CN-121303291-B - Knowledge graph-fused user interest matching network marketing system

Abstract

The invention discloses a user interest matching network marketing system integrating knowledge patterns, and relates to the technical field of computers. The system is used for solving the problem of marketing decision deviation caused by lack of dynamic situation awareness and causal reasoning capability in the existing user interest modeling. Firstly, constructing a dynamic knowledge graph through distributed event stream processing, and expressing a user real-time situation structure as a graph sequence with a timestamp, secondly, extracting user interest evolution characteristics through a time sequence convolution graph network, quantifying influence weights of situation nodes through a multi-head attention mechanism, then, constructing a causal inference model to eliminate exposure and position deviation, extracting user essential interest characterization through a counter-facts learning framework, and finally, dynamically optimizing a knowledge graph structure based on a gradient propagation algorithm to form a closed-loop learning system from decision to perception. The invention realizes the deep understanding and accurate matching of the user interests and improves the accuracy and the interpretability of the marketing system.

Inventors

• LI WEI

Assignees

• 江西师范高等专科学校

Dates

Publication Date

20260512

Application Date

20251017

Claims (6)

1. 1. The user interest matching network marketing system integrating the knowledge graph is characterized by comprising a situation graph construction module, an interest evolution calculation module, a causal depolarization module and an intelligent decision module; The situation map construction module is used for receiving the user behavior event stream and the environmental context data in real time through distributed event stream processing, dynamically maintaining a knowledge map structure based on a flow map updating algorithm of a time window, activating a preset commodity scene association rule when a user is detected to enter a specific geofence, and generating a dynamic knowledge map increment updating sequence with an event time stamp, and the specific process is as follows: Monitoring user position information in real time, identifying whether a user enters a preset interest area through a geofence detection algorithm, triggering a scene identification engine when the user enters a specific geofence, and matching corresponding business scene types; Activating a preset commodity scene association rule base, searching related commodity class and attribute characteristics under the scene, and dynamically constructing an association subgraph of a user-scene-commodity based on the scene association rule; generating an increment updating operation sequence with a time sequence mark for each associated subgraph, and sequencing the increment updating sequence according to event time stamps to ensure time sequence consistency; establishing a spectrum smooth transition mechanism during scene switching, and avoiding abrupt change of association relations through gradual updating, so as to keep the stability of the knowledge spectrum; The interest evolution calculation module is used for extracting the multi-hop neighbor structure change characteristics of the user in the sliding time window through the time sequence convolution graph network according to the dynamic knowledge graph increment updating sequence, quantifying the contribution weights of different situation nodes to the interest state of the user by utilizing the multi-head graph attention mechanism, and outputting the user interest embedding vector with time sequence dependency, and the specific process is as follows: constructing a multi-head diagram attention network, distributing independent characteristic transformation parameters for each attention head, and calculating attention coefficients between user nodes and each situation node to represent the influence degree of the attention coefficients on the user interest state; Designing a hierarchical attention mechanism to calculate attention weights at a node level and a map level respectively, and introducing a time attenuation factor to dynamically adjust contribution weights of historical situation nodes to the current interest state; the comprehensive situation influence weight distribution is obtained through the splicing and fusion of the multi-head attention output, and the neighbor node characteristics are weighted and aggregated based on the attention weight so as to update the embedded representation of the user node; Deep fusion is carried out on the time sequence structural features and the attention weighting features, so that a user interest embedded vector with time sequence dependency is generated; The causal depolarization module is used for constructing a causal inference model containing exposure deviation and position deviation according to the user interest embedded vector, combining sample distribution of a training treatment group and a comparison group through an end-to-end inverse fact learning framework, eliminating confusion influence of selection deviation on the interest vector by using a gradient inversion layer, and outputting a depolarized user intrinsic interest representation; The intelligent decision module is used for carrying out multidimensional association analysis on the user intrinsic interest characterization and the real-time feedback data, and dynamically adjusting node attributes and edge weights in the knowledge graph in the situation graph construction module through a graph structure optimization algorithm based on gradient propagation.
