An e-commerce supply chain anti-out-of-stock decision system based on multi-source data fusion

Abstract

The present invention relates to the field of logistics supply technology, and discloses an e-commerce supply chain anti-out-of-stock decision system based on multi-source data fusion, comprising: a heterogeneous graph network construction and risk prediction module: constructing a heterogeneous graph network representing the e-commerce supply chain relationship based on multi-source data, and using a time-sensitive graph attention network to predict the out-of-stock risk propagation path; an elastic recovery path generation module: based on the constructed heterogeneous graph network, modeling the supply chain recovery as a sequential decision problem on the graph, and generating multiple elastic recovery paths through a variational inference method; by constructing a heterogeneous graph network and implementing a time-sensitive graph attention network, the present invention enables the system to accurately capture the cascade effect and risk propagation path in a complex supply network, thereby improving the accuracy of predicting the out-of-stock risk propagation path and significantly reducing false positives and negatives.

Description

Technical Field

[0001] The present invention relates to the field of logistics supply technology, and more specifically, to an e-commerce supply chain anti-out-of-stock decision system based on multi-source data fusion. Background Art

[0002] Out-of-stock problems in e-commerce supply chains seriously affect user experience and platform revenue. Existing technologies have limitations in solving such problems: existing out-of-stock warning systems mainly rely on a single data source (such as historical sales data or inventory data) for simple statistical analysis and prediction, which cannot fully capture the complex information of each link in the supply chain, resulting in limited prediction accuracy. Existing methods ignore the mutual influence and implicit correlation between multi-source data and cannot effectively cope with the complex and changing e-commerce environment; and traditional supply chain risk analysis methods usually use linear models or simple network analysis, which cannot effectively capture the cascade effects and risk propagation paths in complex supply networks. When a problem occurs in a certain link in the supply chain, it is difficult to accurately predict the scope and extent of its impact on the entire supply network; at the same time, existing supply chain recovery methods are mostly based on deterministic models and lack sufficient modeling of uncertainty, resulting in insufficient robustness of recovery strategies in highly uncertain environments. Traditional methods cannot effectively consider multiple possible scenarios and risk factors and have difficulty adapting to the dynamic changes of e-commerce supply chains. Summary of the Invention

[0003] The present invention provides an e-commerce supply chain anti-out-of-stock decision-making system based on multi-source data fusion, which solves the technical problem of e-commerce supply chain anti-out-of-stock in related technologies.

[0004] The present invention provides an e-commerce supply chain anti-out-of-stock decision-making system based on multi-source data fusion, comprising:

[0005] Heterogeneous graph network construction and risk prediction module: This module builds a heterogeneous graph network representing e-commerce supply chain relationships based on multi-source data, and uses a time-sensitive graph attention network to predict the propagation path of out-of-stock risks.

[0006] Resilient recovery path generation module: Based on the constructed heterogeneous graph network, supply chain recovery is modeled as a sequential decision problem on the graph, and multiple resilient recovery paths are generated through variational inference methods;

[0007] Strategy Optimization and Uncertainty Modeling Module: This module builds a deep probabilistic programming model for multiple recovery paths to model their uncertainties, and uses a multi-agent Monte Carlo tree search algorithm to optimize the recovery strategy.

[0008] Data fusion and policy execution module: Based on the optimized recovery strategy, it realizes the real-time dynamic fusion of multi-source data, adaptively adjusts the weight of data sources, and converts abstract strategies into specific and executable supply chain adjustment instructions;

[0009] Closed-loop monitoring and optimization module: Build a closed-loop feedback algorithm for the executed adjustment instructions, continuously monitor the execution effect and optimize the strategy.

[0010] As a further optimization solution of the present invention, the step of constructing a heterogeneous graph network includes:

[0011] Collect multi-source data in the e-commerce supply chain, including sales data, inventory data, supplier data, logistics data, and market trend data;

[0012] Preprocess the collected multi-source data, including data cleaning, missing value processing, outlier detection and data standardization;

[0013] Construct a heterogeneous graph network representing the e-commerce supply chain relationship, where the nodes include product nodes, warehouse nodes, supplier nodes and logistics nodes, and the edges include supply relationships, inventory relationships and logistics relationships.

[0014] As a further optimization scheme of the present invention, the time-sensitive graph attention network measures the association strength between different nodes by calculating the attention coefficient, which takes into account the temporal characteristics of the edges, including time coding, periodic pattern coding and temporal relationship coding, to capture the characteristics of supply chain relationships changing over time.

[0015] As a further optimization solution of the present invention, the variational inference method generates an elastic recovery path through the following steps:

[0016] Define the posterior distribution of the recovery path and introduce the variational distribution as an approximate alternative to the posterior distribution;

[0017] Optimize parameters by maximizing the variational lower bound;

[0018] Integrate graph network structures into variational inference models and use graph neural networks to process the structural information of supply chain networks;

[0019] Based on Bayesian decision theory, a multi-step reasoning algorithm is implemented to evaluate the costs and effects of different recovery paths.

[0020] As a further optimization solution of the present invention, a set of random variables representing the supply chain state and conditional dependencies between the variables are defined;

[0021] Parameterize conditional probability distributions using neural networks;

[0022] Apply the stochastic variational inference algorithm to estimate the posterior distribution.

[0023] As a further optimization solution of the present invention, the multi-agent Monte Carlo tree search algorithm includes:

[0024] The supply chain recovery problem is modeled as a multi-agent system, where each agent represents a decision-making entity in the supply chain;

[0025] Define system states and joint action spaces;

[0026] Execute the Monte Carlo tree search process, including four steps: selection, expansion, simulation, and backpropagation;

[0027] Implement a collaborative decision-making method among intelligent agents to ensure that the decisions of each intelligent agent form a consistent global recovery strategy.

[0028] As a further optimization scheme of the present invention, a risk-aware strategy optimization objective function is introduced, which includes a reward term and a risk aversion term. The reward term measures the expected return of the strategy, and the risk aversion term controls the degree of risk by calculating the difference between the current strategy and the reference strategy.

[0029] As a further optimization solution of the present invention, the step of real-time dynamic fusion of multi-source data includes:

[0030] Build a unified data access layer to support access to various data sources;

[0031] Conduct real-time quality assessment of incoming multi-source data, including completeness, timeliness, consistency, and accuracy;

[0032] Adaptively adjust data source weights based on out-of-stock risk scenarios and data quality;

[0033] Based on the assigned weights, feature-level fusion, decision-level fusion, and model-level fusion are achieved.

