Building material supplier dynamic evaluation and recommendation system based on big data

By constructing a five-dimensional spatiotemporal tensor and performing causal analysis and reinforcement learning through a big data dynamic evaluation system, the problems of single evaluation dimensions and opaque decision-making for suppliers are solved, enabling dynamic, multi-dimensional evaluation and transparent recommendation of building material suppliers.

CN120912238APending Publication Date: 2025-11-07SHENZHEN YUEXIN DIGITAL TECHNOLOGY GROUP CO LTD

Patent Information

Application Number
CN202511015369.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-23
Publication Date
2025-11-07

AI Technical Summary

Technical Problem

Existing technologies for evaluating building material suppliers have limited dimensions, their recommendations are easily influenced by confounding factors, and their decision-making processes are opaque, making them unable to adapt to complex project environments and dynamic changes.

Method used

We employ a big data-based dynamic evaluation and recommendation system. Through data collection and fusion, spatiotemporal tensor modeling, causal analysis and confounding correction, multi-agent recommendation decision-making and interpretability modules, we construct a five-dimensional spatiotemporal tensor representation of supplier performance, identify and correct confounding variables, and use reinforcement learning to generate transparent recommendation strategies.

Benefits of technology

It enables dynamic, multi-dimensional, and correlated evaluation of supplier performance, improving the comprehensiveness and accuracy of the evaluation, the robustness and transparency of the recommendation results, and enhancing the credibility and decision support capabilities of the system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120912238A_ABST
    Figure CN120912238A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of computer data processing, and discloses a building material supplier dynamic evaluation and recommendation system based on big data, and the system comprises a data fusion module which integrates multi-source heterogeneous data to generate a unified data set; the tensor modeling module is used for constructing and decomposing a five-dimensional space-time tensor to obtain a factor matrix and a dynamic weight; the causal correction module is used for establishing a causal graph based on the network relationship and eliminating hybrid deviation; the recommendation decision module is used for outputting a recommendation list through reinforcement learning in combination with the evaluation weight and the performance distribution; and the interpretable module is used for generating an interpretable report based on the factor and the causal path. According to the method, the technical scheme of multi-source heterogeneous data fusion and five-dimensional space-time tensor decomposition is adopted, and dynamic, multi-dimensional and relevance evaluation of supplier performance is realized by constructing a unified data structure including suppliers, time, static characteristics, context and cooperative relationships.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer data processing, and particularly to a building material supplier dynamic evaluation and recommendation system based on big data. BACKGROUND

[0002] In modern construction engineering practice, the stability and efficiency of the supply chain are the cornerstone of project success, and any link, especially the selection of building material suppliers, will directly affect the cost control, construction progress and final engineering quality of the project. In the face of increasingly complex project environment and massive supplier information, the traditional decision-making method relying on personal experience or simple historical data has been difficult to meet the needs of fine management.

[0003] At present, the supplier management systems on the market have made certain development. These systems can effectively integrate structured data in enterprise internal ERP and project management software, realize the digital archiving of supplier historical transaction records, contract performance, quality inspection reports and other information. Through statistical analysis of these archived data, the system can automatically calculate a series of quantitative performance indicators such as the historical average delivery punctuality rate and historical quality pass rate of suppliers. This way liberates procurement personnel from tedious manual data processing and provides a preliminary, data-supported quantitative reference benchmark for supplier evaluation, marking the preliminary progress of supplier management from pure subjective judgment to data management.

[0004] However, the existing technology still exposes a series of fundamental limitations in depth application. First, the evaluation dimensions of these systems for suppliers are static and isolated, and the performance indicators generated only reflect the historical average level, failing to capture the dynamic performance differences of suppliers at different times, in specific context environments and when working with different partners. This leads to a lack of timeliness and situational adaptability in the evaluation results. Second, the correlation analysis based on observed data by existing systems cannot effectively distinguish the real ability of suppliers from the influence of external mixed factors. Apparently excellent delivery records are not due to the high efficiency of suppliers themselves, but only accidental companionship with specific favorable conditions. This recommendation based on false correlation has significant risks and uncertainties. Third, the decision logic of existing systems is usually opaque. The scores or rankings given by them are closed to users, and the specific reasons for recommending a certain supplier cannot be explained. The unexplainability of this decision-making process seriously restricts its application depth in the engineering procurement field which requires high trust and risk audit. Therefore, the technical personnel in the field propose a building material supplier dynamic evaluation and recommendation system based on big data to solve the above problems. SUMMARY

[0005] In view of the deficiencies of the prior art, the application provides a building material supplier dynamic evaluation and recommendation system based on big data, which solves the technical problems of single supplier evaluation dimension, recommendation results susceptible to interference from mixed factors, and non-transparent decision-making process in the prior art.

[0006] To achieve the above object, the application is implemented by the following technical solutions: a building material supplier dynamic evaluation and recommendation system based on big data, comprising:

[0007] A data acquisition and fusion module is configured to acquire and integrate structured data, unstructured data, context environment data and supply chain network relationship data of building material suppliers from multiple data sources, and generate a unified formatted data set.

