Digital precision traffic diversion and sales management method and system integrating knowledge graph

Abstract

The present invention relates to the technical field of digital precise drainage and consignment management methods. More specifically, it relates to a digital precise drainage and consignment management method and system integrating a knowledge graph. The method includes the following steps: obtaining and preprocessing multi-source heterogeneous data; realizing multi-modal data fusion and entity recognition based on the preprocessed data; performing relationship extraction and attribute mapping according to the fused data and the recognized entities; constructing a dynamic time-series graph embedding based on the extracted relationships and mapped attributes; using the dynamic time-series graph embedding to perform knowledge reasoning and graph completion; performing real-time update and incremental learning on the completed knowledge graph; generating a precise drainage strategy based on the updated knowledge graph; optimizing the consignment process according to the precise drainage strategy; evaluating and providing feedback on the optimized consignment process, and using the feedback results to optimize each step of the method. The implementation of the dynamic time-series graph embedding can capture changes in user interests and market trends in real time, improving the timeliness of recommendations.

Description

Technical Field

[0001] The present invention relates to the technical field of digital precise drainage and consignment management methods, and more specifically, to a digital precise drainage and consignment management method and system integrating a knowledge graph. Background Art

[0002] With the rapid development of e-commerce, digital marketing and precise drainage have become one of the core competitiveness of major e-commerce platforms. In recent years, recommendation systems and data analysis technologies have made remarkable progress in this field, but still face many challenges.

[0003] Traditional recommendation systems mainly rely on collaborative filtering algorithms, such as user-based collaborative filtering and item-based collaborative filtering. Although these methods are simple and effective, they have limitations such as the cold start problem, data sparsity problem, and difficulty in capturing the dynamic changes of user interests. For example, for new users or new products, due to the lack of historical interaction data, traditional collaborative filtering algorithms often cannot provide effective recommendations.

[0004] To solve these problems, researchers have introduced content-based recommendation methods and hybrid recommendation methods. These methods improve the recommendation quality by analyzing product features and user profiles. However, these methods usually can only process structured data and are difficult to fully utilize the rich information contained in a large amount of unstructured data (such as user reviews, product descriptions, etc.).

[0005] In recent years, the application of deep learning technologies has brought new breakthroughs to recommendation systems. For example, collaborative filtering methods based on deep neural networks can automatically learn feature representations and improve the accuracy of recommendations. However, these methods still mainly focus on the direct interaction between users and products and are difficult to model complex multi-entity relationships.

[0006] The introduction of knowledge graph technology provides a new idea for solving the above problems. By representing entities such as users, products, attributes, and their relationships as a graph structure, a knowledge graph can better capture complex semantic relationships. However, the existing knowledge graph-based recommendation methods still have the following problems:

[0007] 1. Difficulty in integrating multi-source heterogeneous data: Existing methods often have difficulty effectively integrating heterogeneous data from different sources, such as user behavior data, product attribute data, social network data, etc.

[0008] 2. Insufficient dynamic modeling: Most methods construct static knowledge graphs and are difficult to reflect the dynamic changes of user interests and market trends in a timely manner.

[0009] 3. Limited reasoning ability: Existing methods mainly rely on simple graph embedding or path mining techniques and lack the in-depth reasoning ability for complex semantic relationships.

[0010] 4. Lack of interpretability: Many deep learning-based methods are "black box" models, making it difficult to provide clear explanations for recommendation results.

[0011] 5. Long-tail problem: Existing methods tend to recommend popular products, with insufficient exploration and utilization of long-tail products.

[0012] 6. Lack of real-time performance and scalability: Facing the massive data and high concurrent requests of large-scale e-commerce platforms, many methods are difficult to meet the requirements of real-time performance and scalability.

[0013] These problems severely limit the application effect of existing technologies in e-commerce precision marketing and consignment management. Therefore, there is an urgent need for a new method that can comprehensively solve the above problems to improve the effect of digital precision drainage and consignment management. Summary of the Invention

[0014] In view of the above problems existing in the prior art, the present invention proposes a digital precision drainage and consignment management method and system integrating a knowledge graph. This method constructs a comprehensive, dynamic, and intelligent knowledge graph system by innovatively combining advanced technologies such as quantum computing ideas, multi-modal data fusion, dynamic time-series graph embedding, and multi-granularity causal reasoning.

[0015] The present invention provides a digital precision drainage and consignment management method integrating a knowledge graph, including the following steps: obtaining and preprocessing multi-source heterogeneous data; realizing multi-modal data fusion and entity recognition based on the preprocessed data; performing relationship extraction and attribute mapping according to the fused data and recognized entities; constructing a dynamic time-series graph embedding based on the extracted relationships and mapped attributes; using the dynamic time-series graph embedding to perform knowledge reasoning and graph completion; performing real-time update and incremental learning on the completed knowledge graph; generating a precision drainage strategy based on the updated knowledge graph; optimizing the consignment process according to the precision drainage strategy; evaluating and providing feedback on the optimized consignment process, and using the feedback results to optimize each step of the method.

[0016] Specifically, in the step of obtaining and preprocessing multi-source heterogeneous data, an adaptive multi-dimensional anomaly detection algorithm is used, and this algorithm includes the following steps:

[0017] a) Using locality-sensitive hashing to map high-dimensional data to a low-dimensional space, where the mapping function is:

[0018]

[0019] where x is the input data, a is a random vector, b is a random offset, and w is the bucket width; b) Applying dynamic clustering within each hash bucket, using an improved DBSCAN algorithm, where the ∈ parameter is adaptively adjusted:

[0020] ∈i = μ i + α·σ i

[0021] wherein, ∈ i is the neighborhood radius of the i-th dimension, μ i and σ i are the mean and standard deviation of the i-th dimension respectively, and α is an adjustable parameter;

[0022] c) Determine the anomaly score:

[0023]

[0024] wherein, AS(x) is the anomaly score of point x, d(x, C) is the distance from point x to the nearest cluster center C, and σ C is the standard deviation of cluster C;

[0025] d) Dynamically determine the anomaly threshold based on extreme value theory.

