Multidimensional scheduling method and system for library literature resources fused with circulation big data

Abstract

The application relates to the technical field of data processing and resource optimization scheduling. A library literature resource multi-dimensional scheduling method and system fusing circulation big data are provided, the method comprises the following steps: constructing a causal diagram based on multi-dimensional circulation data and field knowledge constraints to obtain a dynamic directed acyclic graph; adjusting hybrid factors of the dynamic directed acyclic graph and a target scheduling intervention strategy to generate unbiased causal effect estimation; generating a heterogeneous strategy for the unbiased causal effect estimation to obtain a literature-user dual-granularity scheduling strategy set; and dynamically optimizing the literature-user dual-granularity scheduling strategy set to obtain an optimized scheduling strategy set, so that the technical effects of reducing scheduling bias, improving literature circulation balance and enhancing multi-target coordination capability are achieved.

Description

Technical Field

[0001] This invention relates to the field of data processing and resource optimization scheduling technology, and in particular to a multi-dimensional scheduling method and system for library document resources that integrates circulation big data. Background Technology

[0002] With the deepening of digital transformation in libraries, the multi-dimensional scheduling of document resources has become a core element in improving service efficiency and resource utilization. Especially in scenarios involving multi-branch collaboration and diversified user needs, how to achieve precise scheduling of documents based on circulation big data is a key technological challenge in the construction of smart libraries.

[0003] In traditional technologies, scheduling models based on association rules do not consider the confounding effects of spatiotemporal factors and user attributes, which can lead to distorted estimation of policy causal effects and thus cause scheduling bias. Collaborative filtering algorithms can only generate single-granularity scheduling strategies, ignoring the heterogeneity of document characteristics and user needs, which can easily cause imbalances in document circulation and allocation. In addition, schemes using weighted objective programming are difficult to balance the multi-objective conflicts between document protection, user satisfaction and system load, and often require sacrificing one of the objectives. Summary of the Invention

[0004] Therefore, it is necessary to provide a multi-dimensional scheduling method and system for library document resources that integrates circulation big data to address the above-mentioned technical problems, so as to achieve the technical effects of reducing scheduling deviations, improving the balance of document circulation, and enhancing multi-objective collaborative capabilities.

[0005] Firstly, this application provides a multi-dimensional scheduling method for library document resources that integrates circulation big data, the method comprising:

[0006] A causal graph is constructed based on multi-dimensional circulation data and domain knowledge constraints, resulting in a dynamic directed acyclic graph.

[0007] By adjusting the confounding factors for dynamic directed acyclic graphs and target scheduling intervention strategies, unbiased causal effect estimates are generated.

[0008] Heterogeneous policy generation is performed on unbiased causal effect estimation to obtain a literature-user dual-granularity scheduling policy set;

[0009] Dynamically optimize the literature-user dual-granularity scheduling strategy set to obtain an optimized scheduling strategy set.

[0010] In one embodiment, confounding factor adjustment is applied to the dynamic directed acyclic graph and the target scheduling intervention strategy to generate an unbiased causal effect estimate, including:

[0011] Perform backdoor path blocking analysis on dynamic directed acyclic graphs to generate a set of confounding factors;

[0012] Hierarchical matching of the confounding factor set and the target scheduling intervention strategy yields the adjusted sample group;

[0013] Multiple machine learning estimations are performed on the adjusted sample group to generate unbiased causal effect estimates.

[0014] In one embodiment, hierarchical matching is performed on the confounding factor set and the target scheduling intervention strategy to obtain an adjusted sample group, including:

[0015] The set of confounding factors is divided into multidimensional hierarchical layers to obtain a spatiotemporally homogeneous layer. The set expression of the spatiotemporally homogeneous layer is as follows:

[0016]

[0017] in, Represents a set of spatiotemporally homogeneous layers. Indicates the first A homogeneous layer, Indicates sample The confounding factor eigenvectors Presentation layer Cluster centers Represents Mahalanobis distance, Represents the covariance matrix. Indicates the intra-layer difference threshold. Indicates an indicator function, Represents the set of confounding factors. Indicates the number of homogeneous layers. Indicates the spatiotemporal homogeneous layer number, Indicates the candidate cluster center number. Indicates the sample index. Indicates the first Candidate cluster centers of the layer;

[0018] Within each spatiotemporally homogeneous layer, the target scheduling intervention strategy is matched with a preference score to generate balanced samples within the layer.