[0008] A spatio-temporal tensor modeling and decomposition module is configured to construct a five-dimensional spatio-temporal tensor to represent the dynamic attributes of the suppliers based on the unified formatted data set, and decompose the five-dimensional spatio-temporal tensor to generate a factor matrix and a dynamic evaluation weight.

[0009] A causal analysis and confounding correction module is configured to construct a causal graph model for supplier evaluation based on the supply chain network relationship data, and identify and correct the influence of confounding variables on the evaluation results in the causal graph model using the factor matrix, to generate a causal correction performance distribution of the suppliers.

[0010] A multi-agent recommendation decision module is configured to combine the dynamic evaluation weight and the causal correction performance distribution, train a recommendation decision strategy through reinforcement learning, and output a supplier recommendation list.

[0011] A recommendation result explainability module is configured to generate an explainability report containing factor traceability paths and counterfactual explanation paths for the supplier recommendation list based on the factor matrix and the causal graph model.

[0012] Preferably, the data acquisition and fusion module comprises:

[0013] A structured data acquisition unit is configured to acquire structured data of basic information, historical transaction data, performance time and quality detection records of suppliers from enterprise resource planning systems, project management systems and contract performance records.

[0014] An unstructured data acquisition unit is configured to acquire text information related to the suppliers from Internet public opinion platforms, social media, news websites and policy document databases, and perform natural language processing and sentiment feature extraction on the text information.

[0015] A context environment data acquisition unit is configured to acquire time-sensitive external environment data affecting the operation of the supply chain from meteorological data platforms, public holiday service interfaces and market price index libraries.

[0016] A supply chain network data collection unit is configured to construct a collaboration graph between suppliers and construction units, logistics companies based on records of construction site systems, logistics tracking systems and material distribution platforms;

[0017] A data fusion unit is configured to preprocess, feature align and semantically map the structured data, unstructured data, contextual environment data and supply chain network data, and generate a uniformly formatted data set.

[0018] Preferably, the spatio-temporal tensor modeling and decomposition module comprises:

[0019] A tensor construction unit is configured to construct a five-dimensional tensor according to the supplier dimension, time dimension, static feature dimension, context dimension and collaboration relationship dimension;

[0020] A factor decomposition unit is configured to perform tensor decomposition on the five-dimensional tensor and extract a plurality of latent factor matrices to represent the latent interaction features between the dimensions;

[0021] A weight generation unit is configured to calculate the dynamic weights of each evaluation index at different time points in combination with the time and context feature information.

[0022] Preferably, each dimension in the five-dimensional tensor of the tensor construction unit comprises:

[0023] The first dimension is the supplier dimension, which is used to distinguish between each independent supplier entity;

[0024] The second dimension is the time dimension, which is used to represent the evolution process of data in a continuous time sequence;

[0025] The third dimension is the static feature dimension, which is used to represent the basic unchanging attributes of the supplier;

[0026] The fourth dimension is the context feature dimension, which is used to express the external environmental factors associated with the supplier;

[0027] The fifth dimension is the collaboration relationship dimension, which is used to describe the interaction relationship between the supplier and other subjects in the supply chain network.

[0028] Preferably, the causal analysis and confounding correction module comprises:

[0029] A causal graph construction unit is configured to construct a causal graph model based on the supplier static attributes, context features and collaboration structure, wherein the nodes in the graph represent variables and the edges represent causal relationships;

[0030] A confounding factor identification unit is configured to identify a set of confounding variables that affect the supplier evaluation results according to backdoor path criteria;

[0031] An intervention correction unit is configured to perform intervention modeling based on the identified confounding factors, and generate a distribution of the performance of the supplier after the confounding factors are controlled.

[0032] Preferably, the intervention modeling operation of the intervention correction unit comprises:

[0033] value setting operation is performed on the direct parent nodes of the target variable in the causal graph to simulate the intervention effect;

[0034] Based on the joint probability distribution of each confounding variable, the marginal distribution change of each evaluation index when the target variable is fixed is calculated.

[0035] The distribution of the target evaluation index under the control of the confounding factor is output for the recommendation model to call.

[0036] Preferably, the multi-agent recommendation decision module comprises:

[0037] A state modeling unit is configured to construct a current system state representation including a supplier tensor factor, inventory information and risk parameters;

[0038] A reward function construction unit is configured to construct a reward function corresponding to a supplier recommendation action according to the dynamic weight of the evaluation index, and measure the strategy value of each recommendation combination;

[0039] A reinforcement learning strategy unit is configured to train a strategy model through a multi-agent reinforcement learning method, and optimize the supplier selection action output under the current state;

[0040] A recommendation output unit is configured to output a supplier recommendation list and a ranking result according to the strategy model.

[0041] Preferably, the training strategy model process of the reinforcement learning strategy unit comprises:

[0042] Initialize the parameters of the policy network and the state value function network;

[0043] Collect the state, action and reward sequence in the multi-round interaction process;

[0044] Calculate the gradient update direction based on the policy gradient method and update the policy network parameters;

[0045] Update the state value function using the time difference method.

