Marketing product knowledge graph construction method and device

Abstract

The invention discloses a marketing product knowledge graph construction method and device, and belongs to the technical field of data processing, and the method comprises the steps: constructing a weighted directed graph comprising user nodes, dish nodes, dish category nodes and consumption scene nodes as a basic knowledge graph; collecting time, weather and geographic position signals in real time, calculating a scene influence value through a quantization algorithm, and dynamically updating a relation weight by adopting a multiplication correction model; performing correlation intensity, diversity and novelty optimal path search in parallel on the dynamic map, and outputting an equilibrium path set based on Pareto optimal selection; generating a multi-modal marketing content package in combination with the user request, the real-time scene and the marketing target; the efficiency is evaluated through A / B test, and the click rate and the order-placing conversion rate are used as reward signals to be fed back to the weight correction process. Through the design of dynamic perception, multi-target collaboration and intelligent generation, the accuracy, diversity and real-time performance of the recommendation system are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and apparatus for constructing a marketing product knowledge graph. Background Technology

[0002] With the development of big data and artificial intelligence technologies, knowledge graphs, as an efficient knowledge representation and reasoning tool, have been widely used in marketing and product recommendation. Traditional marketing product knowledge graphs typically structure and connect entities such as product attributes, user profiles, and marketing content, as well as their relationships, and then perform simple association recommendations or content matching based on this.

[0003] In existing technologies, especially in highly dynamic vertical fields such as catering, knowledge graph-based recommendation systems typically suffer from the following drawbacks: After construction, the association weights between nodes in existing knowledge graph systems are usually fixed, making them unable to respond in real-time to drastic changes in contextual factors such as time, weather, and geographical location; existing methods often focus on maximizing the correlation between users' historical preferences and products, leading to highly homogenized recommendation results that lack surprise and diversity; existing technologies typically stop at generating a recommendation list, lacking the ability to transform recommendations into attractive marketing content, resulting in mechanical and rigid outputs that fail to reach users, leading to low marketing conversion rates; core modules such as graph construction, weight calculation, path recommendation, and content generation often operate independently or are simply interconnected, lacking an end-to-end collaborative optimization mechanism, resulting in low overall system efficiency and difficulty in achieving optimal accuracy, real-time performance, and final results.

[0004] Therefore, how to construct a dynamic marketing product knowledge graph system that can deeply integrate real-time scene perception, support multi-objective collaborative optimization, and realize intelligent generation of marketing content, so as to maintain high relevance while taking into account diversity and novelty, and continuously evolve through end-to-end closed-loop feedback, has become a key challenge and a technical problem that needs to be solved by those skilled in the art. Summary of the Invention

[0005] This invention overcomes the shortcomings of the prior art and provides a method and apparatus for constructing a marketing product knowledge graph.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a method and apparatus for constructing a marketing product knowledge graph, comprising: S1, constructing a basic knowledge graph, wherein the basic knowledge graph is a weighted directed graph, including four types of entity nodes: user nodes, dish nodes, dish category nodes, and consumption scenario nodes, wherein the entity nodes are connected through predefined relationship types, and each relationship is assigned an initial basic weight based on historical interaction data statistics; S2. Real-time acquisition of three types of external situational signals: time factor, weather factor, and geographical location factor, and calculation of scene impact values ​​through corresponding quantization algorithm models; dynamic updating of the basic weights of the relationships affected by the external situational signals using a multiplicative correction model to obtain new weights, and arithmetic averaging of the weight values ​​after three consecutive updates using a 30-second sliding window. S3. On the dynamically updated knowledge graph, perform parallel searches for the optimal path in terms of relevance strength, diversity, and novelty to form three types of path candidate sets. Based on the current marketing objective context, perform Pareto optimal selection on the three types of path candidate sets and output an equilibrium path set carrying relevance score, diversity score, and novelty score. S4. Combine the balanced path set with the triplet of user requests, real-time scenarios and marketing objectives to generate a multimodal marketing content package containing structured menu data, visual elements and contextual descriptions; S5. The marketing effectiveness of different path strategy combinations is evaluated through A / B testing, and the click-through rate and order conversion rate are used as reward signals to feed back to the dynamic weight correction process, forming an end-to-end closed-loop optimization.

[0007] In a preferred embodiment of the present invention, in the basic knowledge graph, The initial base weight of the "user-preference-dish" relationship is calculated by weighting the number of times a user clicks on a dish with the number of times they place an order. The relationship between "dish-attribute-dish category" is a one-to-one mapping with a fixed weight of 1.0; The initial weight of the "dish-match-consumption scenario" relationship is determined based on the historical sales proportion of the dish in the corresponding consumption scenario; The initial weight of the "user-occupation-consumption scenario" relationship is inferred based on the user's current timestamp and geographical location.

