Intelligent customer acquisition system and method based on multi-source data fusion

Abstract

The scheme belongs to the technical field of intelligent marketing, and specifically relates to an intelligent customer acquisition system and method based on multi-source data fusion. The intelligent customer acquisition method based on multi-source data fusion comprises the following steps: S10: collecting geographic location information, e-commerce data information and social data information of a user terminal under the condition that authorization of the user terminal is obtained, the e-commerce data information comprising transaction information, browsing information and search information, and the social data information comprising collection clues, collection stores and browsing information; S20: analyzing user interest preferences according to the social data information. Through comprehensive data collection, dynamic weight distribution and real-time feedback mechanism, the scheme significantly improves the accuracy of pushing, user participation and conversion rate, optimizes the allocation of customer resource clues, and enhances user experience, thereby effectively solving the problem of low conversion rate of marketing users into consumption users in the prior art, and significantly improving customer acquisition efficiency and user satisfaction.

Description

TECHNICAL FIELD

[0001] The scheme belongs to the technical field of intelligent marketing, and particularly relates to an intelligent customer acquisition system and method based on multi-source data fusion. BACKGROUND

[0002] In the traditional customer acquisition mode, enterprises or merchants mainly rely on extensive commodity lead delivery, offline activity promotion, telephone sales and other ways to acquire customers. These ways are usually large-scale and non-directional, and it is difficult to accurately locate the target customer group, resulting in waste of commodity lead resources.

[0003] In recent years, with the development of big data and artificial intelligence technology, some precise marketing methods based on multi-feature fusion have appeared. For example, a user online precise marketing method based on multi-feature fusion disclosed in Chinese patent CN111612492A collects multiple types of data related to the user, outputs the multiple types of data to the corresponding type of feature extractor to extract feature vectors, integrates the multiple feature vectors into an integrated vector, outputs the integrated vector to a discriminator, and according to the judgment result of the discriminator, the user is pushed or exploratory pushed, and the online feedback data of the user after receiving the push or exploratory push is received to update the discriminator in real time, so as to achieve the effect of accurately positioning the marketing user.

[0004] However, the prior art can obtain the marketing positioning of the user from the e-commerce data and identify the potential needs of the user, but cannot combine the real-time state of the user terminal (such as time, scene, emotion, etc.) for pushing. This single data source pushing method leads to a low conversion rate of marketing users into consumer users. SUMMARY

[0005] The purpose of the scheme is to provide an intelligent customer acquisition system and method based on multi-source data fusion to solve the problem of low conversion rate of marketing users into consumer users when pushing commodity leads to user terminals.

[0006] In order to achieve the above purpose, the scheme provides an intelligent customer acquisition method based on multi-source data fusion, comprising the following steps:

[0007] S10: Collecting geographic location information, e-commerce data information and social data information of the user terminal under the condition of obtaining authorization of the user terminal, the e-commerce data information including transaction information, viewing information and search information, and the social data information including collection clues, collection stores and browsing information;

[0008] S20: Analyze user interest preferences according to social data information, analyze user consumption habits according to e-commerce data information, obtain goods matching user interest preferences and user consumption habits as first goods according to the current geographic location information of the user terminal, obtain goods clues and strategy information matching the first goods in the e-commerce platform and the social platform, mix the goods clues and strategy information into push content and send it to the user terminal;

[0009] S30: Obtain the number of requests of the user terminal to the goods clues and strategy information in the push content, adjust the mixing ratio of the goods clues and strategy information in the push content according to the number of requests, and send the adjusted push content to the user terminal.

[0010] And a multi-source data fusion-based intelligent customer acquisition system using a multi-source data fusion-based intelligent customer acquisition method.

[0011] The principle and technical effect of the present scheme is that: first, the present scheme collects the geographic location information, e-commerce data information and social data information of the user comprehensively, constructs a multi-dimensional portrait of the user, and the multi-dimensional portrait covers the behavior habits, interest preferences, consumption patterns and real-time state of the user. This comprehensive data collection and fusion can significantly improve the relevance and attractiveness of the push content, ensure that the push content is highly matched with the immediate needs and interests of the user, and thus improve the push accuracy.

[0012] Secondly, the present solution can identify the changes in user needs at different times and scenarios. For example, a user may pay more attention to breakfast-related content in the morning, while in the evening, they may prefer to relax or entertainment-type information. Emotion analysis can further refine the user's preferences at a specific moment, providing more personalized content pushing. This dynamic adjustment of user portraits allows the present solution to reflect changes in user interests and needs in real time, rather than relying solely on historical data. For example, if a user frequently browses a certain type of product recently, the present solution will increase the weight of this type of product information in the user portrait to ensure that the pushed content better matches the user's current interests. Furthermore, the present solution dynamically allocates weights to the pushed information based on the influence of different data sources in the user portrait on the user's consumption intention. This means that the present solution not only considers the user's historical behavior but also analyzes factors such as geographic location information, social interaction, and real-time feedback to determine which information has the greatest impact on the user's purchase decision and adjusts the display priority of product clues and content accordingly. This personalized pushing strategy significantly improves user engagement and satisfaction. Users feel that the content is more tailored to their actual needs and interests, making them more willing to interact with the pushed content, such as clicking, commenting, collecting, and sharing. This positive user engagement not only improves the user experience but also increases user loyalty and activity to the system. At the same time, these factors work together to improve conversion rates, i.e., the proportion of users who change from potential consumers to actual buyers. By accurately pushing product clues and services that users are interested in, the present solution can effectively promote sales and improve the ROI of marketing activities. At the same time, the dynamic weight allocation mechanism helps optimize the allocation of product clues, ensuring that product clue budgets are used on user groups that are most likely to generate conversions, reducing resource waste, and improving marketing efficiency.

[0013] Furthermore, the present solution collects feedback data from users on pushed content (such as click-through rate, conversion rate, dwell time, and close rate), evaluates the effectiveness of the push in real time, and dynamically adjusts the weight and mixing ratio of the pushed content based on user feedback. This feedback mechanism ensures that the pushing strategy always meets the interests and needs of users, optimizes resource allocation, enhances user experience, and improves user acceptance and satisfaction of product clues.