[0026] Specifically, in the steps of implementing multi-modal data fusion and entity recognition, it includes using a quantum-inspired multi-modal fusion network, including the following steps:

[0027] a) Encode the data x of each modality i into a quantum state:

[0028]

[0029] wherein, |ψ i > is the quantum state, α ij is the complex amplitude, |j> is the ground state, and satisfies Σ j |α ij | 2 = 1;

[0030] b) Apply an inter-modal entanglement layer and design an entanglement operator U ent : U ent = exp(-iH ent ),

[0031] wherein, σ x , σ y , σ z are Pauli matrices, and J ij is the learnable entanglement strength;

[0032] c) Use a quantum measurement layer to design a set of learnable measurement operators M k , and obtain a classical output:

[0033] y k = <ψ|M k |ψ>

[0034] where y k is the measurement result and |ψ> is the quantum state;

[0035] d) Process the measurement result using a traditional neural network layer to obtain the final fused representation.

[0036] Specifically, in the step of performing relation extraction and attribute mapping, a graph-enhanced recursive neural inference network is used, including the following steps:

[0037] a) Embed the existing knowledge graph using the TransE algorithm:

[0038] ∥h + r - t∥2≈0

[0039] where h, r, and t are the embedding vectors of the head entity, relation, and tail entity, respectively;

[0040] b) Encode the input sentence using a BiLSTM:

[0041] h t = BiLSTM(x t , h t-1 )

[0042] where h t is the hidden state at time t and x t is the input at time t;

[0043] c) Determine the attention score between the sentence representation and the graph embedding:

[0044] α i = softmax(v T tanh(W[h i ; g j ))

[0045] where α i is the attention score, j i is the sentence hidden state, g j is the graph embedding, and v and W are learnable parameters;

[0046] d) Design a recursive structure to simulate multi-hop reasoning:

[0047] z t = f(W z [h t ; z t-1 ; c t )

[0048] where z t is the inference state at time t, h t is the sentence hidden state, c tis the graph-enhanced context vector, W z is a learnable parameter;

[0049] e) Use the softmax layer to adjust the final z T Categorize relationships.

[0050] Specifically, the step of constructing the dynamic time sequence graph embedding includes using a quantum-inspired space-time graph embedding algorithm, which includes the following steps:

[0051] a) Initialize entity e i and the relationship j The quantum bitmap embedding of:

[0052] Q(e i )=|ψ i >=α i |0>+β i |1>

[0053] Q(r j )=|φ j >=γ j |0>+δ j |1>where |α i | 2 +|β i | 2 =1,|γ j | 2 +|δ j | 2 =1;

[0054] b) Define the nonlinear space-time evolution operator U t :

[0055] U t =exp(-iH t )

[0056] Among them, H t is a time-varying Hamiltonian operator:

[0057]

[0058] σ k is the Pauli matrix, w k (t) and J ij (t) is the time-varying coefficient;

[0059] c) Introducing the quantum walk attention mechanism:

[0060] A QW (i,j,t)=<ψ i (t)|U QW (t)|ψ i(t) >

[0061] Among them, U QW (t) is a quantum walk operator based on a graph structure;

[0062] d) Use fractional quantum field theory for multi-scale time-dependent modeling:

[0063] Φ(x, t) = ∫K α (x - y, t - s)ψ(y, s)dy ds

[0064] Among them, K α is a fractional diffusion kernel, and α ∈ (0, 2) is a fractional parameter;

[0065] e) Introduce a topological entropy regularization term:

[0066]

[0067] Among them, p ij is the transition probability from node i to node j;

[0068] f) Design an optimization objective for non-equilibrium thermodynamics:

[0069]

[0070] Among them, L task is the main task loss, is the change rate of the system entropy;

[0071] g) Design an adaptive quantum measurement strategy:

[0072]

[0073] Among them, M is a learnable measurement operator.

[0074] Specifically, in the steps of performing knowledge reasoning and graph completion, it includes using a multi-granularity causal reasoning network, including the following steps:

[0075] a) Use the spectral clustering algorithm to divide the knowledge graph into multiple granularity levels;

[0076] b) At each granularity level, use the PC algorithm to discover potential causal relationships;

[0077] c) Use the structural equation model (SEM) to estimate the causal strength:

[0078] X j = f j (PA j , U j )

[0079] Among them, X jis the dependent variable, PA j is the set of parent nodes of X j , and U j is the exogenous variable;

[0080] d) Design an attention mechanism to fuse causal information at different granularities:

[0081] α g = softmax(v T tanh(W g h g ))h final = ∑ g α g h g where α g is the attention weight, h g is the representation of causal information at different granularities, and v and W g are learnable parameters;

[0082] e) Perform link prediction based on the final representation h final .

[0083] Specifically, in the steps of generating the accurate drainage strategy, a hierarchical spatio-temporal attention strategy network is used, including the following steps:

[0084] a) Define the state representation:

[0085] s t = [u t ; p t ; c t

[0086] where u t is the user feature, p t is the product feature, and c t is the context feature;

[0087] b) Define a two-level action space: A l is the high-level strategy) and A2 is the specific operation;

[0088] c) Determine the spatio-temporal attention:

[0089] α t = softmax(v T tanh(W s s t + W h h t-1 ))

[0090] where h t-1 is the historical hidden state, and v, W s , W h are learnable parameters;​

[0091] d) Construct the policy network:

[0092] π θ (a|s) = softmax(f θ (s, α))

[0093] where f θ is a parameterized neural network,

[0094] e) Construct the value network:

[0095] V φ (s) = g φ (s, α)

[0096] where g φ is a parameterized neural network;

[0097] f) Use the asynchronous advantage actor-critic algorithm for training.

[0098] Specifically, in the steps of optimizing the consignment process, it further includes using a graph neural network for consignment agent matching and anomaly detection.

[0099] Specifically, in the steps of performing effect evaluation and feedback, dynamic weight multi-objective Bayesian optimization is used, including the following steps:

[0100] a) Define multiple objective functions:

[0101] F(x) = [f1(x), f2(x),..., f m (x)]

[0102] where f i (x) is the i-th objective function,

[0103] b) Establish a Gaussian process model for each objective function:

[0104] f i (x) ~ GP(μ i (x), k i (x, x′))

[0105] where μ i (x) is the mean function, k i (x, x′) is the kernel function;

[0106] c) Determine the dynamic weights:

[0107]

[0108] where λ is the temperature parameter, is the current optimal value of objective i, σi is the standard deviation;

[0109] d) Calculate the acquisition function:

[0110]

[0111] Among them, EI i (x) is the expected improvement of the ith objective;

[0112] e) Use the DIRECT algorithm to maximize EI(x) and select the next evaluation point;

[0113] f) Evaluate the new point, update the Gaussian process model and dynamic weights, and repeat steps ce.

