature matrix, dynamically adjust the weights between nodes, and generate the attention feature vector

[0072]

[0073] in, represents the attention score between nodes i and j, a is the parameter vector of the attention mechanism, || represents the vector connection operation, is the attention weight of node j to node i, σ is the activation function;

[0074] S45. Input the attention feature vector output by the graph attention network layer into the Transformer layer, process the feature vector through the self-attention mechanism, capture long-distance dependencies, and generate a global feature vector Z:

[0075] Q=W Q H (l+1) , K=W K H (l+1) , V=W V H (l+1) ;

[0076]

[0077] Z=Transformer(H (l+1) );

[0078] Among them, Q is the query matrix, K is the key matrix, V is the value matrix, and W Q , W K and W V is the corresponding weight matrix, d k is the dimension of the key vector. The Transformer layer processes long-distance dependencies through the self-attention mechanism to improve the global perception ability of the user behavior prediction model.

[0079] S46. Using dynamic behavior feature dataset F dyn The user behavior prediction model is trained with the training set in , and the optimization objective function is to minimize the prediction error. The Adam optimizer is used to update the parameters:

[0080]

[0081] Where N represents the number of training samples. is the predicted purchase frequency, y i is the actual purchase frequency, is the predicted potential demand, d i is the actual potential demand, λ is the regularization coefficient, θ j Predict model parameters for user behavior;

[0082] S47. Use the validation set to validate the trained user behavior prediction model, calculate the validation error and evaluate the model performance. The validation indicators include mean square error and mean absolute error.

[0083] Optionally, S5 includes the following steps:

[0084] S51. Based on the output of the user behavior prediction model, collect and organize the user's demographic information, including age, gender, occupation and region to form a static profile of the user P static ;

[0085] S52: Generate a dynamic profile of the user P based on the prediction results of the user behavior prediction model, including the user's behavior preference and purchase frequency prediction. dynamic ,Behavioral preferences include users’ interest in different categories of products, and purchase frequency prediction is the user’s future purchase frequency predicted by the user behavior prediction model;

[0086] S53. Combine the user's purchasing motivation and consumption situation to analyze the user's consumption behavior in different situations and form a situational portrait P context ,Purchase motivation reflects the reasons why users buy a certain type of product, and consumption context describes the consumption behavior pattern of users at a specific time and place;

[0087] S54, combining the output results of the user behavior prediction model, the static portrait P static , Dynamic Portrait P dynamic and situational portrait P context Integrate and generate multi-level user portraits P multi :

[0088] P multi = {P static , P dynamic , P context};

[0089] User behavior prediction model M predict The output results are used to guide and optimize the construction process of static portraits, dynamic portraits and situational portraits, forming a multi-level user portrait P multi .

[0090] Optionally, the S6 includes the following steps:

[0091] S61, multi-level user portrait P multi and user behavior prediction model M predict The prediction results are used as input data to generate an initial marketing strategy plan including preliminary product recommendations, exclusive offers and personalized content push strategies;

[0092] S62. Generate a personalized product recommendation list R based on user portraits and prediction results using collaborative filtering algorithm product :

[0093]

[0094] Among them, p represents the product, u represents the current user, and u i represents users similar to the current user, w i represents the weight, similarity(u,u i ) represents the similarity between users, rating(u i , p) represents user u i Rating of product p;

[0095] S63. Generate personalized exclusive discount strategy based on user portrait and purchase frequency prediction results discount , including discounts, coupons and limited-time promotions:

[0096] S discount ={discount(p,u)|p∈R product};

[0097] Where discount(p,u) represents the exclusive discount for product p provided to user u;

[0098] S64. Generate personalized marketing content using natural language processing technology C persinalized , combined with user portraits and current situations, and pushed to target users through a multi-channel reach mechanism:

[0099] C personalized =NLP(P multi , context);

[0100] Among them, context refers to the current situation, and NLP technology is used to generate marketing content that meets user needs;

[0101] S65. By real-time monitoring of user feedback on marketing strategies, the marketing strategies are dynamically adjusted and optimized using reinforcement learning algorithms. The optimization objective function is to maximize user engagement and conversion rate:

[0102]

[0103] Among them, π represents the strategy, γ is the discount factor, and R t is the reward at time t, and the optimization goal is to maximize the cumulative reward under strategy π;

[0104] S66. Regularly evaluate optimized marketing strategies, use key performance indicators including click-through rate, conversion rate and user lifetime value to assess marketing effectiveness and make corresponding adjustments.