[0014] In summary, the present solution significantly improves the accuracy of pushing, user engagement, and conversion rates, optimizes the allocation of product clue resources, and enhances user experience through comprehensive data collection, dynamic weight allocation, and real-time feedback mechanisms, effectively solving the problem of low conversion rates from marketing users to consumer users in existing technologies, significantly improving customer acquisition efficiency and user satisfaction.

[0015] Further, between the S10 step and the S20 step, there is also a vectorization processing of the e-commerce data information and the social data information, which specifically includes the following steps:

[0016] S11: For transaction information, a time sequence embedding method is used to construct a three-dimensional feature vector transaction information vector, and the construction formula is shown in formula (1) as follows:

[0017] (1),

[0018] wherein, is the last transaction amount sequence, is the mean pooling, is the standard deviation calculation, take the Fourier transform before low frequency components;

[0019] For viewing information, viewing information vectors are constructed by attention weighted product embedding, and the embedding formula of the viewing information vector is shown in formula (2) as follows:

[0020] (2),

[0021] wherein, is the page dwell time, is a preset temperature coefficient, is the product description text;

[0022] For search information, search information vectors are constructed by query-click dual channel encoding, and the encoding formula of the search information vector is shown in formula (3) as follows:

[0023] (3),

[0024] wherein, is the search word sequence, is the clicked product picture, denotes vector splicing;

[0025] S12: The location of the collected store is constructed by space-time joint encoding to construct a store location vector, and the encoding formula of the store location vector is shown in formula (4) as follows:

[0026] (4),

[0027] wherein, denotes Hadamard product, is the store location attached text;

[0028] The attention information is aggregated by a graph neural network to construct a collection clue vector, and the aggregation formula of the collection clue vector is shown in formula (5) as follows:

[0029] (5),

[0030] wherein is a user-blogger adjacency matrix, is a node feature matrix;

[0031] The browsing information is modeled by a heterogeneous sequence to build a browsing information vector, and the modeling formula for building the browsing information vector is shown in formula (6) as follows:

[0032] (6),

[0033] wherein is a content type encoding, is a content segment embedding.

[0034] By vectorizing the e-commerce data information and social data information between the S10 step and the S20 step, not only the data processing efficiency can be improved, but also the feature expression capability can be enhanced. This processing mode enables transaction information, viewing information and search information to be converted into three-dimensional feature vectors, so as to be more efficiently analyzed. At the same time, by vectorizing social data such as check-in information, attention information and browsing information, the user's social behavior and preferences can be more accurately captured, and the performance of the present scheme is improved. In addition, the vectorization processing supports complex data analysis, promotes data fusion, and improves the generalization ability of the model, and the present scheme supports real-time analysis and decision-making, which is particularly important for application scenarios that require fast response.

[0035] Further, in the S20 step, when analyzing the user consumption habits according to the e-commerce data information, a time sequence dynamic graph convolution network is used to model the consumption features according to the transaction information vector, and the calculation process includes the following steps:

[0036] S201: Extract the transaction pattern using the following formula (7):

[0037] (7),

[0038] wherein , , , is the trainable parameter of the i-th layer, and the final consumption stability score is calculated according to the transaction pattern extraction result ;

[0039] S202: Model the viewing preference of the viewing information vector through a cross-attention mechanism;

[0040] S203: Use a Gaussian mixture model to cluster the search information vector, and identify the dominant consumption intention according to the clustering result;

[0041] According to the analysis of user interest preferences based on social data information, a multi-modal interest graph is constructed according to the store location vector, the collection clue vector and the browsing information vector, which specifically includes the following steps:

[0042] S204: The store location vector is fused by bilinear for spatiotemporal feature fusion, and the store location scene is classified;

[0043] S205: The collection clue vector is calculated based on the weight diffusion algorithm improved by PageRank, and the social influence propagation is obtained;

[0044] S206: The browsing information vector is calculated by using a hierarchical attention network, and the browsing content representation is obtained.

[0045] In the S20 step, by using a time-series dynamic graph convolution network to model e-commerce data information, using a cross-attention mechanism and a Gaussian mixture model to analyze social data information, and constructing a multi-modal interest graph, the accuracy of user consumption habit and interest preference analysis can be significantly improved. First, the time-series dynamic graph convolution network can extract transaction patterns, calculate consumption stability scores, and model viewing preferences for viewing information vectors through a cross-attention mechanism, and use a Gaussian mixture model to cluster search intent for search information vectors. Second, the construction of the multi-modal interest graph includes the spatiotemporal feature fusion of the store location vector by bilinear, the calculation of the collection clue vector based on the weight diffusion algorithm improved by PageRank, and the calculation of the browsing information vector by using a hierarchical attention network to obtain the browsing content representation. These steps not only can construct accurate user portraits, but also can provide personalized recommendations, predict consumption behavior, analyze social influence, and handle multi-modal data, thereby improving user experience and business results in multiple aspects.

[0046] Further, in the S20 step, when the goods matching the user interest preferences and user consumption habits are obtained as the first goods according to the current geographic location information of the user terminal, A real-time geographic fence is constructed, which is shown in the following formula (14):

[0047] (14),

[0048] Wherein, is the geographic fence associated with the candidate goods k, is a preset reference radius, is a time decay factor, The calculation formula of is shown in the following formula (15):

[0049] (15),

[0050] wherein = 0.3, is a peak period of the product category;

[0051] The user features and the product features are tensor-interacted, and a calculation formula of the tensor interaction is shown in the following formula (16):

[0052] (16),

[0053] wherein, the product-side feature satisfies ; is a product category BERT embedding; is a price logarithm and a user historical average price ratio, ;

[0054] Then, a matching degree is output through a gating scoring network according to a calculation result of the tensor interaction, and a calculation formula of the matching degree is shown in the following formula (17):

[0055] (17),

[0056] wherein, represents a Hadamard product, a calculation formula of which is shown in the following formula (18):

[0057] (18),

[0058] Finally, a first product set is output according to the matching degree, a calculation formula of which is shown in the following formula (19):

[0059] (19),

[0060] wherein, is a threshold value determined through a grid search.