[0046] Preferably, the recommendation result interpretability module comprises:

[0047] A factor path extraction unit is configured to identify the tensor factor that plays a leading role in the current recommendation result, and generate a factor trace path;

[0048] The counterfactual reasoning unit is configured to construct a change scenario of a causal variable and reevaluate a recommended strategy result to generate a counterfactual explanation path.

[0049] The visualization generation unit is configured to generate an explanatory chart and a report file containing factor source descriptions and variable change influence paths.

[0050] Preferably, the counterfactual scenario construction of the counterfactual reasoning unit comprises:

[0051] selecting one or more external variables that can be intervened from the causal graph;

[0052] constructing an intervention scenario input model of variable value changes;

[0053] re-running the recommended strategy model under the intervention input and obtaining a recommended ranking change amount;

[0054] outputting a recommended strategy change path caused by variable changes and an impact analysis report.

[0055] The present application provides a building material supplier dynamic evaluation and recommendation system based on big data.

[0056] The present application has the following beneficial effects:

[0057] 1、The present application adopts a technical scheme of multi-source heterogeneous data fusion and five-dimensional spatiotemporal tensor decomposition, realizes dynamic, multidimensional, and correlation evaluation of supplier performance by constructing a unified data structure containing suppliers, time, static characteristics, context, and collaboration relationship, and solves the problems of single evaluation dimension, inability to reflect time evolution and environmental impact, and neglect of supply chain collaboration relationship in the prior art, thereby significantly improving the comprehensiveness and accuracy of supplier evaluation.

[0058] 2、The present application applies a recommendation decision scheme combining causal graph modeling and multi-agent reinforcement learning, the system first identifies and corrects the interference of mixed variables on the evaluation result by constructing a causal graph, obtains an unbiased supplier performance distribution, and then learns the optimal recommended combination in a dynamic environment using a reinforcement learning strategy, which solves the problems of the prior art that the recommendation result is easily misled by false correlations and cannot adapt to complex project requirements and environmental dynamic changes, making the recommendation decision more robust and scientific.

[0059] 3、The application designs an explainable technical scheme based on factor tracing and counterfactual reasoning, the recommendation system can clearly show the key influence path by backtracking tensor decomposition factors when giving the decision list, and show the sensitivity of the recommendation result by simulating variable changes, compared with the recommendation model in the prior art which has the problems of opaque decision process, difficult to explain the recommendation logic, the application solves the problems of opaque decision process, difficult to trust for users and the result cannot be audited, greatly enhances the transparency, credibility and decision support ability of the system. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 It is a system architecture diagram of the application;

[0061] Figure 2 It is a data acquisition and fusion module architecture diagram of the application;

[0062] Figure 3 It is a space-time tensor modeling and decomposition module architecture diagram of the application;

[0063] Figure 4 It is a causal analysis and confounding correction module architecture diagram of the application;

[0064] Figure 5 It is a multi-agent recommendation decision module architecture diagram of the application;

[0065] Figure 6 It is a recommendation result explainability module architecture diagram of the application. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the accompanying drawings of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0067] Please refer to the accompanying drawings of the application Figure 1 - the accompanying drawings of the application Figure 6 The embodiment of the application provides a building material supplier dynamic evaluation and recommendation system based on big data, which comprises:

[0068] A data acquisition and fusion module is used to acquire and integrate structured data, unstructured data, context environment data and supply chain network relationship data of building material suppliers from multiple data sources, and generate a unified formatted data set;

[0069] Specifically, the data collection and fusion module, as the starting functional unit of the system, plays a fundamental role in data source integration and semantic alignment in the overall system. The design of this module is directly related to the data quality and semantic consistency of downstream modules such as tensor modeling, causal analysis, and recommendation strategy. Therefore, this module not only needs to ensure data diversity and timeliness, but also needs to consider data structure unification, feature association traceability, and multi-source semantic coordination, providing a unified and semantically complete input dataset for subsequent multi-dimensional spatio-temporal tensor construction.

[0070] Generally, the information of building material suppliers is distributed in multiple heterogeneous data sources, including structured information systems such as project management systems, as well as unstructured data sources such as social media and policy texts. Additionally, external time-sensitive factors such as weather and holidays, as well as dynamic collaboration between suppliers, construction units, and logistics nodes, need to be considered.

[0071] In this embodiment, the data collection and fusion module mainly includes four types of data collection units and one data fusion unit, which are used to structure the extraction, semantic mapping, and format standardization of various raw data, thereby generating a unified format dataset for subsequent tensor construction.

[0072] Specifically, in one possible implementation, the structured data collection unit is configured to access existing ERP systems, project management platforms, and contract performance databases of construction enterprises, and collect basic attribute information of suppliers, such as registration information, business scope, qualification level, and historical bid records, transaction amounts, contract performance, and quality inspection data. These data have clear data fields and can be directly input as static feature items of suppliers into downstream models.

[0073] As an option, the unstructured data collection unit accesses internet crawler interfaces, social text scraping APIs, and government disclosure databases to obtain text content related to suppliers, such as news reports, user comments, social dynamics, and policy announcements. In general, the unstructured text needs to go through natural language processing procedures, including word segmentation, part-of-speech tagging, entity recognition, and sentiment analysis, and further mapped to quantitative indicators or labels to form supplementary supplier behavior or reputation features.