[0008] In a preferred embodiment of the present invention, the quantization algorithm model adopts a piecewise linear function or a lookup table mapping mechanism, and the output value range is limited to between zero and one; The time factor includes the current time period and whether it is a statutory holiday or a traditional festival. The weather factors include real-time temperature, precipitation probability, and air quality index; The geographic location factor uses geofencing technology to map the user's location coordinates to one of the following categories: business district, residential area, office area, transportation hub, or tourist attraction.

[0009] In a preferred embodiment of the present invention, the dynamic weight correction process integrates a reinforcement learning model, and the reinforcement learning model adopts the Q-learning framework; The current scenario factor combination serves as the state space, the parameter adjustment direction and step size of the quantification algorithm model serve as the action space, the weighted sum of click-through rate and order conversion rate serves as the reward function, and the strategy network is trained through the experience replay pool.

[0010] In a preferred embodiment of the present invention, the optimal path of association strength is calculated using an improved Dijkstra algorithm, with the dynamically updated reciprocal of the weights as the edge cost, to find the shortest path from the user node to the candidate dish node, and retain the first N paths. The optimal diversity path is obtained by filtering paths that cover at least three categories of dishes and removing redundant paths with a higher overlap of dish sets based on Jaccard similarity. The novelty-optimal path forms a candidate pool of untouched dishes by excluding dish nodes that users have interacted with in the past 30 days, and introduces a popularity decay factor to reverse the original weight of dishes in the candidate pool.

[0011] In a preferred embodiment of the present invention, during the Pareto optimal selection process, the weight ratio of the three types of paths is dynamically adjusted according to the current marketing objective context; if it is detected that the user is in a scenario with a strong exploratory intent, the weight ratio of the novelty path is increased; if it is in an operational cycle with high conversion pressure, the weight ratio of the association strength path is increased; the Pareto front solution set is determined by non-dominated sorting and crowding distance calculation.

[0012] In a preferred embodiment of the present invention, the generation of the multimodal marketing content package includes: matching a template structure that matches the real-time scenario and marketing objectives from a pre-set marketing template library; using path relevance score, diversity score and novelty score as control variables for the natural language generation engine; the natural language generation engine adopts a rule-based and neural network hybrid architecture, and dynamically adjusts the emotional tendency, information density and narrative rhythm of the copy according to the control variables.

[0013] An apparatus for constructing a marketing product knowledge graph, comprising: The basic knowledge graph construction module is used to build a basic knowledge graph for the marketing product domain; The dynamic weight correction module achieves dynamic adjustment of knowledge graph weights through real-time scene perception and closed-loop learning, ensuring that the relationships match the user's real-time needs; it includes a scene factor collection submodule, a weight update submodule, and a reinforcement learning optimization submodule. The multi-target path discovery module is used to generate a set of product paths that balances accuracy and diversity; it includes a path parallel computing submodule and a path fusion submodule. The experiential marketing solution generation module is used to transform an abstract set of paths into marketing content; it includes a template matching submodule, a copywriting adjustment submodule, and a multimodal output submodule. The closed-loop optimization module is used to quantify the marketing effectiveness metrics corresponding to different path strategies through the A / B testing module.

[0014] In a preferred embodiment of the present invention, the scene factor acquisition submodule calculates the scene impact value corresponding to each factor through a predefined quantization algorithm model. The scene factors include time factors, weather factors, and geographical location factors. The weight update submodule uses a multiplication correction model to update the basic weights of the relationships between entities in the basic knowledge graph in seconds, and uses a sliding window mechanism to achieve a smooth transition of weights. The reinforcement learning optimization submodule continuously optimizes the calculation parameters of the scene impact value based on user feedback data, forming a weight correction closed loop. The user feedback data includes click-through rate and order rate.

[0015] This invention addresses the shortcomings of the prior art and has the following beneficial effects: This invention comprehensively enhances the accuracy, diversity, real-time performance, and user experience of marketing recommendation systems. Through the integration of a dynamic weight correction mechanism, a multi-objective path discovery algorithm, experiential marketing solution generation, and end-to-end closed-loop optimization, the system can respond to changes in the external environment in real time, balance multiple user needs, and generate marketing content with emotional resonance.

[0016] By collecting time, weather, and geographic location factors in real time and using a quantitative algorithm model to calculate the scene impact value, a multiplicative correction model is employed to dynamically update the knowledge graph relationship weights, combined with sliding window averaging to suppress noise. A reinforcement learning model is integrated, using click-through rate and order conversion rate as reward signals to adaptively optimize weight parameters. This upgrades the static knowledge graph into a dynamic perception system, enabling the associated weights to respond to scene changes within seconds, and the weight correction mechanism works in conjunction with the closed-loop optimization module to form a continuous evolution capability.

