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

Abstract

The invention belongs to the technical field of intelligent marketing, and particularly relates to an intelligent customer obtaining system and method based on multi-source data fusion. The intelligent customer obtaining method based on multi-source data fusion comprises the following steps that S10, geographical location information, e-commerce data information and social data information of a user terminal are collected under the condition that authorization of the user terminal is obtained, and the e-commerce data information comprises transaction information, browsing information and search information; the social data information comprises collection clues, collection shops and browsing information; and S20, analyzing the interest preference of the user according to the social data information. According to the scheme, through comprehensive data collection, dynamic weight distribution and a real-time feedback mechanism, the pushing accuracy, the user participation degree and the conversion rate are remarkably improved, distribution of customer resource clues is optimized, the user experience is enhanced, the problem that in the prior art, the conversion rate is low when marketing users are converted into consumption users is effectively solved, and the user experience is improved. And the customer obtaining efficiency and the user satisfaction are remarkably improved.

Description

Technical Field

[0001] This solution belongs to the field of intelligent marketing technology, and specifically relates to an intelligent customer acquisition system and method based on multi-source data fusion. Background Art

[0002] In the traditional customer acquisition model, companies or businesses mainly rely on extensive product lead delivery, offline event promotion, telephone sales and other methods to acquire customers. These methods are usually large-scale and non-directional, making it difficult to accurately locate the target customer group, resulting in a waste of product lead resources.

[0003] In recent years, with the development of big data and artificial intelligence technologies, several precision marketing methods based on multi-feature fusion have emerged. For example, Chinese patent CN111612492A discloses a method for online precision marketing based on multi-feature fusion. This method collects multiple types of user-related data, outputs each of these data types to a feature extractor of the corresponding type to extract feature vectors, then integrates these multiple feature vectors into an integrated vector, which is then output to a discriminator. Based on the discriminator's judgment, push notifications or tentative notifications are sent to the user. Online feedback from users following these push notifications or tentative notifications is then received to update the discriminator in real time, thereby achieving the effect of precisely targeting marketing users.

[0004] However, although existing technologies can obtain users' marketing positioning and identify their potential needs from e-commerce data, they fail to push information based on the real-time status of the user terminal (such as time, scene, emotion, etc.). This push method with a single data source results in a low conversion rate from marketing users to consumer users. Summary of the Invention

[0005] The purpose of this solution 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 product leads to user terminals.

[0006] To achieve the above objectives, this solution provides an intelligent customer acquisition method based on multi-source data fusion, which includes the following steps: S10: Collecting geographic location information, e-commerce data information, and social data information of the user terminal with authorization from the user terminal, wherein the e-commerce data information includes transaction information, viewing information, and search information, and the social data information includes favorite leads, favorite stores, and browsing information; S20: Analyze user interests and preferences based on social data information, analyze user consumption habits based on e-commerce data information, obtain a product that matches the user's interests and consumption habits as a first product based on the current geographic location information of the user terminal, obtain product leads and strategy information matching the first product on the e-commerce platform and social platform, and combine the product leads and strategy information into push content and send it to the user terminal; S30: Obtain the number of requests from the user terminal for product clues and strategy information in the push content, adjust the mixing ratio of product clues and strategy information in the push content according to the number of requests, and send the adjusted push content to the user terminal.

[0007] And, an intelligent customer acquisition system based on multi-source data fusion that uses an intelligent customer acquisition method based on multi-source data fusion.

[0008] The principles and technical benefits of this solution are as follows: First, by comprehensively collecting users' geographic location information, e-commerce data, and social media data, it constructs a multi-dimensional user profile that covers their behavioral habits, interests, consumption patterns, and real-time status. This comprehensive data collection and integration significantly improves the relevance and appeal of pushed content, ensuring that pushed content closely matches users' immediate needs and interests, thereby increasing push accuracy.

[0009] Secondly, this solution can identify changing user needs across time and context. For example, a user may be more interested in breakfast-related content in the morning, but more inclined toward relaxation or entertainment-related content in the evening. Sentiment analysis can further refine user preferences at specific moments, providing more personalized content push notifications. This dynamic user profile adjustment allows the solution to reflect evolving user interests and needs in real time, rather than relying solely on historical data. For example, if a user has recently frequently browsed a certain category of products, the solution will increase the weight of that product information in the user profile to ensure that the pushed content is more aligned with the user's current interests. Furthermore, the solution dynamically assigns weight to pushed content based on the influence of different data sources in the user profile on the user's purchase intentions. This means that the solution not only considers historical user behavior but also analyzes factors such as location, social interactions, and real-time feedback to determine which information has the greatest impact on a user's purchase decision and adjusts the display priority of product leads and content accordingly. This personalized push notification strategy significantly improves user engagement and satisfaction. Users perceive content as more relevant to their actual needs and interests, making them more willing to interact with the pushed content, such as clicking, commenting, saving, and sharing. This active user engagement not only improves the user experience but also increases user loyalty and engagement with the system. These factors work together to increase conversion rates—the percentage of users who convert from potential customers to actual buyers. By precisely targeting product leads and services of interest to users, this solution effectively boosts sales and improves the ROI of marketing campaigns. Furthermore, the dynamic weighting mechanism helps optimize the allocation of product leads, ensuring that lead budgets are used by the user groups most likely to convert, reducing resource waste and improving marketing efficiency.