[0105] The beneficial effects of the present invention are:

[0106] (1) The present invention uses embedded user behavior tracking technology to comprehensively collect multi-dimensional behavior data of users on the private domain traffic platform, integrates it with third-party data sources, and applies data enhancement technology to greatly enhance the comprehensiveness and dimensionality of the data. In addition, through improved time series pattern mining algorithms and deep learning technology, user behavior features are dynamically extracted and predicted, significantly improving the precision of marketing strategies and the accuracy of predictions.

[0107] (2) The present invention combines the user's static information, behavioral preferences, purchase motivations and other multi-dimensional data to construct a multi-level user portrait including static portrait, dynamic portrait and situational portrait. This comprehensive user portrait makes the marketing strategy more personalized and refined, and better meets the user's individual needs.

[0108] (3) The present invention uses real-time monitoring of user feedback data on marketing strategies and combines it with reinforcement learning algorithms to dynamically adjust marketing strategies. This not only enables marketing strategies to quickly respond to changes in the market and user behavior, but also enables continuous optimization, thereby maximizing user engagement and conversion rates.

[0109] (4) The present invention significantly improves the user's shopping experience and satisfaction through personalized product recommendations, exclusive discounts and customized content push. This targeted and personalized interactive method not only enhances the user's willingness to buy, but also helps to improve user stickiness and enhance user loyalty to the platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0110] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0111] Figure 1 This is a flow chart of a private domain traffic platform targeted marketing method based on user purchase frequency analysis proposed by the present invention;

[0112] Figure 2 This is a detailed flow chart of the user behavior prediction model construction and training process in a private domain traffic platform targeted marketing method based on user purchase frequency analysis proposed in the present invention. DETAILED DESCRIPTION

[0113] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0114] refer to Figure 1 , a private domain traffic platform targeted marketing method based on user purchase frequency analysis, comprising the following steps:

[0115] S1. On the private domain traffic platform, through embedded user behavior tracking technology, collect multi-dimensional data of users in the private domain traffic platform to form a user behavior data set;

[0116] In this implementation, S1 includes the following steps:

[0117] S11. Embed data collection modules in the front-end and back-end of the private domain traffic platform, collect various types of user behavior data in the private domain traffic platform in real time through API interface and script technology, and build a user behavior data collection matrix M u,t :

[0118]

[0119] Among them, M u,t represents the behavior data collection matrix for user u at time t, m ij Represents user u i The j-th behavior data at time t;

[0120] S12. Collect the user's purchase frequency data by monitoring the user's shopping cart, order history and payment record, and record the timestamp T of each purchase event. i and product IDP i , build a purchase frequency data collection function:

[0121]

[0122] Among them, F purchase (T i , P i ) indicates that at time T i The collected products P i The purchase frequency data, δ is the Dirac function, which is used to represent the occurrence of purchase events, T i Indicates the timestamp of the purchase event, P i Indicates the ID of the purchased product;

[0123] S13. Collect the user's browsing history data by tracking the user's browsing path, page dwell time, and click behavior in the private domain traffic platform, and record the timestamp T of each page browsing event. j and Page IDPC j , construct browsing record data vector:

[0124]

[0125] Among them, B user Represents the user's browsing history data vector, w j is the weight, indicating the importance of the page, f(PG j ) is the user browsing page PGj The frequency function of is the time decay factor, which is used to represent the decreasing importance of data over time;

[0126] S14. Interaction behavior data collection: Collect user interaction behavior data by monitoring user interaction behaviors such as likes, comments, shares, and favorites on the private domain traffic platform, and record the timestamp T of each interaction event. k and interaction type I k , build an interactive behavior scoring model:

[0127]

[0128] Among them, S interact (u) represents the interactive behavior score of user u, α k is the weight, indicating the influence of different interactive behaviors, I k is the number or intensity of the kth interaction behavior, is the time decay factor, β is the time decay coefficient;

[0129] S15. Collect the user's social sharing data by monitoring the user's social sharing activities in the private domain traffic platform, including sharing content, sharing channels and sharing times, and record the timestamp T of each sharing event. l and sharing channels l , build a social sharing influence model:

[0130]

[0131] Among them, I share (u) represents the social sharing influence of user u, γ l is the weight, indicating the influence of different social sharing channels, S l is the weight of the first social sharing event, log(1+n l ) is the logarithmic transformation of the number of shares, which is used to smooth high-frequency data. is the time decay factor;

[0132] S16. Integrate the purchase frequency data, browsing record data, interactive behavior data and social sharing data to form a complete user behavior data set D:

[0133] D={(T i , P i ), (T j , P.G. j ), (T k , I k ), (T l , S l )}.