[0017] The system performs parallel searches for the optimal paths based on association strength, diversity, and novelty on a dynamic graph; it dynamically balances the weights of these three types of paths using a Pareto optimal selection algorithm; and it generates multimodal marketing content packages by combining user requests, real-time scenarios, and marketing objectives. This overcomes the limitations of single-target recommendations in existing technologies by integrating multiple paths to balance relevance, diversity, and novelty. It also addresses the issues of mechanical and unengaging recommendation results in existing technologies, thereby enhancing long-term user engagement.

[0018] The various modules of this invention achieve a synergistic enhancement effect through end-to-end closed-loop optimization. Dynamic weight correction provides a real-time updated graph basis for multi-objective path discovery, while the multi-path results serve as an input source for experiential generation. The closed-loop optimization module continuously adjusts the weight correction parameters and path weight ratios through A / B testing feedback, providing an evolving and implementable solution for intelligent marketing in the high-frequency consumption field. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a flowchart of an embodiment of the present invention; Figure 2 This is a framework diagram of an embodiment of the present invention; Figure 3 A schematic diagram of an electronic device that can be used to implement embodiments of the present invention is shown. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0022] In the description of this application, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting the scope of protection of this application. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, features defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0023] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this application based on the specific circumstances.

[0024] Application Overview: Early marketing rule systems, and later deep learning models trained on historical behavior, all relied on fitting past static patterns. However, restaurant consumption decisions are highly dependent on real-time contexts such as time, weather, and location. Existing technologies primarily make recommendations based on users' historical preferences, resulting in relatively limited content. Furthermore, even if existing marketing methods provide accurate judgments, they struggle to deliver compelling marketing content based on those judgments, thus failing to achieve effective marketing results.

[0025] This invention addresses the static nature of knowledge graphs by constructing real-time contextual quantification, enabling the association weights to respond to changes in the external environment. It achieves a balance of complex user psychological needs within a single model by parallel computing of three paths—association, diversity, and novelty—and utilizing the Pareto frontier for dynamic weighting. Finally, by using the quantified score of the recommendation path as an input variable for natural language generation, it realizes the transformation from data-driven recommendations to marketing content, and forms an end-to-end optimization loop through A / B testing feedback.

[0026] like Figure 1 As shown, a method for constructing a marketing product knowledge graph includes: A basic knowledge graph is constructed, which is a weighted directed graph consisting of a set of vertices and a set of weighted directed edges. Each edge has a clear start point, end point, and corresponding numerical weight. It includes four types of entity nodes: user nodes, dish nodes, dish category nodes, and consumption scenario nodes. Entity nodes are connected through predefined relationship types, and each relationship is assigned an initial basic weight based on historical interaction data statistics. Specifically, user nodes are associated with a summary of their historical behavioral characteristics using a unique identifier as the key. This summary includes the set of dishes they have interacted with in the past ninety days, the average order value range, the distribution of preferred cuisine types, and statistics on dining time preferences.

[0027] The dish node represents a specific salable catering product, including fields such as dish name, standard price, main ingredient composition, calorie value, cooking time, whether it contains allergens, and whether it supports takeout, and is kept synchronized with the central commodity master data system through a standardized coding system.

[0028] The food category nodes are classified at a high level based on flavor, such as Sichuan cuisine, Cantonese cuisine, and Japanese cuisine; cooking method, such as steaming, frying, and stewing; or ingredient source, such as seafood, vegetarian, and organic. Each category node is configured with a unique classification code and semantic tag.

[0029] The consumption scenario nodes define typical dining situations, including six preset types: breakfast, lunch, dinner, late-night snack, holiday gatherings, and business banquets. Each type corresponds to a set of time window rules and social attribute tags.

[0030] Entity nodes are connected through predefined relationship types. In the basic knowledge graph, the initial basic weight of the "user-preference-dish" relationship is calculated by weighting the number of times a user clicks on a dish with the number of times they place an order; the specific formula is as follows: ; in The number of times a user clicks on a dish (p). This represents the total number of clicks for this user, and `corder` represents the number of orders placed. Total number of orders and The preset empirical coefficients satisfy... + =1.

[0031] The relationship between "dish-attribute-dish category" is a one-to-one mapping with a fixed weight of 1.0; The initial weight of the "dish-match-consumption scenario" relationship is determined based on the historical sales proportion of the dish in the corresponding consumption scenario; for example, if the sales proportion of spicy hot pot is very high in the scenario of friends gathering, then its corresponding relationship weight is high.

[0032] The initial weight of the "user-location-consumption scenario" relationship is based on the user's current timestamp and geographical location, inferring the most likely consumption scenario in real time; for example, being near an office building at 18:00 is inferred to be a weekday dinner.

[0033] All entity nodes and relation edges are persistently stored in a distributed graph database. This database adopts an attribute graph model, supports flexible expansion of node and edge attributes, and uses a sharding strategy to evenly distribute data with trillions of edges across multiple storage nodes; the query interface supports the Cypher language, which can efficiently perform multi-hop traversals starting from user nodes.