[0010] Furthermore, this solution collects user feedback on push content (such as click-through rate, conversion rate, dwell time, and close rate) to evaluate push effectiveness in real time and dynamically adjust the weighting and mix of push content based on user feedback. This feedback mechanism ensures that push strategies consistently align with user interests and needs, optimizes resource allocation, enhances user experience, and improves user acceptance and satisfaction with product leads.

[0011] In summary, this solution significantly improves the accuracy of push notifications, user engagement, and conversion rates through comprehensive data collection, dynamic weight allocation, and real-time feedback mechanisms. It also optimizes the allocation of product lead resources and enhances user experience, effectively solving the problem of low conversion rates of marketing users to consumer users in existing technologies and significantly improving customer acquisition efficiency and user satisfaction.

[0012] Furthermore, between step S10 and step S20, vectorization processing of the e-commerce data information and the social data information is also included, which specifically includes the following steps: S11: For transaction information, a time series embedding method is used to construct a transaction information vector of a three-dimensional feature vector. The construction formula is shown in the following formula (1): (1), in, For the recent Transaction amount sequence, is mean pooling, Calculate the standard deviation. Before taking Fourier transform low-frequency components; For viewing information, the viewing information vector is constructed by embedding the product with attention weight. The embedding formula for constructing the viewing information vector is shown in the following formula (2): (2), in, The page dwell time, is the preset temperature coefficient, Product description text; For search information, query-click dual-channel encoding is used to construct the search information vector. The encoding formula for constructing the search information vector is shown in the following formula (3): (3), in, is the search term sequence, To click on the product image, Represents vector concatenation; S12: Construct a store location vector for the locations of the favorite stores through spatiotemporal joint coding. The coding formula for constructing the store location vector is shown in the following formula (4): (4), in, represents the Hadamard product, To attach text to the punch card; The attention information is aggregated using graph neural networks to construct a collection clue vector. The aggregation formula for constructing the collection clue vector is shown in the following formula (5): (5), in is the user-bloggers adjacency matrix, is the node feature matrix; Browsing information is constructed into a browsing information vector through heterogeneous sequence modeling. The modeling formula for constructing the browsing information vector is shown in the following formula (6): (6), in Encode the content type, Embed for content snippets.

[0013] By vectorizing e-commerce data information and social data information between steps S10 and S20, not only can the data processing efficiency be improved, but also the feature expression ability can be enhanced. This processing method enables transaction information, viewing information, and search information to be converted into three-dimensional feature vectors, allowing for more efficient analysis. At the same time, by vectorizing social data such as check-in information, follow-up information, and browsing information, users' social behaviors and preferences can be captured more accurately, improving the performance of this solution. In addition, vectorization processing supports complex data analysis, promotes data fusion, and improves the generalization ability of the model. Moreover, this solution supports real-time analysis and decision-making, which is particularly important for application scenarios that require rapid response.

[0014] Furthermore, in step S20, when analyzing user consumption habits based on e-commerce data information, a time-series dynamic graph convolutional network is used to model consumption characteristics based on transaction information vectors. The calculation process includes the following steps: S201: Extract the transaction pattern using the following formula (7): (7), in , , , For the The layer can be trained with parameters, and then the final consumption stability score is calculated based on the transaction pattern extraction results. ; S202: Modeling viewing preference of viewing information vectors through cross-attention mechanism; S203: clustering the search information vectors into search intents using a Gaussian mixture model, and identifying dominant consumption intentions based on the clustering results; When analyzing user interest preferences based on social data information, a multimodal interest graph is constructed based on store location vectors, collection clue vectors, and browsing information vectors. The specific steps include: S204: Perform spatiotemporal feature fusion on the store location vectors through bilinear analysis to classify the check-in scenes. S205: Calculate the collection clue vector based on the improved weight diffusion algorithm of PageRank to obtain the social influence propagation; S206: Calculate the store location vector using a hierarchical attention network to obtain a representation of the browsed content.

[0015] In step S20, by modeling e-commerce data using a temporal dynamic graph convolutional network, analyzing social data using a cross-attention mechanism and a Gaussian mixture model, and constructing a multimodal interest graph, the accuracy of analyzing user consumption habits and interest preferences can be significantly improved. First, the temporal dynamic graph convolutional network can extract transaction patterns, calculate consumption stability scores, model viewing preferences for viewing information vectors using a cross-attention mechanism, and cluster search intent for search information vectors using a Gaussian mixture model. Second, the construction of the multimodal interest graph involves bilinearly fusing spatiotemporal features of store location vectors, calculating collection clue vectors using a weight diffusion algorithm improved on PageRank, and calculating browsing information vectors using a hierarchical attention network to obtain a representation of browsing content. These steps not only build accurate user profiles, but also provide personalized recommendations, predict consumer behavior, analyze social influence, and process multimodal data, thereby improving user experience and business results in multiple aspects.

[0016] Furthermore, in the step S20, when obtaining a product that matches the user's interest preference and the user's consumption habits as the first product based on the user's current location information, To build a real-time geo-fence, the formula is as follows (14): (14), in, Candidate products associated geofences, is the preset reference radiation radius, is the time decay factor, The calculation formula is shown in the following formula (15): (15), in , It is the peak period for the category; The user features and product features are tensor-interacted. The calculation formula of tensor interaction is shown in the following formula (16): (16), Among them, product side features satisfy ; BERT embedding for product categories; is the ratio of the price logarithm to the user's historical average price, ; Then, the matching degree is output through the gated scoring network based on the calculation results of the tensor interaction. The matching degree calculation formula is shown in the following formula (17): (17), in, represents the Hadamard product, The calculation formula is shown in the following formula (18): (18), Finally, output the first product set based on the matching degree , The calculation formula is shown in the following formula (19): (19), in, is the threshold determined by grid search.