[0114] A digital precise traffic diversion and agency sales management system integrating knowledge graphs includes: a data preprocessing module for acquiring and preprocessing multi-source heterogeneous data; a data fusion module for realizing multimodal data fusion and entity recognition; a relationship extraction module for performing relationship extraction and attribute mapping; a graph embedding module for constructing dynamic time series graph embedding; a knowledge reasoning module for performing knowledge reasoning and graph completion; an update learning module for performing real-time update and incremental learning; a strategy generation module for generating precise traffic diversion strategies; a process optimization module for optimizing agency sales processes; an evaluation feedback module for performing effect evaluation and feedback; wherein the output of the evaluation feedback module is connected to the input of the data preprocessing module to form a closed-loop optimization.

[0115] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0116] 1. Effective integration of multi-source heterogeneous data: Through the AMODA algorithm and the QIMMFN algorithm, this method can effectively process and fuse heterogeneous data from different sources to build a more comprehensive user and product representation.

[0117] 2. Dynamic time series modeling: The QISGE algorithm implements dynamic time series graph embedding, which can capture changes in user interests and market trends in real time and improve the timeliness of recommendations.

[0118] 3. Deep knowledge reasoning: The MGCIN algorithm greatly enhances the system's reasoning ability through multi-granularity causal reasoning, and can discover deeper semantic relationships.

[0119] 4. Personalized strategy generation: The HSASPN algorithm can generate highly personalized traffic strategies based on user characteristics and contextual information to improve user satisfaction.

[0120] 5. Long-tail product mining: By comprehensively considering multiple factors, this method significantly improves the utilization rate of long-tail products and enriches the diversity of recommendations.

[0121] 6. Real-time performance and scalability: By adopting incremental learning and distributed processing technologies, this method can handle large-scale data and high-concurrency requests, meeting the requirements of real-time recommendation.

[0122] 7. Optimization of the consignment process: Through graph neural network technology, this method realizes the intelligence of consignment agent matching and risk management, improving the overall operation efficiency.

[0123] 8. Adaptive optimization: The DWMOBO algorithm realizes the dynamic optimization of multiple objectives, enabling the system to balance different objectives (such as conversion rate, user satisfaction, platform revenue, etc.) and adaptively adjust strategies.

[0124] 9. Interpretability: The method based on the knowledge graph provides good interpretability, which helps to understand the reasons behind the recommendation results and increases user trust.

[0125] 10. Knowledge accumulation: Through continuous learning and graph completion, this method can continuously accumulate and update domain knowledge, forming valuable knowledge assets.

[0126] In summary, the digital precise drainage consignment management method integrating the knowledge graph provided by the present invention comprehensively improves the accuracy, personalization degree and intelligent level of digital marketing by innovatively integrating a variety of advanced technologies. It not only solves many problems faced by the existing technology, but also provides strong technical support for the precise marketing and user service of e-commerce platforms, and is expected to significantly improve the operation efficiency and user experience of the platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0127] Figure 1 It is a flowchart of the digital precise drainage consignment management method integrating the knowledge graph of the present invention.

[0128] Figure 2 It is a schematic diagram of the framework principle of the intelligent market product demand prediction system based on deep learning of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0129] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following combines the drawings and preferred embodiments to describe in detail its specific implementation manner, structure, features and effects. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0130] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs.

[0131] The present invention proposes a digital precise drainage and consignment management method and system integrating knowledge graphs, which is particularly suitable for precision marketing and user service optimization of e-commerce platforms. The following will describe the implementation manner of the present invention in detail in combination with the application scenario of a large e-commerce platform.

[0132] This e-commerce platform has millions of users and hundreds of thousands of products, generating a large amount of user behavior data and transaction data every day. The main challenges faced by the platform include: how to accurately recommend suitable products to users, how to optimize the matching and management of consignment agents, and how to improve the overall marketing efficiency and user satisfaction. The present invention provides an innovative solution by integrating knowledge graph technology, which specifically includes the following steps:

[0133] Data preprocessing step. In this step, the data preprocessing module first obtains and preprocesses heterogeneous data from multiple sources. These data sources include:

[0134] a) User personal information database: contains the basic information of users, such as age, gender, geographical location, etc.

[0135] b) User behavior logs: record the browsing, searching, clicking and other behaviors of users on the platform.

[0136] c) Product information database: contains the detailed information of products, such as category, price, inventory, etc.

[0137] d) Transaction record database: records all completed transaction information.

[0138] e) Social media data: public information of users on major social platforms obtained through APIs.

[0139] The data preprocessing module uses the Adaptive Multi-Dimensional Outlier Detection Algorithm (AMODA) to process this heterogeneous data. The specific steps of the AMODA algorithm are as follows:

[0140] a) Use Locality-Sensitive Hashing (LSH) to map high-dimensional data to a low-dimensional space:

[0141]

[0142] where x is the input data (for example, a feature vector of a user), a is a random vector (usually sampled from a standard normal distribution), b is a random offset (usually uniformly sampled in the range [0, w)), and w is the bucket width (determined according to the data distribution, usually between 1 and 10).

[0143] For example, assume we have a user feature vector x = [25, 1, 10000, 50], representing age, gender, annual income, and the number of logins in the past 30 days.

[0144] We may choose a = [0.1, -0.5, 0.001, 0.2], b = 2, w = 5.

[0145]

[0146] b) Apply dynamic clustering within each hash bucket, using an improved DBSCAN algorithm, where the ∈ parameter is adaptively adjusted:

[0147] ∈ i = μ i + α·σ i

[0148] where ∈ i is the neighborhood radius of the i-th dimension, μ i and σ i are the mean and standard deviation of the i-th dimension respectively, and α is an adjustable parameter (usually between 0.5 and 2).

[0149] Suppose in a certain hash bucket, the mean μ of the age dimension is 30 and the standard deviation σ is 5, and we choose α = 1, then:

[0150] ∈ age = 30 + 1·5 = 35

[0151] c) Calculate the anomaly score:

[0152]

[0153] where AS(x) is the anomaly score of point x, d(x, C) is the distance from point x to the nearest cluster center C, and σ C is the standard deviation of cluster C.

