An integrated platform for tracking omni-channel consumer behavior and market research

Abstract

The application discloses a kind of full-channel consumer behavior tracking and market survey integrated platform, platform uses federal learning and difference privacy technology to construct privacy-friendly multi-feature identity analysis model, realize the high-precision matching of cross-channel user.The platform uses graph attention neural network to represent the full-channel behavior of consumers as a heterogeneous directed graph, proposes a new structural friction index to identify conversion barrier nodes in the decision-making journey.Using auto-encoder to jointly embed explicit survey data and implicit behavior data, and introducing causal forest model to estimate the causal effect of attitude factors on purchasing behavior, a comprehensive value function is constructed to integrate behavior preferences, attitude tendencies, causal effects and journey resistance.The decision support layer automatically triggers personalized marketing interventions based on the friction index and value score, and dynamically optimizes the model based on A / B test feedback.The application realizes the balance between privacy protection and accurate identification, and can deeply understand the consumer decision-making journey and provide causal-driven value prediction.

Description

Technical Field

[0001] This invention relates to the field of market research and big data analysis technology, specifically to an integrated platform for omnichannel consumer behavior tracking and market research. Background Technology

[0002] With the widespread adoption of e-commerce and social media, consumers' decision-making paths span multiple touchpoints, resulting in highly fragmented behavioral data from online browsing, mobile app orders, offline store purchases, and social interactions. Existing technologies have developed cross-platform user identification solutions to address this multi-source data challenge. For example, patent CN104317784A compares user messages on different social media platforms using keywords of interest and word usage to determine if accounts belong to the same user; patent CN110826605A integrates user identity information and published content to cluster users across different platforms. Furthermore, solutions utilizing browser fingerprints and terminal features to track users without relying on cookies have also been proposed. While these technologies can solve some online identity mapping problems, most rely on single features or centralized storage, making it difficult to balance privacy protection with the accuracy of cross-channel (especially offline) identification.

[0003] In terms of behavioral analysis, existing customer journey analysis methods mostly use statistical indicators, such as calculating bounce rate, engagement, and conversion rate for each touchpoint, and then using weighted averages to form relevant indices to identify optimization points. Patent CN116888614A also utilizes a timeline analysis engine to analyze cross-channel event sequences and identify the root causes of customer problems. While these methods are effective in simple path analysis, they cannot capture the complex structural relationships between heterogeneous events, resulting in insufficient precision in identifying deep-seated obstacles.

[0004] In terms of data fusion, existing technologies have attempted to combine attitude surveys with consumer behavior, such as linking survey results with actual purchase data for market segmentation and decision support. However, these fusion methods are mainly simple superposition or correlation, failing to explore the potential relationship between explicit attitudes and implicit behaviors through deep representation learning, and also lacking causal modeling, making it difficult to reveal which attitudinal factors truly influence purchasing behavior.

[0005] In summary, existing technologies have shortcomings in cross-channel identity parsing, complex journey analysis, and cross-modal data fusion, and lack an integrated platform that simultaneously takes into account privacy protection, structured behavior analysis, and causal insights. Summary of the Invention

[0006] Technical Objective: To address the shortcomings of existing technologies, this invention discloses an integrated platform for omnichannel consumer behavior tracking and market research. This platform achieves high-precision identity resolution while ensuring user privacy, utilizes graph neural networks to deeply analyze the consumer decision-making journey, integrates multimodal data, and introduces causal inference to achieve intelligent prediction of consumer value and the timing of marketing interventions.

[0007] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution:

[0008] An integrated platform for omnichannel consumer behavior tracking and market research includes: a data acquisition layer, an identity resolution layer, a decision journey graph analysis engine, a cross-modal joint embedding and causal inference engine, and a decision support and adaptive optimization layer; wherein:

[0009] The data acquisition layer includes a website acquisition SDK, a mobile application acquisition SDK, an offline terminal acquisition unit, a third-party system access gateway, an event standardization and time alignment unit, and a de-identification and compliance processing unit. It is used to collect web browsing, mobile application operation, offline transaction, social interaction and questionnaire data according to a unified event model, add timestamps, channel tags and device identifiers to each event, and perform hashing, encryption or de-identification processing, as well as deduplication and quality verification, to generate a standardized event stream and output it to the identity resolution layer.