[0134] This implementation method uses embedded user behavior tracking technology on the private domain traffic platform to collect multi-dimensional user behavior data in real time, including purchase frequency, browsing history, interactive behavior and social sharing. Compared with traditional static data collection methods, it can capture the dynamic changes of user behavior more comprehensively and in real time, providing more accurate data support for subsequent data analysis and user portrait construction.

[0135] S2, integrating the collected user behavior data set with third-party data sources to form a comprehensive user data set using data enhancement technology;

[0136] In this implementation, S2 includes the following steps:

[0137] S21, preliminarily clean and filter the collected user behavior data set D, remove noise data and duplicate data, and perform missing value filling and data standardization processing;

[0138]

[0139] Where D represents the initial data set, D noise represents the noisy data set, D clean Represents the cleaned user behavior dataset;

[0140]

[0141] Among them, μ is the mean, σ is the standard deviation, and D norm It is a standardized user behavior dataset;

[0142] S22. Obtain relevant user behavior data from social media platforms and e-commerce platforms through API interfaces or data exchange protocols, including social interaction data, purchase history data, and user evaluation data:

[0143] D 3rd ={d|d∈API1∪API2∪…∪API n};

[0144] Among them, D 3rd Represents third-party datasets, API i represents the i-th third-party data source;

[0145] S23, the preprocessed user behavior dataset D norm Align and match data with third-party data sources using a unique user identifier U i To associate and match users from different data sources:

[0146] M aligned ={(u i , t j, d k )|u i ∈U,t j ∈T,d k ∈D 3rd ∪D norm};

[0147] Among them, M aligned represents the aligned dataset, U represents the user set, and T represents the time set;

[0148] S24, using the improved Bayesian inference data fusion technology, the preprocessed user behavior dataset D norm Integrate data from third-party data sources to form a comprehensive user data set D u ,The improved Bayesian inference data fusion technology introduces the temporal dependency and purchase frequency characteristics of user behavior, enhancing the accuracy and reliability of data fusion;

[0149] D u ={f(M aligned )|f:M aligned →D u};

[0150] Among them, the improved Bayesian inference data fusion function f is expressed as:

[0151] f(M aligned )=argmax θ P(M aligned |θ)P(θ);

[0152] Among them, θ represents the model parameters, P(M aligned |θ) represents the aligned dataset M under the condition of given model parameter θ aligned The likelihood of , P(θ) is the prior probability of the model parameter θ, argmax θ It means to find the posterior probability P(M aligned |θ) the largest parameter θ;

[0153] Calculate the fused user data set by maximizing the product of conditional probability and prior probability;

[0154] S25, the user comprehensive data set D u Stored in the user data platform, establish a complete user data archive.

[0155] This implementation method integrates the collected user behavior data set with the third-party data source and uses data enhancement technology to form a more comprehensive user data set. This process not only enhances the integrity of the data, but also improves the diversity of the data and the accuracy of the analysis by integrating data from different sources, so that marketing strategies can be better customized and personalized.

[0156] S3. Based on the improved time series pattern mining algorithm, the short-term behavior fluctuation characteristics, medium-term behavior trend characteristics and long-term behavior pattern characteristics of the user's purchase frequency are dynamically extracted from the user's comprehensive data set to form a dynamic behavior feature data set;

[0157] In this implementation, S3 includes the following steps:

[0158] S31, the user comprehensive data set D u The user purchase frequency data in is segmented by time window to form a segmented data set D seg , where each time window is of length Δt:

[0159] D seg ={d|d∈D u , time period t i ∈[t0+kΔt,t0+(k+1)Δt)};

[0160] Where k represents the index of the time window, t0 is the starting time, and t i is the time point after segmentation, and d represents the user’s purchase frequency data;

[0161] S32. Analyze the user purchase frequency data in each time window to extract the user's behavior fluctuation characteristics in the short term F short , the short-term behavior fluctuation feature extraction formula is:

[0162]

[0163] Among them, f st Indicates the short-term behavior fluctuation characteristics, d i represents the purchase frequency data within the time window, and Δt represents the length of the time window;

[0164] S33, performing trend analysis on the user purchase frequency data in multiple continuous time windows to extract the user's mid-term behavior trend characteristics F mid , the mid-term behavior trend feature extraction formula is:

[0165] F mid ={f mt |f mt =Trend analysis (D seg )};

[0166] Among them, f mt Indicates the mid-term behavioral trend characteristics, D seg Represents a segmented dataset;

[0167] S34, perform pattern recognition on the user purchase frequency data in the entire time period to extract the user's long-term behavior pattern features F long, the long-term behavior pattern feature extraction formula is:

[0168] F long ={f lt |f lt = Pattern recognition (D u )};