[0154] Suppose the distance d(x, C) from a user point x to the nearest cluster center is 10, and the standard deviation σ C of this cluster is 3, then:

[0155]

[0156] d) Dynamically determine the anomaly threshold using Extreme Value Theory (EVT).

[0157] Through the AMODA algorithm, we can effectively detect and process abnormal data, such as identifying abnormal user behaviors or product information, thereby improving the accuracy of subsequent analysis.

[0158] Data fusion step. In this step, the data fusion module uses a Quantum-inspired Multimodal Fusion Network (QIMMFN) to achieve multimodal data fusion and entity recognition. This is particularly effective for integrating various information of users, such as personal profiles, browsing histories, purchase records, social media data, etc. The specific steps of the QIMMFN algorithm are as follows:

[0159] a) Encode the data x of each modality i into a quantum state:

[0160]

[0161] where |ψ i > is the quantum state, α ij is the complex amplitude, |j> is the basis state, satisfying Σ j |α ij | 2 = 1.

[0162] Suppose we have the text description and image data of a user. For the text, we may have |ψ text > = 0.6|0> + 0.8|1>, and for the image, we may have b) Apply the inter-modal entanglement layer and design the entanglement operator U ent :

[0163] U ent = exp(-iH ent

[0164] where is the Pauli matrix and J ij is the learnable entanglement strength. This step allows the information of different modalities to influence each other. For example, the user's browsing history may affect the interpretation of their social media behavior;

[0165] c) Use the quantum measurement layer and design a set of learnable measurement operators M k , to obtain the classical output:

[0166] y k = <ψ|M k |ψ>

[0167] where y k is the measurement result and |ψ> is the quantum state.

[0168] Suppose we design a measurement operator M = |0><0| - |1><1|. For the previous text quantum state, the measurement result will be:

[0169] y = <ψ text |M|ψ text > = 0.6 2 - 0.82 = -0.28

[0170] d) Process the measurement results using traditional neural network layers to obtain the final fused representation. Through the QIMMFN algorithm, we can effectively fuse multi-modal data. For example, integrate the user's text comments, shopping history, and social media activities into a unified user representation, which is crucial for subsequent personalized recommendations and precision marketing.

[0171] Relation extraction step. In this step, the relation extraction module uses the graph-enhanced recursive neural inference network (GERNRN) for relation extraction and attribute mapping. This is particularly important for understanding the complex relationships between users and products, users and users, and products and products. The specific steps of the GERNRN algorithm are as follows:

[0172] a) Embed the existing knowledge graph using the TransE algorithm

[0173] ∥h + r - t∥2 ≈ 0

[0174] where h, r, and t are the embedding vectors of the head entity, relation, and tail entity respectively. Suppose we have a relation "User A likes Product B", we may get: h 用户A + r 喜欢 ≈ t 商品B .

[0175] b) Encode the input sentence using BiLSTM: h t = BiLSTM(x t , h t-1 )

[0176] where h t is the hidden state at time t, and x t is the input at time t. This step can process text data such as user comments or product descriptions and capture sequence information.

[0177] c) Calculate the attention score between the sentence representation and the graph embedding:

[0178] α i = softmax(v T tanh(W[h i ; g j ))

[0179] where α i is the attention score, h i is the sentence hidden state, g j is the graph embedding, and v and W are learnable parameters. This step enables the model to focus on the knowledge graph information most relevant to the current input

[0180] d) Design a recursive structure to simulate multi-hop reasoning:

[0181] z t = f(W z [h t ; z t-1 ; c t )

[0182] where z t is the reasoning state at time t, h t is the sentence hidden state, c t is the context vector enhanced by the knowledge graph, and W z are learnable parameters. This allows the model to perform complex reasoning, such as "User A likes product B, product B belongs to category C, so User A may also like other products in category C".

[0183] e) Use a softmax layer to perform relation classification on the final z T . Through the GERNRN algorithm, we can effectively extract the relationships between entities, such as identifying "products that users often purchase together", "similar user groups", etc., which is crucial for constructing a richer knowledge graph and making more accurate recommendations.

[0184] Graph embedding step. In this step, the graph embedding module uses the Quantum-inspired Spatiotemporal Graph Embedding (QISGE) algorithm to construct dynamic temporal graph embeddings. This is particularly important for capturing the dynamic changes in user interests and market trends. The specific steps of the QISGE algorithm are as follows:

[0185] a) Initialize the quantum bitmap embeddings of entities e i and relations r j :

[0186] Q(e i ) = |ψ i > = α i |0> + β i |1>

[0187] Q(r j ) = |φ j > = γ j |0> + δ j |1>

[0188] where, |α i | 2 +|β i | 2 = 1, |γ j | 2 +|δ j | 2= 1; For a user entity, we may have Q(User A) = 0.6|0> + 0.8|1>, representing a certain characteristic of User A.

[0189] b) Define the non - linear spatio - temporal evolution operator U t :

[0190] U t = exp(-iH t )

[0191] where H t is the time - varying Hamiltonian operator:

[0192]

[0193] σ k is the Pauli matrix, w k (t) and J ij (t) are time - varying coefficients; This allows the representation of entities and relationships to change dynamically over time, for example, to capture seasonal changes in user interests.

[0194] c) Introduce the quantum walk attention mechanism:

[0195] A QW (i,j,t) = <ψ i (t)|U QW (t)|ψ j (t)>

[0196] where U QW (t) is the quantum walk operator based on the graph structure; This helps the model to focus on the most relevant parts of the graph, for example, to focus on other users or products most relevant to the current user in a recommendation system.

[0197] d) Use the fractional quantum field theory for multi - scale time - dependent modeling:

[0198] Φ(x,t) = ∫K α (x - y,t - s)ψ(y,s)dy ds

[0199] where K α is the fractional - order diffusion kernel and α ∈ (0,2) is the fractional - order parameter; This allows the model to capture dependencies at different time scales, such as short - term interests and long - term preferences of users.

[0200] e) Introduce the topological entropy regularization term:

[0201]

[0202] where p ijis the transition probability from node i to node j; this helps to preserve the key features of the graph structure, such as maintaining the key structure of the user social network.

[0203] f) Design the non-equilibrium thermodynamics optimization objective:

[0204]

[0205] where L task is the main task loss, is the change rate of the system entropy; this optimization objective allows the model to adapt to new changes while maintaining stability, such as adapting to new market trends while maintaining the user's basic interests.