[0010] The identity resolution layer includes a local feature extraction unit, a differential privacy noise addition unit, a federated metric learning model training unit, and an embedding clustering unit. Each data provider extracts user identifier features locally and adds noise. It uses federated learning to train and improve the Mahalanobis distance metric model. Through contrastive learning, it maps the features of the same user on different channels to a unified vector space and performs clustering to generate a unified virtual ID.

[0011] The decision journey graph analysis engine includes a heterogeneous graph construction unit, a graph attention neural network training unit, and a friction index calculation unit. It is used to construct a heterogeneous directed graph from user behavior events based on a unified virtual ID, train the graph attention neural network to obtain node embeddings and edge weights, calculate the structural friction index of nodes, and identify conversion barrier nodes based on the friction index.

[0012] The cross-modal joint embedding and causal inference engine includes an autoencoder unit, a causal inference unit, and a comprehensive value calculation unit. The autoencoder unit jointly learns latent variables by combining user behavior embeddings and attitude text embeddings, and captures the correlation between explicit and implicit data through cross-reconstruction loss. The causal inference unit uses a causal forest model to calculate the causal effect of attitude factors on purchasing behavior. The comprehensive value calculation unit outputs a user value score based on the comprehensive value function formula.

[0013] The decision support and adaptive optimization layer is used to automatically generate marketing interventions based on the node's friction index and user value score, perform A / B testing, and use the feedback results to update the identity resolution model, graph attention neural network model, and value function parameters to achieve adaptive optimization.

[0014] Preferably, the differential privacy noise-adding unit adds Laplacian noise to each feature dimension. Where Δ is the sensitivity, To ensure privacy, the contributions of individual users are difficult to identify.

[0015] Preferably, the federated metric learning model training unit employs a contrastive loss function. Where Pos is the positive sample set, i.e., record pairs generated by the same user through different channels, and Neg is the negative sample set, i.e., record pairs from different users. and The multidimensional feature vectors extracted from users i and k across various channels. As the interval threshold, the positive definite weight matrix is ​​updated by aggregating noisy gradients on the aggregation server. And scaling factor.

[0016] Preferably, the graph attention neural network training unit uses the attention weight calculation formula.

[0017] ,

[0018] The node embeddings and edge weights are trained by predicting the next event type and transition probability through a self-supervised task, where p and q are the event node indices in the graph. Let be the attention weights from node p to node q. Let be the set of neighbors of node q. , and Let be the embedding vectors of nodes p, q, and r, respectively. It is a learnable linear transformation matrix. For learnable attention vectors, It is a linear rectification function with leakage coefficient.

[0019] Preferably, the friction index calculation unit calculates the friction index using the following formula.

[0020] ,

[0021] Where j is the index of the node being evaluated, and E is the set of directed edges in the graph. For entering the border, To go out of the boundary, and These are the attention weights from node p to node j and from node j to node q, respectively. , where is the transformation potential coefficient of node p and node q. Let j be the degree of node j; when When the threshold θ is exceeded, node j is determined to be a node that hinders the transformation.

[0022] Preferably, the comprehensive value function formula is as follows:

[0023] ,

[0024] Where i is the user index. Embedded for user-level behavior, For user-level attitude text embedding, and for , The nonlinear scoring function The individual causal effect of user i estimated by the causal inference unit. Let be the friction index along the path traversed by user i. , , , The weighting coefficients and , This is the Sigmoid function.

[0025] Preferably, the autoencoder unit includes two encoders and two decoders, which simultaneously optimize the representations of behavioral embeddings and attitude embeddings through reconstruction loss and cross-reconstruction loss; the causal forest model estimates the average causal effect of attitude factors through dual machine learning.

[0026] Preferably, the decision support and adaptive optimization layer performs the following steps: determining intervention nodes and intervention methods based on node friction index and user value score; collecting intervention effects through online A / B testing and updating potential coefficients, attention weights, and comprehensive value function weights; and dynamically updating the identity resolution model, graph attention neural network model, autoencoder, and causal forest model when new data arrives to achieve adaptive optimization of the platform.