[0061] The scheme can significantly improve the accuracy and personalization of commodity recommendation by constructing real-time geographic fence based on the current location of the user, and combining advanced technologies such as time sequence dynamic graph convolution network, cross attention mechanism and Gaussian mixture model. First, the construction of real-time geographic fence can ensure that the recommended commodities are closely related to the geographical position of the user, and improve the practicability of the recommendation. Second, by extracting transaction patterns, calculating consumption stability scores, and using cross attention mechanism to model viewing preference of viewing information vector, the e-commerce data information of the user can be deeply analyzed, so as to more accurately grasp the consumption habits of the user. In addition, using Gaussian mixture model to cluster search information vector can help identify the dominant consumption intention of the user, further improving the targeting of the recommendation. In terms of social data information analysis, by constructing a multi-modal interest graph, including spatio-temporal feature fusion, social influence propagation analysis and browsing content representation acquisition, the interest preferences of the user can be comprehensively captured. Finally, the user features and commodity features are interacted in the form of tensor, and the matching degree is output through the gate scoring network, to ensure that the recommended commodities are highly matched with the interests and consumption habits of the user.

[0062] Further, in the S20 step, when obtaining commodity clues and strategy information matched with the commodity in the e-commerce platform and the social platform, a heterogeneous content retrieval engine is used to realize accurate matching of commodity clues and strategy information, which includes the following steps:

[0063] S21: Based on the commodity set Generate a unified query vector , The calculation formula is shown in the following formula (20):

[0064] (20),

[0065] Wherein, is a query projection network, and are time encoding isomorphic;

[0066] An incremental index is established for e-commerce commodity clues and social strategies, and the establishment formula is shown in the following formulas (21) and (22):

[0067] (21),

[0068] (22),

[0069] A multi-stage retrieval architecture is used, and the retrieval architecture is shown in the following formula (23):

[0070] (23),

[0071] Wherein, The calculation formula is shown in formula (24) below:

[0072] (twenty four).

[0073] By employing a heterogeneous content retrieval engine to achieve precise matching of product leads and guide information on e-commerce and social media platforms, the relevance and effectiveness of these leads and guides can be significantly improved. Specifically, a unified query vector is first generated based on the product set. Then, incremental indexes are built for e-commerce product leads and social media guides, employing a multi-stage retrieval architecture. These steps ensure the efficiency and accuracy of the retrieval process. Furthermore, by calculating the similarity score between the query vector and the product leads / guide content, the matching results can be further optimized, ensuring that the recommended product leads and guide information are highly relevant to users' interests and needs. This method not only improves the click-through rate and conversion rate of product leads but also enhances the user experience, allowing users to obtain more valuable information and suggestions while browsing products.

[0074] Furthermore, in step S30, when adjusting the mixing ratio of product leads and strategy information in the push content according to the number of requests, the difference in click-through rate between product leads and strategies is calculated by real-time monitoring of the user terminal's request content for the push content. The formula for calculating the difference is shown in the following formula (25):

[0075] (25),

[0076] when The mixing ratio needs to be readjusted.

[0077] The system obtains the user terminal's display progress of the guide, which is determined based on the user terminal's scrolling depth and the total length of the guide content; it calculates the product lead conversion funnel for the user terminal, which is determined by the ratio of add-to-cart requests sent by the user terminal to the number of exposures, and then calculates the product lead conversion funnel. and display progress Adjust the mixing ratio. The formula for adjusting the mixing ratio is shown in formula (26) below:

[0078] (26)

[0079] in, This is the preset learning rate.

[0080] By calculating the click rate difference and the commodity clue conversion funnel, the scheme can adjust the proportion of commodity clues and strategy information in the push content in real time to better meet the user demand and improve the user participation. According to the user's display progress of the strategy and the conversion effect of the commodity clue, the scheme can more accurately push the information most related to the user's current interest and behavior, thereby improving the relevance and attractiveness of the push content. By dynamically adjusting the proportion of commodity clues and strategy information, the display effect of commodity clues can be optimized, and the click rate and conversion rate of commodity clues can be improved, thereby improving the ROI of commodity clue owners. By providing push content that better meets the user's current interest and demand, the user's satisfaction and loyalty can be enhanced, thereby improving the user retention rate and activity. By monitoring and analyzing user behavior data in real time, the scheme can provide data support for the operation team to help them make more accurate decisions to optimize the push strategy and improve the overall marketing effect.

[0081] Further, based on the reinforcement learning framework, a proportion decision model is constructed, and the user real-time feature vector 、 、 environmental context features 、 and behavior indicators in the sliding window 、 、 are taken as state inputs, and the commodity clue weight is output by the deep deterministic policy gradient network.

[0082] A multi-objective optimization function is established to maximize the weighted sum of expected click rate and conversion rate, and the constraint conditions include user experience score threshold and upper limit of commodity clue inventory consumption rate , NSGA-II algorithm is used to generate Pareto optimal solution set, and weighted Chebyshev decomposition is used for online strategy selection;

[0083] A causal reasoning module is constructed, and a double machine learning method is used to estimate the treatment effect of the mixed proportion on the click rate , eliminating the influence of user purchasing power and time period and other confounding variables, wherein the causal graph model includes three core paths: mixed proportion to click rate, user purchasing power to mixed proportion, and user purchasing power to click rate;

[0084] A federated learning architecture is deployed, and after local training of the proportion adjustment model by each terminal device, the global model is updated by weighted aggregation , and Gaussian noise satisfying -differential privacy is added in the gradient update process ;

[0085] An explainable decision report is generated by constructing a feature contribution SHAP value calculation system and a decision tree rule engine, and outputting a logically based mixed proportion adjustment. Key decision rules include: when >0.7 and loc_type is a shopping mall, assign 80% of the product lead weight, when the guide reading completion rate and the time period assign 75% of the guide weight.

