Search

CN-122021828-A - Marketing product knowledge graph construction method and device

CN122021828ACN 122021828 ACN122021828 ACN 122021828ACN-122021828-A

Abstract

The invention discloses a method and a device for constructing a marketing product knowledge graph, which belong to the technical field of data processing and comprise the steps of constructing a weighted directed graph comprising user nodes, dish category nodes and consumption scene nodes as basic knowledge graphs, collecting time, weather and geographic position signals in real time, calculating scene influence values through a quantization algorithm, dynamically updating relation weights through a multiplication correction model, parallelly executing correlation strength, diversity and novelty optimal path searching on the dynamic graph, outputting an equilibrium path set based on pareto optimal selection, generating a multi-mode marketing content package by combining a user request, a real-time scene and a marketing target, evaluating efficiency through an A/B test, and feeding back the click rate and the ordering conversion rate as reward signals to a weight correction process. The invention realizes the remarkable improvement of the recommendation system in the aspects of accuracy, diversity and instantaneity through the design of dynamic perception, multi-target cooperation and intelligent generation.

Inventors

  • LIANG YU

Assignees

  • 奥琦玮科技(江苏)有限公司

Dates

Publication Date
20260512
Application Date
20260128

Claims (9)

  1. 1. A method for constructing a knowledge graph of a marketing product is characterized by comprising the following steps: S1, constructing a basic knowledge graph, wherein the basic knowledge graph is a weighted directed graph and comprises four entity nodes, namely a user node, a dish category node and a consumption scene node, the entity nodes are connected through a predefined relation type, and each relation is endowed with an initial basic weight obtained based on historical interaction data statistics; S2, acquiring three types of external situation signals, namely a time factor, a weather factor and a geographic position factor in real time, and calculating a scene influence value through corresponding quantization algorithm models respectively; s3, searching the correlation strength optimal path, the diversity optimal path and the novelty optimal path on the dynamically updated knowledge graph in parallel to form three types of path candidate sets; performing pareto optimal selection on the three types of path candidate sets based on the current marketing target context, and outputting balanced path sets carrying relevance scores, diversity scores and novelty scores; s4, combining the balanced path set with a user request, a real-time scene and a marketing target triplet to generate a multi-mode marketing content package containing structured dish data, visual elements and contextualized description; S5, evaluating marketing efficiency of different path strategy combinations through an A/B test, and feeding back click rate and order conversion rate as reward signals to a dynamic weight correction process to form end-to-end closed loop optimization.
  2. 2. The method for constructing a knowledge graph of a marketing product of claim 1, wherein in the basic knowledge graph, The initial basic weight of the relation of the user-preference-dishes is calculated according to the weight of the clicking times and the ordering times of the dishes by the user; The relation of dish-attribution-dish category is one-to-one mapping and the weight is fixed to be 1.0; The initial weight of the relation of dish-adaptation-consumption scene is determined according to the historical sales occupancy rate of the dish in the corresponding consumption scene; the initial weight of the "user-in-consumption scenario" relationship is inferred based on the user's current timestamp and geographic location.
  3. 3. The method of claim 1, wherein the quantization algorithm model uses piecewise linear function or table look-up mapping mechanism, and the output value range is limited between zero and one; the time factors comprise the current time interval and whether the time factor is in legal holidays or traditional festival days; the weather factors comprise real-time air temperature, precipitation probability and air quality index; The geographic location factor maps the user location coordinates to one of a business, residential, office, transportation hub, or tourist attraction through a geofence technique.
  4. 4. The method for constructing a knowledge graph of a marketing product of claim 1, wherein the dynamic weight correction process integrates a reinforcement learning model, and the reinforcement learning model adopts a Q-learning framework; The current scene factor combination is used as a state space, the parameter adjustment direction and the step length of the quantization algorithm model are used as an action space, the sum of the weighted click rate and the ordering conversion rate is used as a reward function, and the strategy network training is carried out through an experience playback pool.
  5. 5. The method for constructing a knowledge graph of a marketing product according to claim 1, wherein the optimal path of the association strength is calculated by adopting an improved Dijkstra algorithm, the dynamically updated inverse weight is used as a side cost, the shortest path reaching a candidate dish node is searched from a user node, and the first N paths are reserved; The optimal diversity paths are used for screening paths covering at least three dish categories, and redundant paths with the overlapping degree of the dish sets higher than a threshold value are removed based on Jaccard similarity; The novel optimal path forms a non-contact dish candidate pool by excluding dish nodes interacted by a user in the past thirty days, and introduces a popularity decay factor to reversely adjust the original weight of dishes in the candidate pool.
  6. 6. The method for constructing the marketing product knowledge graph according to claim 1 is characterized in that in the pareto optimal selection process, the scoring weight proportions of three paths are dynamically adjusted according to the context of a current marketing target, the weight proportion of a novel path is improved if a user is detected to be in a scene with stronger exploration intention, the weight proportion of a related strength path is improved if the user is in an operation period with higher conversion pressure, and the pareto front edge solution set is determined through non-dominant sequencing and crowding distance calculation.
  7. 7. The method for constructing the knowledge graph of the marketing product according to claim 1, wherein the generation of the multi-mode marketing content package comprises the steps of matching a template structure matched with a real-time scene and a marketing target from a preset marketing template library, taking relevance scores, diversity scores and novelty scores of paths as control variables of a natural language generation engine, and dynamically adjusting emotion tendencies, information density and narration rhythms of a document according to the control variables by adopting a framework based on mixing rules and a neural network.
  8. 8. A device for constructing a knowledge graph of a marketing product, which is used for realizing the method for constructing the knowledge graph of the marketing product according to any one of claims 1 to 7, and is characterized by comprising: the basic map construction module is used for constructing a basic knowledge map in the field of marketing products; the dynamic weight correction module realizes dynamic adjustment of the weight of the knowledge graph through real-time scene perception and closed-loop learning, and ensures that the association relation meets the real-time requirement of a user; The multi-target path discovery module is used for generating a product path set which is accurate and multi-element, and comprises a path parallel computing sub-module and a path fusion sub-module; The system comprises an experience type marketing scheme generation module, a template matching sub-module, a document adjustment sub-module and a multi-mode output sub-module, wherein the experience type marketing scheme generation module is used for converting an abstract path set into marketing contents; and the closed loop optimization module is used for quantifying marketing effect indexes corresponding to different path strategies through the A/B test module.
  9. 9. The device for constructing a knowledge graph of a marketing product of claim 8, wherein the scene factor acquisition sub-module calculates scene impact values corresponding to each factor through a predefined quantization algorithm model, and the scene factors comprise time factors, weather factors and geographic location factors; The weight updating submodule adopts a multiplication correction model to update the basic weight of the incidence relation between the entities in the basic knowledge graph in a second level, and realizes the smooth transition of the weight through a sliding window mechanism; and the reinforcement learning optimization submodule continuously optimizes the calculation parameters of the scene influence value according to the user feedback data to form a weight correction closed loop, wherein the user feedback data comprises a click rate and a ordering rate.