[0027] Beneficial Effects: The integrated platform for omnichannel consumer behavior tracking and market research provided by this invention has the following beneficial effects:

[0028] By employing federated learning and differential privacy mechanisms, this invention achieves high-precision matching of multi-source identifiers without disclosing the original data, overcoming the privacy issues and recognition biases caused by centralized identity resolution. This invention uses an improved Mahalanobis distance metric and contrastive learning to adaptively adjust weights across different feature dimensions, achieving high-precision mapping between online and offline identities.

[0029] Compared to existing statistical indicators that only focus on dropout rate or conversion rate, the structural friction index proposed in this invention not only considers the conversion potential of a node itself, but also the conversion and churn impact of its neighboring nodes. By characterizing the heterogeneous relationship between events through the weights of graph attention learning, it can identify obstructive nodes that are difficult to find using traditional methods, and provide precise entry points for marketing intervention.

[0030] By fusing explicit and implicit data through an autoencoder and introducing a causal forest to estimate the causal effects of attitude factors, this invention integrates behavioral preferences, subjective attitudes, causal influences, and journey resistance into the comprehensive value function, achieving accurate prediction of consumer value. This method overcomes the shortcomings of traditional simple weighting methods, which cannot distinguish between correlation and causation.

[0031] The identity resolution, graph analysis, and data fusion modules are coupled together. The accuracy of the behavioral graph is ensured by the unified ID, the friction index provides structural constraints for the value function, and the value function guides the parameter adjustment of the identity and graph models, realizing the platform's collaborative optimization in multiple dimensions and forming an unpredictable overall technical effect. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0033] Figure 1 This is a schematic diagram of the overall system architecture of the platform of this invention;

[0034] Figure 2 This is a schematic diagram of the federated learning process of the identity resolution layer of this invention;

[0035] Figure 3 This is a schematic diagram illustrating the construction and analysis of the decision journey diagram of this invention;

[0036] Figure 4 This is a schematic diagram of the cross-modal joint embedding and causal inference module of the present invention. Detailed Implementation

[0037] The present invention will now be described more clearly and completely by way of a preferred embodiment in conjunction with the accompanying drawings, but this does not limit the invention to the scope of the described embodiment.

[0038] like Figure 1 As shown, an integrated platform for omnichannel consumer behavior tracking and market research includes: a data acquisition layer, an identity resolution layer, a decision journey graph analysis engine, a cross-modal joint embedding and causal inference engine, and a decision support and adaptive optimization layer; wherein:

[0039] The data acquisition layer includes a website acquisition SDK, a mobile application acquisition SDK, an offline terminal acquisition unit, a third-party system access gateway, an event standardization and time alignment unit, and a de-identification and compliance processing unit. It is used to collect web browsing, mobile application operation, offline transaction, social interaction and questionnaire data according to a unified event model, add timestamps, channel tags and device identifiers to each event, and perform hashing, encryption or de-identification processing, as well as deduplication and quality verification, to generate a standardized event stream and output it to the identity resolution layer.

[0040] The identity resolution layer includes a local feature extraction unit, a differential privacy noise addition unit, a federated metric learning model training unit, and an embedding clustering unit. Each data provider extracts user identifier features locally and adds noise. Federated learning is used to train and improve the Mahalanobis distance metric model. Through contrastive learning, the features of the same user on different channels are mapped to a unified vector space and clustered to generate a unified virtual ID.

[0041] Each data provider (website, mobile application, offline terminal, social platform) extracts user identifiers locally to form a multi-dimensional feature vector, including account ID encoding, device fingerprint vector, dynamic POS device signature vector, temporal and geographical features, and interaction pattern statistical features. To protect user privacy, a differential privacy mechanism is used to add noise to each feature vector.