[0034] As a further optimization solution of the present invention, the step of converting the abstract strategy into a specific executable supply chain adjustment instruction includes:

[0035] Parse the optimization recovery strategy into a specific set of execution instructions, including procurement instructions, inventory adjustment instructions, logistics instructions, and pricing instructions;

[0036] Generate a detailed execution plan based on the parsed instruction set, including task decomposition, task sequencing, resource allocation, and time planning;

[0037] Perform constraint checks on the generated execution plan to ensure its feasibility;

[0038] Build standardized interfaces with various supply chain systems to achieve automated execution of instructions.

[0039] A storage medium includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the above-mentioned e-commerce supply chain anti-out-of-stock decision system based on multi-source data fusion.

[0040] The beneficial effects of the present invention are as follows: by constructing a heterogeneous graph network and realizing a time-sensitive graph attention network, the present invention enables the system to accurately capture the cascade effects and risk propagation paths in complex supply networks, thereby improving the accuracy of out-of-stock risk propagation path prediction and significantly reducing false positives and negatives. At the same time, based on graph structure variational inference and deep probabilistic programming methods, the system can generate robust recovery strategies in highly uncertain environments, thereby improving the robustness of the system's recovery strategies in highly uncertain environments and enabling it to cope with a variety of complex scenarios and disturbances. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a module diagram of an e-commerce supply chain anti-out-of-stock decision system based on multi-source data fusion according to the present invention. DETAILED DESCRIPTION

[0042] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0043] At least one embodiment of the present invention discloses an e-commerce supply chain anti-out-of-stock decision system based on multi-source data fusion, such as Figure 1 Shown, including:

[0044] Heterogeneous graph network construction and risk prediction module: This module builds a heterogeneous graph network representing e-commerce supply chain relationships based on multi-source data, and uses a time-sensitive graph attention network to predict the propagation path of out-of-stock risks.

[0045] The specific steps include:

[0046] Step 1.1: Multi-source data collection and preprocessing;

[0047] Specifically, we collect multi-source data in the e-commerce supply chain, including but not limited to:

[0048] Product sales data: including product ID, sales volume, sales time, price and other information;

[0049] Inventory data: including product ID, inventory quantity, inventory location, minimum safety stock, etc.

[0050] Supplier data: including supplier ID, supply capacity, historical supply records, supply cycle and other information;

[0051] Logistics data: including logistics routes, transportation time, logistics costs and other information;

[0052] Market trend data: including seasonal factors, market competition, consumer behavior and other information.

[0053] Furthermore, the collected multi-source data are preprocessed, including data cleaning, missing value processing, outlier detection, data standardization, etc., to form a standardized data set in preparation for subsequent analysis.

[0054] Step 1.2: Heterogeneous graph network construction;

[0055] Specifically, based on the preprocessed multi-source data, a heterogeneous graph network representing the e-commerce supply chain relationships is constructed:

[0056]

[0057] Specifically, Represents the constructed heterogeneous graph network;

[0058] Represents a collection of nodes;

[0059] in, Represents the set of nodes in the entire graph network; 、 、 Respectively represent 、 、 nodes; is the total number of nodes;

[0060] represents an edge set;

[0061] in, Represents the set of edges in the entire graph network; Represents a slave node To Node An edge of Indicates the starting node of the edge; Indicates the end node of the edge; Indicates the relationship type of the edge; Represents a collection of nodes; Represents a set of relationship types;

[0062] Represents a set of node attributes;

[0063] in, Represents the set of node attributes in the entire graph network; Representation node The attribute set of Represents a node; Represents a collection of nodes;

[0064] Represents a set of relationship types;

[0065] in, Represents the set of relationship types in the entire graph network; 、 、 Respectively represent 、 、 Types of relationships, is the total number of relationship types;

[0066] Furthermore, various types of data are mapped into the graph network: commodity data is mapped to commodity nodes and their attributes; warehouse data is mapped to warehouse nodes and their attributes; supplier data is mapped to supplier nodes and their attributes; logistics data is mapped to logistics nodes and their attributes as well as logistics relationship edges; supply chain relationship data is mapped to different types of edges and their attributes.

[0067] Step 1.3: Implementing a time-sensitive graph attention network.

[0068] Specifically, we build a time-sensitive graph attention network to learn important relationships between nodes. The implementation is as follows:

[0069] First, define the calculation formula of the graph attention layer:

[0070]

[0071] in, Indicates the Nodes in the layer Feature representation; represents a nonlinear activation function; Indicates the node All neighbor nodes of To sum, the node Perform weighted aggregation on the features of all neighboring nodes; Indicates the index of the neighbor node; Representation node The set of neighbor nodes of Representation node For Node The attention coefficient; Indicates the The weight matrix of the layer; represents the feature representation of node j in layer l.

[0072] Attention coefficient The calculation method is:

[0073]

[0074] in, Representation node For Node The attention coefficient; represents the exponential function; represents the leaky linear rectification activation function; represents the attention vector; Represents vector transpose; Indicates the The weight matrix of the layer; Representation node Feature representation; Representation node Feature representation; Representation node Feature representation; Represents vector concatenation operation; Represents an edge The temporal characteristics of Represents an edge The temporal characteristics of Indicates the node All neighbor nodes of Perform summation; Indicates the index of the neighbor node; Representation node The set of neighbor nodes of

[0075] It should be noted that by introducing the time series feature ,This application can effectively capture the dynamic changes of supply chain ,relationships over time, which is crucial for accurately predicting the ,propagation path of out-of-stock risks.

[0076] According to one embodiment of the present application, a time-sensitive graph attention network specifically includes an input layer, multiple graph attention layers, and an output layer. The input layer receives a node feature matrix and an adjacency matrix. The node feature matrix contains attribute information for various types of nodes (such as products, warehouses, and suppliers), and the adjacency matrix contains the connection relationships between nodes. The graph attention layer is responsible for calculating the attention coefficients between nodes and updating the node representations. The output layer generates the final node representation for subsequent risk prediction tasks.

[0077] In the specific implementation, according to the characteristics of e-commerce supply chain scenarios, time series characteristics Constructed by:

[0078] Time encoding: For the time information of each edge (such as the establishment time of the supply relationship, the last interaction time, etc.), sine-cosine position encoding is applied:

[0079]

[0080] in, represents the position encoding function; Indicates a point in time; Represents dimension index; represents the sine function; represents the encoding dimension; Indicates even-numbered dimension positions; Indicates odd-numbered dimension positions; represents the cosine function; Represents the scaling factor.

[0081] Periodic pattern encoding: Capturing periodic patterns in the supply chain (e.g., weekly, monthly, seasonal) through Fourier feature transform:

[0082]

[0083] in, represents the Fourier transform function; Indicates a point in time; represents the sine function; represents the cosine function; represents pi; Indicates the Length of cycle; Indicates the Length of cycle;

[0084] Temporal relationship encoding: Captures the temporal relationship of interactions between nodes, such as the time since the last interaction, the frequency of interaction, etc.

