CN-121235749-B - Whole channel consumer behavior tracking and market investigation integrated platform

CN121235749BCN 121235749 BCN121235749 BCN 121235749BCN-121235749-B

Abstract

The invention discloses a full-channel consumer behavior tracking and market investigation integrated platform, which adopts federal learning and differential privacy technology to construct a privacy-friendly multi-feature identity analysis model so as to realize high-precision matching of cross-channel users. The platform uses the graph-annotation-force neural network to represent consumer full-channel behavior as a heterogeneous directed graph, proposing a new structural friction index to identify conversion barrier nodes in the decision-making journey. The self-encoder is utilized to jointly embed dominant investigation data and recessive behavior data, and a causal forest model is introduced to estimate causal effects of attitude factors on purchasing behavior, so that comprehensive cost function fusion behavior preference, attitude tendency, causal effects and journey resistance are constructed. The decision support layer automatically triggers personalized marketing intervention according to the friction index and the value score, and feeds back a dynamic optimization model in combination with the A/B test. The invention realizes the balance of privacy protection and accurate identification, and can deeply insight into the decision-making journey of consumers and provide causal driven value prediction.

Inventors

JIANG MING
WANG SHENGYONG
LIU SIZHU
WU KAIWEI

Assignees

三明医学科技职业学院

Dates

Publication Date: 20260508
Application Date: 20251201

Claims (6)

1. The all-channel consumer behavior tracking and market investigation integrated platform is characterized by comprising a data acquisition layer, an identity analysis layer, a decision trip map analysis engine, a cross-modal joint embedding and causal inference engine and a decision support and self-adaptive optimization layer, wherein: The data acquisition layer comprises a website acquisition SDK, a mobile application acquisition SDK, an off-line terminal acquisition unit, a third party system access gateway, an event standardization and time alignment unit and a de-identification and compliance processing unit, and is used for acquiring webpage browsing, mobile terminal application operation, off-line transaction, social interaction and questionnaire data according to a unified event model, adding a time stamp, a channel label and a device identifier to each event, performing hash, encryption or desensitization processing and de-duplication and quality verification, generating a standardized event stream and outputting the standardized event stream to the identity analysis layer; The identity analysis layer comprises a local feature extraction unit, a differential privacy noise adding unit, a federal metric learning model training unit and an embedded clustering unit, wherein each data provider locally extracts user identification features and adds noise, the federal learning training is utilized to improve a mahalanobis distance metric model, features of the same user in different channels are mapped into a unified vector space through comparison learning and clustered, and a unified virtual ID is generated; The decision trip map analysis engine comprises a heterogeneous map construction unit, a map attention neural network training unit and a friction index calculation unit, wherein the heterogeneous map construction unit is used for constructing user behavior events into heterogeneous directed maps according to unified virtual IDs, training the map attention neural network to obtain node embedding and edge weights, calculating the structural friction index of the nodes, and identifying transformation obstacle nodes according to the friction index; The friction index calculating unit calculates the friction index using the following formula , Where j is the node index being evaluated, E is the directed edge set of the graph, In order to be in-line with the edge, In order to form the edge of the steel plate, And The attention weight from node p to node j and the attention weight from node j to node q, The transformation potential coefficients of the node p and the node q, Degree of node j when When the preset threshold value theta is exceeded, judging the node j as a transformation obstructing node; The cross-modal joint embedding and causal inference engine comprises a self-encoder unit, a causal inference unit and a comprehensive value calculation unit, wherein the self-encoder unit acquires hidden variables by joint learning of user behavior embedding and attitude text embedding, and captures dominant and hidden data association through cross-reconstruction loss; The formula of the comprehensive cost function is as follows , Where i is the index of the user and, For the embedding of the user-level behavior, For user-level attitude text embedding, And Is that , Is a non-linear scoring function of (c), Individual causal effects of user i estimated for the causal inference unit, For the friction index in the path experienced by user i, , , , Is a weight coefficient and , Is a Sigmoid function; The decision support and self-adaptive optimization layer is used for automatically generating marketing intervention measures according to friction indexes of the nodes and user value scores, executing A/B tests, and using feedback results for updating identity analysis models, graph annotation meaning neural network models and cost function parameters to realize self-adaptive optimization.
2. The full channel consumer behavior tracking and market research integration platform of claim 1 wherein the differential privacy noise adding unit adds laplacian noise to each feature dimension Where delta is the sensitivity, where, For privacy budgets, it is guaranteed that the contributions of individual users are difficult to identify.
3. The integrated platform for full channel consumer behavior tracking and market research of claim 1 wherein the federal metric learning model training unit employs a contrast loss function Where Pos is a positive set of samples, i.e. pairs of records generated by the same user in different channels, neg is a negative set of samples, i.e. pairs of records of different users, And The multidimensional feature vectors extracted for the users i, k in each channel, Updating a positive weighting matrix for interval threshold by aggregating noisy gradients at an aggregation server And scale factors.
4. The full channel consumer behavior tracking and market research integration platform of claim 1, wherein the attention neural network training unit adopts an attention weight calculation formula , And predicting the next event type and transition probability through the self-supervision task to train node embedding and edge weights, wherein p and q are event node indexes in the graph, For the attention weight of node p to node q, As a set of neighbors of node q, 、 And The embedded vectors of nodes p, q and r respectively, As a matrix of linear transformations that can be learned, As a learnable attention vector, Is a linear rectification function with leakage coefficients.
5. The all-channel consumer behavior tracking and market research integration platform of claim 1 wherein the self-encoder unit comprises two encoders and two decoders, the representation of behavior embedding and attitude embedding is optimized simultaneously by reconstruction loss and cross reconstruction loss, and the causal forest model estimates the average causality of attitude factors by double machine learning.
6. The full channel consumer behavior tracking and market research integrated platform according to claim 1, wherein the decision support and adaptive optimization layer determines intervention nodes and intervention modes according to node friction indexes and user value scores, collects intervention effects through online A/B tests, updates potential coefficients, attention weights and comprehensive cost function weights, and dynamically updates an identity analysis model, a schematic neural network model, a self-encoder and a causal forest model when new data arrives, so that the adaptive optimization of the platform is realized.