[0019] By aggregating the intralayer equilibrium samples of all spatiotemporally homogeneous layers, an adjusted sample group is obtained.

[0020] In one embodiment, backdoor path blocking analysis is performed on a dynamic directed acyclic graph to generate a set of confounding factors, including:

[0021] Backdoor path identification is performed on a dynamic directed acyclic graph to obtain a set of candidate paths;

[0022] Perform a conditional independence test on each path in the candidate path set to generate a blocking node set. The expression for the blocking node set is:

[0023]

[0024] in, Represents the set of blocking nodes. Represents the set of candidate paths. This represents the function for testing conditional independence. Indicates the significance threshold. Indicator of node hybridity intensity Represents the complete set of nodes in a dynamic directed acyclic graph. Indicates candidate paths, Represents a condition set. Indicates the starting point of the path. Indicates the endpoint of the path. This represents a candidate node in a dynamic directed acyclic graph.

[0025] By filtering the set of blocking nodes, nodes that simultaneously affect both the intervention variable and the outcome variable are obtained, resulting in a set of confounding factors.

[0026] In one embodiment, heterogeneous policy generation is performed on the unbiased causal effect estimation to obtain a literature-user dual-granularity scheduling policy set, including:

[0027] Feature interaction enhancement is performed on the unbiased causal effect estimate to generate a heterogeneous effect tensor. The expression for the heterogeneous effect tensor is as follows:

[0028]

[0029] in, Represents the heterogeneity effect tensor. Documents For users Unbiased causal effect Represents the feature interaction function, Represents the feature vector of a document. Represents the user feature vector. Represents the environmental feature vector. Represents the regularization coefficient. Represents the regularization term;

[0030] Multi-granularity clustering analysis was performed on the heterogeneous effect tensor to obtain the document cluster strategy prototype and the user cluster strategy prototype.

[0031] The document cluster strategy prototype and the user cluster strategy prototype are fused and optimized to generate a document-user dual-granularity scheduling strategy set.

[0032] In one embodiment, multi-granularity clustering analysis is performed on the heterogeneity effect tensor to obtain document cluster strategy prototypes and user cluster strategy prototypes, including:

[0033] By slicing the heterogeneous effects tensor along the literature dimension, a literature-environment response matrix is ​​obtained.

[0034] Multi-scale spectral clustering of the literature-environment response matrix is ​​performed to obtain a prototype of the literature cluster strategy;

[0035] Slice the heterogeneous effects tensor along the user dimension to generate a user-environment response matrix;

[0036] Multi-scale spectral clustering of the user-environment response matrix is ​​performed to obtain a prototype of the user cluster strategy.

[0037] In one embodiment, the document-user dual-granularity scheduling strategy set is dynamically optimized to obtain an optimized scheduling strategy set, including:

[0038] A digital twin environment simulation was performed on the document-user dual-granularity scheduling strategy set to obtain a strategy execution effect map;

[0039] Detect drift in real-time environmental variables and generate policy update trigger signals;

[0040] In response to the policy update trigger signal, multi-objective reinforcement learning is used to optimize the policy execution effect graph, generating an optimized scheduling policy set.

[0041] Secondly, this application also provides a multi-dimensional scheduling system for library document resources that integrates circulation big data, the system including:

[0042] The causal graph construction module is used to construct causal graphs based on multi-dimensional flow data and domain knowledge constraints, resulting in a dynamic directed acyclic graph.

[0043] The confounding adjustment module is used to adjust the confounding factors of dynamic directed acyclic graphs and target scheduling intervention strategies to generate unbiased causal effect estimates.