[0017] This solution significantly improves the accuracy and personalization of product recommendations by building real-time geofencing based on the user's current location and combining advanced technologies such as temporal dynamic graph convolutional networks, cross-attention mechanisms, and Gaussian mixture models. First, real-time geofencing ensures that recommended products are closely aligned with the user's location, enhancing the practicality of recommendations. Second, by extracting transaction patterns, calculating consumption stability scores, and modeling viewing preferences based on viewing information vectors using a cross-attention mechanism, it enables in-depth analysis of users' e-commerce data, enabling a more accurate understanding of their consumption habits. Furthermore, using a Gaussian mixture model to cluster search information vectors based on search intent helps identify users' dominant consumption intentions, further enhancing the targeting of recommendations. Regarding social data analysis, a multimodal interest graph is constructed, encompassing spatiotemporal feature fusion, social influence propagation analysis, and the extraction of browsed content representations, to comprehensively capture users' interests and preferences. Finally, user and product features are tensor-interacted, and a gated scoring network outputs a matching score, ensuring that recommended products closely match users' interests and consumption habits.

[0018] Furthermore, in step S20, when obtaining product clues and strategy information matching the product on the e-commerce platform and the social platform, a heterogeneous content search engine is used to achieve accurate matching of the product clues and strategy information, which specifically includes the following steps: S21: Based on product collection Generate unified query vector , The calculation formula is shown in the following formula (20): (20), in, To query the projection network, and The temporal coding isomorphism of Incremental indexes are established for e-commerce product leads and social strategies respectively, and the formulas are as shown in the following formulas (21) and (22): (twenty one), (twenty two), A multi-stage retrieval architecture is adopted, and the retrieval architecture is shown in the following formula (23): (twenty three), in, The calculation formula is shown in formula (24): (twenty four).

[0019] By employing a heterogeneous content retrieval engine to accurately match product leads and guide information across 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 collection. Then, incremental indexes are established for both e-commerce product leads and social media guides, respectively, and retrieval is performed using a multi-stage retrieval architecture. These steps ensure the efficiency and accuracy of the retrieval process. Furthermore, by calculating similarity scores between the query vector and the product leads and guide content, matching results can be further optimized, ensuring that recommended product leads and guide information are highly relevant to users' interests and needs. This approach 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 recommendations while browsing products.

[0020] 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 strategy information is calculated by real-time monitoring of the user terminal's request content for the push content. The difference calculation formula is shown in the following formula (25): (25), when Readjust the mixing ratio when Obtain the display progress of the strategy by the user terminal, which is determined by the scrolling depth of the user terminal and the total length of the strategy content; calculate the product lead conversion funnel of the user terminal, which is determined by the ratio of the add-to-cart request sent by the user terminal to the number of exposures, and calculate the product lead conversion funnel according to the product lead conversion funnel. and display progress Adjust the mixing ratio. The adjustment formula of the mixing ratio is shown in the following formula (26): (26), in, is the preset learning rate.

[0021] By calculating click-through rate differences and product lead conversion funnels, this solution can adjust the ratio of product leads and strategy information in push content in real time to better meet user needs and increase user engagement. Based on the user's progress in displaying strategy information and the conversion effect of product leads, this solution can more accurately push information most relevant to the user's current interests and behaviors, thereby improving the relevance and appeal of pushed content. By dynamically adjusting the ratio of product leads and strategy information, the display effect of product leads can be optimized, and the click-through rate and conversion rate of product leads can be increased, thereby improving the ROI of product lead owners. By providing push content that better suits users' current interests and needs, user satisfaction and loyalty can be enhanced, thereby improving user retention and activity. By monitoring and analyzing user behavior data in real time, this solution can provide data support to the operations team, helping them make more accurate decisions to optimize push strategies and improve overall marketing effectiveness.

[0022] Furthermore, a proportional decision model is constructed based on the reinforcement learning framework to transform the user's real-time feature vector 、 、 Environmental context features 、 and behavioral indicators within the sliding window 、 、 As state input, the product clue weight is output through the deep deterministic policy gradient network ; Establish a multi-objective optimization function to maximize the weighted sum of expected click-through rate and conversion rate, with constraints including user experience score threshold and the upper limit of product lead inventory consumption rate , using the NSGA-II algorithm to generate the Pareto optimal solution set and perform online strategy selection through weighted Chebyshev decomposition; building a causal reasoning module and using a dual machine learning method to estimate the treatment effect of the mixing ratio on the click-through rate , eliminating the influence of confounding variables such as user purchasing power and time period. The causal graph model includes three core paths: mixing ratio to click-through rate, user purchasing power to mixing ratio, and user purchasing power to click-through rate; deploying a federated learning architecture, each terminal device locally trains the ratio adjustment model, and then updates the global model through weighted aggregation. , the gradient update process adds -Gaussian noise for differential privacy Generate an explainable decision report, and output the logical basis for adjusting the mixing ratio by building a feature contribution SHAP value calculation system and a decision tree rule engine. The key decision rules include: When loc_type is shopping mall, 80% of product lead weight is assigned. And time period Assign 75% strategy weight.