[0169] Among them, f lt Indicates the long-term behavior pattern characteristics, D u Represents a comprehensive user data set;

[0170] S35. When extracting short-term behavior trend features, medium-term behavior trend features, and long-term behavior trend features, an improved temporal pattern mining algorithm based on topic drive is used. The improved temporal pattern mining algorithm integrates the temporal dependency of user behavior and topic information to form topic behavior features F topic :

[0171] F topic ={f tp |f tp = Theme Relationship (F short , F mid , F long , T)};

[0172] Among them, f tp represents the behavioral characteristics after combining the topic information, and T represents the topic set;

[0173] S36, short-term behavior fluctuation characteristics F short 、Medium-term behavioral trend characteristics F mid 、Long-term behavior pattern characteristics F long and the subject behavior characteristics F topic Integrate to form a dynamic behavior feature dataset F dyn :

[0174] F dyn ={F short , F mid , F long , F topic}.

[0175] In this implementation, the improved time series pattern mining algorithm introduces the time dependency modeling of user behavior and adopts the time series variational autoencoder method:

[0176]

[0177] Among them, d t is the purchase frequency data at time t, μ and σ are the mean and standard deviation respectively;

[0178] Combined with the topic information, the LDA topic model is used for association analysis:

[0179]

[0180] Among them, α i represents the weight of the i-th topic, t i is the i-th topic, the relevance (f tp , t i ) represents the relevance between the feature and the topic;

[0181] Based on the short-term behavior trend characteristics, medium-term behavior trend characteristics, and long-term behavior trend characteristics, the final topic behavior characteristics are obtained by maximizing the joint probability of the characteristics under all conditions:

[0182] f tp = argmax f P(f|F short , F mid , F long , T).

[0183] This implementation method dynamically extracts the behavioral characteristics of purchase frequency from the user data set through an improved time series pattern mining algorithm, which not only handles short-term fluctuations, but also analyzes medium- and long-term trends and patterns, revealing the deep-level trends and cyclical patterns of user behavior.

[0184] S4. Build a user behavior prediction model and use the GNN-Transformer hybrid model to train the user behavior prediction model based on the dynamic behavior feature dataset;

[0185] In this implementation, S4 includes the following steps:

[0186] S41, converting the dynamic behavior feature dataset Fdyn into graph structure data G, where nodes represent users and products, and edges represent interactive behaviors;

[0187] S42. Build a user behavior prediction model, add a graph neural network layer, and use the adjacency matrix A and the node feature matrix H 0 Perform graph feature extraction and generate graph feature vector H (l+1) :

[0188]

[0189] Among them, H (l) represents the node feature matrix of the lth layer, W (l) is the weight matrix, σ is the activation function, A ij is an element in the adjacency matrix, c ij is the normalized coefficient between nodes i and j, N(i) represents the set of neighbor nodes of node i;

[0190] S43, embedding time information into the graph feature vector to form a time-enhanced feature matrix. The time information enables the user behavior prediction model to capture the time dependency of user behavior and the changing trend of purchase frequency:

[0191]

[0192] in, represents the feature after embedding time information, α k and β k is the time embedding parameter, t ij is the time interval between nodes i and j, and T is the time period;

[0193] S44. Use the graph attention network to apply the attention mechanism on the time-enhanced feature matrix, dynamically adjust the weights between nodes, and generate the attention feature vector

[0194]

[0195]

[0196] in, represents the attention score between nodes i and j, a is the parameter vector of the attention mechanism, and represents the vector connection operation. is the attention weight of node j to node i, σ is the activation function;

[0197] S45. Input the attention feature vector output by the graph attention network layer into the Transformer layer, process the feature vector through the self-attention mechanism, capture long-distance dependencies, and generate a global feature vector Z:

[0198] Q=W Q H (l+1) , K=W K H (l+1) , V=W V H (l+1) ;

[0199]

[0200] Z=Transformer(H (l+1) );

[0201] Among them, Q is the query matrix, K is the key matrix, V is the value matrix, and W Q , W K and W V is the corresponding weight matrix, d k is the dimension of the key vector. The Transformer layer processes long-distance dependencies through the self-attention mechanism to improve the global perception ability of the user behavior prediction model.

[0202] S46. Using dynamic behavior feature dataset F dyn The user behavior prediction model is trained with the training set in , and the optimization objective function is to minimize the prediction error. The Adam optimizer is used to update the parameters:

[0203]

[0204] Where N represents the number of training samples. is the predicted purchase frequency, y i is the actual purchase frequency, is the predicted potential demand, d i is the actual potential demand, λ is the regularization coefficient, θ j Predict model parameters for user behavior;

[0205] S47. Use the validation set to validate the trained user behavior prediction model, calculate the validation error and evaluate the model performance. The validation indicators include mean square error and mean absolute error.