[0044] The dual-granularity generation module is used to generate heterogeneous strategies for unbiased causal effect estimation, resulting in a literature-user dual-granularity scheduling strategy set.

[0045] The dynamic optimization module is used to dynamically optimize the document-user dual-granularity scheduling strategy set to obtain an optimized scheduling strategy set.

[0046] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods in the first aspect of this application.

[0047] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods in the first aspect of this application.

[0048] The method and system for multidimensional scheduling of library document resources that integrates circulation big data provided in this application include: constructing a dynamic directed acyclic graph based on multidimensional circulation data and domain knowledge constraints, which can present the causal relationship between documents, users and environment; adjusting confounding factors for the dynamic directed acyclic graph and target scheduling intervention strategy on this basis, which can effectively reduce the interference of confounding factors on the estimation of causal effects, thereby generating an unbiased causal effect estimate that is more in line with the actual relationship, and reducing the scheduling deviation caused by confounding effects from the root.

[0049] Heterogeneity strategy generation based on unbiased causal effect estimation can fully adapt to the differentiated characteristics of document features and user needs, and construct a document-user dual-granularity scheduling strategy set that takes into account different dimensions of needs. Then, dynamic optimization of this dual-granularity scheduling strategy set can better coordinate multiple objectives such as document protection, user satisfaction and system load in real-time environmental changes, and achieve the technical effect of improving the balance of document circulation and enhancing multi-objective collaborative capabilities. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A flowchart of a multi-dimensional scheduling method for library document resources integrating circulation big data in one embodiment of the present invention;

[0052] Figure 2 The flowchart shows a process for dynamically optimizing a document-user dual-granularity scheduling strategy set in one embodiment of the present invention.

[0053] Figure 3 This is a structural diagram of a multi-dimensional scheduling system for library document resources that integrates circulation big data, according to one embodiment of the present invention. Detailed Implementation

[0054] To make the above-mentioned objects, features, and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0055] First, the application scenarios of the embodiments of this application are described. In the embodiments of this application, a multi-dimensional scheduling method and system for library document resources that integrates circulation big data is provided, applicable to but not limited to scenarios involving collaborative operation of multiple branch libraries and diversified user borrowing needs.

[0056] In illustrative terms, the multi-dimensional scheduling method and system for library document resources that integrates circulation big data provided in this application embodiment can also be applied to other application scenarios that require intelligent scheduling of resources based on multi-source data, such as cross-university document allocation scenarios of university library alliances, regional resource sharing service scenarios of public libraries, and targeted allocation scenarios of characteristic documents of professional libraries. This is only an example and does not limit the specific application scenarios.

[0057] like Figure 1 As shown, this application provides a multi-dimensional scheduling method for library document resources that integrates circulation big data. The method includes:

[0058] S101: Construct a causal graph based on multi-dimensional circulation data and domain knowledge constraints to obtain a dynamic directed acyclic graph.

[0059] For example, the scheduling terminal collects multi-dimensional circulation data and obtains domain knowledge constraints. The multi-dimensional circulation data includes document borrowing records, user behavior data, library building distribution information, resource inventory status and circulation time information. The domain knowledge constraints include library document management professional rules, resource scheduling industry standards and document circulation practice experience.

[0060] The scheduling terminal preprocesses the collected multi-dimensional circulation data, ensuring the integrity and standardization of the data through data cleaning, missing data completion, and data format standardization. It deeply integrates the preprocessed circulation data with domain knowledge constraints, sorts out the causal relationships between literature resources, user needs, environmental factors, and scheduling behavior, and clarifies the influence direction and correlation strength of each element.

[0061] The scheduling terminal builds an initial causal graph based on the causal relationships formed by sorting out the causal relationships. It then performs logical verification and optimization on the initial causal graph, eliminates conflicting causal links, supplements missing reasonable relationships, and corrects logical contradictions. After completing the construction and optimization of the causal graph, a dynamic directed acyclic graph is obtained.