[0042] A federated learning framework is adopted, where multi-party local feature vectors are uploaded to the aggregation server via a secure channel with added noisy gradients. The aggregation server uses contrastive learning loss to train a deep metric learning model, making the feature embedding vectors of the same user closer together across different channels, and wider apart across different users. Specifically, an improved identity similarity metric is defined:

[0043]

[0044] in Index for users, For users Multi-source fusion feature vector For Markov normal, , It is a symmetric positive definite metric matrix (obtained through metric learning). This is the scaling parameter for similarity (controlling the rate of exponential decay). The similarity score is assigned to two user records indicating that the user is the same person. The parameters M and σ are obtained by training a contrastive loss function that minimizes the embedding distance between positive samples (same user) and maximizes the distance between negative samples (different users).

[0045] After training, each party maps its local features to a shared vector space using a unified metric. Density clustering or graph community detection algorithms are then used to group feature vectors with distances within a threshold into the same group, generating a unified virtual ID. This approach allows the model to achieve high-precision identity matching without revealing the original identifiers, overcoming the shortcomings of existing solutions that rely on single features or centralized storage.

[0046] The decision journey graph analysis engine includes a heterogeneous graph construction unit, a graph attention neural network training unit, and a friction index calculation unit. It is used to construct a heterogeneous directed graph based on user behavior events according to a unified virtual ID, train the graph attention neural network to obtain node embeddings and edge weights, calculate the structural friction index of nodes, and identify conversion barrier nodes based on the friction index.

[0047] Based on the unified ID, each consumer's behavioral events are represented as a heterogeneous directed graph in chronological order. Node set Represents the event type, such as browsing nodes. Collection Nodes Add to cart node Offline purchase nodes Social interaction nodes Etc. Each edge Indicates an event In channels Transfer to the event after occurrence The order relationship, For time difference.

[0048] A heterogeneous graph is input into a graph attention neural network to learn low-dimensional embeddings for nodes and edges. Attention weights. Calculated by a graph attention neural network, it measures the strength of influence between events. Furthermore, a node potential coefficient is defined to capture the transformation potential of nodes. The attention weight is a normalized value representing the historical average conversion rate or revenue contribution of this node. The formula for calculating attention weight is:

[0049]

[0050] The node embeddings and edge weights are trained by predicting the next event type and transition probability through a self-supervised task, where p and q are the event node indices in the graph. Let be the attention weights from node p to node q. Let be the set of neighbors of node q. and Let be the embedding vectors of nodes p and q. It is a learnable linear transformation matrix. For learnable attention vectors, It is a linear rectification function with leakage coefficient.

[0051] To overcome the limitations of traditional lagging rates or simple weighted indices, this invention proposes an improved structural friction index. This is used to identify obstructing nodes. It is defined as:

[0052]

[0053] Where j is the index of the node being evaluated, and E is the set of directed edges in the graph. For entering the border, To go out of the boundary, and These are the attention weights from node p to node j and from node j to node q, respectively. , where is the transformation potential coefficient of node p and node q. Let j be the degree of node j; when When the threshold θ is exceeded, node j is determined to be a node that hinders the transformation. The higher the value, the more complex the relationship structure of the node, and the greater the resistance to transformation, making it a node that can be prioritized for intervention. Compared to existing indicators based solely on the dropout rate, this indicator introduces the potential of adjacent nodes and attention weights, enabling the discovery of non-obvious obstacles.

[0054] When the identity resolution model is updated or new events enter the graph, the graph attention neural network model incrementally updates node embeddings and friction indices in real time. This module assigns nodes with high friction indices to the decision engine for processing and feeds the intervention effect back to the graph model, adjusting attention weights and potential coefficients in the next round of training to achieve cross-layer coupling.

[0055] The cross-modal joint embedding and causal inference engine includes an autoencoder unit, a causal inference unit, and a comprehensive value calculation unit. The autoencoder unit jointly learns latent variables by combining user behavior embeddings and attitude text embeddings, and captures the correlation between explicit and implicit data through cross-reconstruction loss. The causal inference unit uses a causal forest model to calculate the causal effect of attitude factors on purchasing behavior. The comprehensive value calculation unit outputs a user value score based on the comprehensive value function formula.