[0206] g) Design the adaptive quantum measurement strategy:

[0207]

[0208] where M is the learnable measurement operator. This step converts the quantum representation into a classical representation for subsequent processing.

[0209] Knowledge reasoning step, in this step, the knowledge reasoning module uses the multi-granularity causal inference network (MGCIN) to perform knowledge reasoning and graph completion. This is particularly important for understanding the reasons behind user behavior and predicting future trends. The specific steps of the MGCIN algorithm are as follows:

[0210] a) Use the spectral clustering algorithm to divide the knowledge graph into multiple granularity levels. For example, we can divide the graph into user level, commodity category level, and overall market level. This allows reasoning at different levels of abstraction.

[0211] b) At each granularity level, use the PC algorithm to discover potential causal relationships. The PC algorithm constructs a causal graph through conditional independence tests. For example, at the user level, we may discover a causal chain such as "user age" → "purchase category" → "purchase frequency".

[0212] c) Use the structural equation model (SEM) to estimate the causal strength:

[0213] X j = f j (PA j , U j )

[0214] where X j is the dependent variable, PA j is the set of parent nodes of X j , and U jis an exogenous variable. For example, we might obtain an equation like: Purchase Frequency = 0.3 · Age + 0.5 · Income + 0.2 · Random Factor;

[0215] d) Design an attention mechanism to fuse causal information at different granularities:

[0216] α g = softmax(v T tanh(W g h g ))

[0217]

[0218] where α g is the attention weight, h g is the representation of causal information at different granularities, and v and W g are learnable parameters. This allows the model to dynamically adjust the attention to different granularity information according to the specific situation. For example, when predicting the purchase behavior of a certain user, it may pay more attention to user-level causal relationships; while when predicting the overall market trend, it may pay more attention to market-level causal relationships.

[0219] e) Perform link prediction based on the final representation h final This step can be used to predict the missing relationships in the knowledge graph, such as predicting the products that a user may be interested in but has not interacted with yet.

[0220] Through the MGCIN algorithm, we can gain a deeper understanding of user behavior and market dynamics, such as identifying complex patterns like "young users are more likely to purchase electronic products during festivals", thus providing more powerful support for precision marketing.

[0221] Update the learning step. In this step, the update learning module is responsible for real-time updating and incremental learning of the knowledge graph. This is crucial for maintaining the timeliness and adaptability of the system. It mainly includes the following aspects:

[0222] a) Use the Reservoir Sampling algorithm to maintain a sample pool of a fixed size to ensure the balance between new and old data. For example, we can maintain a sample pool containing the most recent 1 million user behaviors. When new data enters, it replaces old data with a certain probability.

[0223] b) Apply ElasticWeightConsolidation (EWC) technology to prevent catastrophic forgetting. EWC prevents important parameters from changing too much by adding additional loss terms to them. For example, when updating the user interest model, we will protect those parameters representing the long-term interests of users from changing drastically due to short-term behaviors.

[0224] c) Implement an adaptive-size sliding window, where the window size is dynamically adjusted according to the data stream velocity. For example, during the shopping peak period, we may reduce the window size to capture changes in user interests more quickly; while during the quiet period, we may expand the window to obtain a more stable pattern.

[0225] d) Use the Hoeffding tree algorithm for real-time decision tree updates. This allows us to quickly update classification or regression models when new data arrives, such as adjusting user segmentation or product classification in real time.

[0226] e) Adopt Apache Flink for distributed stream processing to ensure the scalability of the system. This enables the system to handle large-scale real-time data streams, such as processing clickstream data of millions of users simultaneously.

[0227] f) Implement a distributed state management mechanism based on the Raft consensus algorithm. This ensures that in a distributed environment, the state of the knowledge graph remains consistent, preventing data inconsistencies caused by node failures.

[0228] Through these technologies, the system can continuously learn and adapt to new user behaviors and market trends, maintaining the timeliness of recommendation and marketing strategies.

[0229] Policy generation step, in this step, the policy generation module uses a hierarchical spatio-temporal attention strategy network (HSASPN) to generate precise traffic-driving strategies. This is crucial for recommending the right products to the right users at the right time. The specific steps of the HSASPN algorithm are as follows:

[0230] a) Define the state representation: s t =[u t ;p t ;c t

[0231] Where u t is the user feature, p t is the product feature, and c t is the context feature. The state may include user demographic information, recent browsing history, features of the currently viewed product, time, location, etc.

[0232] b) Define a two-level action space: A1 (high-level strategy) and A2 (specific operations). For example, A1 may include strategies such as "recommend similar products", "recommend complementary products", "recommend popular products", etc., while A2 is the specific product ID or promotion plan.

[0233] c) Calculate spatio-temporal attention: α t =softmax(v T tanh(W s s t ​+W h h t-1 ))

[0234] where h t-1 is the historical hidden state, and v, W s , W h are learnable parameters. This allows the model to dynamically adjust its focus based on the current state and historical information. For example, when a user browses mobile phones, the model may pay more attention to the user's budget and brand preferences.

[0235] d) Construct the policy network:

[0236] π θ (a|s) = softmax(f θ (s, α))

[0237] where f θ is a parameterized neural network. This network outputs the probabilities of taking various actions in a given state.

[0238] e) Construct the value network:

[0239] V φ (s) = g φ (s, α)

[0240] where g φ is a parameterized neural network. This network estimates the long-term value of a given state, helping the system to balance short-term rewards and long-term user value.

[0241] f) Use the A3C (Asynchronous Advantage Actor-Critic) algorithm for training. This allows the system to learn efficiently in a large-scale parallel environment and adapt to the high-concurrency characteristics of the e-commerce platform.

[0242] Through the HSASPN algorithm, the system can generate fine-grained personalized marketing strategies. For example, for a user who often browses sports equipment on weekends, the system may push a time-limited discount on sports shoes on Friday afternoon to maximize the conversion probability.

[0243] Process optimization step. In this step, the process optimization module mainly uses graph neural networks (GNNs) to optimize the consignment process. This is crucial for improving the overall operational efficiency and user satisfaction of the platform. It mainly includes the following aspects:

[0244] a) Use Graph Attention Networks (GAT) to construct a consignor-item bipartite graph. This allows the system to capture the complex relationships between consignors and items. For example, some consignors may perform better in specific categories of items.