[0062] S102: Adjust the confounding factors for the dynamic directed acyclic graph and the target scheduling intervention strategy to generate an unbiased causal effect estimate.

[0063] For example, the scheduling terminal acquires a dynamic directed acyclic graph and a target scheduling intervention strategy, performs path analysis on the dynamic directed acyclic graph, identifies potential correlation paths that may interfere with the causal effect judgment, clarifies key nodes in the path, and filters out key nodes that simultaneously affect the target scheduling intervention strategy and literature scheduling result variables, thus obtaining a set of confounding factors.

[0064] The scheduling terminal correlates the set of confounding factors with the target scheduling intervention strategy, constructs a homogeneous analysis sample group through hierarchical partitioning, balances the distribution differences of confounding factors in different samples to eliminate interference, and uses a scientific causal effect estimation method to perform calculation and analysis based on the processed analysis sample group, fully removes the influence of confounding factors, and generates an unbiased causal effect estimate.

[0065] S103: Heterogeneous policy generation is performed on the unbiased causal effect estimation to obtain a literature-user dual-granularity scheduling policy set.

[0066] For example, the scheduling terminal performs feature interaction enhancement processing on the unbiased causal effect estimation, integrates document features, user features and environmental features, mines heterogeneous response relationships under different feature combinations, performs multi-dimensional cluster analysis on the resulting heterogeneous response results, divides feature clusters according to document attributes and user attributes respectively, and generates corresponding document cluster strategy prototypes and user cluster strategy prototypes.

[0067] The scheduling terminal performs policy fusion optimization on the document cluster policy prototype and the user cluster policy prototype, matches the resource scheduling requirements and adaptation rules of different clusters, and obtains a document-user dual-granularity scheduling policy set.

[0068] S104: Dynamically optimize the literature - user dual-granularity scheduling strategy set to obtain the optimized scheduling strategy set.

[0069] For example, the scheduling terminal places the document-user dual-granularity scheduling strategy set into the digital twin environment for simulation execution, generates a strategy execution effect map, performs drift detection on real-time environmental variables, identifies changes in environmental state and operating parameters, and generates a strategy update trigger signal.

[0070] The scheduling terminal responds to the policy update trigger signal, and optimizes the policy execution effect graph with document protection, user satisfaction and system load as optimization guides. It then performs multi-objective reinforcement learning optimization on the policy execution effect graph, adjusts the scheduling logic and configuration parameters of the document-user dual-granularity scheduling policy set, and obtains the optimized scheduling policy set.

[0071] An embodiment of this application provides a multi-dimensional scheduling method for library document resources that integrates circulation big data. The method includes: constructing a dynamic directed acyclic graph based on multi-dimensional circulation data and domain knowledge constraints, which can present the causal relationship between documents, users and the environment; adjusting confounding factors for the dynamic directed acyclic graph and the target scheduling intervention strategy, which can effectively reduce the interference of confounding factors on the estimation of causal effects, thereby generating an unbiased causal effect estimate that is more in line with the actual relationship, and reducing the scheduling deviation caused by confounding effects from the root.

[0072] Heterogeneity strategy generation based on unbiased causal effect estimation can fully adapt to the differentiated characteristics of document features and user needs, and construct a document-user dual-granularity scheduling strategy set that takes into account different dimensions of needs. Then, dynamic optimization of this dual-granularity scheduling strategy set can better coordinate multiple objectives such as document protection, user satisfaction and system load in real-time environmental changes, and achieve the technical effect of improving the balance of document circulation and enhancing multi-objective collaborative capabilities.

[0073] In one embodiment, confounding factor adjustment is applied to the dynamic directed acyclic graph and the target scheduling intervention strategy to generate an unbiased causal effect estimate, including:

[0074] (1) Perform backdoor path blocking analysis on the dynamic directed acyclic graph to generate a set of confounding factors.