[0023] By applying a reinforcement learning framework to a proportional decision-making model and integrating real-time user feature vectors, environmental contextual features, and behavioral indicators within a sliding window, the accuracy and user engagement of product lead push notifications can be significantly improved. A deep deterministic policy gradient network (DDPG) is used 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 an upper bound on product lead inventory depletion rates, the NSGA-II algorithm is used to generate a Pareto-optimal solution set. Online policy selection is performed through weighted Chebyshev decomposition, resulting in a more optimal product lead placement strategy. Furthermore, a causal inference module is constructed, utilizing a dual machine learning approach to estimate the effect of mixing ratio on click-through rate, eliminating the influence of confounding variables such as user purchasing power and time of day, ensuring accurate evaluation of product lead effectiveness. A federated learning architecture is deployed. After each terminal device locally trains a proportional adjustment model, the global model is updated through weighted aggregation. Gaussian noise that satisfies differential privacy is added to the gradient update process, protecting user privacy while improving model generalization. Finally, an explainable decision report is generated. By building a feature contribution SHAP value calculation system and a decision tree rule engine, the logical basis for adjusting the mixed ratio is output, making key decision rules clearer. For example, the weights of product leads and strategies are dynamically adjusted according to user location and time period, thereby protecting user privacy while improving the personalization and effectiveness of product lead push.

[0024] Further, obtain the moving speed of the user terminal , collect the voice information recorded by the user terminal with permission , obtain the destination of the user terminal through voice information and moving speed and travel experience , specifically including the following steps: A10: Voice messages Perform sentiment analysis and extract sentiment feature vectors , formula (27) is constructed as follows: (27), in, is the speech spectrum feature, is the semantic embedding vector, is the sentiment classification weight matrix; A20: Combined movement speed and geographic location change rate , calculate the travel scenario classification score , The calculation formula is shown in formula (28): (28), in, is the scene classifier, is the time decay factor; A30: Computing the travel experience through multimodal fusion , The calculation formula is shown in formula (29): (29), in, is the fusion weight, ReLU is the normalization function; A40: Based on travel experience and destination POI features , predict consumption intention vector , The prediction formula is shown in the following formula (30): (30), Among them, LSTM is the intention prediction network. Embedding for destination features; A50: Based on consumption intention vector , filter the target product set through the gated matching network , The screening formula is shown in the following formula (31): (31), in, is the product feature matrix, is the matching threshold; A60: Uses a heterogeneous content retrieval engine to obtain product leads and strategy information, and generates push content based on a mixed ratio adjustment method.

[0025] Furthermore, the travel experience The calculation also includes the following steps: The travel scenario-emotion joint space is established as shown in the following formula (32): (32), in, is the scene emotion correlation matrix, is the spatiotemporal context feature; The consumption intention is updated through the temporal attention mechanism, as shown in the following formula (33): (33), in, For the historical intention sequence, is the attention weight; Use geo-fencing formulas to optimize the search scope of destination products and combine multimodal interest graphs for product matching; Construct the consumption intention-product feature tensor interaction space shown in the following formula (34): (34), in, is the feature interaction operator; Use reinforcement learning framework to optimize push strategy and integrate travel experience characteristics As the state input, it is shown in the following formula (35): (35), in, For the strategy network, is the travel scenario feature; According to formula (25) and formula (26), the mixing ratio is adjusted dynamically. When increasing the weight of instant product clues, the formula is as follows: (36), in, is the speed adaptation coefficient, is the base ratio.

[0026] This solution creatively combines real-time identification of user travel scenarios with analysis of long-term consumption habits, which can not only grasp the user's stable preferences, but also keenly capture immediate changes in demand. This dual cognitive mechanism allows the pushed content to maintain respect for the user's historical preferences while flexibly adapting to various sudden scenario needs, realizing truly personalized services. It is particularly worth noting that this solution has built an adaptive intelligent push ecosystem through the deep integration of environmental perception and content decision-making. This solution can automatically optimize the information presentation method and content combination strategy based on the user's mobility status, surrounding environment, and emotional changes. This dynamic adjustment mechanism not only greatly improves the relevance of information, but also significantly improves business conversion efficiency while protecting the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a flowchart of an intelligent customer acquisition method based on multi-source data fusion in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The following will clearly and completely describe the concept and technical effects of the present invention in conjunction with the embodiments to fully understand the purpose, features and effects of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative work are all within the scope of protection of the present invention: like Figure 1 As shown in FIG, the intelligent customer acquisition method based on multi-source data fusion includes the following steps: S10: Collecting geographic location information, e-commerce data information, and social data information of the user terminal with authorization from the user terminal, wherein the e-commerce data information includes transaction information, viewing information, and search information, and the social data information includes favorite leads, favorite stores, and browsing information; S20: Analyze user interests and preferences based on social data information, analyze user consumption habits based on e-commerce data information, obtain a product that matches the user's interests and consumption habits as a first product based on the current geographic location information of the user terminal, obtain product leads and strategy information matching the first product on the e-commerce platform and social platform, and combine the product leads and strategy information into push content and send it to the user terminal; S30: Obtain the number of requests from the user terminal for product clues and strategy information in the push content, adjust the mixing ratio of product clues and strategy information in the push content according to the number of requests, and send the adjusted push content to the user terminal.