[0034] The system collects three types of external situational signals in real time: time factor, weather factor, and geographical location factor, and calculates the scene impact value through the corresponding quantization algorithm model. The quantization algorithm model adopts a piecewise linear function or lookup table mapping mechanism, and the output value range is limited to between zero and one. The time factor includes the current time period and whether it is a statutory holiday or a traditional festival. The quantification model of the time factor adopts a piecewise linear function: if it is breakfast time and not a holiday, the output value is 0.3; if it is dinner time and a weekend, the output value is 0.7; if it is the Spring Festival, the output value is 0.9 regardless of the time period.

[0035] Weather factors are obtained by calling third-party meteorological APIs, including real-time temperature, precipitation probability, and air quality index. The quantification model of weather factors adopts a lookup table mapping mechanism. For example, when the temperature is below 5°C and the air quality index is greater than 150, the output value is 0.6, indicating that the user tends to choose hot soup or high-calorie dishes. When the precipitation probability is higher than 80%, the output value is 0.5, indicating that the user may prefer convenient takeaway options.

[0036] The geolocation factor uses geofencing technology to map the user's location coordinates to one of the following categories: business district, residential area, office area, transportation hub, or tourist attraction. Each category is assigned a unique area code. The quantification model of the geolocation factor is also based on a lookup table. The office area outputs 0.8 at noon on weekdays, and the tourist attraction outputs 0.75 on holidays.

[0037] The multiplicative correction model is used to dynamically update the basic weights of the relationships affected by external contextual signals, resulting in new weights. The new weights are calculated as follows: ; in The original base weights, This represents the scene impact value of the time factor. Similarly, the weather factor affects the "dish-suitability-consumption scene" relationship, and the geographic location factor affects the "user-current-consumption scene" relationship. All update operations are performed in the memory cache layer of the graph database to ensure low latency. To suppress transient noise, the system introduces a sliding window mechanism with a length of 30 seconds, and the weight values ​​of the same relationship after three consecutive updates within the window are arithmetically averaged as the final effective weight.

[0038] The dynamic weight adjustment process integrates a reinforcement learning model, which adopts the Q-learning framework. The core of Q-learning is based on the Bellman equation, and the learning objective is to estimate a Q-value function.

[0039] The current scenario factor combination serves as the state space, the parameter adjustment direction and step size of the quantification algorithm model serve as the action space, the weighted sum of click-through rate and order conversion rate serves as the reward function, and the strategy network is trained through the experience replay pool.

[0040] The action space is defined as the fine-tuning of key thresholds or slopes in the quantization algorithm model; for example, adjusting the output value during breakfast from 0.3 to 0.32. The reward function is defined as... ; in For click-through rate, To improve order conversion rate, The balancing factor is set to 0.6. Each time a user completes a recommendation interaction, the system records the user's state, selected action, and reward, storing these in the experience replay pool. During the training phase, batches of samples are randomly sampled from the pool, and the Bellman equation is used to update the Q-value table, thereby adjusting the quantization model parameters to make the calculation of scene impact values ​​more closely reflect real user behavior.

[0041] The search for the optimal path based on association strength, diversity, and novelty is performed in parallel on the dynamically updated knowledge graph, forming three types of path candidate sets. The optimal path for association strength is calculated using an improved Dijkstra's algorithm. The algorithm uses the reciprocal of the dynamically updated weights as the edge cost, finding the shortest path from the user node to the candidate dish node and retaining the top N paths. This algorithm uses the reciprocal of the edge weights as the path cost, i.e., the cost... Starting from the current user node, calculate the shortest path to all candidate dish nodes. Due to the large graph size, the algorithm uses bidirectional search optimization and limits the maximum number of hops to 3. Finally, the top N paths with the lowest cost are retained as the candidate set, with N being 20.

[0042] The diversity-optimal path selection involves filtering paths that cover at least three dish categories. The system iterates through all three possible paths, counts the number of categories to which the dishes at the end of each path belong, and retains only paths with ≥3 categories. Furthermore, redundant paths with a dish set overlap exceeding a threshold are removed based on Jaccard similarity. ; Here, A and B represent the sets of dishes contained in the two paths. If the similarity exceeds a preset threshold of 0.4, it is considered redundant, and the path with the stronger association is retained. This ultimately forms a diverse subset of paths with high coverage and low repetition.

[0043] The novelty-optimal path excludes dish nodes that the user has interacted with within the past 30 days from the full dish database. The remaining nodes form a candidate pool of untouched dishes. A popularity decay factor is then introduced to inversely adjust the original weights of the dishes in the candidate pool. The original weight of each dish in the pool. Perform reverse adjustment: ; in The overall popularity score for the dish, within a certain range. , Set to 0.5. Preferably, the top 15 paths with the highest weights after decay are selected as novelty candidates.