2. 2. The system for marketing the user interest matching network with the integrated knowledge graph of claim 1, wherein the system for marketing the user interest matching network with the integrated knowledge graph is characterized by receiving the user behavior event stream and the environmental context data in real time through distributed event stream processing, and dynamically maintaining the knowledge graph structure based on a flow graph updating algorithm of a time window comprises the following specific processes: establishing a distributed event stream processing pipeline to continuously receive multi-source data streams, eliminating event disorder influence caused by network transmission through event time alignment processing, and dividing the continuous data streams into time slices with fixed duration based on a sliding time window mechanism of event time; Analyzing user behavior events in real time in each time window, extracting entity and relation information in the events, incrementally updating the extracted entity and relation into a knowledge graph through a flow graph updating algorithm, maintaining version management of the knowledge graph, and marking each updating operation with a time stamp; And based on the newly arrived data, the association relation among the nodes is adjusted in real time, so that the dynamic evolution of the graph structure is realized, and the knowledge graph can accurately reflect the latest state of the user.
3. 3. The knowledge-graph-fused user interest matching network marketing system of claim 1, wherein the specific process of extracting the multi-hop neighbor structure change characteristics of the user in the sliding time window through the time sequence convolution graph network according to the dynamic knowledge-graph increment updating sequence is as follows: constructing a time sequence convolution graph network architecture, designing a graph roll stacking layer with time sequence sensing capability, and performing time sequence slicing processing on the incremental updating sequence of the knowledge graph in a sliding time window; Capturing local neighbor structural features by applying graph convolution operation to each time sequence slice, realizing the propagation and aggregation of multi-hop neighbor information through a multi-layer graph convolution network, and extracting deep-layer graph structural features; the design time sequence dependence modeling module captures the structure evolution rule between adjacent time slices, and the expanded time sequence convolution is utilized to expand the receptive field so as to capture the long-distance time sequence dependence; and comprehensively analyzing the pattern structure change modes of the user in the continuous time segments to form the user behavior characterization containing the multi-scale time sequence features.
4. 4. The knowledge-graph-fused user interest matching network marketing system of claim 1, wherein the construction logic for constructing a causal inference model comprising exposure bias and position bias from the user interest embedding vector is as follows: establishing a causal graph structure based on the user interest embedded vector, taking an exposure mechanism and position information as intervention variables into a causal graph model, and defining a causal relationship path among user characteristics, commodity attributes and situation factors; Designing a structured causal model framework, characterizing the influence mechanism of different deviation factors on the user interests through a learnable parameter matrix, and establishing a causal conduction chain from exposure events to user feedback; constructing a potential result predictor based on a deep neural network, respectively estimating user interest states under the condition of exposure intervention or not, quantifying confusion effect of exposure deviation and position deviation on interest vectors, designing a dynamic causal graph updating mechanism, adjusting causal graph structural parameters according to real-time feedback data, and ensuring that a causal inference model can adapt to a continuously-changing user behavior mode.
5. 5. The knowledge graph-fused user interest matching network marketing system of claim 4, wherein the sample distribution of the training treatment group and the comparison group is combined through an end-to-end inverse fact learning framework, the confusion influence of the selection deviation on the interest vector is eliminated by using a gradient inversion layer, and the specific process of outputting the unbiased user intrinsic interest representation is as follows: constructing an end-to-end inverse fact learning network architecture, designing a double-branch structure to respectively process sample data of a processing group and sample data of a comparison group, and learning a characteristic representation with unchanged deviation through a parameter sharing mechanism; introducing a gradient inversion layer into the feature extraction layer, and performing inversion operation on the gradient of the bias related features in the back propagation process, so as to force the network to learn the user interest characterization irrelevant to the bias factors; Designing an antagonism regularization loss function, eliminating the influence of selection deviation by minimizing the difference of characteristic distribution of a treatment group and a control group, and simultaneously keeping the discrimination capability of user interest characterization; the multi-task learning strategy is adopted to jointly optimize the inverse fact prediction task and the deviation elimination task, the intrinsic interests and the deviation factors of the user are gradually separated through iterative training, and the user interest characterization vector is output.
6. 6. The knowledge-graph-fused user interest matching network marketing system of claim 1, wherein the intelligent decision module comprises the steps of: Deep association mining is carried out on the user intrinsic interest characterization and the real-time feedback data, and key factors influencing the user interest change are identified through correlation analysis and causal inference; designing a graph structure optimization algorithm based on gradient propagation, constructing a differentiable graph neural network calculation graph, and guiding the optimization process of the knowledge graph by taking a user feedback signal as a supervision signal; And calculating the gradients of the node attributes and the edge weights of the knowledge graph through a back propagation algorithm, updating the graph structure parameters through a self-adaptive optimization algorithm, and transmitting the optimized knowledge graph parameters to an environment graph construction module in real time.