[0029] The process between step S10 and step S20 also includes vectorizing the e-commerce data information and the social data information, which specifically includes the following steps: S11: For transaction information, a time series embedding method is used to construct a transaction information vector of a three-dimensional feature vector. The construction formula is shown in the following formula (1): (1), in, For the recent Transaction amount sequence, is mean pooling, Calculate the standard deviation. Before taking Fourier transform low-frequency components; For viewing information, the viewing information vector is constructed by embedding the product with attention weight. The embedding formula for constructing the viewing information vector is shown in the following formula (2): (2), in, The page dwell time, is the preset temperature coefficient, Product description text; For search information, query-click dual-channel encoding is used to construct the search information vector. The encoding formula for constructing the search information vector is shown in the following formula (3): (3), in, is the search term sequence, To click on the product image, Represents vector concatenation; S12: Construct a store location vector for the locations of the favorite stores through spatiotemporal joint coding. The coding formula for constructing the store location vector is shown in the following formula (4): (4), in, represents the Hadamard product, To attach text to the punch card; The attention information is aggregated using graph neural networks to construct a collection clue vector. The aggregation formula for constructing the collection clue vector is shown in the following formula (5): (5), in is the user-bloggers adjacency matrix, is the node feature matrix; Browsing information is constructed into a browsing information vector through heterogeneous sequence modeling. The modeling formula for constructing the browsing information vector is shown in the following formula (6): (6), in Encode the content type, Embed for content snippets.

[0030] In step S20, when analyzing user interest preferences based on social data information, a time-series dynamic graph convolutional network is used to model consumption characteristics based on transaction information vectors. The calculation process includes the following steps: S201: Extract the transaction pattern using the following formula (7): (7), in , , , For the The layer can be trained with parameters, and then the final consumption stability score is calculated based on the transaction pattern extraction results. , The calculation formula is as follows: (8); S202: View preference modeling is performed on the view information vector through the cross attention mechanism. The modeling formula is shown in the following formula (9): (9), in, , and is the projection matrix, is the preset vector dimension; S203: Use the Gaussian mixture model to cluster the search information vector into search intents. The clustering formula is shown in the following formula (10): (10), According to the clustering results Identify the dominant consumer intention; When analyzing user interest preferences based on social data information, a multimodal interest graph is constructed based on store location vectors, collection clue vectors, and browsing information vectors. The specific steps include: S204: Perform spatiotemporal feature fusion on the store location vector through bilinear method to classify the check-in scene. The fusion formula is shown in the following formula (11): (11), in and is the spatiotemporal interaction matrix; S205: Calculate the collection clue vector based on the improved weight diffusion algorithm of PageRank to obtain the social influence propagation. The calculation formula is shown in the following formula (12): (12), in , For users A collection of bloggers you follow; S206: Use the hierarchical attention network to calculate the store location vector and obtain the browsing content representation. The calculation formula is shown in the following formula (13): (13), in, For the The first browse sequence Fragments are embedded.

[0031] Among them, in step S20, when obtaining a product that matches the user's interest preference and user consumption habits as the first product based on the user's current location information, To build a real-time geo-fence, the formula is as shown in the following formula (14): (14), in, Candidate products associated geofences, is the preset reference radiation radius, is the time decay factor, The calculation formula is shown in the following formula (15): (15), in , It is the peak period for the category; The user features and product features are tensor-interacted. The calculation formula of tensor interaction is shown in the following formula (16): (16), Among them, product side features satisfy ; BERT embedding for product categories; is the ratio of the price logarithm to the user's historical average price, ; Then, the matching degree is output through the gated scoring network based on the calculation results of the tensor interaction. The matching degree calculation formula is shown in the following formula (17): (17), in, represents the Hadamard product, The calculation formula is shown in the following formula (18): (18), Finally, output the first product set based on the matching degree , The calculation formula is shown in the following formula (19): (19), in, is the threshold determined by grid search.

[0032] In step S20, when obtaining product clues and strategy information matching the product on the e-commerce platform and the social platform, a heterogeneous content search engine is used to achieve accurate matching of product clues and strategy information, which specifically includes the following steps: S21: Based on product collection Generate unified query vector , The calculation formula is shown in the following formula (20): (20), in, To query the projection network, and The temporal coding isomorphism of Incremental indexes are established for e-commerce product leads and social strategies respectively, and the formulas are as shown in the following formulas (21) and (22): (twenty one), (twenty two), A multi-stage retrieval architecture is adopted, and the retrieval architecture is shown in the following formula (23): (twenty three), in, The calculation formula is shown in formula (24): (twenty four).

[0033] Among them, in step S30, when adjusting the mixing ratio of product clues and strategy information in the push content according to the number of requests, the difference in click-through rate between product clues and strategy information is calculated by real-time monitoring of the user terminal's request content for the push content. The difference calculation formula is shown in the following formula (25): (25), when Readjust the mixing ratio when Obtain the display progress of the strategy by the user terminal, which is determined by the scrolling depth of the user terminal and the total length of the strategy content; calculate the product lead conversion funnel of the user terminal, which is determined by the ratio of the add-to-cart request sent by the user terminal to the number of exposures, and calculate the product lead conversion funnel according to the product lead conversion funnel. and display progress Adjust the mixing ratio. The adjustment formula of the mixing ratio is shown in the following formula (26): (26), in, is the preset learning rate.