[0245] b) Apply the Hungarian algorithm for optimal matching. Based on the representations learned by GAT, the system can find the most suitable reseller for each product or assign the most suitable product combination to each reseller.

[0246] c) Implement graph-based anomaly detection algorithms such as GraphAutoencoder. This can help identify abnormal reselling behaviors, such as sudden spikes in sales volume or abnormal return rates.

[0247] d) Use the LocalOutlierFactor (LOF) algorithm to detect local anomalies. This helps to discover abnormal reselling behaviors in specific product categories or specific time periods.

[0248] e) Design an anomaly pattern recognition method based on temporal graphs. This can capture abnormal patterns in reselling behaviors over time, such as the difference between normal and abnormal sales fluctuations of seasonal products.

[0249] f) Construct a dynamic Bayesian network model to evaluate risks in the reselling process in real time. This allows the system to dynamically adjust risk management strategies, such as adjusting credit limits based on the reseller's historical performance and current market conditions.

[0250] g) Implement a risk mitigation strategy generator based on Monte Carlo Tree Search (MCTS). This can help the system generate a series of possible mitigation strategies and evaluate their effectiveness when potential risks are discovered.

[0251] Through these technologies, the system can continuously optimize the reselling process, improve matching efficiency, reduce risks, and thus enhance the overall platform performance.

[0252] The evaluation and feedback step, in which the evaluation and feedback module uses the Dynamic Weighted Multi-Objective Bayesian Optimization (DWMOBO) algorithm for effectiveness evaluation and feedback. This is crucial for continuously optimizing the system performance and adapting to the changing market environment. The specific steps of the DWMOBO algorithm are as follows:

[0253] a) Define multiple objective functions: F(x) = [f1(x), f2(x),..., f m (x)]

[0254] where f i (x) is the i-th objective function. For example, f1 may be the conversion rate, f2 may be user satisfaction, f3 may be platform revenue, etc.

[0255] b) Establish a Gaussian process model for each objective function:

[0256] f i (x) ∼ GP(μi (x), k i (x, x′))

[0257] Among them, μ i (x) is the mean function, k i (x, x′) is the kernel function. This allows the system to model the uncertainty of each objective function

[0258] c) Calculate the dynamic weight:

[0259]

[0260] Among them, λ is the temperature parameter (usually set between 0.1 and 1), f imax is the current optimal value of objective i, σ i is the standard deviation. This allows the system to dynamically adjust the attention to different objectives. For example, if the conversion rate is already very high, the system may pay more attention to improving user satisfaction.

[0261] d) Calculate the acquisition function:

[0262]

[0263] Among them, EI i (x) is the expected improvement of the i-th objective. This function guides the system to balance between exploration (trying new strategies) and exploitation (using known good strategies)

[0264] e) Use the DIRECT algorithm to maximize EI(x)\text{EI}(x)EI(x) and select the next evaluation point. The DIRECT algorithm is a deterministic global optimization algorithm, especially suitable for dealing with multi-objective problems.

[0265] f) Evaluate the new point, update the Gaussian process model and the dynamic weight, and repeat steps c - e.

[0266] Through the DWMOBO algorithm, the system can find the best balance among multiple potentially conflicting objectives. For example, the system may find that slightly reducing the recommendation accuracy for some high-value users but significantly increasing their exploratory recommendations can improve the overall user satisfaction and platform revenue in the long run.

[0267] Finally, the output of the evaluation feedback module is fed back to the data preprocessing module to form a complete closed-loop optimization system. For example, if it is found that the satisfaction of certain user groups continues to decline, the system may conduct a more detailed analysis of the data of these users in the data preprocessing stage, so as to pay more attention in the subsequent steps.

[0268] Through the above nine steps, the digital precision traffic generation and consignment sales management method integrated with knowledge graph of the present invention can comprehensively improve the operational efficiency and user experience of the e-commerce platform. From data processing to strategy generation, to effect evaluation and feedback, each step uses advanced algorithms and technologies to ensure the efficiency, accuracy and adaptability of the system. This method can not only achieve accurate personalized recommendations and marketing, but also optimize the entire consignment sales process, while adapting to the ever-changing e-commerce environment through continuous learning and optimization.

[0269] The digital precision traffic diversion and sales management system integrated with knowledge graph of the present invention is a highly integrated intelligent platform, which includes the following core modules:

[0270] The data preprocessing module 1 is responsible for acquiring and preprocessing multi-source heterogeneous data and using the AMODA algorithm to handle outliers and noise.

[0271] Data fusion module 2 uses the QIMMFN algorithm to achieve multimodal data fusion and entity recognition, integrating data from different sources and types into a unified representation.

[0272] Relation extraction module 3 uses the GERNRN algorithm to perform relationship extraction and attribute mapping to construct a preliminary knowledge graph structure.

[0273] The graph embedding module 4 uses the QISGE algorithm to construct dynamic temporal graph embeddings to capture the dynamic relationships between entities.

[0274] The knowledge reasoning module 5 performs knowledge reasoning and graph completion through the MGCIN algorithm to enrich the content of the knowledge graph.

[0275] The update learning module 6 is responsible for real-time updating and incremental learning to ensure that the system can adapt to the latest data and trends.

[0276] Strategy generation module 7 uses the HSASPN algorithm to generate precise traffic diversion strategies to achieve personalized recommendations and marketing.

[0277] Process optimization module 8 uses graph neural networks to optimize the sales process and improve overall operational efficiency.

[0278] The evaluation and feedback module 9 uses the DWMOBO algorithm to perform effect evaluation and feedback to continuously optimize system performance.

[0279] These modules work closely together to form a closed-loop intelligent system. Data starts with preprocessing, and then goes through steps such as fusion, relationship extraction, and graph embedding to build a rich knowledge graph. Based on this graph, the system generates accurate marketing strategies and optimizes the consignment sales process. Finally, the performance of the entire system is continuously optimized through the evaluation feedback module.

[0280] To verify the superiority of the present invention, we conducted a series of comparative experiments. The following is a detailed comparison between Example 1 of the present invention and three comparative examples:

[0281] Example 1: Implement the method and system of the present invention completely

[0282] Comparative Example 1: Traditional collaborative filtering recommendation system

[0283] Comparative Example 2: Recommendation system based on a simple graph neural network

[0284] Comparative Example 3: System that uses a traditional knowledge graph but does not include dynamic updates

[0285] Test metrics and their detection criteria and methods:

[0286] 1. Recommendation accuracy (Precision@K): The proportion of correct recommendations in the Top-K recommendations.