[0085] The time series features constructed in this way can effectively capture the temporal dynamics in the supply chain and improve the accuracy of out-of-stock risk propagation prediction.

[0086] In the specific application scenario of preventing out-of-stock situations in the e-commerce supply chain, the time-sensitive graph attention network can be used in the following typical situations:

[0087] Seasonal product out-of-stock risk prediction: For products with strong seasonal characteristics (such as holiday items and seasonal clothing), we use time series features to capture historical seasonal patterns and identify potential out-of-stock risks in advance. For example, before the Spring Festival, we identified that suppliers of certain New Year products may be unable to meet the surge in demand, causing the out-of-stock risk to spread to related products.

[0088] Promotional Event Impact Analysis: Before large-scale promotional events on e-commerce platforms (such as Singles' Day), analyze time series patterns in historical promotional data to predict which products may experience inventory shortages due to promotional events and identify how such shortages propagate along the supply chain.

[0089] Supplier stability assessment: By analyzing the timing characteristics of suppliers' historical deliveries (such as whether there are periodic delays, quality fluctuations, etc.), we can evaluate the stability of suppliers and predict the possible risk of supply disruptions and their impact on downstream products.

[0090] Step 1.4: Predict the out-of-stock risk transmission path;

[0091] Specifically, based on the constructed time-sensitive graph attention network, the out-of-stock risk propagation path is predicted:

[0092] Initial risk identification: Identify initial out-of-stock risk points by analyzing factors such as inventory levels, sales trends, and supplier status.

[0093] Propagation path learning: Using self-supervised learning methods, based on historical out-of-stock event data, we learn the propagation pattern of out-of-stock risk. The specific implementation is as follows:

[0094] Construct training samples:

[0095]

[0096] in, Indicates time The graph structure of Indicates time The risk distribution, Indicates time risk distribution.

[0097] Training objective: Minimize the difference between predicted risk and actual risk:

[0098]

[0099] in, represents the loss function; represents the square of the two norm; Indicates that the parameter is The prediction function of Indicates time The graph structure of Indicates time risk distribution; Indicates time risk distribution; Indicates a time interval; Represents model parameters.

[0100] Use historical data to train model parameters .

[0101] Identification of high-risk propagation substructures: Apply subgraph mining algorithms to identify substructures in the graph network that are most likely to cause cascading out-of-stock risks. Specific implementation is as follows:

[0102] Define the risk propagation subgraph scoring function:

[0103]

[0104] in, represents the subgraph scoring function; Represents a subgraph; Representing a subgraph The sum of the risk probabilities of all nodes in ; Indicates the subgraph nodes; Representation node the risk probability; represents the equilibrium parameter; Representing a subgraph The sum of the risk propagation probabilities of all edges in ; Represents an edge the probability of risk transmission; Represents a slave node To Node An edge represents the supply chain relationship between two nodes;

[0105] Use a heuristic search algorithm to find the set of subgraphs with the highest scores.

[0106] Risk propagation path generation: Based on the identified high-risk substructures and learned propagation patterns, the most likely out-of-stock risk propagation paths are generated and the risk probability of each path is calculated.

[0107] The output includes:

[0108] Initial list of out-of-stock risk points and their risk levels;

[0109] Risk propagation path diagram, including propagation direction, propagation probability and time estimate;

[0110] Identification results of key risk nodes (nodes that may cause large-scale cascading risks).

[0111] It should be noted that through heterogeneous graph networks and time-sensitive graph attention networks, the cascade effects and risk propagation paths in complex supply chains can be effectively captured, which has significant advantages over traditional methods.

[0112] Resilient recovery path generation module: Based on the constructed heterogeneous graph network, supply chain recovery is modeled as a sequential decision problem on the graph, and multiple resilient recovery paths are generated through variational inference methods;

[0113] The specific steps include:

[0114] Step 2.1, recovery path modeling;

[0115] Supply chain recovery is modeled as a sequential decision problem on a graph as follows:

[0116] Define the state space : Represents the state set of each node in the supply chain network, including inventory level, production capacity, logistics status, etc.

[0117] Defining the action space : Represents a set of possible recovery operations, such as adding orders, adjusting inventory, changing suppliers, adjusting logistics routes, etc.

[0118] Define the transfer function :Indicates the current state Next execution action After transfer to state probability;

[0119] Define the reward function :Indicates that the state Next execution action The benefits are determined by comprehensively considering factors such as recovery efficiency, cost, and time.

[0120] Step 2.2, variational inference model construction;

[0121] Construct a path generation model based on variational inference to generate robust recovery paths in high uncertainty environments; the specific implementation is as follows:

[0122] Posterior distribution modeling: defining the posterior distribution of the recovery path Directly computing the posterior distribution is usually intractable, so variational methods are used to approximate it.

[0123] Variational lower bound construction: introducing variational distribution As the posterior distribution The parameters are optimized by maximizing the variational lower bound (ELBO):

[0124]

[0125] in, represents the variational lower bound loss function; represents the parameters of the generative model; represents the parameters of the inferred model; represents the expectation operation; Indicates that the parameter is The variational distribution of represents potential recovery path variables; represents the observed supply chain status; represents the logarithmic function; Indicates that the parameter is Generative model distribution of ; represents KL divergence; represents the prior distribution;

[0126] Specifically:

[0127] Represents the reconstruction term, encouraging the model to be based on latent variables Reconstructing observation data ;

[0128] Represents the KL divergence term, ensuring the variational distribution Will not deviate from the prior distribution Too far.

[0129] Graph structure integration: Integrate the graph network structure into the variational inference model and use graph neural networks to process the structural information of the supply chain network:

[0130] Encoder network:

[0131]

[0132] in: Indicates that the parameter is The variational distribution of represents potential recovery path variables; represents the observed supply chain status; represents a normal distribution; represents the mean function; represents the variance function; Represents observation data The corresponding graph structure.

[0133] The decoder network transforms the latent variables Mapping back to the supply chain state space, its expression is:

[0134]

[0135] in, Indicates that the parameter is Generative model distribution of ; represents the observed supply chain status; represents potential recovery path variables; Indicates that the parameter is The decoder function of A graph structure representing a supply chain network.

[0136] It should be noted that by integrating graph structure information into the variational inference process, this application can more accurately capture the complex dependencies in the supply chain network, thereby generating a recovery path that is more in line with the actual situation.

[0137] According to one embodiment of the present application, the specific implementation of the graph structure variational inference model includes the following network structure:

[0138] Graph Structure Encoder: Using Graph Convolutional Network ( Graph Convolutional Network , GCN) or Graph Attention Network (GAT) processes graph structure information and generates node representation and graph representation .