Description

Marketing product knowledge graph construction method and device Technical Field The invention relates to the technical field of data processing, in particular to a method and a device for constructing a marketing product knowledge graph. Background With the development of big data and artificial intelligence technology, the knowledge graph is used as an efficient knowledge representation and reasoning tool and is widely applied to the marketing and product recommendation fields. Conventional marketing product knowledge maps typically structurally connect entities and their relationships such as product attributes, user portraits, marketing content, and the like, and based thereon, make simple associative recommendations or content matches. In the prior art, particularly in the vertical field with extremely strong dynamic performance such as catering, a recommendation system based on a knowledge graph generally has the defects that after the prior knowledge graph system is constructed, the association weight among nodes of the prior knowledge graph system is usually fixed and cannot respond to severe changes of context factors such as time, weather, geographic position and the like in real time, the prior method mostly takes the maximized association strength of user historical preference and commodities as a core target, so that the recommendation result is seriously homogenized and lacks surprise feeling and diversity, the prior art generally stops generating a recommendation list, the capability of converting the recommendation result into attractive marketing content is lacking, the output scheme is mechanical and hard, so that high-quality recommendation cannot reach users, the marketing conversion rate is low, core modules such as graph construction, weight calculation, path recommendation and content generation independently run or are simply connected in series, the overall efficiency of the system is low, and the accuracy, the instantaneity and the final effect of the recommendation result are often difficult to reach the optimal. Therefore, how to construct a dynamic marketing product knowledge graph system which can deeply fuse real-time scene perception, support multi-objective collaborative optimization and realize intelligent generation of marketing content, so that the dynamic marketing product knowledge graph system has high correlation, meanwhile, diversity and novelty are considered, and continuous evolution is realized through end-to-end closed loop feedback, so that the dynamic marketing product knowledge graph system becomes a key challenge and a technical problem to be solved urgently for a person skilled in the art. Disclosure of Invention The invention overcomes the defects of the prior art and provides a method and a device for constructing a marketing product knowledge graph. S1, constructing a basic knowledge graph, wherein the basic knowledge graph is a weighted directed graph and comprises four entity nodes, namely a user node, a dish category node and a consumption scene node, the entity nodes are connected through a predefined relation type, and each relation is endowed with an initial basic weight obtained based on historical interaction data statistics; S2, acquiring three types of external situation signals, namely a time factor, a weather factor and a geographic position factor in real time, and calculating a scene influence value through corresponding quantization algorithm models respectively; s3, searching the correlation strength optimal path, the diversity optimal path and the novelty optimal path on the dynamically updated knowledge graph in parallel to form three types of path candidate sets; performing pareto optimal selection on the three types of path candidate sets based on the current marketing target context, and outputting balanced path sets carrying relevance scores, diversity scores and novelty scores; s4, combining the balanced path set with a user request, a real-time scene and a marketing target triplet to generate a multi-mode marketing content package containing structured dish data, visual elements and contextualized description; S5, evaluating marketing efficiency of different path strategy combinations through an A/B test, and feeding back click rate and order conversion rate as reward signals to a dynamic weight correction process to form end-to-end closed loop optimization. In a preferred embodiment of the present invention, in the basic knowledge graph, The initial basic weight of the relation of the user-preference-dishes is calculated according to the weight of the clicking times and the ordering times of the dishes by the user; The relation of dish-attribution-dish category is one-to-one mapping and the weight is fixed to be 1.0; The initial weight of the relation of dish-adaptation-consumption scene is determined according to the historical sales occupancy rate of the dish in the corresponding consumption scene; the initial weight of the "use