[0287] Detection method: Using the hold-out method, divide the user interaction data into a training set and a test set, and calculate the proportion of correctly predicted items in the test set.

[0288] 2. User satisfaction: The degree of satisfaction of users with the recommendation results.

[0289] Detection method: Through user questionnaires, use a Likert scale of 1-5 for scoring.

[0290] 3. Average processing time: The average time for the system to process one request.

[0291] Detection method: Record the system response time and calculate the average value.

[0292] 4. Long-tail product coverage rate: The proportion of long-tail products (with low sales volume) in the recommendation results.

[0293] Detection method: Define long-tail products (such as products with sales volume lower than 20% of the average), and calculate the proportion of long-tail products in the recommendation results.

[0294] 5. Knowledge graph completeness: The richness of entities and relationships in the knowledge graph.

[0295] Detection method: Calculate the number of entities and relationships in the graph and compare with the benchmark dataset.

[0296] The test results are shown in the following table:

[0297] Index Example 1 Comparative Example 1 Comparative Example 2 Comparative Example 3 Recommended Accuracy (P@10) 0.82 0.65 0.73 0.7 User Satisfaction (1 - 5 points) 4.5 3.2 3.8 3.6 Average Processing Time (ms) 150 80 200 180 Coverage Rate of Long-Tail Products (%) 35 10 20 25 Completeness of Knowledge Graph (%) 95 N / A 70 85

[0298] According to the test results, Example 1 (i.e., the complete implementation of the present invention) shows significant advantages in most metrics and can therefore be considered the best example.

[0299] The analysis and explanation are as follows:

[0300] 1. Recommendation accuracy: The method of the present invention achieves a high accuracy of 0.82, far exceeding other comparative methods. This is mainly due to the combination of the QISGE algorithm and the MGCIN algorithm, enabling the system to capture more complex and dynamic user-item relationships.

[0301] 2. User satisfaction: The high score of 4.5 reflects the high recognition of users for the recommendation results. This is not only related to the high accuracy but also related to the fact that the HSASPN algorithm can generate more personalized and diverse recommendations.

[0302] 3. Average processing time: Although the processing time of the present invention (150 ms) is higher than that of simple collaborative filtering methods, considering the complexity of the system and the performance improvement, this processing time is acceptable and still meets the requirements of real-time recommendations.

[0303] 4. Coverage rate of long-tail items: The present invention performs outstandingly in this indicator (35%), which proves that the system can effectively tap the potential of long-tail items. This is mainly due to the effective fusion of multi-modal data by the QIMMFN algorithm, enabling the system to more comprehensively understand item characteristics.

[0304] 5. Completeness of the knowledge graph: The high completeness of 95% reflects the advantages of the present invention in constructing and maintaining the knowledge graph. This is mainly attributed to the combination of the GERNRN algorithm and the real-time update mechanism.

[0305] These test results fully illustrate the following superiority of the present invention:

[0306] 1. High-precision recommendation: By fusing multi-source data and adopting advanced graph embedding technology, this system can provide more accurate personalized recommendations.

[0307] 2. Optimization of user experience: The high user satisfaction indicates that the system not only provides accurate recommendations but also performs well in the diversity and timeliness of recommendations.

[0308] 3. Long-tail discovery ability: The system significantly improves the exposure rate of long-tail items, which is of great significance for enhancing the overall item utilization rate of the platform and the diversity of user choices.

[0309] 4. Knowledge accumulation and utilization: The highly complete knowledge graph enables the system to perform more in-depth reasoning and prediction, providing a solid foundation for precision marketing.

[0310] 5. Real-time performance and adaptability: Despite the high complexity of the system, it can still complete processing within an acceptable time and adapt to the changing user needs and market environment through continuous learning.

[0311] Generally speaking, these results verify the innovations of the present invention in aspects such as knowledge graph integration, multimodal data processing, and dynamic time series modeling, as well as the significant performance improvement brought by these innovations in practical applications. This system not only improves the accuracy of recommendations and user satisfaction, but also performs excellently in the utilization of long-tail products and knowledge accumulation, providing strong technical support for the precision marketing and user services of e-commerce platforms.

[0312] It should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A digital precision sales agent management method integrating knowledge graph, characterized in that: The following steps are involved: Acquire and preprocess multi-source heterogeneous data; Based on the preprocessed data, multimodal data fusion and entity recognition are realized; relationship extraction and attribute mapping are performed based on the fused data and the recognized entities; dynamic time sequence graph embedding is constructed based on the extracted relationship and the mapped attributes; knowledge reasoning and graph completion are performed using the dynamic time sequence graph embedding; Real-time update and incremental learning of the completed knowledge graph; Generate accurate traffic diversion strategies based on the updated knowledge graph; Optimize the sales process based on the precise traffic diversion strategy; Evaluate and provide feedback on the optimized sales process, and use the feedback results to optimize each step of the method; In the step of acquiring and preprocessing multi-source heterogeneous data, an adaptive multi-dimensional anomaly detection algorithm is used, and the algorithm includes the following steps: a) Use locality sensitive hashing to map high-dimensional data to low-dimensional space, where the mapping function is: Where x is the input data, a is a random vector, b is a random offset, and w is the bucket width; b) Apply dynamic clustering within each hash bucket using a modified DBSCAN algorithm where the ∈ parameter is adaptively adjusted: ∈ i = μ i + α·σ i where, ∈ i is the neighborhood radius of the i-th dimension, μ i and σ i are the mean and standard deviation of the i-th dimension respectively, and α is an adjustable parameter; c) Determine the anomaly score: where AS(x) is the anomaly score of point x, d(x, C) is the distance from point x to the nearest cluster center C, and σ C is the standard deviation of cluster C; d) Dynamically determine the abnormal threshold based on extreme value theory; In the step of realizing multimodal data fusion and entity recognition, a quantum-inspired multimodal fusion network is used, which includes the following steps: a) Quantum state encoding is performed on the data x of each modality i : where, |ψ i > is a quantum state, α ij is a complex amplitude, |j> is a ground state, satisfying ∑ j |α ij | 2 = 1; b) Apply the inter-modal entanglement layer to design the entanglement operator U ent :U ent =exp(-iH ent ), Among them, σ x , σ y , σ z are Pauli matrices, and J ij is the learnable entanglement strength; c) Design a set of learnable measurement operators M using the quantum measurement layer k , and obtain a classical output: y k = <ψ|M k |ψ> where y k is the measurement result and |ψ> is the quantum state; d) Process the measurements using traditional neural network layers to obtain the final fused representation.