[0056] To integrate explicit survey data (questionnaire responses, comment texts, ratings) and implicit behavioral data, a dual-path autoencoder was constructed. Behavioral sequences were embedded into vectors. and attitude text vector The input is fed into two encoders, which encode the latent variable z. i By using reconstruction loss and cross-modal alignment loss, z i Simultaneously capture common characteristics of behavior and attitude.

[0057] User attitudinal factors (such as price sensitivity and brand preference) are collected as potential treatment variables, and causal inference models such as causal forests are used to estimate the causal effects of these factors on purchase conversion or repeat purchase. .

[0058] Combining embedded representation and causal effects, we define a consumer comprehensive value function.

[0059] ,

[0060] Where i is the user index. Embedded for user-level behavior, For user-level attitude text embedding, and for , The nonlinear scoring function The individual causal effect of user i estimated by the causal inference unit. Let be the friction index along the path traversed by user i. , , , The weighting coefficients and , This is the Sigmoid function. This value function comprehensively considers user behavioral preferences, attitudinal tendencies, causal effects, and journey resistance to predict the user's future value or purchase probability.

[0061] The decision support and adaptive optimization layer is used to automatically generate marketing interventions based on the node's friction index and user value score, perform A / B testing, and use the feedback results to update the identity resolution model, graph attention neural network model, and value function parameters to achieve adaptive optimization.

[0062] The platform outputs intervention strategies based on the friction index and comprehensive value function. For users with both high friction index and high value function, the decision support module prioritizes pushing coupons, customer service assistance, or personalized suggestions at obstacle nodes. For groups with significant attitudinal causal effects, targeted questionnaires or content pushes are implemented. Marketing investment is reduced for users with low value function. The system uses A / B testing to monitor the intervention effect in real time. The collected behavioral data is fed back to the identity resolution model and graph model, and weights are adjusted through meta-learning to achieve closed-loop optimization. Internal testing shows that the platform's identity matching accuracy is improved by approximately 30% compared to existing methods based on single features, the accuracy of key node identification is improved by 35%, and the accuracy of predicting consumer value function is improved by more than 50%. It has also significantly improved conversion rates and customer satisfaction in actual marketing activities.

[0063] Example

[0064] Data Collection and Preprocessing: User behavior logs, transaction records, comment information, and questionnaires are collected through website SDK, mobile app SDK, offline POS terminals, middleware API, membership system, and social platform API. Timestamps, channel tags, and device identifiers are added to each data entry. Accounts, phone numbers, email addresses, and geolocations are anonymized using hashing or encryption techniques to meet personal information protection requirements.

[0065] Local Feature Extraction and Noise Addition: Each data provider maps user-related information into feature vectors through a feature extraction unit. .in Encode the encrypted account ID. Embedding fingerprints on devices, Embed signatures for offline POS devices. For user behavior frequency statistics, This includes geographic-time vectors, etc.; each dimension can be expanded according to business needs. Differential privacy techniques are used to add noise to each vector dimension. ,in For sensitivity, To ensure privacy, the contributions of individual users are difficult to trace back.

[0066] Federated Metric Learning Model Training: Employing a federated framework of contrastive learning, each party locally trains its embedding network and computes gradients using a loss function incorporating Mahalanobis distance; a central server aggregates the gradients and updates the global model parameters. Positive sample pairs consist of records generated by the same user from different channels, while negative sample pairs consist of records from different users. The objective function for training is...

[0067]

[0068] Where Pos is the set of positive samples, i.e., record pairs generated by the same user on different channels, and Neg is the set of negative samples, i.e., record pairs from different users. and The multidimensional feature vectors extracted from users i and k across various channels. As the interval threshold, the positive definite weight matrix is ​​updated by aggregating noisy gradients on the aggregation server. And scaling coefficients. Original features are not exchanged locally; only noisy gradients are uploaded to protect privacy. After iterative training, the model can group vectors from the same user together.