[0034] Among them, a proportional decision model is built based on the reinforcement learning framework to transform the user's real-time feature vector 、 、 Environmental context features 、 and behavioral indicators within the sliding window 、 、 As state input, the product clue weight is output through the deep deterministic policy gradient network ; Establish a multi-objective optimization function to maximize the weighted sum of expected click-through rate and conversion rate, with constraints including user experience score threshold and the upper limit of product lead inventory consumption rate ,NSGA-II algorithm is used to generate the Pareto optimal solution set, and online strategy selection is performed through weighted Chebyshev decomposition; Construct a causal inference module and use a dual machine learning approach to estimate the treatment effect of mixture ratio on click-through rate , eliminating the influence of confounding variables such as user purchasing power and time period. The causal graph model includes three core paths: mixing ratio to click-through rate, user purchasing power to mixing ratio, and user purchasing power to click-through rate; Deploy a federated learning architecture, and each terminal device adjusts the model through local training ratio, and then updates the global model through weighted aggregation. , the gradient update process adds -Gaussian noise for differential privacy ; Generate an explainable decision report, and output the logical basis for adjusting the mixing ratio by building a feature contribution SHAP value calculation system and a decision tree rule engine. The key decision rules include: When loc_type is shopping mall, 80% of product lead weight is assigned. And time period Assign 75% strategy weight.

[0035] Among them, obtain the moving speed of the user terminal , collect the voice information recorded by the user terminal with permission , obtain the destination of the user terminal through voice information and moving speed and travel experience , specifically including the following steps: A10: Voice messages Perform sentiment analysis and extract sentiment feature vectors , formula (27) is constructed as follows: (27), in, is the speech spectrum feature, is the semantic embedding vector, is the sentiment classification weight matrix; A20: Combined movement speed and geographic location change rate , calculate the travel scenario classification score , The calculation formula is shown in formula (28): (28), in, is the scene classifier, is the time decay factor; A30: Computing the travel experience through multimodal fusion , The calculation formula is shown in formula (29): (29), in, is the fusion weight, ReLU is the normalization function; A40: Based on travel experience and destination POI features , predict consumption intention vector , The prediction formula is shown in the following formula (30): (30), Among them, LSTM is the intention prediction network. Embedding for destination features; A50: Based on consumption intention vector , filter the target product set through the gated matching network , The screening formula is shown in the following formula (31): (31), in, is the product feature matrix, is the matching threshold; A60: Uses a heterogeneous content retrieval engine to obtain product leads and strategy information, and generates push content based on a mixed ratio adjustment method.

[0036] Specifically, the travel experience The calculation also includes the following steps: The travel scenario-emotion joint space is established as shown in the following formula (32): (32), in, is the scene emotion correlation matrix, is the spatiotemporal context feature; The consumption intention is updated through the temporal attention mechanism, as shown in the following formula (33): (33), in, For the historical intention sequence, is the attention weight; Use geo-fencing formulas to optimize the search scope of destination products and combine multimodal interest graphs for product matching; Construct the consumption intention-product feature tensor interaction space shown in the following formula (34): (34), in, is the feature interaction operator; Use reinforcement learning framework to optimize push strategy and integrate travel experience characteristics As the state input, it is shown in the following formula (35): (35), in, For the strategy network, is the travel scenario feature; According to formula (25) and formula (26), the mixing ratio is adjusted dynamically. When increasing the weight of instant product clues, the formula is as follows: (36), in, is the speed adaptation coefficient, is the base ratio.

[0037] This embodiment also includes an intelligent customer acquisition system based on multi-source data fusion that uses an intelligent customer acquisition method based on multi-source data fusion.

[0038] In specific implementation, an intelligent customer acquisition system was deployed within the app of Gouyi.com, a large e-commerce platform. The following describes the personalized push notification process using user Zhang as an example. Zhang is a 25-year-old office worker in a first-tier city.

[0039] Through authorization, the system obtained Zhang's real-time location, which was located in the Guomao business district in Chaoyang District, Beijing, at 20:30 on a weekday. In terms of e-commerce data, Zhang's average monthly consumption in the past three months reached 3,800 yuan, and the high-frequency purchase categories were concentrated in "fitness equipment" and "fast food". On that day, Zhang looked at the yoga mat for 120 seconds, and also looked at protein powder for 45 seconds. In addition, Zhang searched for "portable fitness equipment" and clicked on three folding dumbbell products. Judging from social data, Zhang posted a check-in information on the social platform that day, "Day 5 of checking in at the Guomao gym." He follows 5 fitness bloggers and 3 food bloggers, and 70% of the content he browsed recently is fitness tutorials and 30% is reviews of new products in convenience stores.

[0040] In the vectorization processing (S11-S12) phase, the transaction information vector Generated by formula (1), it is found that Zhang's periodic consumption peak is every Friday night. View information vector Calculated by formula (2), the attention weight of yoga mat is as high as 0.73. Store location vector By encoding with formula (4), the “gym” scene label is successfully identified.

[0041] In the product matching process, the geo-fencing formula (14) was used to activate merchants within 3 km of the Guomao business district, and then the "smart folding dumbbells" with a matching degree of 0.82 and the "low-calorie protein bar" with a matching degree of 0.76 were selected. After that, the features were fused through the tensor interaction formula (16), and the final matching product set was output through the gated scoring network formula (17). .

[0042] In terms of content retrieval, the heterogeneous search engine obtains relevant content based on formula (23). The product clue is "Smart dumbbells 20% off for a limited time", with an estimated CTR of 5.2%; the guide is "Gym Equipment Buying Guide for Beginners", with an interaction index of 4.8 / 5.0.