[0075] For example, the scheduling terminal performs a full-link traversal scan of the dynamic directed acyclic graph, sorts out the node relationships node by node, extracts all potential related paths, and filters out non-causal related paths connecting the target scheduling intervention strategy node and the literature scheduling result variable node from the potential related paths. It clarifies the starting point, passing nodes and the ending point of each backdoor path, performs a combined traversal of the passing nodes of each backdoor path, performs a conditional independence test on each group of nodes, calculates the significance level of the association and determines whether the non-causal related path can be cut off, and filters out the key nodes that can block each backdoor path based on the test results, and summarizes and integrates them to obtain the blocking node set.

[0076] The scheduling terminal filters nodes from the blocking node set that have a direct or indirect impact on both the target scheduling intervention strategy and the literature scheduling result variables. The selected nodes are then subject to attribute verification and association confirmation to ensure that the nodes have an impact on both variables. The nodes that pass the verification are then integrated to obtain the mixed factor set.

[0077] Among them, the potential association path is the directed association link formed between any two nodes in the dynamic directed acyclic graph, covering causal and non-causal association paths; the backdoor path is the non-causal association path between the intervention variable and the outcome variable, which is the core path that produces confounding effects; the conditional independence test is a statistical analysis method to determine whether there is an independent association between the node combination and the target variable; the blocking node set is the set of key nodes that can cut off all backdoor paths; and the confounding factor set is the set of nodes that simultaneously affect the intervention variable and the outcome variable, which is the core factor that leads to the distortion of causal effect estimation.

[0078] (2) Perform hierarchical matching of the mixed factor set and the target scheduling intervention strategy to obtain the adjusted sample group.

[0079] For example, the scheduling terminal performs multi-dimensional feature decomposition on the set of confounding factors, clarifies the core feature attributes, level of action and scope of influence of each confounding factor, and sorts out a complete feature system of confounding factors. Based on this system, it formulates hierarchical division rules, performs multi-dimensional hierarchical division on the overall data containing samples related to the target scheduling intervention strategy, and groups samples with consistent feature attributes into the same category to obtain multiple homogeneous layers. Through feature verification within each layer, it ensures that the samples in each homogeneous layer maintain a high degree of consistency in the features of confounding factors.

[0080] Within each homogeneous layer, the scheduling terminal calculates the propensity score of the target scheduling intervention strategy based on the confounding factor characteristics (this score represents the probability of a sample accepting a specific intervention strategy). Based on the propensity score, matching rules are used to match samples corresponding to different intervention strategies, so that the samples corresponding to each intervention strategy achieve a distribution balance in terms of confounding factor characteristics, eliminating the sample difference interference caused by confounding factors within the layer. Then, all balanced samples that have completed matching within the homogeneous layers are summarized and integrated. Duplicate matching samples and invalid samples are removed through sample validity verification to obtain the adjustment sample group.

[0081] Among them, the confounding factor characteristic system is a complete sorting of the core characteristic attributes, hierarchical level and scope of influence of confounding factors, and is the core basis for stratification; the homogeneous layer is a sample set with consistent characteristic attributes, used to control the heterogeneity of confounding factors; the propensity score is the probability value of a sample accepting a specific intervention strategy, and is the core indicator for sample matching; the adjusted sample group is the sample set after eliminating the differences in the distribution of confounding factors, providing an unbiased analytical basis for subsequent causal effect estimation.

[0082] (3) Perform multiple machine learning estimation on the adjusted sample group to generate unbiased causal effect estimates.

[0083] For example, the scheduling terminal selects multiple machine learning algorithms with different principles and applicable scenarios as basic estimation models based on the feature types of the confounding factor set and the sample structure of the adjustment sample group. It clarifies the core functions and applicable boundaries of each model, determines the training parameters, iteration rules and validation criteria, and constructs a complete multi-estimation model system. The adjustment sample group is divided into training sample subsets and validation sample subsets according to a preset ratio. The training sample subsets are input into each basic estimation model for training. By continuously iterating and adjusting the internal parameters to optimize the fitting effect, each model gradually fits the correlation between the target scheduling intervention strategy and the literature scheduling result variables. The validation sample subsets are input into each trained basic estimation model to obtain the preliminary estimation results of the causal effect output by each model.