[0086] By applying the reinforcement learning framework to the proportion decision model, and integrating the user's real-time feature vector, environmental context features, and behavior indicators within the sliding window, the accuracy of product lead pushing and user engagement can be significantly improved. Using the Deep Deterministic Policy Gradient Network (DDPG) to output product lead weights, combined with a multi-objective optimization function, the goal is to maximize the weighted sum of expected click-through rate and conversion rate, while considering user experience score thresholds and product lead inventory consumption rates. The NSGA-II algorithm is used to generate a Pareto optimal solution set, and the weighted Chebyshev decomposition is used for online strategy selection, thereby achieving a better product lead placement strategy. In addition, a causal reasoning module is constructed to estimate the processing effect of mixed proportions on click-through rate using a dual machine learning method, eliminating the influence of user purchasing power and time period, and ensuring the accuracy of product lead effectiveness evaluation. Deploy a federated learning architecture, each terminal device locally trains the proportion adjustment model, and then updates the global model through weighted aggregation. The gradient update process adds Gaussian noise that satisfies differential privacy, which not only protects user privacy but also improves the generalization ability of the model. Finally, an explainable decision report is generated by constructing a feature contribution SHAP value calculation system and a decision tree rule engine, outputting the logical basis for mixed proportion adjustment, making key decision rules more explicit, such as dynamically adjusting the weights of product leads and guides based on user location and time period, thereby protecting user privacy while improving the personalization and effectiveness of product lead pushing.

[0087] Further, the moving speed of the user terminal is obtained , the voice information entered by the user terminal is collected under the condition of obtaining permission , the destination of the user terminal is obtained through the voice information and the moving speed and the travel experience , specifically including the following steps:

[0088] A10: Perform sentiment analysis on the voice information to extract a sentiment feature vector , and construct formula (27) as follows:

[0089] (27),

[0090] wherein, For the speech spectrum feature, For the semantic embedding vector, For the emotion classification weight matrix;

[0091] A20: Combined with the moving speed And the rate of change of geographical position , the travel scene classification score is calculated , The calculation formula is shown in the following formula (28):

[0092] (28),

[0093] Wherein, The scene classifier, The time attenuation factor;

[0094] A30: Calculate the travel experience , The calculation formula is shown in the following formula (29):

[0095] (29),

[0096] Wherein, The fusion weight, ReLU is the activation function;

[0097] A40: According to the travel experience And the POI features of the destination , predict the consumption intention vector , The prediction formula is shown in the following formula (30):

[0098] (30),

[0099] Wherein, LSTM is the intention prediction network, The destination feature embedding;

[0100] A50: Based on the consumption intention vector , the target commodity set is screened through the gate matching network , The screening formula is shown in the following formula (31):

[0101] (31),

[0102] Wherein, The commodity feature matrix, The matching degree threshold;

[0103] ​A60: Obtain commodity clues and strategy information by using a heterogeneous content retrieval engine, and generate push content according to a mixed proportion adjustment method.

[0104] Further, the travel experience The calculation further includes the following steps:

[0105] The travel scene-emotion joint space is established as shown in the following formula (32):

[0106] (32),

[0107] Wherein, is a scene-emotion association matrix, is a spatio-temporal context feature;

[0108] The consumption intention is updated through a time sequence attention mechanism, as shown in the following formula (33):

[0109] (33),

[0110] Wherein, is a historical intention sequence, is an attention weight;

[0111] The destination commodity retrieval range is optimized by using a geographic fence formula, and the commodities are matched by combining a multi-modal interest graph;

[0112] The consumption intention-commodity feature tensor interaction space is constructed as shown in the following formula (34):

[0113] (34),

[0114] Wherein, is a feature interaction operator;

[0115] The push strategy is optimized by using a reinforcement learning framework, and the travel experience features are used as state inputs, as shown in the following formula (35):

[0116] (35),

[0117] Wherein, is a strategy network, is a travel scene feature;

[0118] The mixed proportion is dynamically adjusted according to formula (25) and formula (26), and when > The real-time commodity clue weight is increased, and the increase formula is as shown in the following formula (36):

[0119] (36),

[0120] wherein, is a speed adaptation coefficient, is a reference ratio.

[0121] The scheme creatively combines real-time recognition of user travel scenarios with long-term consumption habit analysis, which can grasp the user's stable preference and sensitively capture the immediate demand changes. This dual cognitive mechanism makes the push content not only respect the user's historical preference, but also flexibly adapt to various sudden scene demands, realizing truly personalized service. It is particularly worth pointing out that the scheme integrates environmental perception and content decision-making in depth to build an adaptive intelligent push ecosystem. The scheme can automatically optimize the information presentation mode and content combination strategy according to the user's moving state, surrounding environment and emotional changes. This dynamic adjustment mechanism not only greatly improves the relevance of information, but also significantly improves the business conversion efficiency while protecting the user experience. BRIEF DESCRIPTION OF DRAWINGS

[0122] Figure 1 is a flowchart of the intelligent customer acquisition method based on multi-source data fusion in the embodiment of the present application. DETAILED DESCRIPTION

[0123] The concept and technical effects of the present application will be described below in conjunction with the embodiments to fully understand the purpose, features and effects of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all embodiments, and other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative labor are within the scope of the present application:

[0124] As Figure 1 shown, the intelligent customer acquisition method based on multi-source data fusion includes the following steps:

[0125] S10: Collecting the geographic location information, e-commerce data information and social data information of the user terminal under the condition of obtaining the authorization of the user terminal, the e-commerce data information including transaction information, viewing information and search information, the social data information including collection clues, collection stores and browsing information;

[0126] S20: Analyzing the user interest preference according to the social data information, analyzing the user consumption habit according to the e-commerce data information, obtaining the goods matching the user interest preference and the user consumption habit as the first goods according to the current geographic location information of the user terminal, obtaining the goods clues and strategy information matching the first goods in the e-commerce platform and the social platform, mixing the goods clues and strategy information as the push content and sending it to the user terminal;

[0127] S30: Obtain the number of requests of the user terminal for the commodity clues and the strategy information in the push content, adjust the mixing ratio of the commodity clues and the strategy information in the push content according to the number of requests, and send the adjusted push content to the user terminal.