[0206] This implementation uses the GNN-Transformer hybrid model to build a user behavior prediction model. By combining LSTM, Bi-LSTM and attention mechanism, the user behavior prediction model can more effectively process and learn long-term dependencies in time series data, thereby improving the accuracy of the user behavior prediction model in predicting future user behavior.

[0207] S5. Combine the output results of the user behavior prediction model to build a multi-level user portrait, including static portrait, dynamic portrait and situational portrait;

[0208] In this implementation, S5 includes the following steps:

[0209] S51. Based on the output of the user behavior prediction model, collect and organize the user's demographic information, including age, gender, occupation and region to form a static profile of the user P static ;

[0210] S52: Generate a dynamic profile of the user P based on the prediction results of the user behavior prediction model, including the user's behavior preference and purchase frequency prediction. dynamic ,Behavioral preferences include users’ interest in different categories of products, and purchase frequency prediction is the user’s future purchase frequency predicted by the user behavior prediction model;

[0211] S53. Combine the user's purchasing motivation and consumption situation to analyze the user's consumption behavior in different situations and form a situational portrait P context,Purchase motivation reflects the reasons why users buy a certain type of product, and consumption context describes the consumption behavior pattern of users at a specific time and place;

[0212] S54, combining the output results of the user behavior prediction model, the static portrait P static , Dynamic Portrait P dynamic and situational portrait P context Integrate and generate multi-level user portraits P multi :

[0213] P multi = {P static , P dynamic , P context};

[0214] User behavior prediction model M predict The output results are used to guide and optimize the construction process of static portraits, dynamic portraits and situational portraits, forming a multi-level user portrait P multi .

[0215] This implementation utilizes the output of the constructed user behavior prediction model to generate a multi-level user portrait. This approach creates a comprehensive, multi-dimensional representation for each user by comprehensively considering static information and dynamic behavior data.

[0216] S6. Dynamically optimize marketing strategies based on the prediction results of user portraits and user behavior prediction models;

[0217] In this implementation, S6 includes the following steps:

[0218] S61, multi-level user portrait P multi and user behavior prediction model M predict The prediction results are used as input data to generate an initial marketing strategy plan including preliminary product recommendations, exclusive offers and personalized content push strategies;

[0219] S62. Generate a personalized product recommendation list R based on user portraits and prediction results using collaborative filtering algorithm product :

[0220]

[0221] Among them, p represents the product, u represents the current user, and u i represents users similar to the current user, w i represents the weight, similarity(u,u i ) represents the similarity between users, rating(u i , p) represents user u i Rating of product p;

[0222] S63. Generate personalized exclusive discount strategy based on user portrait and purchase frequency prediction results discount , including discounts, coupons and limited-time promotions:

[0223] S discount ={discount(p,u)|p∈R product};

[0224] Where discount(p,u) represents the exclusive discount for product p provided to user u;

[0225] S64. Generate personalized marketing content using natural language processing technology C personalized , combined with user portraits and current situations, and pushed to target users through a multi-channel reach mechanism:

[0226] C personalized =NLP(P multi , context);

[0227] Among them, context refers to the current situation, and NLP technology is used to generate marketing content that meets user needs;

[0228] S65. By real-time monitoring of user feedback on marketing strategies, the marketing strategies are dynamically adjusted and optimized using reinforcement learning algorithms. The optimization objective function is to maximize user engagement and conversion rate:

[0229]

[0230] Among them, π represents the strategy, γ is the discount factor, and R t is the reward at time t, and the optimization goal is to maximize the cumulative reward under strategy π;

[0231] S66. Regularly evaluate optimized marketing strategies, use key performance indicators including click-through rate, conversion rate and user lifetime value to assess marketing effectiveness and make corresponding adjustments.

[0232] S7. Combine user portraits and current situations to automatically generate marketing content that meets user needs, and push it to target users through the multi-channel reach mechanism of the private domain traffic platform;

[0233] S8. Establish a closed-loop feedback system to monitor and evaluate the implementation effect of marketing strategies in real time, dynamically adjust and optimize marketing strategies based on user interaction data and feedback information, and optimize marketing effects and user satisfaction.

[0234] Embodiment 1:

[0235] In order to verify the feasibility and effectiveness of the present invention, the method of the present invention is periodically applied to an e-commerce platform, which is carrying out a large-scale promotion activity to increase user purchase frequency and platform sales. The platform hopes to improve marketing accuracy and user satisfaction through the method of the present invention, thereby increasing sales and user stickiness.