[0084] The scheduling terminal performs cross-validation and consistency checks on the preliminary estimation results of each basic estimation model, sets a reasonable error threshold, calculates the error value of each estimation result, removes model results that exceed the threshold, and then uses a weighted fusion method to integrate and calculate the effective preliminary estimation results. It also assigns weights based on the fitting accuracy of each model to generate the final causal effect estimate that has eliminated the interference of confounding factors, thus obtaining the unbiased causal effect estimate.

[0085] Among them, the basic estimation model consists of various machine learning algorithm models based on different principles, used to fit the association between intervention variables and outcome variables; cross-validation and consistency verification are analytical methods to verify the reliability of model estimation results, used to eliminate model results with excessive errors; unbiased causal effect estimation is an accurate estimation result of the true causal effect between the target scheduling intervention strategy and the literature scheduling outcome variables after eliminating the interference of confounding factors, which can truly reflect the actual impact of the intervention strategy.

[0086] In one embodiment, hierarchical matching is performed on the confounding factor set and the target scheduling intervention strategy to obtain an adjusted sample group, including:

[0087] (1) The set of mixed factors is divided into multidimensional hierarchical layers to obtain a spatiotemporal homogeneous layer. The set expression of the spatiotemporal homogeneous layer is:

[0088]

[0089] in, Represents a set of spatiotemporally homogeneous layers. Indicates the first A homogeneous layer, Indicates sample The confounding factor eigenvectors Presentation layer Cluster centers Represents Mahalanobis distance, Represents the covariance matrix. Indicates the intra-layer difference threshold. Indicates an indicator function, Represents the set of confounding factors. Indicates the number of homogeneous layers. Indicates the spatiotemporal homogeneous layer number, Indicates the candidate cluster center number. Indicates the sample index. Indicates the first Candidate cluster centers for the layer.

[0090] For example, the scheduling terminal performs multi-dimensional hierarchical partitioning of the confounding factor set, extracts the confounding factor feature vectors of each sample in the confounding factor set, calculates the Mahalanobis distance between the sample confounding factor feature vectors and the candidate cluster centers based on the covariance matrix, determines whether the Mahalanobis distance exceeds the intra-layer difference threshold by using an indicator function, determines the optimal cluster center with the goal of minimizing the number of samples exceeding the threshold, and assigns each sample to the category to which the corresponding cluster center belongs, forming multiple spatiotemporally homogeneous layers, thus obtaining a set of spatiotemporally homogeneous layers.

[0091] Among them, the spatiotemporal homogeneous layer set is the summation set of all spatiotemporal homogeneous layers, the spatiotemporal homogeneous layer is the sample set whose confounding factor feature differences are within the intra-layer difference threshold, the confounding factor feature vector is the vector form data representing the confounding factor features of the sample, and the candidate cluster center is the cluster center candidate object to be verified.

[0092] Mahalanobis distance is a distance metric that measures the difference between a sample and its cluster center. The covariance matrix is ​​the characteristic covariance data used to calculate Mahalanobis distance. The intra-layer difference threshold is the critical value for determining whether a sample belongs to the same homogeneous layer. The indicator function is a logical function for determining whether the difference exceeds the threshold. The confounding factor set is the set of nodes that simultaneously affect the intervention variable and the outcome variable.

[0093] (2) Within each spatiotemporally homogeneous layer, the target scheduling intervention strategy is matched with the tendency score to generate balanced samples within the layer.

[0094] For example, within each spatiotemporally homogeneous layer, the scheduling terminal calculates the tendency score of each sample to accept the target scheduling intervention strategy based on the confounding factor characteristics of the samples within the spatiotemporally homogeneous layer. Based on the tendency score, the terminal matches the samples that accept different intervention strategies, so that the matched samples maintain a balance in the distribution of confounding factor characteristics, eliminate the sample distribution differences caused by confounding factors within the homogeneous layer, and generate balanced samples within the layer.