[0044] Based on the current marketing objective context, Pareto optimal selection is performed on the three types of path candidate sets. Pareto optimality refers to an ideal state of resource allocation, outputting an equilibrium set of paths carrying relevance scores, diversity scores, and novelty scores. The relevance score is taken from the normalized reciprocal of its Dijkstra path cost, the diversity score is based on a comprehensive index of the number of categories covered and the Jaccard distance, and the novelty score is... The standardized value.

[0045] For example, if the operations backend is configured for a new product promotion week, the novelty weight increases to 0.5, and the relevance weight decreases to 0.3; if it is a major promotion conversion period, the relevance weight increases to 0.6. The Pareto front is determined through non-dominated ranking: if path A is better than path B in at least one score and worse than B in any other score, then A dominates B. Paths not dominated by any path constitute the front solution set. To further ensure the uniformity of the solution set distribution, the system calculates the crowding distance and prioritizes retaining solutions with larger distances. The final output contains no more than 10 paths, each carrying three score labels.

[0046] The balanced path set is combined with a triplet of user requests, real-time scenarios, and marketing objectives to generate a multimodal marketing content package containing structured menu data, visual elements, and contextualized descriptions. User requests can be explicit keywords, such as "spicy lunch," or implicit intent signals, such as continuously browsing Sichuan cuisine product detail pages. Real-time scenarios are provided synchronously by a dynamic weight correction module, including structured data such as current time period, weather, and location. Marketing objectives are configured by the operations backend, such as increasing average order value, clearing inventory, or attracting new customers.

[0047] The system first matches the most suitable template from a pre-set marketing template library. The template library is divided into twelve categories based on specific scenarios; for example, templates for replenishing energy during overtime work emphasize high protein, quick preparation, and energizing elements, while templates for creating a festive atmosphere highlight cultural allusions, exquisite plating, and sharing attributes. The matching logic is based on a rule engine: if the real-time scenario is "weekday dinner" and the weather is cold, then "warm and comforting" templates are prioritized.

[0048] The system extracts three scores from the path set output by the multi-target path discovery module and uses them as control variables for the natural language generation engine. This engine employs a hybrid architecture: the bottom layer is a template-based rule system to ensure accurate embedding of key information; the upper layer is a finely tuned Transformer language model responsible for generating fluent and emotionally resonant text. Control variables are injected into the model through a conditional embedding layer.

[0049] For example, when the novelty score is greater than 0.7, the model activates the exploration style, using words such as "first time trying something new" and "original creation"; when the relevance score is greater than 0.8, the model activates the familiarity style, using expressions such as "your usual order" and "classic return".

[0050] The final multimodal marketing content package consists of three parts: structured menu data encapsulated in JSON format, with fields including menu ID, name, price, main ingredient list, calories, and whether it's a new product; visual elements provided by the content management system, including high-resolution main menu images, suggested pairing images, and scene / atmosphere images; and a contextualized description of a marketing copy of no more than 150 characters, which is finally pushed to the front-end app or mini-program display layer via a unified API interface.

[0051] The marketing effectiveness of different path strategy combinations is evaluated through A / B testing, and the click-through rate and order conversion rate are used as reward signals to feed back to the dynamic weight correction process, forming an end-to-end closed-loop optimization.

[0052] The system evenly divides user traffic into multiple experimental groups based on hash IDs, applying different strategy combinations to each group: Group 1 uses only the association strength path, Group 2 combines association strength and diversity paths, and Group 3 uses a three-objective balanced mode. During the testing period, four core metrics are collected in real time: page click-through rate, time spent on the menu details page, add-to-cart ratio, and final order conversion rate.

[0053] At a fixed time each day, the module performs a full performance attribution analysis. First, a t-test is conducted on each metric to determine if the differences between groups are statistically significant. If the three-objective balanced model significantly outperforms other groups in the final order conversion rate, its path selection logic and weight allocation parameters are solidified as the new default configuration. Simultaneously, the daily aggregated page click-through rate and final order conversion rate data are used as reward signals, fed back to the reinforcement learning model of the dynamic weight correction module via a message queue to update the Q-value table. This closed loop ensures the system continuously evolves in a real business environment.

[0054] like Figure 2 As shown, a device for constructing a marketing product knowledge graph includes: The basic knowledge graph construction module is used to build a basic knowledge graph for the marketing product domain; to build a structured knowledge framework for the marketing product domain, providing basic support for subsequent dynamic optimization and path discovery; to construct key entities in the marketing product domain, clarify the attribute information of each entity, mine the semantic relationships between entities, and assign initial quantitative weights to the relationships between entities based on historical data, which serve as the benchmark values ​​for subsequent dynamic correction.