[0043] In the mixed push link, according to Zhang’s high consumption habits, the initial ratio is set to By generating adversarial networks to output fused content, the top shows product clues with a "Buy Now" button, and the bottom embeds a selected guide paragraph.

[0044] After receiving the push notification, user Zhang clicked on the product clue but did not purchase it. Instead, he read the guide in its entirety, with a scroll depth of 92%. 0.25 lower than the user average. Click rate difference This value triggers the adjustment of formula (26). After calculation, the new ratio Based on this, the guide section was increased when the content was pushed the next day, and new video content on "fitness diet matching" was added.

[0045] The above is only an embodiment of the present invention, and the common knowledge such as the specific structure and characteristics of the scheme is not described in detail here. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. An intelligent customer acquisition method based on multi-source data fusion, characterized by: The following steps are involved: S10: Collecting geographic location information, e-commerce data information, and social data information of the user terminal with authorization from the user terminal, wherein the e-commerce data information includes transaction information, viewing information, and search information, and the social data information includes favorite leads, favorite stores, and browsing information; S20: Analyze user interests and preferences based on social data information, analyze user consumption habits based on e-commerce data information, obtain a product that matches the user's interests and consumption habits as a first product based on the current geographic location information of the user terminal, obtain product leads and strategy information matching the first product on the e-commerce platform and social platform, and combine the product leads and strategy information into push content and send it to the user terminal; S30: Obtain the number of requests from the user terminal for product clues and strategy information in the push content, adjust the mixing ratio of product clues and strategy information in the push content according to the number of requests, and send the adjusted push content to the user terminal.

2. The intelligent customer acquisition method based on multi-source data fusion according to claim 1 is characterized by: Between step S10 and step S20, the e-commerce data information and the social data information are vectorized, which specifically includes the following steps: S11: For transaction information, a time series embedding method is used to construct a transaction information vector of a three-dimensional feature vector. The construction formula is shown in the following formula (1): （1）， in, For the recent Transaction amount sequence, is mean pooling, Calculate the standard deviation. Before taking Fourier transform low-frequency components; For viewing information, the viewing information vector is constructed by embedding the product with attention weight. The embedding formula for constructing the viewing information vector is shown in the following formula (2): （2）， in, The page dwell time, is the preset temperature coefficient, Product description text; For search information, query-click dual-channel encoding is used to construct the search information vector. The encoding formula for constructing the search information vector is shown in the following formula (3): （3）， in, is the search term sequence, To click on the product image, Represents vector concatenation; S12: Construct a store location vector for the locations of the favorite stores through spatiotemporal joint coding. The coding formula for constructing the store location vector is shown in the following formula (4): （4）， in, represents the Hadamard product, Accompanying text for store location; The attention information is aggregated using graph neural networks to construct a collection clue vector. The aggregation formula for constructing the collection clue vector is shown in the following formula (5): （5）， in is the user-bloggers adjacency matrix, is the node feature matrix; Browsing information is constructed into a browsing information vector through heterogeneous sequence modeling. The modeling formula for constructing the browsing information vector is shown in the following formula (6): （6）， in Encode the content type, Embed for content snippets.

3. The intelligent customer acquisition method based on multi-source data fusion according to claim 2 is characterized by: In step S20, when analyzing user consumption habits based on e-commerce data information, a time-series dynamic graph convolutional network is used to model consumption characteristics based on transaction information vectors. The calculation process includes the following steps: S201: Extract the transaction pattern using the following formula (7): （7）， in , , , For the The layer can be trained with parameters, and then the final consumption stability score is calculated based on the transaction pattern extraction results. ; S202: Modeling viewing preference of viewing information vectors through cross-attention mechanism; S203: Using a Gaussian mixture model to cluster the search information vector into search intents, and identifying dominant consumption intentions based on the clustering results; When analyzing user interest preferences based on social data information, a multimodal interest graph is constructed based on store location vectors, collection clue vectors, and browsing information vectors. The specific steps include: S204: Performing spatiotemporal feature fusion on the store location vectors through bilinear analysis to classify the store location scenes; S205: Calculate the collection clue vector based on the improved weight diffusion algorithm of PageRank to obtain the social influence propagation; S206: Calculate the store location vector using a hierarchical attention network to obtain a representation of the browsed content.

4. The intelligent customer acquisition method based on multi-source data fusion according to claim 3 is characterized by: In the step S20, when obtaining a product that matches the user's interest preferences and consumption habits as the first product based on the user's current location information, To build a real-time geo-fence, the formula is as follows (14): （14）， in, Candidate products associated geofences, is the preset reference radiation radius, is the time decay factor, The calculation formula is shown in the following formula (15): （15）， in , It is the peak period for the category; The user features and product features are tensor-interacted. The calculation formula of tensor interaction is shown in the following formula (16): （16）， Among them, product side features satisfy ; BERT embedding for product categories; is the ratio of the price logarithm to the user's historical average price, ; Then, the matching degree is output through the gated scoring network based on the calculation results of the tensor interaction. The matching degree calculation formula is shown in the following formula (17): （17）， in, represents the Hadamard product, The calculation formula is shown in the following formula (18): （18）， Finally, output the first product set based on the matching degree , The calculation formula is shown in the following formula (19): （19）， in, is the threshold determined by grid search.