[0074] In some embodiments, the contextual environment data collection unit is used to provide time-sensitive external influencing factors, such as weather conditions (temperature, rainfall, humidity, etc.), market price indices (raw materials such as steel and cement), and holiday factors, which can be obtained from national meteorological platforms, price index service systems, and calendar APIs. These data are mainly indexed by time and can be used as contextual feature dimensions input on the supply chain.

[0075] In another implementation, the supply chain network data collection unit extracts the interaction event records between suppliers and construction units, logistics companies by accessing the construction site material tracking system, logistics path recording system and material scheduling platform, and constructs the collaboration relationship triplets between heterogeneous entities. The general form of the triplet is (v i , k , j ), where v i and v j represent the suppliers and the cooperation objects (such as construction units or logistics nodes), respectively, and r k represents the collaboration relationship type between them, such as "material supply", "transportation support", etc.

[0076] After all the data is preprocessed, missing data is filled, features are standardized and encoded, the data fusion unit will perform unified formatting operations. The unified formatted dataset is defined as:

[0077]

[0078] where: represents the static feature vector of the i-th supplier; represents the context environment feature vector at time point t; represents the collaboration relationship feature vector between supplier i and cooperation object j at time t; N is the total number of suppliers; T is the length of the sampling time sequence; M, K, and Q represent the static feature dimension, the context feature dimension, and the collaboration feature dimension, respectively.

[0079] In a specific implementation, the structured data can be extracted in bulk through database connection or RESTful API interface; the unstructured data is extracted by text extraction and directional semantic model; the context data is extracted by sliding window with time index; the collaboration relationship data is represented in graph structure and encoded as sparse vector in the fusion stage.

[0080] As an extension, in some implementations, a time window sliding strategy can also be introduced to aggregate and calculate historical data by window, extract trend change features, such as normalizing the fluctuation amplitude of material price in each time period, for subsequent dynamic modeling.

[0081] Further, in a specific embodiment, the fused data adopts a unified key index strategy when writing to the data warehouse, for example, the primary key is (i, t), to ensure that the subsequent tensor construction stage can be read in order by dimension.

[0082] The spatio-temporal tensor modeling and decomposition module is configured to construct a five-dimensional spatio-temporal tensor representing dynamic attributes of the supplier based on the uniformly formatted data set, and decompose the five-dimensional spatio-temporal tensor to generate a factor matrix and dynamic evaluation weights.

[0083] Specifically, after the data acquisition and fusion module completes data processing, a uniformly formatted data set The uniformly formatted data set is input to the spatio-temporal tensor modeling and decomposition module of the application. The module is configured to construct a multi-dimensional tensor structure to comprehensively represent time evolution characteristics, static attributes, context environmental factors and interaction states of the supplier in the supply chain collaboration network, and extract potential factors and dynamic weights by a tensor decomposition method, so as to provide a formal standard, semantic compression and dynamically adjustable index expression mode for subsequent state modeling of causal evaluation and recommendation strategy.

[0084] In the embodiment, the spatio-temporal tensor modeling and decomposition module includes a tensor construction unit, a factor decomposition unit and a weight generation unit, which are configured to complete tasks such as high-dimensional data structuring, potential factor learning and dynamic index weight assignment. The core objective of the module is to map information from different dimensions to a unified tensor space and extract coupling relationships between the dimensions.

[0085] In a possible implementation, the tensor construction unit is configured to construct a five-dimensional tensor based on the uniformly formatted data set The five-dimensional tensor is constructed as Wherein:

[0086] N represents the number of suppliers;

[0087] T represents the total number of time points sampled;

[0088] M represents the number of static feature dimensions, such as qualification level, registered capital, etc.;

[0089] K represents the number of context feature dimensions, such as weather, market index, holiday, etc.;

[0090] Q represents the number of collaboration relationship feature dimensions, such as delivery frequency, response time delay, number of cooperation nodes, etc.

[0091] represents the interaction value of the mth static feature and the kth context variable and the qth collaboration index of the supplier i at time t.

[0092] In general, the above tensor data can be filled into the five-dimensional tensor by using a dimension-by-dimension mapping method after tensorizing information from different dimensions.

[0093] In the embodiment, to extract the potential low-rank structure inside the tensor, the factor decomposition unit is configured to decompose the five-dimensional tensor into a factor matrix and a weight matrix. Canonical Polyadic (CP) decomposition is performed, and its mathematical expression is as follows:

[0094]

[0095] wherein R is the rank of decomposition, that is, the number of latent factors; is a supplier factor vector; is a time factor vector; is a static feature factor vector; is a context factor vector; is a collaboration relationship factor vector; and denotes an outer product operation.

[0096] In the decomposition process, each dimension of the tensor is projected into the corresponding factor matrix, thereby compressing the high-dimensional interaction features and obtaining the interpretable latent variable structure.