[0069] like Figure 2As shown in the diagram, this illustrates the interaction process between the various participants and the aggregation server within the identity resolution layer. Each data provider's side sequentially includes: a local feature extraction unit, which cleans and performs feature engineering on the local raw logs, outputting user-side features; a differential privacy noise-adding unit, which adds noise to features or local gradients according to a preset privacy budget; and a federated metric learning model training unit, which, after completing several rounds of gradient calculation and parameter updates locally, sends the encrypted or masked gradients / parameters to the aggregation server. The aggregation server securely aggregates and merges the updates from each participant, and distributes new global model weights to each participant to initiate a new round of local training and convergence iteration. Through multiple alternating rounds, the platform obtains a unified identity metric model without aggregating the original data; subsequent embedding clustering can be implemented offline or near real-time based on the updated global embedding space to generate a unified virtual ID. The arrows in the diagram only indicate the upload / download direction of parameters and gradients, without specifying the specific encryption protocol or transmission standard.

[0070] Clustering and Unified ID Generation: Each party uses an updated metric model to map local features to a unified vector space. Through density-based clustering (such as DBSCAN) or graph community algorithms, groups nodes with distances less than a threshold within the embedding space. Vectors are grouped into the same cluster, and a unique virtual ID is generated for each cluster. If fuzzy matching still exists within the same cluster, it can be calibrated through offline or manual verification.

[0071] Heterogeneous Graph Construction: User behavior events with a unified ID are constructed into a heterogeneous directed graph. Node types include browsing nodes. Collection Nodes Add to cart node Checkout Node Offline purchase nodes Social interaction nodes After-sales feedback nodes The edge type represents the sequential relationship and channel attribute between events, such as "online browsing → adding to cart" or "app favorites → offline in-store shopping". The edge attribute also includes the time interval and channel identifier.

[0072] like Figure 3As shown in the diagram, this illustrates the process of constructing and analyzing a journey graph based on a unified virtual ID. First, the heterogeneous graph construction unit instantiates cross-channel events as nodes (such as "browse," "add to cart," "favorite," "place order," "arrive at store," etc.) and represents the sequence and dependencies between events with directed edges. Edges can include attributes such as type and timestamp. Subsequently, the graph attention neural network training unit trains this heterogeneous directed graph to obtain node embeddings and categorized edge weights. Edge weights characterize the importance and transmission strength of adjacent nodes to the target node. Based on this, the friction index calculation unit calculates the structural friction index of each node according to the weights of its in / out edges and the conversion potential of adjacent nodes, quantifying the degree to which the node may hinder the conversion path. The platform can then sort and locate nodes with significant friction indices and prioritize their optimization in subsequent intervention strategies. The node and edge types in the diagram are only examples; in actual deployment, node categories and edge attributes can be dynamically expanded based on business data.

[0073] Graph Attention Training: A graph attention neural network is used to train the heterogeneous graph, calculating the node embeddings. Attention weights of edges The model input includes node type encoding and temporal difference features; the model output consists of node embedding vectors and edge weights, satisfying...

[0074]

[0075] in and For learnable weights, This represents vector concatenation. For nodes The set of neighbors. During training, self-supervised tasks, such as predicting the type of the next event or the transition probability, can be used to optimize the embedding.

[0076] Friction index calculation and key node identification: For each node in the graph Calculate its structural friction index

[0077]

[0078] in It is a node The conversion potential can be calculated from the conversion rate in historical data; Represents a node The probability of churn; It is a node The number of adjacent edges. The larger the value, the stronger the node. Located in a high-friction region, this may be a conversion barrier. The platform follows... Arrange the nodes in descending order and select the nodes with the highest friction index as priority intervention targets.

[0079] Dynamic updates to the graph model: When the identity resolution model is updated or new behavioral events occur, the heterogeneous graph incrementally updates nodes and edges, and uses an online learning method to update the parameters of the graph attention neural network. The friction index is adjusted synchronously with the embedding, enabling the model to adapt to new behavioral patterns.

[0080] Autoencoder joint embedding: For each user, collect behavioral embeddings. and attitude text vector Attitude text is encoded using a natural language model (e.g., BERT). A two-path autoencoder is constructed: the first path encodes and reconstructs the text. The second pathway is encoded and reconstructed. Introducing cross-reconstruction loss allows the two pathways to share latent variables. And they can predict each other, ensuring It includes both behavioral and attitudinal characteristics. The training objective is...