2. The method according to claim 1, wherein In the step of performing relationship extraction and attribute mapping, a graph-enhanced recursive neural reasoning network is used, which includes the following steps: a) Use the TransE algorithm to embed the existing knowledge graph: ||h+rt||2≈0 Among them, h, r, t are the embedding vectors of the head entity, relation, and tail entity respectively; b) Use BiLSTM to encode the input sentence: h t = BiLSTM(x t , h t-1 ) where h t is the hidden state at time t, and x t is the input at time t; c) Determine the attention score of sentence representation and graph embedding: α i = softmax(v T tanh(W[h i ; g j )) Among them, α i is the attention score, h i is the sentence hidden state, g j is the graph embedding, and v and W are learnable parameters; d) Design a recursive structure to simulate multi-hop reasoning: z t = f(W z [h t ; z t-1 ; c t ) Among them, z t is the inference state at time t, h t is the sentence hidden state, c t is the context vector enhanced by the knowledge graph, W z is a learnable parameter; e) Use the softmax layer to perform relation classification on the final z T ​ 3. The method according to claim 1, wherein The step of constructing a dynamic time-series graph embedding includes using a quantum-inspired space-time graph embedding algorithm, which includes the following steps: a) Initialize entity e i and relationship r j for quantum bitmap embedding: Q(e i ) = |ψ i > = α i |0> + β i |1> Q(r j ) = |φ j > = γ j |0> + δ j |1> where |α i | 2 + |β i | 2 = 1, |γ j | 2 + |δ j | 2 = 1; b) Define the non - linear space - time evolution operator U t : U t = exp(-iH t ) where, H t is a time-varying Hamiltonian operator: σ k is the Pauli matrix, w k (t) and J ij (t) are time-varying coefficients; c) Introducing the quantum walk attention mechanism: A QW (i,j,t) = <ψ i (t)|U QW (t)|ψ j (t)> Among them, U QW (t) is the quantum walk operator based on the graph structure; d) Multiscale time-dependent modeling using fractional quantum field theory: Φ(x,t) = ∫K α (x - y,t - s)ψ(y,s)dy ds where K α is the fractional diffusion kernel, and α ∈ (0, 2) is the fractional order parameter; e) Introduce the topological entropy regularization term: where p ij is the transition probability from node i to node j; f) Design non-equilibrium thermodynamic optimization objectives: Among them, L task is the main task loss, is the change rate of the system entropy; g) Design of adaptive quantum measurement strategies: Among them, M is a learnable measurement operator.

4. The method according to claim 1, wherein In the steps of performing knowledge reasoning and knowledge graph completion, a multi-granularity causal reasoning network is used, including the following steps: a) Using a spectral clustering algorithm to divide the knowledge graph into multiple granularity levels; b) At each granularity level, using the PC algorithm to discover potential causal relationships; c) Using a structural equation model to estimate the causal strength: X j = f j (PA j , U j ) Among them, X j is the dependent variable, PA j is the set of parent nodes of X j and U j is the exogenous variable; d) Design an attention mechanism to integrate causal information of different granularities: α g = softmax(v T tanh(W g h g ))h final = ∑ g α g h g Among them, α g is the attention weight, h g is the causal information representation of different granularities, and v and W g are learnable parameters; e) Based on the final representation h final Perform link prediction.

5. The method according to claim 1, characterized in that, In the step of generating a precise traffic diversion strategy, a hierarchical spatiotemporal attention strategy network is used, including the following steps: a) Define state representation: s t = [u t ; p t ; c t ​ Among them, u t is the user feature, p t is the product feature, c t is the context feature; b) Define a two-level action space: A l is the high-level policy), and A2 is the specific operation; c) Determine spatiotemporal attention: α t = softmax(v T tanh(W s s t + W h h t-1 )) where h t-1 is the historical hidden state, and v, W s , W h are learnable parameters; d) Build a policy network: π θ (a|s) = softmax(f θ (s, α)) where f θ is a parameterized neural network, e) Build a value network: V φ (s) = g φ (s, α) where, g φ is a parameterized neural network; f) Training using an asynchronous advantage actor-critic algorithm.

6. The method according to claim 1, wherein The step of optimizing the agency sales process includes using graph neural networks for agency matching and anomaly detection.

7. The method according to claim 1, wherein In the step of effect evaluation and feedback, dynamic weight multi-objective Bayesian optimization is used, including the following steps: a) Define multiple objective functions: F(x) = [f1(x), f2(x),..., f m (x) where f i (x) is the i-th objective function, b) Establish a Gaussian process model for each objective function: f i (x) ~ GP(μ i (x), k i (x, x′)) where, μ i (x) is the mean function, k i (x, x′) is the kernel function; c) Determine dynamic weights: where λ is the temperature parameter, is the current optimal value for target i, and σ i is the standard deviation; d) Calculate the acquisition function: Among them, EI i (x) is the expected improvement of the i-th target; e) Maximize EI(x) using the DIRECT algorithm and select the next evaluation point; f) Evaluate the new point, update the Gaussian process model and the dynamic weights, and repeat steps c - e.

8. A digital precision traffic diversion and sales management system integrating knowledge graph, used to execute the method described in any one of claims 1 to 7, characterized in that: Including: A data preprocessing module for acquiring and preprocessing multi-source heterogeneous data; A data fusion module for implementing multi-modal data fusion and entity recognition; A relation extraction module for performing relation extraction and attribute mapping; A graph embedding module for constructing dynamic temporal graph embeddings; A knowledge reasoning module for performing knowledge reasoning and graph completion; An update learning module for performing real-time updates and incremental learning; A strategy generation module for generating precise drainage strategies; A process optimization module for optimizing the consignment process; An evaluation feedback module for performing effect evaluation and feedback; wherein, the output of the evaluation feedback module is connected to the input of the data preprocessing module to form a closed-loop optimization.

Images