[0097] Specifically, if the factor matrices are denoted as:

[0098]

[0099] then any element in the tensor can be approximately expressed as:

[0100]

[0101] wherein R represents the number of latent factors extracted after tensor decomposition; U ir represents an element in the supplier factor matrix U; tr represents an element in the time factor matrix V; mr represents an element in the static feature factor matrix W; kr represents an element in the context factor matrix X; qr represents an element in the collaboration relationship factor matrix Z.

[0102] In the above formula, the rows of each factor matrix represent the embedding representation of the entity in the corresponding dimension, and the columns represent the projection value of each latent factor in different dimensions. This representation form provides a structured basis for subsequent dynamic modeling and interpretable analysis.

[0103] As an option, to further improve the ability of the model to depict dynamics, a dynamic weight calculation mechanism based on the time factor and the context factor is designed in the embodiment. The mechanism defines the dynamic weight w i (t) of the index at time point t to reflect the influence of the change of the context environment on the relative importance of the evaluation index. The calculation method is as follows:

[0104]

[0105] wherein: represents the projection coefficient of the rth latent factor on the ith evaluation index; v tr is the factor value of the rth factor at time t; x k(t)r is the value of the context feature index k(t) corresponding to time t on the rth factor; w i (t) represents the weighted importance of the evaluation index i at time t.

[0106] In some embodiments, to ensure weight stability and interpretability, the weight vector w(t) can be normalized so that for all i, ∑ i w i (t) = 1.

[0107] In addition, the calculated dynamic weight will be cached in the dynamic index database as a time-varying index weight benchmark called by subsequent modules for causal assessment and strategy modeling.

[0108] The causal analysis and confounding correction module is used to construct a causal graph model of supplier evaluation according to the supply chain network relationship data, and identify and correct the influence of confounding variables on the evaluation results in the causal graph model using the factor matrix, to generate a causal corrected performance distribution of the supplier;

[0109] Specifically, after completing the tensor modeling and factor decomposition of the multi-dimensional attributes of the supplier, the invention further introduces a causal analysis and confounding correction module as an important bridge between modeling and recommendation. This module is used to mine the causal relationship between variables from the structured factor space, to clearly understand the actual influence mechanism of different feature dimensions on the evaluation results of the supplier, and to effectively identify and control the estimation bias caused by interference variables, thereby providing more reliable and unbiased performance estimation data for recommendation strategy training.

[0110] Generally, in a multi-source heterogeneous data driven evaluation system, since the context variables and collaboration structure are often influenced by external mechanisms together with the supplier behavior, there is serious confounding factor interference, which will cause bias in the scoring system based on observed data, affecting the effectiveness of the recommendation system. Therefore, a causal model needs to be established on the observed data, and the confounding effect needs to be eliminated through intervention modeling.

[0111] In this embodiment, the causal analysis and confounding correction module includes a causal graph construction unit, a confounding factor identification unit, and an intervention correction unit, which correspond to core functions such as causal structure modeling, confounding variable identification, and causal effect correction, respectively.

[0112] In one possible implementation, the causal graph construction unit constructs a directed acyclic graph structure in accordance with factor matrices extracted from the tensor modeling module, in combination with context variables and collaboration network information wherein: V = {X1, X2, …, X d , Y} represents a node set, containing input variables X such as supplier static attributes, temporal context variables, collaboration structure features, and evaluation result variables Y; i represents a set of causal edges between variables; edges between nodes represent direct causal influence relationships.

[0113] Alternatively, the preliminary structure of the causal graph can be generated by a structure learning method (such as the PC algorithm, the NOTEARS method) in combination with domain knowledge, while maintaining the interpretability of the graph and controlling the modeling error.

[0114] Specifically, after the causal graph is constructed, the confounder identification unit will analyze the potential confounder set in the graph based on the “backdoor path criterion” The backdoor path criterion is a standard for judging whether there is a non-causal path from variable X to result variable Y, and its core idea is that if there is a path from X to Y through a non-causal direction of a confounder variable, the path needs to be blocked by conditioning z.

[0115] In the present application, the confounder set is defined as the smallest set of variables that satisfy the following conditions:

[0116]

[0117] This condition ensures that all non-causal paths from variable X to Y are closed, thereby achieving an unbiased estimate of the causal effect.

[0118] In some embodiments, the intervention correction unit eliminates the influence of confounders by simulating “intervention operations”. This process uses the do-operator formalization framework proposed by Pearl to calculate the distribution of the result Y under external intervention on the variable X:

[0119]

[0120] wherein: P(Y|do(X=x)) represents the conditional distribution of the result variable Y after intervention on X; is the set of identified confounders; is the joint conditional probability under observation; is the marginal distribution of the confounder variable.

[0121] The above formula realizes the modeling of the causal effect of X on Y under intervention by marginalizing the confounder variable.​

[0122] As a supplement mechanism, in actual system implementation, to avoid the calculation overhead brought by high-dimensional conditional probability estimation, the embodiment can introduce a propensity score approximation method to convert the high-dimensional adjustment problem into a one-dimensional matching problem. The propensity score e(x) function is defined as:

[0123]

[0124] Wherein, X is the judgment variable; is the identified confounding variable set.