[0081]

[0082] in and The outputs are respectively the path reconstruction outputs. , This is for cross-modal reconstruction output.

[0083] Causal effect estimation: Collect attitude factor vectors for each user (For example, ratings based on dimensions such as price, brand, and service), using a causal forest model as a non-parametric causal inferrer. The model will... As a treatment variable, user purchase or repeat purchase metrics are used as outcome variables, while controlling for covariates (such as demographic attributes, historical behavior, etc.) to estimate the average causal effect of each factor on the outcome. Causal forests calculate double overlap effects using random forests, thus avoiding modeling bias.

[0084] Comprehensive value function: Calculates the comprehensive value for each user.

[0085]

[0086] in and It can be a fully connected neural network that outputs user behavior value and attitude value scores; The average friction index of the nodes experienced by the user during the journey; parameter , , , The system adapts to historical prediction results. Used to predict future purchase probability, average order value, or lifetime value.

[0087] Intervention strategy generation: when the friction index Exceeding the threshold And user value Above the threshold At that time, the system automatically at the node This allows for the generation of intervention strategies for users, such as pushing limited-time offers, reminding them to add items to their cart, or providing customer support. If High and attitudinal causal effects A negative value function indicates that the user's intent is influenced by negative attitudes, which can be corrected through questionnaires or content marketing. For users with low value functions, the system reduces marketing resource allocation.

[0088] A / B Testing and Feedback Learning: The platform conducts A / B testing to evaluate the intervention measures, collecting user metrics such as conversion rates and dwell time before and after the intervention as feedback. This feedback data is used to update the identity resolution threshold, graph model weights, and causal model parameters. A meta-learning strategy is employed for adjustment. This makes value function predictions more accurate.

[0089] The system's adaptive optimization process involves periodically or in real-time retraining of the federated identity model, graph attention neural network model, autoencoder, and causal forest. Through closed-loop feedback, each module continuously iterates and improves its performance. For example, if A / B testing shows that the friction index of a certain node has a weakened effect on conversion prediction, the node potential coefficient and the causal model are updated to make the model in the next cycle more closely reflect reality.

[0090] like Figure 4 As shown in the diagram, this illustrates the fusion analysis process of behavioral data and attitude data such as surveys / text. The autoencoder unit performs joint representation learning on user behavior embeddings and attitude embeddings, improving the alignment and complementarity of the two types of data through loss constraints of same-modal reconstruction and cross-modal reconstruction. The causal inference unit estimates the individual causal effect of attitude factors on target behaviors (such as purchase / in-store / renewal) based on processing variables, confounding features, and outcome variables, used to distinguish between correlation and causation. The comprehensive value calculation unit weighted and fused behavioral scores, attitude scores, causal effects, and journey friction information, outputting a user value or response probability score, which is then used by the decision support and adaptive optimization layers to generate personalized interventions, conduct A / B testing, and perform continuous iteration. The module boundaries in the diagram are only for illustrating functional attribution and do not limit the specific model structure or inference engine implementation.