[0128] Among them, between S10 step and S20 step also includes vectorization processing to e-commerce data information and social data information, including the following steps:

[0129] S11: For transaction information, a three-dimensional feature vector transaction information vector is constructed by using a time sequence embedding method, and the construction formula is shown in the following formula (1):

[0130] (1),

[0131] Among them, is the last transaction amount sequence, is the mean pooling, is the standard deviation calculation, take the Fourier transform before low frequency component;

[0132] For viewing information, a viewing information vector is constructed by using attention weighted product embedding, and the embedding formula of the viewing information vector is shown in the following formula (2):

[0133] (2),

[0134] Among them, is the page dwell time, is a preset temperature coefficient, is the product description text;

[0135] For search information, a search information vector is constructed by using query-click dual-channel encoding, and the encoding formula of the search information vector is shown in the following formula (3):

[0136] (3),

[0137] Among them, is the search word sequence, is the clicked product picture, represents vector splicing;

[0138] S12: The position of the collection store is constructed by space-time joint encoding to construct a store location vector, and the encoding formula of the store location vector is shown in the following formula (4):

[0139] (4),

[0140] Among them, It represents the Hadamah accumulation. Include text with the store location;

[0141] The collection clue vector is constructed by a graph neural network, and the aggregation formula for constructing the collection clue vector is shown in formula (5) below:

[0142] (5),

[0143] in For user-blogger adjacency matrix, The node feature matrix;

[0144] Browsing information is modeled using heterogeneous sequences to construct browsing information vectors. The modeling formula for constructing browsing information vectors is shown in formula (6) below:

[0145] (6),

[0146] in Encode the content type. Embed content fragments.

[0147] In step S20, when analyzing user interests and preferences based on social data, a temporal dynamic graph convolutional network is used to model consumption features based on transaction information vectors. The calculation process includes the following steps:

[0148] S201: Extract the transaction pattern using the following formula (7):

[0149] (7),

[0150] in , , , For the first The layer can be trained parameters, and the final consumption stability score is calculated based on the results extracted from the transaction pattern. , The calculation formula is shown in formula (8) below:

[0151] (8);

[0152] S202: Viewing preference modeling is performed on the viewing information vector through cross-attention mechanism. The modeling formula is shown in formula (9) below:

[0153] (9),

[0154] in, , and Let be the projection matrix. preset vector dimension;

[0155] S203: search intent clustering of the search information vector using a Gaussian mixture model, and the clustering formula is shown in the following formula (10):

[0156] (10),

[0157] According to the clustering result identify the dominant consumption intent;

[0158] When analyzing user interest preferences according to social data information, a multi-modal interest graph is constructed according to the store location vector, the collection clue vector and the browsing information vector, and the construction includes the following steps:

[0159] S204: time-space feature fusion of the store location vector is performed through a bilinear, and a check-in scene is classified, and the fusion formula is shown in the following formula (11):

[0160] (11),

[0161] wherein and are time-space interaction matrices;

[0162] S205: the collection clue vector is calculated based on a PageRank improved weight diffusion algorithm, and social influence propagation is obtained, and the calculation formula is shown in the following formula (12):

[0163] (12),

[0164] wherein =0.85 , is a set of bloggers followed by the user

[0165] S206: the browsing information vector is calculated using a hierarchical attention network, and browsing content representation is obtained, and the calculation formula is shown in the following formula (13):

[0166] (13),

[0167] wherein, is the jth segment embedding of the ith browsing sequence.

[0168] In the S20 step, when the goods matching the user interest preferences and the user consumption habits are obtained as the first goods according to the current geographic position information of the user terminal, the real-time geographic fence is constructed based on the current position of the user, and the construction formula is shown in the following formula (14): ​

[0169] (14),

[0170] wherein, is a candidate commodity associated with a geographical fence, is a preset reference radiation radius, is a time decay factor, The calculation formula of is shown in the following formula (15):

[0171] (15),

[0172] wherein , is a category peak period;

[0173] The user features and the commodity features are tensor interacted, and the calculation formula of the tensor interaction is shown in the following formula (16):

[0174] (16),

[0175] wherein, the commodity side feature satisfies ; is a commodity category BERT embedding; is a price logarithm and a user historical average price ratio, ;

[0176] The matching degree is outputted according to the calculation result of the tensor interaction through a gating scoring network, and the calculation formula of the matching degree is shown in the following formula (17):

[0177] (17),

[0178] wherein, represents a Hadamard product, The calculation formula of is shown in the following formula (18):

[0179] (18),

[0180] Finally, the first commodity set is outputted according to the matching degree The calculation formula of is shown in the following formula (19):

[0181] (19),

[0182] wherein, is a threshold value determined through a grid search.

[0183] In the S20 step, when obtaining the commodity clues and strategy information matched with the commodity in the e-commerce platform and the social platform, a heterogeneous content retrieval engine is used to realize accurate matching of the commodity clues and strategy information, and the specific steps include the following steps:

[0184] S21: Based on the commodity set generate a unified query vector , The calculation formula is shown in the following formula (20):

[0185] (20),

[0186] wherein, is a query projection network, and are time encoding isomorphic;

[0187] An incremental index is established for e-commerce commodity clues and social strategies, and the establishment formula is shown in the following formulas (21) and (22):

[0188] (21),

[0189] (22),

[0190] A multi-stage retrieval architecture is used, and the retrieval architecture is shown in the following formula (23):

[0191] (23),

[0192] wherein, The calculation formula is shown in the following formula (24):

[0193] (24).

[0194] In the S30 step, when adjusting the mixed ratio of commodity clues and strategy information in the push content according to the number of requests, the click rate difference of commodity clues and strategies is calculated by real-time monitoring of the request content of the user terminal to the push content, and the difference calculation formula is shown in the following formula (25):

[0195] (25),

[0196] When , the mixed ratio is readjusted;

[0197] obtain a display progress of the guide by the user terminal, the display progress being determined according to a scroll depth of the user terminal and a total length of the guide content; calculate a commodity lead conversion funnel of the user terminal, the commodity lead conversion funnel being determined by a ratio of a shopping cart joining request sent by the user terminal and an exposure frequency, and adjust the mixing ratio according to the commodity lead conversion funnel and the display progress and the display progress Adjust the mixing ratio, and the adjustment formula of the mixing ratio is shown in the following formula (26):

[0198] (26),

[0199] wherein, is a preset learning rate.