[0125] Then, the propensity score can be used for weighted regression or matching estimation to indirectly control the confounding bias.

[0126] In addition, to ensure the efficiency of subsequent module calls, the intervention correction unit will re-model the current variable distribution after updating the tensor factor each time, and store the corresponding causal correction evaluation results. The corrected performance estimation value will replace the original evaluation label as the target distribution in the training process of the recommendation strategy module.

[0127] In some embodiments, intervention modeling can also be extended to handle problems such as covariate shift and context switching. For example, the context heterogeneity existing in different geographical areas or construction stages can be handled through subgraph segmentation and local causal reasoning, thereby improving the regional adaptability of the model.

[0128] The multi-agent recommendation decision module is used to combine the dynamic evaluation weight and the causal correction performance distribution, train a recommendation decision strategy through reinforcement learning method, and output a supplier recommendation list.

[0129] Specifically, after the performance estimation after causal correction is completed, the system of the present application enters the multi-agent recommendation decision module to realize supplier combination recommendation based on the comprehensive historical evaluation, environmental context and causal adjustment results. The module takes the dynamic evaluation weight and the causal correction performance distribution as input, combines the current project demand state, constructs the state space of the recommendation environment, and designs a recommendation strategy based on the reinforcement learning theory to generate an optimal supplier set.

[0130] Generally, due to the characteristics of frequent state changes, high-dimensional decision space, complex collaborative constraints, etc. in the supply chain system, traditional rule-based or heuristic methods are difficult to effectively deal with the recommendation optimization problem. Therefore, the present application introduces a multi-agent reinforcement learning-based method to dynamically learn the recommendation strategy in the state-action space to obtain the optimal supplier selection result.

[0131] In this embodiment, the multi-agent recommendation decision module mainly includes a state modeling unit, a reward function construction unit, a reinforcement learning strategy unit, and a recommendation output unit. The above functional units work together to realize the complete decision chain from state representation, strategy training to recommendation result output.

[0132] In one possible implementation, the state modeling unit is configured to construct the system state s t of the current decision cycle.Specifically, the state consists of three parts:

[0133]

[0134] Wherein: denotes the normalized representation matrix of the supplier spatiotemporal tensor factor at time point t, where M ′ is the factor dimension after tensor compression; denotes the current inventory state vector of each supplier; denotes the resource demand feature vector of the current project, and the dimension L is set according to the project size and material type.

[0135] In this state, the action space is defined as a set of supplier combination selection strategies:

[0136]

[0137] Wherein: denotes whether supplier i is selected to participate in the supply task at time t.

[0138] The dimension of the action vector a t is N, corresponding to the full set of suppliers.

[0139] As an option, the reward function construction unit combines the dynamic weight w i (t) output by the tensor module and the performance distribution provided by the causal module to construct the reward function for action evaluation:

[0140]

[0141] Wherein: R(s t ,a t ) represents the reward function in reinforcement learning; action selection of supplier i; w j (t) is the weight of the jth evaluation index at time t; the causal correction score of the ith supplier under the jth index; D is the total number of evaluation indexes; λ is a regulation factor, which controls the trade-off between benefits and costs; Cost(a tThis indicates the procurement or logistics costs resulting from the current combination of actions, and constraints such as transportation distance and response time can be further considered.

[0142] Specifically, the reinforcement learning policy unit uses the policy gradient method to construct the policy network π. θ (a t |s t ), where θ are the parameters of the strategy model. The strategy objective function J(θ) is defined as the expected cumulative return:

[0143]

[0144] Where: γ∈(0,1) is the time discount factor; T is the time length of the strategy evaluation.

[0145] By analyzing the policy gradient The estimation is performed using a sampling-based gradient update method to optimize the training of the policy network. In some embodiments, a state-value function V can be further introduced. φ (s t Using this as a benchmark, an advantage function is constructed for strategy improvement.

[0146] In one implementation, the training process employs a multi-agent parallel interaction mechanism, with each agent corresponding to a project context or geographical region to enhance the generalization performance of the policy. Each training round collects the state-action-reward sequence {(s)}. t ,a t ,R t The state value function is optimized using the time difference (TD) update method.

[0147] The recommendation output unit uses the finally converged policy model for actual recommendations, given the current state s. t At that time, output a list of recommended suppliers and their sorting results.

[0148]

[0149] The above recommended list can be used as input to the procurement management system to generate material procurement plans or scheduling tasks.

[0150] As an extension mechanism, this embodiment can also introduce constraint strategy optimization methods, such as the Lagrange multiplier method, to meet business rules such as project budget caps and cooperation frequency thresholds. For example, the constraint item Budget(a t )≤B is transformed into a constraint penalty term embedded in the objective function.

[0151] The recommendation result interpretability module is used to generate an interpretability report containing factor tracing paths and counterfactual explanation paths for the supplier recommendation list, based on the factor matrix and the causal graph model.