[0236] Before the promotion begins, e-commerce platforms use embedded user behavior tracking technology to comprehensively collect multi-dimensional behavior data of users on the platform, including purchase frequency, browsing history, interactive behavior, and social sharing. In order to enhance the comprehensiveness of the data, the platform integrates this data with third-party data sources (such as social media data and other e-commerce platform data) to form a more comprehensive user data set.

[0237] The platform uses an improved time series pattern mining algorithm to dynamically extract user purchase frequency data, generate short-term behavior fluctuation characteristics, medium-term behavior trend characteristics, and long-term behavior pattern characteristics, and form a dynamic behavior feature data set. Subsequently, based on the dynamic behavior feature data set, the platform builds a user behavior prediction model, which is trained through an improved deep learning algorithm to predict users' future purchase frequency and potential demand.

[0238] Based on the output of the user behavior prediction model, the platform builds a multi-level user portrait, including static portrait (demographic information), dynamic portrait (behavioral preference and purchase frequency prediction) and situational portrait (purchase motivation and consumption situation). Combining these user portraits and prediction results, the platform uses collaborative filtering algorithms to generate personalized product recommendation lists, and formulates personalized exclusive discount strategies and personalized content push strategies. Through natural language processing technology, the platform generates personalized marketing content that meets user needs and pushes it to target users through a multi-channel reach mechanism.

[0239] During the promotional activities, the platform monitors the user feedback data on the marketing strategy in real time and uses reinforcement learning algorithms to dynamically adjust and optimize the marketing strategy. Through continuous optimization, the platform ensures that the marketing strategy can maximize user engagement and conversion rate.

[0240] In order to verify the effectiveness of the method of the present invention, the platform carried out a one-month promotion activity and compared and analyzed the method of the present invention with the traditional targeted marketing method. The following are the specific data:

[0241] Table 1 Data from May 1, 2024 to May 31, 2024

[0242]

[0243] As can be seen from Table 1 above, the method of the present invention is significantly superior to the traditional method in multiple key performance indicators. For example, in terms of user click-through rate, the method of the present invention reached 8.7%, an increase of 67% compared to 5.2% of the traditional method. In terms of user conversion rate, the method of the present invention is 4.3%, an increase of 139% compared to 1.8% of the traditional method. In addition, the average order value increased from 45.30 to 60.75, the user lifetime value increased from 15,000 to 22,000, the user retention rate increased from 35% to 58%, the marketing ROI increased from 3.5 to 6.8, and the user satisfaction score also increased from 3.8 / 5 to 4.5 / 5.

[0244] From these data, it can be seen that the method of the present invention has significant advantages in improving marketing accuracy, user participation and user satisfaction. During the promotion period, the platform not only achieved sales growth through the method of the present invention, but also significantly increased the purchase frequency and stickiness of users, achieving the expected promotion effect.

[0245] The method of the present invention effectively solves the problems in the prior art such as difficulty in capturing dynamic changes in user behavior, low prediction accuracy, and insufficient optimization of marketing strategies through comprehensive data collection, sophisticated data analysis, multi-level user portrait construction, and intelligent marketing strategy optimization, thus providing strong support for the precision marketing of private domain traffic platforms.

[0246] The present invention uses embedded user behavior tracking technology to comprehensively collect multi-dimensional behavior data of users on private traffic platforms, integrates it with third-party data sources, and applies data enhancement technology to greatly enhance the comprehensiveness and dimensionality of the data. In addition, through improved time series pattern mining algorithms and deep learning technology, user behavior characteristics are dynamically extracted and predicted, significantly improving the precision of marketing strategies and the accuracy of predictions.

[0247] The present invention combines the user's static information, behavioral preferences, purchase motivation and other multi-dimensional data to construct a multi-level user portrait including static portrait, dynamic portrait and situational portrait. This comprehensive user portrait makes the marketing strategy more personalized and refined, and better meets the user's individual needs.

[0248] The present invention utilizes real-time monitoring of user feedback data on marketing strategies and combines it with reinforcement learning algorithms to dynamically adjust marketing strategies. This not only enables marketing strategies to quickly respond to changes in the market and user behavior, but also enables continuous optimization, thereby maximizing user engagement and conversion rates.

[0249] The present invention significantly improves the user's shopping experience and satisfaction through personalized product recommendations, exclusive discounts and customized content push. This targeted and personalized interactive method not only enhances the user's willingness to buy, but also helps to improve user stickiness and enhance user loyalty to the platform.