[0055] The dynamic weight correction module achieves dynamic adjustment of knowledge graph weights through real-time scene perception and closed-loop learning, ensuring that the relationships match the user's real-time needs; it includes a scene factor collection submodule, a weight update submodule, and a reinforcement learning optimization submodule. The scene factor acquisition submodule calculates the scene impact value corresponding to each factor through a predefined quantification algorithm model. Scene factors include time factors, weather factors, and geographical location factors. Time factors include time periods and holidays; weather factors include temperature, precipitation, and humidity; and geographical location factors include business districts, residential areas, and office areas. This transforms unstructured scene data into calculable quantifiable scene impact values. The weight update submodule uses a multiplicative correction model to update the basic weights of the relationships between entities in the basic knowledge graph at the second level, and achieves smooth weight transition through a sliding window mechanism. The reinforcement learning optimization submodule continuously optimizes the calculation parameters of the scene impact value based on user feedback data, forming a weight correction closed loop. User feedback data includes click-through rate and order rate.

[0056] The multi-target path discovery module is used to generate a set of product paths that balances accuracy and diversity. It includes a path parallel computing submodule and a path fusion submodule. The path parallel computing submodule executes the calculation of three target paths in parallel. The path fusion submodule merges the three target paths using the Pareto optimal selection algorithm, assigns a dynamic weight ratio to each path, and generates a balanced set of product paths.

[0057] The experiential marketing solution generation module transforms abstract sets of paths into marketing content. It includes a template matching submodule, a copywriting adjustment submodule, and a multimodal output submodule. The template matching submodule automatically selects a suitable template from a pre-set marketing template library. The copywriting adjustment submodule uses natural language generation technology to inject extracted factors into the matching template to generate personalized copy. The multimodal output submodule integrates structured product data, visual elements, and contextualized descriptive copy, which can be directly used for merchant promotion.

[0058] The closed-loop optimization module quantifies the marketing performance metrics corresponding to different path strategies through the A / B testing module. The quantified performance data is then fed back to the dynamic weight correction module to adjust parameters such as scenario factor calculation parameters, weight update coefficients, and path fusion weight ratios, continuously improving recommendation accuracy, user satisfaction, and marketing conversion rates. Example 1

[0059] The basic knowledge graph construction module uses Janus Graph as a distributed graph database, deployed on a Kubernetes cluster, responsible for automatically deploying, scaling, and managing running applications. It is configured with SSD storage and memory caching, supporting millisecond-level traversal of billions of nodes and trillions of edges. The dynamic weight correction module is deployed on edge computing nodes, closer to users or data sources, significantly reducing the physical distance and time of data transmission and ensuring real-time recommendations. Located close to the user access layer, the end-to-end latency from environmental factor collection to weight updates is controlled within 500 milliseconds. The multi-objective path discovery module employs a heterogeneous computing architecture: association strength path calculation tasks are submitted to a GPU cluster, utilizing NVIDIA GPU's CUDA parallel computing platform to accelerate Dijkstra's algorithm; diversity and novelty path calculations are handled by a CPU multi-threaded task queue, with concurrent requests managed through a thread pool. The experiential marketing solution generation module is deeply integrated with the enterprise content management system (CMS) via a RESTful API. It requests the latest templates, and the CMS can also notify your system of new material releases, achieving loose coupling and standardized integration between the two independent systems. It supports online editing of marketing templates and uploading of visual materials by operations personnel, automatically triggering version release processes. The end-to-end closed-loop optimization module establishes an incremental synchronization channel with the data warehouse, using Apache Kafka to transmit real-time behavior logs. Apache Kafka reliably buffers massive amounts of data and allows multiple downstream systems to consume it simultaneously. It uses Spark Streaming to calculate streaming metrics and writes the results to a Hive table for subsequent analysis.

[0060] Specifically, the system was deployed on the intelligent recommendation platform of a national fast-food chain brand. The testing period was four consecutive weeks, covering weekdays and weekends, including a Mid-Autumn Festival promotional event. The system adopted the complete five-module architecture described in this invention. The dynamic weight correction module enabled a reinforcement learning adaptive mechanism, the multi-objective path discovery module was configured with a three-objective equilibrium mode with an initial weight ratio of 4:3:3, and the experiential marketing plan generation module enabled all twelve types of templates. During the testing period, the system served more than 500,000 unique users, processing an average of two million recommendation requests per day. Example 2

[0061] Figure 3A schematic diagram of an electronic device that can be used to implement embodiments of the present invention is shown. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0062] like Figure 3 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., which is communicatively connected to the at least one processor 11. The memory stores a computer program that can be executed by the at least one processor 11, and the computer program is executed by the at least one processor 11 to enable the at least one processor 11 to perform the method provided by the present invention.

[0063] The processor 11 can perform various appropriate actions and processes based on a computer program stored in the read-only memory (ROM) 12 or a computer program loaded from the storage unit 18 into the random access memory (RAM) 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0064] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0065] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as a method for constructing a marketing product knowledge graph.