[0095] Among them, the propensity score is a numerical value representing the probability of a sample accepting a specific target scheduling intervention strategy, the target scheduling intervention strategy is the literature resource scheduling intervention scheme to be evaluated, and the intra-layer balanced sample is a sample set in which the distribution of mixed factor characteristics is balanced within the same spatiotemporal homogeneous layer.

[0096] (3) Aggregate the intralayer equilibrium samples of all spatiotemporally homogeneous layers to obtain the adjusted sample group.

[0097] For example, the scheduling terminal traverses all spatiotemporally homogeneous layers, extracts the intra-layer balance samples generated in each spatiotemporally homogeneous layer, performs duplicate sample removal and validity verification on the extracted intra-layer balance samples to ensure the uniqueness and usability of the samples, and integrates and summarizes all the intra-layer balance samples that pass the verification to obtain the adjustment sample group.

[0098] The adjusted sample group is a collection of effective intra-layer equilibrium samples within all spatiotemporally homogeneous layers, which forms the basis for subsequent analysis to estimate unbiased causal effects.

[0099] In one embodiment, backdoor path blocking analysis is performed on a dynamic directed acyclic graph to generate a set of confounding factors, including:

[0100] (1) Perform backdoor path identification on dynamic directed acyclic graph to obtain a set of candidate paths.

[0101] For example, the scheduling terminal performs a full-link traversal of the dynamic directed acyclic graph, checks the association direction and function type of each node in the dynamic directed acyclic graph one by one, identifies non-causal association paths connecting intervention variable nodes and result variable nodes, and summarizes and integrates all identified non-causal association paths to obtain a candidate path set.

[0102] Among them, the dynamic directed acyclic graph is a directed acyclic graph structure that represents the causal relationship between literature, users and environment. The intervention variable node is a node representing the target scheduling intervention strategy, the outcome variable node is a node representing the literature scheduling result, the non-causal relationship path is the backdoor path between the intervention variable and the outcome variable, and the candidate path set is the summary set of all identified backdoor paths.

[0103] (2) Perform conditional independence tests on each path in the candidate path set to generate a blocking node set. The expression for the blocking node set is:

[0104]

[0105] in, Represents the set of blocking nodes. Represents the set of candidate paths. This represents the function for testing conditional independence. Indicates the significance threshold. Indicator of node hybridity intensity Represents the complete set of nodes in a dynamic directed acyclic graph. Indicates candidate paths, Represents a condition set. Indicates the starting point of the path. Indicates the endpoint of the path. This represents a candidate node in a dynamic directed acyclic graph.

[0106] For example, the scheduling terminal traverses each candidate path in the candidate path set. For the starting point and ending point of each candidate path, different condition sets are selected from the complete set of nodes in the dynamic directed acyclic graph. Condition independence is tested, and the degree of matching between the test results and the significance threshold is calculated. The smallest condition set that makes the starting point and ending point of the path meet the significance requirement is selected. From this condition set, candidate nodes that meet the node mixing intensity index are further selected. All candidate nodes that meet the index are merged and integrated to generate a blocking node set.

[0107] Here, a candidate path is a single non-causal path in the candidate path set, the path start point is the starting node of the candidate path, the path end point is the ending node of the candidate path, the condition set is a subset of nodes used to conduct conditional independence tests, and the conditional independence test function is an analytical function that determines whether the path start point and the path end point are independent under a given condition set.

[0108] The significance threshold is the critical value for determining whether the starting point and ending point of a path satisfy the independence condition. The node confounding intensity index is a numerical value that characterizes the intensity of node confounding. The total set of nodes in a dynamic directed acyclic graph is the sum of all nodes in the graph. The blocking node set is the sum of candidate nodes that can block all candidate paths.

[0109] (3) Filter the nodes that simultaneously affect the intervention variable and the outcome variable from the blocking node set to obtain the confounding factor set.