[0250] In summary, the present invention realizes the refined management and dynamic optimization of marketing strategies through in-depth analysis and application of user purchase frequency and behavior data, which not only significantly improves the accuracy and effectiveness of marketing activities, but also improves the user's shopping experience and satisfaction, bringing substantial competitive advantages and commercial value to private domain traffic platforms.

[0251] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A private domain traffic platform targeted marketing method based on user purchase frequency analysis, characterized in that: The following steps are involved: S1. On the private domain traffic platform, through embedded user behavior tracking technology, collect multi-dimensional data of users in the private domain traffic platform to form a user behavior data set; S2, integrating the collected user behavior data set with third-party data sources to form a comprehensive user data set using data enhancement technology; S3. Based on the improved time series pattern mining algorithm, the short-term behavior fluctuation characteristics, medium-term behavior trend characteristics and long-term behavior pattern characteristics of the user's purchase frequency are dynamically extracted from the user's comprehensive data set to form a dynamic behavior feature data set; S4. Build a user behavior prediction model and use the GNN-Transformer hybrid model to train the user behavior prediction model based on the dynamic behavior feature dataset; S5. Combine the output results of the user behavior prediction model to build a multi-level user portrait, including static portrait, dynamic portrait and situational portrait; S6. Dynamically optimize marketing strategies based on the prediction results of user portraits and user behavior prediction models; S7. Combine user portraits and current situations to automatically generate marketing content that meets user needs, and push it to target users through the multi-channel reach mechanism of the private domain traffic platform; S8. Establish a closed-loop feedback system to monitor and evaluate the implementation effect of marketing strategies in real time, dynamically adjust and optimize marketing strategies based on user interaction data and feedback information, and optimize marketing effects and user satisfaction; The S1 comprises the following steps: S11. Embed data collection modules in the front-end and back-end of the private domain traffic platform, collect various types of user behavior data in the private domain traffic platform in real time through API interface and script technology, and build a user behavior data collection matrix : ； in, Represents the behavioral data collection matrix for user u at time t; S12. Collect the user's purchase frequency data by monitoring the user's shopping cart, order history and payment record, and record the timestamp of each purchase event and product ID ; S13. Collect users’ browsing history data by tracking their browsing paths, page dwell time and click behaviors within the private traffic platform, and record the timestamp of each page browsing event. and the page ID ; S14. Interaction behavior data collection: Collect user interaction behavior data by monitoring user interaction behaviors such as likes, comments, shares and favorites on the private domain traffic platform, and record the timestamp of each interaction event and interaction type ; S15. Collect users’ social sharing data by monitoring their social sharing activities on the private traffic platform, including sharing content, sharing channels, and sharing times, and record the timestamp of each sharing event. and sharing channels ; S16. Integrate the purchase frequency data, browsing record data, interactive behavior data and social sharing data to form a complete user behavior data set D: ； The S2 comprises the following steps: S21, preliminarily clean and filter the collected user behavior data set D, remove noise data and duplicate data, and perform missing value filling and data standardization processing; ； Where D represents the initial data set, represents a noisy dataset, Represents the cleaned user behavior dataset; ； in, is the mean, is the standard deviation, It is a standardized user behavior dataset; S22. Obtain relevant user behavior data from social media platforms and e-commerce platforms through API interfaces or data exchange protocols, including social interaction data, purchase history data, and user evaluation data: ； in, Represents a third-party dataset, represents the pth third-party data source; S23. Preprocessed user behavior dataset Align and match data with third-party data sources using unique user identifiers To associate and match users from different data sources: ； in, represents the aligned dataset, U represents the user set, and T represents the time set; S24, using the improved Bayesian inference data fusion technology, the pre-processed user behavior data set Integrate data from third-party data sources to form a comprehensive user data set , the improved Bayesian inference data fusion technology introduces the temporal dependency and purchase frequency characteristics of user behavior to enhance the accuracy and reliability of data fusion; ； Among them, the improved Bayesian inference data fusion function f is expressed as: ； in, represents the model parameters, Represents a given model parameter Under the condition of The likelihood of is the model parameter The prior probability of Represents the search for the posterior probability The largest parameter ; Calculate the fused user data set by maximizing the product of conditional probability and prior probability; S25. Comprehensive user data set Store in the user data platform to establish a complete user data archive; The S3 comprises the following steps: S31. Comprehensive user data set The user purchase frequency data in the data is segmented by time window to form a segmented data set , where the length of each time window is : ； Where k represents the index of the time window, is the starting time, is the time point after segmentation, and d represents the user’s purchase frequency data; S32. Analyze the user purchase frequency data in