[0066] In some embodiments, a method for constructing a marketing product knowledge graph can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the method for constructing a marketing product knowledge graph described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to perform a method for constructing a marketing product knowledge graph by any other suitable means (e.g., by means of firmware).

[0067] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard parts (ASSPs), systems-on-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0068] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0069] In the context of this invention, a computer-readable storage medium stores computer instructions that, when executed by a processor, implement a method for constructing a marketing product knowledge graph provided by this invention. The computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. The computer-readable storage medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, the computer-readable storage medium can be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0070] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a cathode ray tube (CRT) or a liquid crystal display (LCD)) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0071] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0072] A computing system can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is established by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system. It addresses the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0073] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0074] Comparative Example 1: Under the same limited scenario as Example 1, a static knowledge graph system is used; the same basic knowledge graph construction module is used; the multi-objective path discovery module only retains the calculation of the optimal path with the strongest association strength; the marketing plan generation uses a unified template and does not distinguish between scenario features; the end-to-end closed-loop optimization function is turned off, and there is no A / B testing or reinforcement learning mechanism.

[0075] Comparative Example 2: Under the same limited scenario as Example 1, a pure association strength recommendation system is adopted; the same basic knowledge graph construction module is used; the dynamic weight correction module is enabled, but only used for association strength calculation; the calculation of diversity optimal path and novelty optimal path is completely disabled; the marketing plan generation is based on a simple sorting output of association strength score; the basic A / B testing function is retained, but only the association strength parameter is optimized.

[0076] Comparative Example 3: Under the same limited scenario as Example 1, the dynamic recommendation system without closed-loop optimization includes a complete dynamic weight correction and multi-objective path discovery module; the end-to-end closed-loop optimization module is disabled, and there is no reinforcement learning feedback mechanism; the dynamic weight parameters are fixed at the initial experience value and cannot be adaptively adjusted; the A / B testing function is only used for monitoring and does not trigger automatic strategy updates; the marketing plan generation module is fully retained.

[0077] Table 1: Average values ​​of key metrics for the examples and comparative examples.

[0078]

[0079] The data shows that: Example 1 exhibits the highest page click-through rate (CTR) and conversion rate (CVR), showing an improvement of over 50% compared to Comparative Example 1. This demonstrates that dynamic weight adjustment and multi-objective path collaboration effectively match users' real-time needs. The leading time spent on the product detail page and the higher add-to-cart ratio indicate that the experiential marketing content package enhances user engagement. The outstanding performance in diversity index and new product exposure ratio validates that the multi-objective algorithm breaks down information silos. Although the response latency of this invention is slightly higher than Comparative Example 1, edge computing optimization still meets real-time requirements and is far superior to the degradation observed in Comparative Example 3.

[0080] Comparative Example 1 suffers from low click-through rate (CTR) due to static and fixed weights, failing to adapt to changing scenarios. Comparative Example 3, while dynamic, lacks closed-loop optimization, resulting in unadaptable weight parameters and a difference in conversion rate (CVR). This invention demonstrates how reinforcement learning adjusts scenario factors in real-time, enabling recommendations to accurately match the environment. Comparative Example 2 only optimizes association strength, leading to homogenized recommendations, such as users seeing the same dish for seven consecutive days, resulting in low diversity. This invention balances association strength, diversity, and novelty, effectively increasing average order value. Comparative Example 3 lacks a feedback mechanism, causing system parameters to stagnate and slow growth in new product exposure. A / B testing drives Q-learning model iteration, achieving continuous evolution.

[0081] Based on the preferred embodiments of the present invention described above, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for constructing a marketing product knowledge graph, characterized in that: include: S1. Construct a basic knowledge graph, which is a weighted directed graph containing four types of entity nodes: user nodes, dish nodes, dish category nodes, and consumption scenario nodes. The entity nodes are connected through predefined relationship types, and each relationship is assigned an initial basic weight based on the statistics of historical interaction data. S2. Real-time acquisition of three types of external situational signals: time factor, weather factor, and geographical location factor, and calculation of scene impact values ​​through corresponding quantization algorithm models; dynamic updating of the basic weights of the relationships affected by the external situational signals using a multiplicative correction model to obtain new weights, and arithmetic averaging of the weight values ​​after three consecutive updates using a 30-second sliding window. S3. On the dynamically updated knowledge graph, perform parallel searches for the optimal path in terms of relevance strength, diversity, and novelty to form three types of path candidate sets. Based on the current marketing objective context, perform Pareto optimal selection on the three types of path candidate sets and output an equilibrium path set carrying relevance score, diversity score, and novelty score. S4. Combine the balanced path set with the triplet of user requests, real-time scenarios and marketing objectives to generate a multimodal marketing content package containing structured menu data, visual elements and contextual descriptions; S5. The marketing effectiveness of different path strategy combinations is evaluated through A / B testing, and the click-through rate and order conversion rate are used as reward signals to feed back to the dynamic weight correction process, forming an end-to-end closed-loop optimization.