[0139] Node embedding layer: maps node features to latent space;

[0140] Multi-layer graph convolution: aggregate neighbor node information and update node representation;

[0141] Graph pooling layer: generates a global representation of the entire graph;

[0142] Variational Encoders: Generating Latent Variables Based on Graph Representations The distribution parameters of

[0143] Mean Network:

[0144]

[0145] in, Indicates that the parameter is The mean function of Represents input data; Represents a multilayer perceptron network used to calculate the mean; A global representation of the representation graph;

[0146] Variance Network:

[0147]

[0148] in, represents the logarithmic function; Indicates that the parameter is The variance function of Represents input data; represents a multilayer perceptron network used to calculate variance; Represents a global representation of the graph.

[0149] Sampling operation:

[0150]

[0151] in, represents the latent variable obtained by sampling; Indicates that the parameter is The mean function of the input The mapping result of Indicates that the parameter is The standard deviation function of the input The mapping result of Represents the element-wise multiplication operator; represents random noise sampled from a standard normal distribution; represents a standard normal distribution with a mean of 0 and a covariance matrix equal to the identity matrix;

[0152] Decoder network: transforms latent variables Decoded as recovery path;

[0153] State decoding layer: Mapping to the initial state representation;

[0154] Sequence generation layer: uses a recurrent neural network (RNN) or a Transformer decoder to generate a recovery path sequence;

[0155] Output layer: generates specific recovery operations for each time step;

[0156] In the e-commerce supply chain out-of-stock prevention scenario, application examples of graph structure variational inference models include:

[0157] Seasonal Product Out-of-Stock Recovery: For products significantly impacted by seasonality (such as holiday items and seasonal clothing), when out-of-stock risks are detected, the system first represents the current supply chain status through a heterogeneous graph network, including inventory levels, supplier capacity, and logistics status. A graph-structured variational inference model then generates multiple possible recovery paths, factoring in seasonal uncertainty. Ultimately, a robust recovery strategy is generated, such as pre-emptively adjusting safety stock, adding alternative suppliers, and optimizing logistics routes.

[0158] Multi-regional Collaborative Recovery: When a warehouse in a region faces the risk of out-of-stock, the system uses a graph-structured variational inference model to generate a coordinated recovery path involving warehouses in multiple regions. The model considers factors such as regional inventory levels, logistics costs, and time constraints to generate an optimal cross-regional inventory allocation plan without impacting supply in other regions.

[0159] Response to supplier disruptions: When a major supplier experiences a production disruption, the system uses a graph-structured variational inference model to generate a recovery strategy based on the supply chain network structure and historical data, including enabling alternative suppliers, adjusting logistics routes, and reallocating production orders, to minimize the impact of the disruption on commodity supply.

[0160] Step 2.3: Implement the multi-step reasoning algorithm;

[0161] Based on Bayesian decision theory, a multi-step reasoning algorithm is implemented to evaluate the costs and effects of different recovery paths:

[0162] Path sampling: from variational distributions Sampling multiple possible recovery paths ;

[0163] in, Indicates that the parameter is The variational distribution function of represents potential recovery path variables; represents the observed supply chain status; 、 、 Respectively represent 、 、 The recovery path of the samples; Indicates the total number of sampling paths;

[0164] Path evaluation: Evaluate each sample path and calculate its expected utility:

[0165]

[0166] in, represents the expected utility function; Indicates the sampling recovery paths; Represents the expectation operator; Indicates the current state Perform joint actions Then transfer to the next state probability; Indicates the current state Perform joint actions Instant rewards received; represents the discount factor; Indicates the next state The value function of .

[0167] Bayesian risk analysis: Considering the decision risk, calculate the Bayesian risk of each path:

[0168]

[0169] in: represents the Bayesian risk function; Indicates the sampling recovery paths; Represents the expectation operator; Indicates the current state Perform joint actions Then transfer to the next state The probability distribution of represents the loss function; Indicates the next state of the system; Indicates the current state of the system; Indicates a joint action Indicates that the status Next execution path The loss function of .

[0170] Multi-objective trade-off: Comprehensively consider multiple objectives such as recovery cost, time, and risk, and calculate a comprehensive score:

[0171]

[0172] in: represents the comprehensive scoring function; Indicates the sampling recovery paths; 、 、 、 Respectively represent the weight coefficients of utility term, risk term, cost term, and time term; represents the expected utility function of the path; represents the Bayesian risk function of the path; represents the cost function of the path; represents the recovery time function of the path;

[0173] In addition, the multi-step reasoning algorithm, through the Bayesian decision theory framework, can make more robust decisions while considering uncertainty, which is particularly important for out-of-stock recovery in e-commerce supply chains.

[0174] Step 2.4, generating the optimal elastic recovery path;

[0175] Based on the multi-step reasoning results, the optimal elastic recovery path is generated:

[0176] Path sorting: sort all sampled paths according to the comprehensive score and select the K paths with the highest scores as candidates;

[0177] Robustness analysis: Perform robustness analysis on candidate paths to test their performance under different disturbance conditions:

[0178] Supplier delay disturbance: simulates supplier delivery delay;

[0179] Demand fluctuation disturbance: simulates sudden changes in market demand;

[0180] Logistics disruption: simulates the situation of logistics channel disruption;

[0181] Path optimization: Based on the results of robustness analysis, fine-tune and optimize candidate paths to enhance their ability to cope with uncertainty;

[0182] Optimal path determination: Comprehensively consider the score and robustness to determine the final optimal elastic recovery path;

[0183] It should be understood that the core innovation of this step lies in modeling the supply chain recovery problem as a sequential decision problem on a graph and generating a highly robust recovery path through variational inference. Compared to traditional methods, the method provided in this application is more capable of addressing out-of-stock recovery in highly uncertain environments.

[0184] The output includes:

[0185] The optimal recovery path, including specific steps, execution time, and expected results;

[0186] Risk assessment and response recommendations for path execution;

[0187] resource requirements and cost estimates for pathway implementation;

[0188] Strategy Optimization and Uncertainty Modeling Module: This module builds a deep probabilistic programming model for multiple recovery paths to model their uncertainties, and uses a multi-agent Monte Carlo tree search algorithm to optimize the recovery strategy.