[0152] Specifically, after the optimization output of the multi-agent recommendation strategy is completed, to improve the transparency of the system, enhance the trust of users in the recommendation decision, and support decision auditing, the application further sets up a recommendation result explainability module. As a terminal component of the system, the module is closely connected after the output of the strategy recommendation, and combines the intermediate structural information extracted in the tensor modeling and the causal graph construction process to perform explanatory enhancement on the recommendation list.

[0153] Generally, a recommendation model based on deep policy learning or high-dimensional tensor modeling is difficult to directly understand the generation logic of the recommendation result due to its complex structure and non-transparent parameters, which can easily cause users to question the legality and scientificity of the recommendation basis. Therefore, the module takes “factor path backtracking” and “counterfactual reasoning” as the core design concept, and provides an output form with traceability and causal explanation through path-level factor contribution analysis and variable-level scenario comparison.

[0154] In this embodiment, the recommendation result explainability module includes a factor path extraction unit, a counterfactual reasoning unit, and a visualization generation unit, which correspond to traceability analysis of recommendation factors, simulation of changes in recommendation results under variable changes, and generation of final user-side explainable results, respectively.

[0155] In one possible implementation, the factor path extraction unit calculates which potential factors mainly affect each supplier in the recommendation ranking based on the CP decomposition factor matrix generated in the tensor modeling module. The influence of the rth potential factor on the recommendation score of supplier i is defined as:

[0156]

[0157] wherein: is the element of the supplier factor matrix in the ith row and the rth column; is the element of the time factor matrix in the jth row and the rth column; are the elements of the context and collaboration factor matrices, respectively; K and Q are the context dimension number and the collaboration relationship dimension number, respectively.

[0158] The above formula calculates the importance of the rth potential factor to the ranking of the ith supplier. In some embodiments, the contributions of all potential factors can be further normalized to output a factor path chart for identifying the key path combination of the recommendation basis.

[0159] Specifically, the counterfactual reasoning unit performs scenario simulation on the variable dependence of the current recommendation result based on the causal graph model and the strategy network. In one implementation, one or more external variables with intervenable are first selected, such as “holiday status” or “logistics interaction intensity”, denoted as Zk . The variable Z k is modified to construct counterfactual inputs The strategy model is then re-run:

[0160]

[0161] The original recommendation result a t is compared with the counterfactual result The ranking change is obtained:

[0162]

[0163] Wherein: represents the ranking position of the supplier i in the recommendation list a; is the position of the supplier in the counterfactual recommendation result; Δ rank (i) is the recommendation ranking change amount caused by the variable change.

[0164] As an option, the counterfactual deduction result can form a counterfactual influence path diagram to assist decision makers in identifying the sensitivity area of key variables and the stability of the recommendation result.

[0165] In some embodiments, the visualization generation unit also supports embedding the factor contribution degree and the counterfactual change curve into an interactive chart, and forming a natural language report containing recommendation reason explanation, factor influence summary, variable sensitivity analysis, etc.

[0166] To improve the readability and regulatory auditability of the output, the output result of this module supports presentation in a graph structure, including a “factor-indicator-recommendation” ternary causal path view, and a “variable-recommendation change” curve chart, etc.

[0167] In addition, in a specific implementation scheme, a causal inversion explanation mechanism can be introduced, that is, by reversely deducing whether a certain recommendation result is necessarily based on a specific factor combination, the qualitative audit of the recommendation rationality is realized.

[0168] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A big data based dynamic assessment and recommendation system for building material suppliers, characterized in that, The method comprises the following steps: a data acquisition and fusion module for acquiring and integrating structured data, unstructured data, context environment data and supply chain network relationship data of building material suppliers from multiple data sources to generate a unified formatted data set; a spatio-temporal tensor modeling and decomposition module for constructing a five-dimensional spatio-temporal tensor to represent the dynamic attributes of the suppliers based on the unified formatted data set, and decomposing the five-dimensional spatio-temporal tensor to generate factor matrices and dynamic evaluation weights; a causal analysis and confounding correction module for constructing a causal graph model of supplier evaluation based on the supply chain network relationship data, identifying and correcting the influence of confounding variables on the evaluation results in the causal graph model by using the factor matrices, and generating a causal correction performance distribution of the suppliers; a multi-agent recommendation decision module for combining the dynamic evaluation weights and the causal correction performance distribution, training a recommendation decision strategy by a reinforcement learning method, and outputting a supplier recommendation list; a recommendation result interpretability module for the supplier recommendation list, and generating an interpretability report containing factor traceability paths and counterfactual explanation paths based on the factor matrices and the causal graph model.

2. The dynamic assessment and recommendation system for building material suppliers based on big data according to claim 1, wherein, The data acquisition and fusion module comprises: a structured data acquisition unit for acquiring structured data of basic information, historical transaction data, performance time and quality detection records of suppliers from enterprise resource planning systems, project management systems and contract performance records; an unstructured data acquisition unit for acquiring text information related to suppliers from Internet public opinion platforms, social media, news websites and policy document databases, and performing natural language processing and sentiment feature extraction on the text information; a context environment data acquisition unit for acquiring time-sensitive external environment data affecting the operation of the supply chain from meteorological data platforms, public holiday service interfaces and market price index libraries; a supply chain network data acquisition unit for constructing a collaboration relationship graph between suppliers and construction units, logistics companies based on records of construction site systems, logistics tracking systems and material distribution platforms; a data fusion unit for preprocessing, feature alignment and semantic mapping of the structured data, unstructured data, context environment data and supply chain network data, and generating a unified formatted data set.