[0200] wherein, a proportion decision model is constructed based on a reinforcement learning framework, real-time feature vectors of users 、 、 environmental context features ), and behavior indicators in a sliding window 、 、 are taken as state inputs, and a commodity lead weight is output by a deep deterministic policy gradient network;

[0201] A multi-objective optimization function is established to maximize the weighted sum of expected click rate and conversion rate, and the constraint conditions include a user experience score threshold and an upper limit of commodity lead inventory consumption rate , an NSGA-II algorithm is used to generate a Pareto optimal solution set, and a weighted Chebyshev decomposition is used for online strategy selection;

[0202] A causal reasoning module is constructed, and a double machine learning method is used to estimate the treatment effect of the mixing ratio on the click rate , so as to eliminate the influence of user purchasing power and time period and other confounding variables, wherein the causal graph model includes three core paths of mixing ratio to click rate, user purchasing power to mixing ratio, and user purchasing power to click rate;

[0203] A federated learning architecture is deployed, and after local training of the proportion adjustment model by each terminal device, a global model is updated by weighted aggregation , and Gaussian noise satisfying differential privacy is added in the gradient update process ;

[0204] An interpretable decision report is generated, a feature contribution SHAP value calculation system and a decision tree rule engine are constructed, and the logical basis for adjusting the mixing ratio is output, wherein the key decision rules include: when When the value is >0.7 and loc_type is "shopping mall", allocate 80% of the product clue weight. This applies when the strategy guide completion rate is... and time period 75% of the strategy weight is allocated to time.

[0205] Among them, obtaining the user terminal's movement speed Collect voice information recorded by the user terminal with the necessary permissions. The destination of the user terminal is obtained through voice information and mobile speed. and travel experience Specifically, it includes the following steps:

[0206] A10: Regarding voice information Perform sentiment analysis and extract sentiment feature vectors. Formula (27) is constructed as follows:

[0207] (27)

[0208] in, For speech spectrum features, For semantic embedding vectors, The sentiment classification weight matrix;

[0209] A20: Combining movement speed and the rate of change of geographical location Calculate travel scenario classification scores , The calculation formula is shown in the following formula (28):

[0210] (28)

[0211] in, For scene classifiers, This is the time decay factor;

[0212] A30: Enhancing the travel experience through multimodal fusion computing , The calculation formula is shown in formula (29) below:

[0213] (29)

[0214] in, ReLU is used as the activation function for the fusion weights;

[0215] A40: Based on travel experience and destination POI features Predicting consumer intention vectors , The prediction formula is shown in the following formula (30):

[0216] (30),

[0217] wherein LSTM is an intention prediction network, is a destination feature embedding;

[0218] A50: screening target commodity set based on consumption intention vector by a gated matching network , The screening formula is shown in the following formula (31):

[0219] (31),

[0220] wherein, is a commodity feature matrix, is a matching degree threshold;

[0221] A60: obtaining commodity clues and strategy information by using a heterogeneous content retrieval engine, and generating push content according to a mixed proportion adjustment method.

[0222] Specifically, the travel experience The calculation further includes the following steps:

[0223] The travel scene-emotion joint space is established as shown in the following formula (32):

[0224] (32),

[0225] wherein, is a scene-emotion correlation matrix, is a spatio-temporal context feature;

[0226] The consumption intention is updated through a time sequence attention mechanism, which is specifically shown in the following formula (33):

[0227] (33),

[0228] wherein, is a historical intention sequence, is an attention weight;

[0229] The destination commodity retrieval range is optimized by using a geographic fence formula, and the commodities are matched in combination with a multi-modal interest graph;

[0230] The consumption intention-commodity feature tensor interaction space is constructed as shown in the following formula (34):

[0231] (34),

[0232] wherein, is a characteristic interaction operator;

[0233] The push strategy is optimized using a reinforcement learning framework, and the travel experience characteristics As state input, as shown in the following formula (35):

[0234] (35),

[0235] wherein, is a policy network, is a travel scene characteristic;

[0236] According to formula (25) and formula (26), the mixing ratio is dynamically adjusted, and when it is detected > The real-time product clue weight is increased when, and the increase formula is as shown in the following formula (36):

[0237] (36),

[0238] wherein, is a speed adaptation coefficient, is a reference ratio.

[0239] The embodiment also includes a multi-source data fusion-based intelligent customer acquisition system using a multi-source data fusion-based intelligent customer acquisition method.

[0240] In specific implementation, in the App of a large e-commerce platform "Gouyi Network", a set of intelligent customer acquisition system is deployed, and the following will take user Zhang as an example to describe the personalized push process. Zhang is a 25-year-old white-collar worker in a first-tier city.

[0241] Through authorization, the system obtains the real-time location of Zhang, which is located in the Guomao business district of Beijing Chaoyang District, and the time is 20:30 in the evening on weekdays. In terms of e-commerce data, Zhang's monthly average consumption in the past 3 months reaches 3800 yuan, and the high-frequency purchase categories are concentrated in "fitness equipment" and "instant food". On that day, Zhang viewed a yoga mat for 120 seconds, and also viewed protein powder for 45 seconds. In addition, Zhang searched for "portable fitness equipment" and clicked on 3 folding dumbbell products. From the social data, Zhang posted a check-in information of "Guomao gym check-in for the 5th day" on the social platform. He followed 5 fitness bloggers and 3 food bloggers, and 70% of the content he browsed recently was fitness tutorials, and 30% was new product reviews of convenience stores.

[0242] In the vectorization process (S11-S12), the transaction information vector Generated by formula (1), from which the periodic consumption peak of Zhang is captured on Friday night every week. View information vector Calculated by formula (2), where the attention weight of yoga mat is as high as 0.73. Store location vector Encoded by formula (4), the "gym" scene label is successfully identified.

[0243] In the process of commodity matching, the formula (14) of the geographic fence is used to activate the merchants within 3km of the Guomao business circle, and then the "smart folding dumbbell" with a matching degree of 0.82 and the "low-carb protein bar" with a matching degree of 0.76 are screened out. Then, the features are fused by the tensor interaction formula (16), and the final matching commodity set is output by the gate scoring network formula (17) .