Description

Knowledge graph-fused user interest matching network marketing system Technical Field The invention relates to the technical field of computers, in particular to a user interest matching network marketing system integrating knowledge patterns. Background In the current digital living environment, users are in a continuous information overload state and are exposed to massive commodity and marketing information at all times. At the same time, consumer decision logic is moving from passive reception to active seeking, and they expect the service platform to be able to understand its explicit historical preferences, as well as to insight into its immediate intent and deep needs in different situations, as a close-fitting intelligent advisor. The marketing paradigm from 'thousand people' to 'thousand people' and then to 'one thousand people' is updated, and unprecedented challenges are presented for enterprises to intelligently realize accurate user access and efficient resource allocation by utilizing data. However, there are several substantial technical bottlenecks with existing mainstream user interest modeling and matching methods. First, the traditional approach, represented by collaborative filtering, is severely limited by data sparseness and cold start problems, with recommendation logic built on the statistical correlation of the user-item co-occurrence matrix, e.g., the semantic logic behind "why users who purchased alpenstocks still need waterproof jacket" cannot be understood, resulting in poor interpretation of the recommendation results and difficulty in covering long tail interests. Secondly, although the content or label-based model can alleviate cold start, the interest characterization is flattened and isolated, the interest of the user is simplified into a set of discrete labels, and complex networked relations between interest points, such as dynamic excitation relations between 'weekends', 'fine weather' and 'outdoor sports equipment', cannot be described. In addition, most of the existing methods are built on static data snapshots, the interest model is updated with hysteresis, real-time interest drift generated by the change of the situation (such as geographic position, time and social hot spot) of the user is difficult to capture and respond, and finally marketing decisions and the actual state of the user are disjointed, so that resource release waste and user experience reduction are caused. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a user interest matching network marketing system integrating knowledge patterns, which solves the problems of the background technology. The user interest matching network marketing system integrating the knowledge patterns comprises a situation pattern construction module, an interest evolution calculation module, a causal depolarization module and an intelligent decision module, wherein the situation pattern construction module is used for receiving user behavior event streams and environmental context data in real time through distributed event stream processing, dynamically maintaining a knowledge pattern structure based on a flow pattern updating algorithm of a time window, activating a preset commodity scene association rule when a user enters a specific geofence is detected, generating a dynamic knowledge pattern increment updating sequence with an event time stamp, the interest evolution calculation module is used for extracting multi-hop neighbor structure change characteristics of the user in a sliding time window through a time sequence convolution graph network, quantifying contribution weights of different border nodes to the user interest states through a multi-head graph attention mechanism, outputting user interest embedding vectors with time sequence dependency, the causal depolarization module is used for constructing a training frame association processing group comprising exposure deviation and position deviation and contrast facts, generating a dynamic knowledge pattern increment updating sequence with an event time stamp, optimizing the situation index profile by using a time sequence convolution graph, and optimizing the situation index rule, and the situation index profile is used for carrying out decision-making feature analysis by the decision-making module, and the situation-making the decision-making module have a real-time sequence dependency on the user interest state by using the influence on the user interest state by using the decision-making module. The method comprises the steps of receiving user behavior event stream and environment context data in real time through distributed event stream processing, establishing a distributed event stream processing pipeline to continuously receive multi-source data stream, eliminating event disorder influence caused by network transmission through event time alignment processing, dividing the continuous data stream into time slices with