3. The dynamic assessment and recommendation system for building material suppliers based on big data according to claim 1, wherein, The spatio-temporal tensor modeling and decomposition module comprises: a tensor construction unit for constructing a five-dimensional tensor according to the dimensions of suppliers, time, static features, context and collaboration relationships; a factor decomposition unit for performing tensor decomposition operation on the five-dimensional tensor and extracting multiple latent factor matrices to represent the latent interaction features between dimensions; a weight generation unit for calculating the dynamic weights of each evaluation index at different time points in combination with time and context feature information.

4. The dynamic assessment and recommendation system for building material suppliers based on big data according to claim 3, wherein, Each dimension in the five-dimensional tensor of the tensor construction unit comprises: the first dimension is the supplier dimension for distinguishing between independent supplier entities; the second dimension is the time dimension for representing the evolution process of data in continuous time series; the third dimension is the static feature dimension for representing the basic invariable attributes of the suppliers; The fourth dimension is a context feature dimension, used to express external environmental factors associated with the supplier; The fifth dimension is a cooperation relationship dimension, used to describe the interaction between the supplier and other subjects in the supply chain network.

5. The dynamic assessment and recommendation system for building material suppliers based on big data according to claim 1, wherein, The causal analysis and confounding correction module comprises: A causal graph construction unit for constructing a causal graph model based on the static attributes of the supplier, the context features and the cooperation structure, wherein the nodes in the graph represent variables and the edges represent causal relationships; A confounding factor identification unit for identifying a set of confounding variables that affect the evaluation results of the supplier according to the backdoor path criterion; An intervention correction unit for intervention modeling based on the identification of confounding factors, generating the supplier evaluation performance distribution after controlling the influence of confounding factors.

6. The dynamic assessment and recommendation system for building material suppliers based on big data according to claim 5, wherein, The intervention modeling operation of the intervention correction unit comprises: Value setting operation on the direct parent nodes of the target variable in the causal graph to simulate the intervention effect; Based on the joint probability distribution of each confounding variable, the marginal distribution change of each evaluation index when the target variable is fixed is calculated; Output the target evaluation index distribution under the control of confounding factors for the recommendation model to call.

7. The dynamic assessment and recommendation system for building material suppliers based on big data according to claim 1, wherein, The multi-agent recommendation decision module comprises: A state modeling unit for constructing a current system state representation comprising a supplier tensor factor, inventory information and risk parameters; A reward function construction unit for constructing a reward function corresponding to the supplier recommendation action according to the dynamic weight of the evaluation index, measuring the strategy value of each recommendation combination; A reinforcement learning strategy unit for training a strategy model through a multi-agent reinforcement learning method to optimize the supplier selection action output under the current state; A recommendation output unit for outputting a supplier recommendation list and a ranking result according to the strategy model.

8. The dynamic assessment and recommendation system for building material suppliers based on big data according to claim 7, wherein, The training strategy model process of the reinforcement learning strategy unit comprises: Initialize the parameters of the policy network and the state value function network; Collect the state, action and reward sequence in the multi-round interaction process; Calculate the gradient update direction based on the policy gradient method and update the policy network parameters; Update the state value function using the time difference method.

9. The dynamic assessment and recommendation system for building material suppliers based on big data according to claim 1, wherein, The recommendation result interpretability module comprises: A factor path extraction unit for identifying the tensor factor that plays a dominant role in the current recommendation result and generating a factor traceability path; An counterfactual reasoning unit for constructing a change scenario of the causal variable and reevaluating the recommendation strategy result to generate a counterfactual explanation path; A visualization generation unit for generating an explanatory chart and report file containing factor source explanation and variable change impact path.

10. The dynamic assessment and recommendation system for building material suppliers based on big data according to claim 9, wherein, The counterfactual scenario construction of the counterfactual reasoning unit comprises: Select one or more external variables that can be intervened from the causal graph; Construct an intervention scenario input model for variable value change; Re-run the recommendation strategy model under the intervention input and obtain the recommendation ranking change amount; Output the recommendation strategy change path caused by variable change and impact analysis report.

Citation Information

Patent Citations

  • Commodity tracing method and system for cross-border logistics

    CN119359320A

  • Dynamic equipment quality evaluation method based on multi-agent collaborative reinforcement learning

    CN119671383A

  • Method for intelligently generating exhaustion report of financial unfavorable assets

    CN120198232A

  • Chronic disease risk assessment and intervention strategy generation system based on data analysis

    CN120221116A

  • Enterprise big data mining method and system based on artificial intelligence

    CN120296158A

Cited By

  • Commodity association rule intelligent analysis method and system based on knowledge mining

    CN121480728A

  • Supply chain resource configuration optimization method and system based on multi-source heterogeneous data fusion

    CN121684224A