[0244] In terms of content retrieval, the heterogeneous retrieval engine obtains relevant content according to formula (23). Among them, the commodity clue is "smart dumbbell limited time 8 discount", and the CTR estimate is 5.2%; the guide is "gym white equipment selection guide", and the interaction index is 4.8 / 5.0.

[0245] In the mixed push link, according to the high consumption habit of Zhang, the initial proportion is set to 0.6. The output fusion content is generated by the generative adversarial network, the top shows the commodity clue with the "buy now" button, and the lower part embeds the selected paragraphs of the guide.

[0246] After receiving the push, user Zhang clicked the commodity clue but did not purchase, but read the guide completely, and the scroll depth reached 92%. At this time, the conversion funnel value =0.15 is lower than the user average value 0.25. The click rate difference This value triggers the adjustment of formula (26). After calculation, the new proportion is 0.53. Based on this, the guide length is increased in the next day's push, and a "gym diet matching" video content is also added.

[0247] The above is only an embodiment of the present application, and the specific structure and characteristics of the scheme are not described in detail. It should be noted that for those skilled in the art, without departing from the structure of the present application, a number of modifications and improvements can be made, which should be regarded as the protection scope of the present application, and these will not affect the effect and practicality of the patent. The protection scope claimed in this application should be subject to the content of its claims, and the specific implementation mode and the like recorded in the specification can be used to explain the content of the claims.