[0189] This module uses deep probabilistic programming and multi-agent Monte Carlo tree search methods to further optimize the generated recovery strategy and improve its robustness in highly uncertain environments. Specifically, it includes the following steps:

[0190] Step 3.1: Deep probabilistic programming model construction;

[0191] Build a deep probabilistic programming model to accurately model the uncertainty of the recovery strategy. The specific implementation is as follows:

[0192] Probabilistic Program Definition: Constructing a probabilistic program that represents a recovery strategy , which contains the following components:

[0193] A collection of random variables: , represents the supply chain state variable;

[0194] in, is a set of random variables; 、 、 Respectively 、 、 supply chain state random variables; is the total number of random variables;

[0195] Conditional dependencies , represents the probability dependence graph between variables;

[0196] Conditional probability distribution: ;

[0197] in, represents the conditional probability distribution function; Indicates the supply chain state random variables; Representing variables The parent node set of

[0198] Neural Network Parameterization: Parameterizing conditional probability distributions using neural networks:

[0199]

[0200] in, represents the conditional probability distribution function; No. supply chain state random variables; Representing variables The parent node set of Indicates that the parameter is Neural network function; Represents the parameters of the neural network;

[0201] Stochastic Variational Inference: Apply the stochastic variational inference algorithm to estimate the posterior distribution:

[0202]

[0203] in, represents the variational objective function; represents the variational parameter; Represents the expectation operator; Indicates that the parameter is The variational distribution of Represents a path variable; represents observation data; represents the logarithmic function; represents a probability distribution; Represents KL divergence, which is used to measure the difference between two probability distributions;

[0204] Optimize variational parameters via stochastic gradient ascent :

[0205]

[0206] in, Indicates the The variational parameters of the iteration; Indicates the The variational parameters of the iteration; Indicates the The learning rate of the iteration; Indicates the gradient of the variational parameter λ; shows the variational objective function;

[0207] The gradient estimation uses the reparameterization technique:

[0208]

[0209] in, Indicates the parameters Find the gradient; Represents the expectation operator; Indicates that the parameter is The variational distribution of Represents a path variable; represents observation data; In variational inference, it represents the evaluation function to be optimized, which is used to calculate the performance of the model under given path variables; Indicates the number of samples; Indicates from arrive The summation operation is about to The results of the samples are summed up; Indicates the sampling index; Indicates that the parameter is The generating function of Indicates the random noise samples; Indicates the random noise samples Obey the noise distribution ; Indicates the generated path samples;

[0210] It is worth noting that the deep probabilistic programming model, by combining the advantages of deep learning and probabilistic programming, can more accurately model the uncertainty in the supply chain recovery process, thereby generating more robust recovery strategies.

[0211] According to one embodiment of the present application, the specific structure of the deep probabilistic programming model includes:

[0212] Probability variable definition layer: defines the key random variables of the supply chain status, including:

[0213] Inventory level variable: represents the inventory status of each node;

[0214] Supply capacity variable: represents the supplier's production capacity;

[0215] Logistics state variable: represents the state of the logistics channel;

[0216] Demand variable: represents the market demand for goods;

[0217] Risk variable: represents the risk status of each link.

[0218] Probabilistic dependency network: defines the conditional dependency relationship between variables and constructs it into a directed graph:

[0219] Direct dependence: For example, inventory levels depend on upstream supply and downstream demand;

[0220] Indirect dependence: such as substitution effect between different commodities;

[0221] Temporal dependency: such as the impact of the current state on the future state.

[0222] Neural Network Parameterization Module: Use different types of neural networks to parameterize different types of conditional probability distributions:

[0223] Continuous variables: using Gaussian mixture networks;

[0224] Discrete variables: use categorical distribution networks;

[0225] Time series variants: using recurrent neural networks.

[0226] In the specific application of e-commerce supply chain out-of-stock decision-making, deep probabilistic programming models can be used in the following scenarios:

[0227] Demand uncertainty modeling: Product demand is highly uncertain during promotions or new product launches. Deep probabilistic programming models learn demand patterns from historical data, constructing a probability distribution for demand variables and establishing dependencies with other supply chain variables to generate inventory strategies that can address a variety of demand scenarios.

[0228] Supplier risk assessment: For products that rely on multiple suppliers, the deep probabilistic programming model evaluates the risk levels of different supply options by modeling the probability distribution of variables such as supplier reliability and production capacity, and generates a robust multi-supplier collaboration strategy.

[0229] Inventory-cost balancing decisions: In inventory management, excessive inventory increases costs, while insufficient inventory increases stock-out risk. Deep probabilistic programming models model the joint probability distribution of variables such as inventory, cost, and demand to generate optimal inventory strategies that balance cost and stock-out risk under various scenarios.

[0230] Optionally, deep probabilistic programming models can also incorporate external knowledge (such as seasonality patterns and market expert experience) as prior information to further enhance model accuracy and interpretability. In some implementations, the model can continuously optimize itself by continuously updating the probability distribution based on new data using an active learning framework.

[0231] Step 3.2: Implementation of the multi-agent Monte Carlo tree search algorithm;

[0232] Based on a deep probabilistic programming model, a multi-agent Monte Carlo tree search algorithm is implemented to explore a better strategy space by sampling and evaluating multiple possible recovery paths:

[0233] Multi-agent definition: Modeling the supply chain recovery problem as a multi-agent system ,Each agent represents a decision-making entity in the supply chain (e.g., the purchasing department, inventory management department, logistics department, etc.);

[0234] in, 、 、 Respectively represent 、 、 an agent; represents the total number of agents;

[0235] State Space Definition: System State Represents the current state of the supply chain network, including the status information of each node

[0236] Joint Action Space: Joint Action ,in Representing an agent actions;

[0237] in, Indicates joint action; 、 、 Respectively represent 、 、 The actions of an agent; represents the total number of agents; Indicates the The actions of an agent; represents the joint action space; Indicates the The action space of each agent; Represents the index of the agent;

[0238] Monte Carlo tree search process:

[0239] Selection: Starting from the root node, select child nodes according to the UCT (Upper Confidence Bound for Trees) formula:

[0240]

[0241] in, Represents the UCT function, which is used to select nodes in tree search; Indicates the current state; Indicates joint action; represents the valuation of the state-action pair; represents the exploration constant, which is used to balance exploration and exploitation; Indicates status Number of visits; represents the number of visits to the state-action pair; represents the natural logarithm function;

[0242] Expansion: Expand new nodes based on transition probability;

[0243] Simulation: Starting from the expansion node, simulation is performed using a random strategy or a heuristic strategy until the termination state is reached;

[0244] Backpropagation: propagate the simulation results back to all nodes in the tree and update the node statistics;

[0245] Multi-agent collaborative approach: Implement collaborative decision-making among agents to ensure that the decisions of each agent form a consistent global recovery strategy;

[0246] Information sharing: Building an information sharing protocol between agents, allowing them to share key state information;

[0247] Collaborative decision-making: building a consensus-based decision-making algorithm to coordinate the action choices of each intelligent agent;

[0248] Conflict resolution: Implement conflict detection and resolution algorithms to handle possible decision conflicts between agents;

[0249] In addition, according to the embodiments of the present application, the advantage of the multi-agent Monte Carlo tree search algorithm is that it can more comprehensively evaluate the possible decision space through the collaborative exploration of multiple agents and find the globally optimal or near-optimal recovery strategy.