each time window to extract the user's behavior fluctuation characteristics in the short term ; S33: Perform trend analysis on user purchase frequency data in multiple continuous time windows to extract the user's mid-term behavior trend characteristics ; S34. Perform pattern recognition on the user purchase frequency data within the entire time period to extract the user's long-term behavior pattern characteristics ; S35. When extracting short-term behavior trend features, medium-term behavior trend features, and long-term behavior trend features, an improved temporal pattern mining algorithm based on topic drive is used. The improved temporal pattern mining algorithm integrates the temporal dependency of user behavior and topic information to form topic behavior features. : ； in, Indicates the behavioral characteristics after combining the subject information. Represents a collection of topics; S36. Short-term behavioral fluctuation characteristics , Mid-term behavioral trend characteristics , long-term behavioral pattern characteristics and subject behavior characteristics Integrate to form a dynamic behavior feature dataset : ； The improved time series pattern mining algorithm introduces the time dependency modeling of user behavior and adopts the time series variational autoencoder method: ； in, is the purchase frequency data at time t, and are the mean and standard deviation, respectively; Combined with the topic information, the LDA topic model is used for association analysis: ； in, Indicates The weight of the topic, For the Themes, Indicates the relevance of the feature to the topic; Based on the short-term behavior trend characteristics, medium-term behavior trend characteristics, and long-term behavior trend characteristics, the final topic behavior characteristics are obtained by maximizing the joint probability of the characteristics under all conditions: ； The S4 comprises the following steps: S41. Dynamic behavior feature dataset Converted into graph structure data G, where nodes represent users and products, and edges represent interaction behaviors; S42. Build a user behavior prediction model, add a graph neural network layer, and use the adjacency matrix A and the node feature matrix Perform graph feature extraction and generate graph feature vector : ； in, represents the node feature matrix of the lth layer, is the weight matrix, is the activation function, is an element in the adjacency matrix, For Node and The normalization coefficient between Representation Node The set of neighbor nodes of S43, embedding time information into the graph feature vector to form a time-enhanced feature matrix. The time information enables the user behavior prediction model to capture the time dependency of user behavior and the changing trend of purchase frequency: ； in, represents the feature after embedding time information, and is the time embedding parameter, For Node and The time interval between is the time period; S44. Use the graph attention network to apply the attention mechanism on the time-enhanced feature matrix, dynamically adjust the weights between nodes, and generate the attention feature vector : ； ； ； in, Representation Node and The attention score between them, a is the parameter vector of the attention mechanism, || represents the vector connection operation, For Node For Node The attention weight, is the activation function; S45. Input the attention feature vector output by the graph attention network layer into the Transformer layer, process the feature vector through the self-attention mechanism, capture long-distance dependencies, and generate a global feature vector Z: ； ； ； Among them, Q is the query matrix, K is the key matrix, and V is the value matrix. , and is the corresponding weight matrix, is the dimension of the key vector. The Transformer layer processes long-distance dependencies through the self-attention mechanism to improve the global perception ability of the user behavior prediction model. S46. Using dynamic behavior feature datasets The user behavior prediction model is trained with the training set in , and the optimization objective function is to minimize the prediction error. The Adam optimizer is used to update the parameters: ； Where N represents the number of training samples. is the predicted purchase frequency, is the actual purchase frequency, To predict potential demand, For actual potential needs, is the regularization coefficient, Predict model parameters for user behavior; S47. Use the validation set to validate the trained user behavior prediction model, calculate the validation error and evaluate the model performance. The validation indicators include mean square error and mean absolute error.

2. According to claim 1, a method for targeted marketing on a private domain traffic platform based on user purchase frequency analysis is characterized in that: The S5 comprises the following steps: S51. Based on the output of the user behavior prediction model, collect and organize the user's demographic information, including age, gender, occupation and region to form a static profile of the user ; S52. Generate a dynamic profile of the user based on the prediction results of the user behavior prediction model, including the user's behavior preference and purchase frequency prediction. ,Behavioral preferences include users’ interest in different categories of products, and purchase frequency prediction is the user’s future purchase frequency predicted by the user behavior prediction model; S53. Combine the user's purchasing motivation and consumption situation to analyze the user's consumption behavior in different situations and form a situational portrait ,Purchase motivation reflects the reasons why users buy a certain type of product, and consumption context describes the consumption behavior pattern of users at a specific time and place; S54, combined with the output results of the user behavior prediction model, the static portrait , Dynamic Portrait and situational portraits Integrate and generate multi-level user portraits : ； User behavior prediction model The output results are used to guide and optimize the construction process of static portraits, dynamic portraits and situational portraits to form multi-level user portraits .

Images