Claims

1. A method for intelligent customer acquisition based on multi-source data fusion, characterized in that, The method comprises the following steps: S10: collecting geographical position information, e-commerce data information and social data information of the user terminal in the case of obtaining authorization of the user terminal, the e-commerce data information comprising transaction information, viewing information and search information, and the social data information comprising collection clues, collection stores and browsing information; S20: analyzing user interest preferences according to the social data information, analyzing user consumption habits according to the e-commerce data information, obtaining goods matching the user interest preferences and the user consumption habits as first goods according to the current geographical position information of the user terminal, obtaining goods clues and strategy information matching the first goods in the e-commerce platform and the social platform, mixing the goods clues and the strategy information into push content and sending the push content to the user terminal; S30: obtaining a number of requests of the user terminal for the goods clues and the strategy information in the push content, adjusting a mixing ratio of the goods clues and the strategy information in the push content according to the number of requests, and sending the adjusted push content to the user terminal; In the step S20, when analyzing the user consumption habits according to the e-commerce data information, a time sequence dynamic graph convolution network is used to model consumption features according to a transaction information vector, and the calculation process comprises the following steps: S201: extracting a transaction mode by using the following formula (7): （7）， Wherein , , , is the first layer trainable parameter, and then according to the transaction mode extraction result, the final consumption stability score is calculated ; S202: modeling viewing preferences of a viewing information vector through a cross-attention mechanism; S203: using a Gaussian mixture model to perform search intention clustering on a search information vector, and identifying a dominant consumption intention according to a clustering result; When analyzing the user interest preferences according to the social data information, a multi-modal interest graph is constructed according to a store location vector, a collection clue vector and a browsing information vector, and specifically comprises the following steps: S204: performing time-space feature fusion on the store location vector through a bilinear method, and classifying a store location scene; S205: calculating the collection clue vector based on a weight diffusion algorithm improved based on PageRank, and obtaining social influence propagation; S206: calculating the browsing information vector by using a hierarchical attention network, and obtaining a browsing content representation. 2.The intelligent customer acquisition method based on multi-source data fusion according to claim 1, characterized in that: Between the steps S10 and S20, the e-commerce data information and the social data information are further subjected to vectorization processing, and specifically comprises the following steps: S11: for the transaction information, a time sequence embedding method is used to construct a three-dimensional feature vector transaction information vector, and a construction formula is shown in the following formula (1): （1）， wherein, is the most recent transaction amount sequence, is mean-pooling, is standard deviation calculation, before Fourier transform low-frequency components; For the viewing information, a viewing information vector is constructed through attention-weighted product embedding, and an embedding formula of the viewing information vector is shown in the following formula (2): （2）， wherein, is a page dwell time, is a preset temperature coefficient, is a product description text; For the search information, a search information vector is constructed by using a query-click dual-channel coding method, and a coding formula of the search information vector is shown in the following formula (3): （3）， wherein, is a search term sequence, is a click product image, denotes vector concatenation; S12: a store location vector is constructed by using time-space joint coding on a collection store location, and a coding formula of the store location vector is shown in the following formula (4): （4）， wherein, denotes a Hadamard product, textual description of the store location; A collection clue vector is constructed by using graph neural network aggregation on attention information, and an aggregation formula of the collection clue vector is shown in the following formula (5): （5）， wherein is a user-blogger adjacency matrix, is a node feature matrix; A browsing information vector is constructed by using heterogeneous sequence modeling on browsing information, and a modeling formula of the browsing information vector is shown in the following formula (6): （6）， wherein is encoded for the content type, is embedded for the content segment. 3.The intelligent customer acquisition method based on multi-source data fusion according to claim 2, characterized in that: In the S20 step, when the goods matching the user interest preference and the user consumption habit are obtained as the first goods according to the current geographic position information of the user terminal, the first goods are obtained based on the current position of the user The real-time geographic fence is constructed, and a formula is constructed as shown in the following formula (14). （14）， wherein, a geographic fence associated with the candidate item k, a preset reference radiation radius, a time decay factor, The calculation formula is shown in the following formula (15). （15）， wherein = 0.3, is a peak time for the category; The user features and the product features are tensor interacted, and a calculation formula of the tensor interaction is shown in the following formula (16): （16）， wherein the commodity-side features satisfy ; are commodity category BERT embeddings; are the price log and user historical average price ratio, ; Then, the matching degree is output by the gating scoring network according to the calculation result of the tensor interaction, and a calculation formula of the matching degree is shown in the following formula (17): （17）， wherein, denotes a Hadamard product, The calculation formula of the product is shown in the following formula (18). （18）， Finally, the first set of products is output based on the matching degree. , The calculation formula is shown in formula (19) below: （19）， wherein, is a threshold value determined by grid search. 4.The intelligent customer acquisition method based on multi-source data fusion according to claim 3, characterized in that: In the S20 step, when the product clues and the strategy information matched with the product are obtained in the e-commerce platform and the social platform, a heterogeneous content retrieval engine is used to realize accurate matching of the product clues and the strategy information, and the specific steps include the following steps: S21: generating the commodity set based on the query vector generating a unified query vector , The calculation formula is shown in the following formula (20). （20）， wherein, is a query projection network, with temporal encoding isomorphism; Incremental indexes are respectively established for the e-commerce product clues and the social strategy, and the establishment formulas are shown in the following formulas (21) and (22): （21）， （22）， A multi-stage retrieval architecture is used, and the retrieval architecture is shown in the following formula (23): （23）， wherein The calculation formula of the above-mentioned formula (24) is shown in the following formula (24). （24）。 5.The intelligent customer acquisition method based on multi-source data fusion according to claim 4, characterized in that: In the S30 step, when the mixed proportion of the product clues and the strategy information in the push content is adjusted according to the number of requests, the difference in the click rate of the product clues and the strategy is calculated by real-time monitoring of the request content of the user terminal to the push content, and a difference calculation formula is shown in the following formula (25): （25）， When the mixing ratio is readjusted; The display progress of the user terminal to the guide is acquired, and the display progress is determined according to the scroll depth of the user terminal and the total length of the guide content; a commodity lead conversion funnel of the user terminal is calculated, the commodity lead conversion funnel is determined by the ratio of the shopping cart joining request sent by the user terminal and the exposure times, and the commodity lead conversion funnel is adjusted according to the display progress and the display progress The mixing ratio is adjusted, and the adjustment formula of the mixing ratio is shown in the following formula (26): （26）， wherein, is a preset learning rate. 6.The intelligent customer acquisition method based on multi-source data fusion according to claim 5, characterized in that: A proportion decision model is constructed based on a reinforcement learning framework, real-time feature vectors of users are input into the model , , environmental context features , and behavior indicators in a sliding window , , are taken as state inputs, and commodity clue weights are output by a deep deterministic policy gradient network ; A multi-objective optimization function is established to maximize the weighted sum of expected click-through rate and conversion rate, with constraints including user experience score threshold and upper limit of commodity clue inventory consumption rate NSGA-II algorithm is used to generate a set of Pareto optimal solutions, and online strategy selection is performed through weighted Chebyshev decomposition A causal inference module is constructed, and a dual machine learning method is used to estimate the processing effect of the mixing ratio on the click rate , and the influence of the mixed variables of user purchasing power and time period is eliminated, wherein the causal graph model contains three core paths of mixing ratio to click rate, user purchasing power to mixing ratio, and user purchasing power to click rate; Deploy a federated learning architecture, each terminal device locally trains a proportion adjustment model, and then updates the global model through weighted aggregation , the gradient update process adds Gaussian noise satisfying - differential privacy ; Generate an interpretable decision report by building a feature contribution SHAP value calculation system and a decision tree rule engine, and output the logical basis for adjusting the mixed proportion, where the key decision rules include: > 0.7 and loc_type is shopping mall, assign 80% of the product lead weight, when the strategy reading completion rate is greater than 0.7 and the time period is between 00:00-06:00, assign 75% of the strategy weight. 7.The intelligent customer acquisition method based on multi-source data fusion according to claim 6, characterized in that: Acquiring a moving speed of the user terminal Collecting voice information input by the user terminal in a case where the authority is acquired Acquiring a destination of the user terminal by the voice information and the moving speed And a travel experience Specifically comprising the following steps: A10: on voice information performing sentiment analysis, extracting a sentiment feature vector , and constructing equation (27) as shown below: （27）， wherein, is a voice spectrum feature, is a semantic embedding vector, is a sentiment classification weight matrix; A20: Binding to the speed of movement and the rate of change of geographical position , calculate the travel scenario classification score , The calculation formula is shown in the following formula (28): （28）， wherein, is a scene classifier, is a time decay factor; A30: Calculate the travel experience by multi-modal fusion , The calculation formula is shown in the following formula (29): （29）， wherein, is the fusion weight, and ReLU is the activation function. A40: According to the travel experience and destination POI features , the predicted consumption intention vector , The prediction formula is shown in the following formula (30): （30）， wherein LSTM is an intent prediction network, destination feature embedding; A50: consumption intention vector , filtering the target commodity set through the gating matching network , The filtering formula of the gating matching network is shown in the following formula (31). （31）， wherein, is a product feature matrix, is a matching degree threshold value; A60: A heterogeneous content retrieval engine is used to obtain product clues and strategy information, and push content is generated by adjusting the mixed proportion. 8.The intelligent customer acquisition method based on multi-source data fusion according to claim 7, characterized in that: The travel experience The calculations also include the following steps: An out-of-travel-scenario-emotion joint space is established as shown in the following formula (32): （32）， wherein, is a scene sentiment correlation matrix, is a spatio-temporal context feature; The consumption intention is updated through a time sequence attention mechanism, and the specific formula is shown in the following formula (33): （33）， wherein, is a historical intent sequence, is an attention weight; The destination product retrieval range is optimized by using a geographic fence formula, and product matching is performed in combination with a multi-modal interest graph; A consumption intention-product feature tensor interaction space is constructed as shown in the following formula (34): （34）， wherein is a characteristic interaction operator; Optimizing push strategies using a reinforcement learning framework, travel experience features As state input, as shown in the following equation (35): （35）， wherein, is a policy network, is a travel scenario feature; The mixing ratio is dynamically adjusted according to formula (25) and formula (26), and when the instant commodity clue weight is increased when the detection result is < / < / The instant commodity clue weight is increased as shown in the following formula (36): （36）， wherein is a speed adaptation coefficient, is a reference ratio.

9. An intelligent customer acquisition system based on multi-source data fusion, characterized in that, The intelligent customer acquisition method based on multi-source data fusion of any one of claims 1-8 is used.

Images