[0250] According to one embodiment of the present application, the detailed implementation of the multi-agent Monte Carlo tree search algorithm includes the following specific components:

[0251] Tree structure design:

[0252] Node representation: Each node contains state information , cumulative rewards , visit count and parent node reference;

[0253] While expressing: while expressing action , connects the parent node and the child node, and stores action related information;

[0254] Tree depth control: set a maximum depth limit to prevent the search space from exploding;

[0255] Decision-making time allocation:

[0256] Fixed iteration allocation: Allocate a fixed number of MCTS iterations to each decision point;

[0257] Adaptive allocation: dynamically allocate computing resources based on state complexity and uncertainty;

[0258] Real-time constraints: Optimize search quality within a limited time;

[0259] Multi-agent collaboration mechanism:

[0260] Hierarchical decision-making framework: divides decision-making into two levels: strategy level and execution level;

[0261] Shared value network: All agents share the same value evaluation network, but maintain their own search trees;

[0262] Message passing mechanism: defines a standardized message format to enable information exchange between agents;

[0263] Conflict resolution strategies: Use priority-based conflict resolution or auction-based resource allocation;

[0264] In the specific application of e-commerce supply chain out-of-stock decision-making, the multi-agent Monte Carlo tree search algorithm can be used in the following scenarios:

[0265] Multi-warehouse collaborative replenishment: When a product faces the risk of being out of stock in multiple warehouses at the same time, the system needs to coordinate the replenishment behavior of multiple warehouses.

[0266] By modeling each warehouse as an agent, the multi-agent MCTS system takes into account factors such as inventory status, service area, and replenishment costs, generating a coordinated replenishment plan to avoid blind competition and resource waste among warehouses. For example, the system might prioritize replenishing warehouses in core areas while adjusting order allocations to other warehouses to achieve overall optimization.

[0267] Supplier-Logistics-Warehousing Collaborative Decision-Making: Facing potential stock-out risks requires coordination across multiple links, including supplier production, logistics and transportation, and warehouse management. Multi-Agent MCTS models supplier management, logistics scheduling, and warehouse management as separate agents. Through collaborative search, it generates a coordinated recovery strategy that considers multiple constraints, including production capacity, shipping timeliness, and inventory capacity. For example, before a holiday promotion, the system might generate a collaborative strategy that includes pre-production, batched shipping, and temporary expansion of warehousing.

[0268] Multi-category inventory balance: For interrelated multi-category products (such as main products and accessories), when one category faces out-of-stock risk, it is necessary to comprehensively consider the inter-category relationships. Multi-agent MCTS models inventory management for different categories as separate agents. Through collaborative search, it generates a recovery strategy that balances demand across categories, preventing a shortage of one category from impacting sales of other related categories.

[0269] Alternatively, in some implementations, multi-agent MCTS can be combined with deep reinforcement learning, using a policy network to guide the tree search process to improve search efficiency, and using a value network to replace random simulation to improve evaluation accuracy. In other implementations, a distributed implementation can be used to distribute search tasks across multiple computing nodes, improving the ability to handle complex supply chain networks.

[0270] Step 3.3: Optimize the risk perception strategy;

[0271] Introducing risk-aware strategic optimization objectives to balance recovery efficiency and risk exposure:

[0272] Risk perception objective function definition:

[0273]

[0274] in, represents the objective function for risk-aware strategy optimization; Indicates the current strategy; represents the expectation operator; Indicates the current state; Represents the state distribution under the current strategy; Indicates joint action; represents the reward function; represents the reference strategy; To balance parameters and control the degree of risk aversion; Represents KL divergence, which measures the degree of deviation between the current strategy and the reference strategy;

[0275] Risk metrics: Define multiple risk metrics to comprehensively assess the risk characteristics of the strategy:

[0276] Expected risk: ;

[0277] in, represents the expectation operator; represents the risk loss function; Indicates the current state; Indicates joint action;

[0278] Conditional Value at Risk (CVaR):

[0279]

[0280] in, represents the conditional value at risk; represents the risk probability threshold; represents the risk loss function; represents the expectation operator; represents the value at risk;

[0281] Biggest risk:

[0282]

[0283] It indicates represents the maximum operator; Indicates the current state; Represents the state distribution under the current strategy; represents the risk loss function; Indicates the current strategy;

[0284] Policy optimization algorithm: Based on the risk-aware objective function, the policy gradient method is applied to optimize the policy parameters:

[0285]

[0286] in, Indicates the Strategy parameters for iterations; Indicates the Strategy parameters for iterations; represents the learning rate coefficient; Indicates the parameters Find the gradient; Representation Strategy The objective function

[0287] The gradient is estimated as:

[0288]

[0289] in, Indicates the policy parameters Find the gradient; Representation Strategy The objective function of Indicates from arrive The summation operator; Indicates the number of samples; represents a parameterized policy function; Indicates the The joint action of samples; Indicates the The current status of each sample; Indicates the sample index; Reward function representing a state-action pair; represents the risk weight coefficient; represents KL divergence; represents the reference strategy;

[0290] It should be understood that risk-aware policy optimization methods enable the generated policies to achieve a better balance between efficiency and risk by explicitly considering risk factors in the objective function, which is particularly important for out-of-stock recovery decisions in e-commerce supply chains.

[0291] According to one embodiment of the present application, the specific implementation of risk-aware strategy optimization includes the following core components:

[0292] Risk Measurement Framework:

[0293] Multidimensional risk representation: decomposing risk into multiple dimensions, such as time risk, cost risk, quality risk, etc.

[0294] Risk dependency modeling: capturing the interdependencies between different risk dimensions;

[0295] Risk propagation model: simulates the risk propagation process in the supply chain network.

[0296] Reference strategy construction:

[0297] Expert strategy extraction: extract expert strategies from historical successful cases as reference;

[0298] Conservative strategy generation: Constructing a conservative baseline strategy to ensure basic security;

[0299] Hybrid reference strategy: Dynamically combine multiple reference strategies according to different situations.

[0300] Risk-Aware Optimizer:

[0301] Constraint processing: converting risk thresholds into optimization constraints;

[0302] Multi-objective balance: seeking the best balance between efficiency, cost and risk;

[0303] Adaptive weight adjustment: Dynamically adjust risk weights based on environmental feedback .

[0304] In the specific application of out-of-stock prevention decision-making in e-commerce supply chains, risk-aware strategy optimization can be used in the following scenarios:

[0305] Risk control of key commodities: For core or high-value commodities, the risk of out-of-stock may bring serious reputation and economic losses. Risk perception strategy optimization is achieved by setting higher risk aversion parameters. , generate more conservative but lower-risk recovery strategies, such as maintaining higher safety stocks, establishing multi-level alternative supply channels, etc., to ensure the supply stability of such goods.

[0306] Seasonal product strategy adjustments: For highly seasonal products, both out-of-stock risks and excess inventory risks need to be considered. Risk-aware strategy optimization dynamically adjusts risk parameters, prioritizing out-of-stock risk control during peak sales seasons and excess inventory risk control during off-seasons, achieving year-round risk balance.