[0091] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An integrated platform for omnichannel consumer behavior tracking and market research, characterized in that, include: The system comprises a data acquisition layer, an identity resolution layer, a decision journey graph analysis engine, a cross-modal joint embedding and causal inference engine, and a decision support and adaptive optimization layer; among which: The data acquisition layer includes a website acquisition SDK, a mobile application acquisition SDK, an offline terminal acquisition unit, a third-party system access gateway, an event standardization and time alignment unit, and a de-identification and compliance processing unit. It is used to collect web browsing, mobile application operation, offline transaction, social interaction and questionnaire data according to a unified event model, add timestamps, channel tags and device identifiers to each event, and perform hashing, encryption or de-identification processing, as well as deduplication and quality verification, to generate a standardized event stream and output it to the identity resolution layer. The identity resolution layer includes a local feature extraction unit, a differential privacy noise addition unit, a federated metric learning model training unit, and an embedding clustering unit. Each data provider extracts user identifier features locally and adds noise. It uses federated learning to train and improve the Mahalanobis distance metric model. Through contrastive learning, it maps the features of the same user on different channels to a unified vector space and performs clustering to generate a unified virtual ID. The decision journey graph analysis engine includes a heterogeneous graph construction unit, a graph attention neural network training unit, and a friction index calculation unit. It is used to construct a heterogeneous directed graph from user behavior events based on a unified virtual ID, train the graph attention neural network to obtain node embeddings and edge weights, calculate the structural friction index of nodes, and identify conversion barrier nodes based on the friction index. The friction index calculation unit uses the following formula to calculate the friction index. ， Where j is the index of the node being evaluated, and E is the set of directed edges in the graph. For entering the border, To go out of the boundary, and These are the attention weights from node p to node j and from node j to node q, respectively. , where is the transformation potential coefficient for nodes p and q. Let j be the degree of node j; when When the threshold θ is exceeded, node j is determined to be a node that hinders the transformation. The cross-modal joint embedding and causal inference engine includes an autoencoder unit, a causal inference unit, and a comprehensive value calculation unit. The autoencoder unit jointly learns latent variables by combining user behavior embeddings and attitude text embeddings, and captures the correlation between explicit and implicit data through cross-reconstruction loss. The causal inference unit uses a causal forest model to calculate the causal effect of attitude factors on purchasing behavior. The comprehensive value calculation unit outputs a user value score based on the comprehensive value function formula. The formula for the comprehensive value function is as follows: ， Where i is the user index. Embedded for user-level behavior, For user-level attitude text embedding, and for , The nonlinear scoring function The individual causal effect of user i estimated by the causal inference unit. Let be the friction index along the path traversed by user i. , , , The weighting coefficients and , For the Sigmoid function; The decision support and adaptive optimization layer is used to automatically generate marketing interventions based on the node's friction index and user value score, perform A / B testing, and use the feedback results to update the identity resolution model, graph attention neural network model, and value function parameters to achieve adaptive optimization.

2. The integrated platform for omnichannel consumer behavior tracking and market research according to claim 1, characterized in that, The differential privacy noise-adding unit adds Laplacian noise to each feature dimension. Where Δ is the sensitivity, To ensure privacy, the contributions of individual users are difficult to identify.

3. The integrated platform for omnichannel consumer behavior tracking and market research according to claim 1, characterized in that, The federated metric learning model training unit uses a contrastive loss function. Where Pos is the positive sample set, i.e., record pairs generated by the same user through different channels, and Neg is the negative sample set, i.e., record pairs from different users. and The multidimensional feature vectors extracted from users i and k across various channels. As the interval threshold, the positive definite weight matrix is ​​updated by aggregating noisy gradients on the aggregation server. And scaling factor.

4. The integrated platform for omnichannel consumer behavior tracking and market research according to claim 1, characterized in that, The graph attention neural network training unit uses the attention weight calculation formula. ， The node embeddings and edge weights are trained by predicting the next event type and transition probability through a self-supervised task, where p and q are the event node indices in the graph. Let be the attention weights from node p to node q. Let be the set of neighbors of node q. , and Let be the embedding vectors of nodes p, q, and r, respectively. It is a learnable linear transformation matrix. For learnable attention vectors, It is a linear rectification function with leakage coefficient.

5. The integrated platform for omnichannel consumer behavior tracking and market research according to claim 1, characterized in that, The autoencoder unit includes two encoders and two decoders, which simultaneously optimize the representations of behavioral embeddings and attitude embeddings through reconstruction loss and cross-reconstruction loss; the causal forest model estimates the average causal effect of attitude factors through dual machine learning.

6. The integrated platform for omnichannel consumer behavior tracking and market research according to claim 1, characterized in that, The decision support and adaptive optimization layer performs the following steps: determining intervention nodes and intervention methods based on node friction index and user value score; collecting intervention effects through online A / B testing and updating potential coefficients, attention weights, and comprehensive value function weights; and dynamically updating the identity resolution model, graph attention neural network model, autoencoder, and causal forest model when new data arrives to achieve adaptive optimization of the platform.

Images