Description

Whole channel consumer behavior tracking and market investigation integrated platform Technical Field The invention relates to the technical field of market research and big data analysis, in particular to a full-channel consumer behavior tracking and market investigation integrated platform. Background With the popularity of electronic commerce and social media, consumer decision paths are spread over multiple contacts, and behavioral data such as online browsing, mobile application ordering, offline store consumption, and social interactions are highly fragmented. In the prior art, in order to open up multi-source data, some cross-platform user identification schemes are presented. For example, patent CN104317784A judges whether the account belongs to the same user by comparing interest keywords and word habits of user messages on different social platforms, and patent CN110826605A performs cluster recognition on users of different platforms by integrating user identity information and release content. In addition, a scheme of tracking a user without depending on cookies by utilizing browser fingerprints and terminal features is also proposed. The techniques can solve the problem of partial online identity mapping, but most rely on single characteristics or centralized storage, so that the privacy protection and the accuracy of cross-channel (especially offline) identification are difficult to be considered. In the aspect of behavior analysis, the existing customer journey analysis mostly adopts statistical indexes, such as calculating the jump rate, participation degree and conversion rate of each contact, and forming a correlation index through weighted average to identify an optimization link. Also patent CN116888614a utilizes a timeline analysis engine to comb across channel event sequences to identify the root cause that led to the customer problem. These methods are effective in simple path analysis, but cannot capture complex structural relationships between heterogeneous events, resulting in insufficient refinement of deep-level hindering node identification. In terms of data fusion, the prior art attempts to combine attitude surveys with consumption behavior, such as linking questionnaire results with actual purchase data, for market segments and decision support. However, the fusion modes are mainly simple superposition or association, potential association of explicit attitudes and implicit behaviors cannot be explored through deep characterization learning, modeling of causal relations is lacking, and it is difficult to reveal which attitudes factors really influence purchasing behaviors. In conclusion, the prior art has the defects in the aspects of cross-channel identity analysis, complex journey analysis and cross-modal data fusion, and lacks an integrated platform which simultaneously combines privacy protection, structured behavior analysis and causal insight. Disclosure of Invention Aiming at the defects of the prior art, the invention discloses a full-channel consumer behavior tracking and market investigation integrated platform which realizes high-precision identity analysis on the premise of ensuring user privacy, utilizes a graph neural network to deeply analyze consumer decision trip, fuses multi-mode data and introduces causal inference to realize intelligent prediction of consumer value and marketing intervention opportunity. The technical scheme is that in order to achieve the technical purpose, the invention adopts the following technical scheme: The all-channel consumer behavior tracking and market investigation integrated platform comprises a data acquisition layer, an identity analysis layer, a decision trip map analysis engine, a cross-modal joint embedding and causal inference engine and a decision support and self-adaptive optimization layer, wherein: The data acquisition layer comprises a website acquisition SDK, a mobile application acquisition SDK, an off-line terminal acquisition unit, a third party system access gateway, an event standardization and time alignment unit and a de-identification and compliance processing unit, and is used for acquiring webpage browsing, mobile terminal application operation, off-line transaction, social interaction and questionnaire data according to a unified event model, adding a time stamp, a channel label and a device identifier to each event, performing hash, encryption or desensitization processing and de-duplication and quality verification, generating a standardized event stream and outputting the standardized event stream to the identity analysis layer; The identity analysis layer comprises a local feature extraction unit, a differential privacy noise adding unit, a federal metric learning model training unit and an embedded clustering unit, wherein each data provider locally extracts user identification features and adds noise, the federal learning training is utilized to improve a mahalanobis distance metric model, features o