[0307] New product launch risk management: New product launches face the high risk of inaccurate demand forecasts. Risk-aware strategy optimization compares new product strategies with conservative strategies for similar historical products to generate a supply strategy that balances innovation and safety. For example, this strategy combines small-batch trial sales with rapid replenishment to manage the risks associated with uncertainty.

[0308] In some implementations, risk-aware strategy optimization can also incorporate online learning techniques to adjust risk weights based on real-time market feedback, adapting to the ever-changing market environment. Optionally, the system can also incorporate a risk warning mechanism that automatically triggers a review process or switches to a more conservative strategy when strategy risk exceeds a preset threshold.

[0309] Step 3.4: Generate optimized recovery strategy;

[0310] By integrating the results of deep probabilistic programming and multi-agent Monte Carlo tree search, we generate an optimized recovery strategy:

[0311] Strategy fusion: The generated elastic recovery path is integrated with the strategy explored by the multi-agent Monte Carlo tree search to combine the advantages of both.

[0312] Risk Adjustment: Based on the results of risk perception optimization, the strategy is risk-adjusted to ensure that risk exposure is controlled while maintaining efficient recovery;

[0313] Strategy Verification: Verify the performance of the optimization strategy in a variety of simulation scenarios to ensure that it performs well in various situations;

[0314] Final strategy generation: Based on the verification results, the final optimized recovery strategy is generated, including a detailed execution plan and emergency measures.

[0315] Therefore, this step combines deep probabilistic programming with multi-agent Monte Carlo tree search to generate a more robust and efficient recovery strategy while fully considering uncertainty and risk factors.

[0316] The output includes:

[0317] Optimized recovery strategies with higher robustness and lower risk;

[0318] Uncertainty estimation and risk assessment during strategy execution;

[0319] Predictions of strategy performance under various possible scenarios.

[0320] Data fusion and policy execution module: Based on the optimized recovery strategy, it realizes the real-time dynamic fusion of multi-source data, adaptively adjusts the weight of data sources, and converts abstract strategies into specific and executable supply chain adjustment instructions;

[0321] The specific steps include:

[0322] Step 4.1: Real-time dynamic fusion of multi-source data;

[0323] Realize real-time dynamic integration of multi-source data in the e-commerce supply chain, and adaptively adjust data source weights and integration strategies for different out-of-stock risk scenarios:

[0324] Data source access: Build a unified data access layer to support access to various data sources:

[0325] Transaction data sources: sales orders, refund orders, evaluation data, etc.

[0326] Inventory data source: product inventory level, inventory location, inventory status, etc.;

[0327] Supplier data source: supplier production capacity, delivery cycle, quality rating, etc.

[0328] Logistics data sources: transportation status, logistics timeliness, transportation capacity, etc.;

[0329] External data sources: market trends, competitor activities, seasonal factors, etc.

[0330] Data quality assessment: Real-time quality assessment of the incoming multi-source data:

[0331] Completeness assessment: detect missing data and make appropriate supplements;

[0332] Timeliness assessment: Evaluate the time delay of data and give priority to the latest data;

[0333] Consistency assessment: Detect and resolve conflicts between different data sources;

[0334] Accuracy assessment: Evaluate data accuracy through historical data verification;

[0335] Adaptive weight allocation: Adaptively adjust data source weights based on out-of-stock risk scenarios and data quality:

[0336]

[0337] in, Represents a data source The weight of Represents a data source Quality rating of Represents a data source Score the relevance of the current risk scenario; Express from arrive Sum all the items of ; Represents the sum variable, from arrive integer; Indicates the total number of data sources; Represents a data source Quality rating of Represents a data source Score the relevance of the current risk scenario; Indicates the index of the data source;

[0338] Multimodal data fusion is based on the assigned weights to achieve the fusion of multimodal data: feature-level fusion: weighted fusion of features from different data sources; decision-level fusion: making decisions based on different data sources, and then performing weighted integration; model-level fusion: training different models with different data sources, and then performing model integration.

[0339] It should be understood that the adaptive data fusion method provided in this application can dynamically adjust the importance of data sources according to different out-of-stock risk scenarios, thereby providing more accurate decision support information.

[0340] Step 4.2: Build the execution engine after out-of-stock recovery.

[0341] Build an out-of-stock recovery execution engine to convert abstract strategies into specific and executable supply chain adjustment instructions: Instruction parsing: parse the generated optimized recovery strategy into a specific set of execution instructions: Procurement instructions: including order quantity, supplier selection, delivery time, etc.; Inventory adjustment instructions: including inventory transfer, safety stock adjustment, etc.; Logistics instructions: including logistics route optimization, transportation method selection, etc.; Pricing instructions: including promotion strategies, purchase limit strategies, etc.

[0342] Execution plan generation: Generate a detailed execution plan based on the parsed instruction set: Task decomposition: decompose complex instructions into atomic tasks; Task sorting: sort tasks according to dependencies and priorities; Resource allocation: allocate required resources to each task; Time planning: formulate a schedule for task execution.

[0343] Constraint checking: perform constraint checking on the generated execution plan to ensure its feasibility: resource constraints: check whether it exceeds the available resource limit; time constraints: check whether the time requirements are met; business rule constraints: check whether it complies with business rules and policies; conflict check: check whether there are conflicts between instructions.

[0344] Instruction execution interface, builds standardized interfaces with various supply chain systems to achieve automated execution of instructions: procurement system interface: automatically generates and sends purchase orders; inventory management system interface: automatically adjusts inventory allocation and transfer; logistics system interface: automatically optimizes and adjusts logistics routes; pricing system interface: automatically adjusts product prices and promotion strategies.

[0345] In addition, the out-of-stock recovery execution engine can convert abstract recovery strategies into specific executable operational instructions, achieving seamless connection from decision-making to execution.

[0346] Closed-loop monitoring and optimization module: Builds a closed-loop feedback algorithm for executed adjustment instructions, continuously monitors execution results, and optimizes strategies;

[0347] The specific steps include:

[0348] Step 5.1, closed-loop feedback algorithm implementation;

[0349] Build a closed-loop feedback algorithm to continuously monitor execution results and optimize strategies: Execution effect monitoring: Real-time monitoring of the execution effect of the recovery strategy: Key indicator monitoring: out-of-stock rate, inventory level, customer satisfaction, etc.; Deviation analysis: Compare the deviation between actual execution results and expected results; Anomaly detection: Real-time detection of anomalies during the execution process.

[0350] Real-time adjustment: Based on monitoring results, the execution plan is adjusted in real time: parameter fine-tuning: fine-tuning the strategy parameters according to the execution effect; re-planning: re-planning the execution plan when major deviations occur; emergency handling: launching the emergency plan when an anomaly is detected.