5. The intelligent customer acquisition method based on multi-source data fusion according to claim 4 is characterized by: In step S20, when obtaining product clues and strategy information matching the product on the e-commerce platform and the social platform, a heterogeneous content search engine is used to achieve accurate matching of the product clues and strategy information, which specifically includes the following steps: S21: Based on product collection Generate unified query vector , The calculation formula is shown in the following formula (20): （20）， in, To query the projection network, and The temporal coding isomorphism of Incremental indexes are established for e-commerce product leads and social strategies respectively, and the formulas are as shown in the following formulas (21) and (22): （21）， （22）， A multi-stage retrieval architecture is adopted, and the retrieval architecture is shown in the following formula (23): （23）， in, The calculation formula is shown in formula (24): （24）。 6. The intelligent customer acquisition method based on multi-source data fusion according to claim 5 is characterized by: In step S30, when adjusting the mixing ratio of product clues and strategy information in the push content according to the number of requests, the difference in click-through rate between product clues and strategy information is calculated by real-time monitoring of the user terminal's request content for the push content. The difference calculation formula is shown in the following formula (25): （25）， when Readjust the mixing ratio when Obtain the display progress of the strategy by the user terminal, which is determined by the scrolling depth of the user terminal and the total length of the strategy content; calculate the product lead conversion funnel of the user terminal, which is determined by the ratio of the add-to-cart request sent by the user terminal to the number of exposures, and calculate the product lead conversion funnel according to the product lead conversion funnel. and display progress Adjust the mixing ratio. The adjustment formula of the mixing ratio is shown in the following formula (26): （26）， in, is the preset learning rate.

7. The intelligent customer acquisition method based on multi-source data fusion according to claim 6 is characterized by: Based on the reinforcement learning framework, a proportional decision model is built to transform the user's real-time feature vector 、 、 Environmental context features 、 and behavioral indicators within the sliding window 、 、 As state input, the product clue weight is output through the deep deterministic policy gradient network ; Establish a multi-objective optimization function to maximize the weighted sum of expected click-through rate and conversion rate, with constraints including user experience score threshold and the upper limit of product lead inventory consumption rate ,NSGA-II algorithm is used to generate the Pareto optimal solution set, and online strategy selection is performed through weighted Chebyshev decomposition; Construct a causal inference module and use a dual machine learning approach to estimate the treatment effect of mixture ratio on click-through rate , eliminating the influence of confounding variables such as user purchasing power and time period. The causal graph model includes three core paths: mixing ratio to click-through rate, user purchasing power to mixing ratio, and user purchasing power to click-through rate; Deploy a federated learning architecture, and each terminal device adjusts the model through local training ratio, and then updates the global model through weighted aggregation. , the gradient update process adds -Gaussian noise for differential privacy ; Generate an explainable decision report, and output the logical basis for adjusting the mixing ratio by building a feature contribution SHAP value calculation system and a decision tree rule engine. The key decision rules include: When loc_type is shopping mall, 80% of product lead weight is assigned. And time period Assign 75% strategy weight.

8. The intelligent customer acquisition method based on multi-source data fusion according to claim 7 is characterized by: Get the moving speed of the user terminal , collect the voice information recorded by the user terminal with permission , obtain the destination of the user terminal through voice information and moving speed and travel experience , specifically including the following steps: A10: Voice messages Perform sentiment analysis and extract sentiment feature vectors , formula (27) is constructed as follows: （27）， in, is the speech spectrum feature, is the semantic embedding vector, is the sentiment classification weight matrix; A20: Combined movement speed and geographic location change rate , calculate the travel scenario classification score , The calculation formula is shown in formula (28): （28）， in, is the scene classifier, is the time decay factor; A30: Computing the travel experience through multimodal fusion , The calculation formula is shown in formula (29): （29）， in, is the fusion weight, ReLU is the normalization function; A40: Based on travel experience and destination POI features , predict consumption intention vector , The prediction formula is shown in the following formula (30): （30）， Among them, LSTM is the intention prediction network. Embedding for destination features; A50: Based on consumption intention vector , filter the target product set through the gated matching network , The screening formula is shown in the following formula (31): （31）， in, is the product feature matrix, is the matching threshold; A60: Uses a heterogeneous content retrieval engine to obtain product leads and strategy information, and generates push content based on a mixed ratio adjustment method.

9. The intelligent customer acquisition method based on multi-source data fusion according to claim 8, characterized in that: The travel experience The calculation also includes the following steps: The travel scenario-emotion joint space is established as shown in the following formula (32): （32）， in, is the scene emotion correlation matrix, is the spatiotemporal context feature; The consumption intention is updated through the temporal attention mechanism, as shown in the following formula (33): （33）， in, For the historical intention sequence, is the attention weight; Use geo-fencing formulas to optimize the search scope of destination products and combine multimodal interest graphs for product matching; Construct the consumption intention-product feature tensor interaction space shown in the following formula (34): （34）， in, is the feature interaction operator; Use reinforcement learning framework to optimize push strategy and integrate travel experience characteristics As the state input, it is shown in the following formula (35): （35）， in, For the strategy network, is the travel scenario feature; According to formula (25) and formula (26), the mixing ratio is adjusted dynamically. When increasing the weight of instant product clues, the formula is as follows: （36）， in, is the speed adaptation coefficient, is the base ratio.

10. Intelligent customer acquisition system based on multi-source data fusion, characterized by: The intelligent customer acquisition method based on multi-source data fusion described in any one of claims 1 to 9 is used.

Images