2. The method for constructing a marketing product knowledge graph according to claim 1, characterized in that: In the aforementioned basic knowledge graph The initial base weight of the "user-preference-dish" relationship is calculated by weighting the number of times a user clicks on a dish with the number of times they place an order. The "dish-attribute-dish category" relationship is a one-to-one mapping with a fixed weight of 1.0; The initial weight of the "dish-match-consumption scenario" relationship is determined based on the historical sales proportion of the dish in the corresponding consumption scenario; The initial weight of the "user-in-consumption scenario" relationship is inferred based on the user's current timestamp and geographical location.

3. The method for constructing a marketing product knowledge graph according to claim 1, characterized in that: The quantization algorithm model adopts a piecewise linear function or a lookup table mapping mechanism, and the output value range is limited to between zero and one. The time factor includes the current time period and whether it is a statutory holiday or a traditional festival. The weather factors include real-time temperature, precipitation probability, and air quality index; The geographic location factor uses geofencing technology to map the user's location coordinates to one of the following categories: business district, residential area, office area, transportation hub, or tourist attraction.

4. The method for constructing a marketing product knowledge graph according to claim 1, characterized in that: The dynamic weight correction process integrates a reinforcement learning model, which adopts the Q-learning framework. The current scenario factor combination serves as the state space, the parameter adjustment direction and step size of the quantification algorithm model serve as the action space, the weighted sum of click-through rate and order conversion rate serves as the reward function, and the strategy network is trained through the experience replay pool.

5. The method for constructing a marketing product knowledge graph according to claim 1, characterized in that: The optimal path for association strength is calculated using an improved Dijkstra algorithm. The inverse of the dynamically updated weight is used as the edge cost. The shortest path from the user node to the candidate dish node is found, and the first N paths are retained. The optimal diversity path is obtained by filtering paths that cover at least three categories of dishes and removing redundant paths with a higher overlap of dish sets based on Jaccard similarity. The novelty-optimal path forms a candidate pool of untouched dishes by excluding dish nodes that users have interacted with in the past 30 days, and introduces a popularity decay factor to reverse the original weight of dishes in the candidate pool.

6. The method for constructing a marketing product knowledge graph according to claim 1, characterized in that: In the Pareto optimal selection process, the scoring weight ratio of the three types of paths is dynamically adjusted according to the current marketing objective context; If a user is detected to be in a scenario with a strong exploratory intent, the weight of the novelty path is increased; if the user is in an operational cycle with high conversion pressure, the weight of the association path is increased; the Pareto front solution set is determined by non-dominated sorting and crowding distance calculation.

7. The method for constructing a marketing product knowledge graph according to claim 1, characterized in that: The generation of the multimodal marketing content package includes: matching template structures that fit the real-time scenario and marketing objectives from a pre-set marketing template library; using path relevance scores, diversity scores, and novelty scores as control variables for the natural language generation engine; the natural language generation engine adopts a rule-based and neural network hybrid architecture, and dynamically adjusts the emotional tendency, information density, and narrative rhythm of the copy according to the control variables.

8. An apparatus for constructing a marketing product knowledge graph, used to implement the method for constructing a marketing product knowledge graph according to any one of claims 1-7, characterized in that: include: The basic knowledge graph construction module is used to build a basic knowledge graph for the marketing product domain; The dynamic weight correction module achieves dynamic adjustment of knowledge graph weights through real-time scene perception and closed-loop learning, ensuring that the relationships match the user's real-time needs. It includes a scene factor acquisition submodule, a weight update submodule, and a reinforcement learning optimization submodule; The multi-target path discovery module is used to generate a set of product paths that balances accuracy and diversity; it includes a path parallel computing submodule and a path fusion submodule. The experiential marketing solution generation module is used to transform abstract sets of paths into marketing content; It includes a template matching submodule, a copywriting adjustment submodule, and a multimodal output submodule; The closed-loop optimization module is used to quantify the marketing effectiveness metrics corresponding to different path strategies through the A / B testing module.

9. The apparatus for constructing a marketing product knowledge graph according to claim 8, characterized in that: The scene factor acquisition submodule calculates the scene impact value corresponding to each factor through a predefined quantization algorithm model. The scene factors include time factors, weather factors, and geographical location factors. The weight update submodule uses a multiplication correction model to update the basic weights of the relationships between entities in the basic knowledge graph in seconds, and uses a sliding window mechanism to achieve a smooth transition of weights. The reinforcement learning optimization submodule continuously optimizes the calculation parameters of the scene impact value based on user feedback data, forming a weight correction closed loop. The user feedback data includes click-through rate and order rate.

Images