[0351] Strategy evaluation and optimization: Regularly evaluate the effectiveness of strategies and optimize them: Effectiveness evaluation: Comprehensively evaluate the execution effect of strategies; Experience accumulation: Add execution data and results to the historical database; Strategy optimization: Optimize the strategy generation model based on accumulated execution experience.

[0352] It should be noted that the closed-loop feedback algorithm can continuously improve the performance and adaptability of the anti-out-of-stock decision-making system by continuously monitoring the execution results and making real-time adjustments.

[0353] Step 5.2: System integration and coordination;

[0354] Integrate various components into a unified system to achieve collaborative work:

[0355] Component integration: Integrate the components in each module into a unified system architecture;

[0356] Interface standardization: Define standardized interfaces between components to ensure smooth transmission of data and control flow.

[0357] Collaborative workflow: Define the system's workflow to ensure that all components work together:

[0358] Risk monitoring process: Continuously monitor out-of-stock risks;

[0359] Strategy generation process: Generates recovery strategies when risks are detected;

[0360] Strategy execution process: converting strategy into action and executing it;

[0361] Feedback optimization process: Optimize strategies based on feedback from execution results.

[0362] Therefore, through system integration and collaboration, the method provided in this application can form a complete closed-loop anti-out-of-stock decision-making system, providing comprehensive out-of-stock risk management capabilities from risk prediction, recovery strategy generation to execution monitoring and continuous optimization.

[0363] The above data can verify that the e-commerce supply chain out-of-stock prevention decision-making method based on multi-source data fusion provided in this application has significant technical effects in practical applications, and can effectively improve the supply chain resilience and operational efficiency of the e-commerce platform.

[0364] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. An e-commerce supply chain anti-out-of-stock decision system based on multi-source data fusion, characterized by: include: Heterogeneous graph network construction and risk prediction module: This module builds a heterogeneous graph network representing e-commerce supply chain relationships based on multi-source data, and uses a time-sensitive graph attention network to predict the propagation path of out-of-stock risks. Resilient recovery path generation module: Based on the constructed heterogeneous graph network, supply chain recovery is modeled as a sequential decision problem on the graph, and multiple resilient recovery paths are generated through variational inference methods; Strategy Optimization and Uncertainty Modeling Module: This module builds a deep probabilistic programming model for multiple recovery paths to model their uncertainties, and uses a multi-agent Monte Carlo tree search algorithm to optimize the recovery strategy. Data fusion and policy execution module: Based on the optimized recovery strategy, it realizes real-time dynamic fusion of multi-source data, adaptively adjusts the weight of data sources, and converts abstract strategies into specific and executable supply chain adjustment instructions; Closed-loop monitoring and optimization module: Build a closed-loop feedback algorithm for the executed adjustment instructions, continuously monitor the execution effect and optimize the strategy.

2. The e-commerce supply chain anti-out-of-stock decision system based on multi-source data fusion according to claim 1 is characterized by: The steps of constructing a heterogeneous graph network include: Collect multi-source data in the e-commerce supply chain, including sales data, inventory data, supplier data, logistics data, and market trend data; Preprocess the collected multi-source data, including data cleaning, missing value processing, outlier detection and data standardization; Construct a heterogeneous graph network representing the e-commerce supply chain relationship, where the nodes include product nodes, warehouse nodes, supplier nodes and logistics nodes, and the edges include supply relationships, inventory relationships and logistics relationships.

3. The e-commerce supply chain anti-out-of-stock decision system based on multi-source data fusion according to claim 1 is characterized by: The time-sensitive graph attention network measures the association strength between different nodes by calculating the attention coefficient, which takes into account the temporal characteristics of the edges, including time coding, periodic pattern coding and temporal relationship coding, to capture the characteristics of supply chain relationships that change over time.

4. The e-commerce supply chain anti-out-of-stock decision system based on multi-source data fusion according to claim 1 is characterized by: The variational inference method generates an elastic recovery path through the following steps: Define the posterior distribution of the recovery path and introduce the variational distribution as an approximate alternative to the posterior distribution; Optimize parameters by maximizing the variational lower bound; Integrate graph network structures into variational inference models and use graph neural networks to process the structural information of supply chain networks; Based on Bayesian decision theory, a multi-step reasoning algorithm is implemented to evaluate the costs and effects of different recovery paths.

5. The e-commerce supply chain anti-out-of-stock decision system based on multi-source data fusion according to claim 1 is characterized by: Define the set of random variables representing the supply chain status and the conditional dependencies between the variables; Parameterize conditional probability distributions using neural networks; Apply the stochastic variational inference algorithm to estimate the posterior distribution.

6. The e-commerce supply chain anti-out-of-stock decision system based on multi-source data fusion according to claim 1 is characterized by: The multi-agent Monte Carlo tree search algorithm includes: The supply chain recovery problem is modeled as a multi-agent system, where each agent represents a decision-making entity in the supply chain; Define system states and joint action spaces; Execute the Monte Carlo tree search process, including four steps: selection, expansion, simulation, and backpropagation; Implement a collaborative decision-making method among intelligent agents to ensure that the decisions of each intelligent agent form a consistent global recovery strategy.

7. The e-commerce supply chain anti-out-of-stock decision system based on multi-source data fusion according to claim 1 is characterized by: The objective function of the risk-aware strategy optimization includes a reward term and a risk aversion term. The reward term measures the expected return of the strategy, and the risk aversion term controls the degree of risk by calculating the difference between the current strategy and the reference strategy.

8. The e-commerce supply chain anti-out-of-stock decision system based on multi-source data fusion according to claim 1 is characterized by: The step of real-time dynamic fusion of multi-source data includes: Build a unified data access layer to support access to various data sources; Conduct real-time quality assessment of incoming multi-source data, including completeness, timeliness, consistency, and accuracy; Adaptively adjust data source weights based on out-of-stock risk scenarios and data quality; Based on the assigned weights, feature-level fusion, decision-level fusion, and model-level fusion are achieved.

9. The e-commerce supply chain anti-out-of-stock decision system based on multi-source data fusion according to claim 1 is characterized by: The steps of converting the abstract strategy into specific executable supply chain adjustment instructions include: Parse the optimization recovery strategy into a specific set of execution instructions, including procurement instructions, inventory adjustment instructions, logistics instructions, and pricing instructions; Generate a detailed execution plan based on the parsed instruction set, including task decomposition, task sequencing, resource allocation, and time planning; Perform constraint checks on the generated execution plan to ensure its feasibility; Build standardized interfaces with various supply chain systems to achieve automated execution of instructions.

10. A storage medium, characterized in that: It includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement an e-commerce supply chain anti-out-of-stock decision system based on multi-source data fusion as described in any one of claims 1 to 9.

Images