[0110] For example, the scheduling terminal traverses each candidate node in the blocking node set, verifies the association link between the candidate node and the intervention variable and the outcome variable one by one, determines whether the candidate node has a direct or indirect impact on both the intervention variable and the outcome variable, retains the candidate nodes that have both effects, and summarizes and integrates the retained candidate nodes to obtain the confounding factor set.

[0111] Among them, the candidate nodes in the blocking node set are node objects in the dynamic directed acyclic graph that can block backdoor paths, the intervention variable is the variable representing the target scheduling intervention strategy, the outcome variable is the variable representing the literature scheduling result, and the confounding factor set is the set of nodes that simultaneously affect the intervention variable and the outcome variable.

[0112] In one embodiment, heterogeneous policy generation is performed on the unbiased causal effect estimation to obtain a literature-user dual-granularity scheduling policy set, including:

[0113] (1) Feature interaction enhancement is performed on the unbiased causal effect estimate to generate a heterogeneity effect tensor. The expression of the heterogeneity effect tensor is:

[0114]

[0115] in, Represents the heterogeneity effect tensor. Documents For users Unbiased causal effect Represents the feature interaction function, Represents the feature vector of a document. Represents the user feature vector. Represents the environmental feature vector. Represents the regularization coefficient. This represents the regularization term.

[0116] For example, the scheduling terminal extracts the unbiased causal effect of the corresponding literature on the user in the unbiased causal effect estimation, and at the same time obtains the literature feature vector, user feature vector and environment feature vector. The literature feature vector, user feature vector and environment feature vector are fused through the feature interaction function to obtain the feature interaction result. The result is then used to perform tensor operation with the unbiased causal effect of the literature on the user. A regularization term is then introduced and the constraint strength is controlled by the regularization coefficient to generate the heterogeneity effect tensor.

[0117] Among them, the heterogeneity effect tensor is tensor data that characterizes the differences in causal effects under the multi-dimensional interaction of literature, users and environment features; the unbiased causal effect is the numerical value of the true causal effect of literature on users after eliminating the interference of confounding factors; and the feature interaction function is a computational function that integrates the features of literature, users and environment.

[0118] The document feature vector is a vector data representing the attribute characteristics of the document; the user feature vector is a vector data representing the user demand characteristics; the environment feature vector is a vector data representing the scheduling environment characteristics; the regularization coefficient is a coefficient that controls the strength of the regularization constraint; and the regularization term is a constraint term used to avoid model overfitting.

[0119] (2) Perform multi-granularity clustering analysis on the heterogeneity effect tensor to obtain the document cluster strategy prototype and the user cluster strategy prototype.

[0120] For example, the scheduling terminal performs multi-granularity traversal analysis on the heterogeneity effect tensor, sets clustering rules according to the document dimension and the user dimension respectively, calculates the feature similarity between each element in the heterogeneity effect tensor, and groups the elements whose similarity meets the preset standard into the same cluster. For each document dimension cluster, the corresponding scheduling strategy logic is extracted to obtain the document cluster strategy prototype, and for each user dimension cluster, the corresponding scheduling strategy logic is extracted to obtain the user cluster strategy prototype.

[0121] Among them, the heterogeneity effect tensor is tensor data that characterizes the differences in causal effects under the interaction of multi-dimensional features of documents, users and environment; multi-granularity clustering analysis is a clustering analysis operation carried out simultaneously from the document dimension and the user dimension; the document cluster strategy prototype is a preliminary scheduling strategy for the corresponding document feature clusters; and the user cluster strategy prototype is a preliminary scheduling strategy for the corresponding user feature clusters.

[0122] (3) Perform policy fusion optimization on the document cluster policy prototype and the user cluster policy prototype to generate a document-user dual-granularity scheduling policy set.

[0123] For example, the scheduling terminal matches the document cluster strategy prototype and the user cluster strategy prototype one by one, verifies the compatibility of scheduling rules between different strategy prototypes, eliminates conflicting scheduling rule entries, adjusts the execution priority and scope of application of compatible rules, integrates and optimizes the compatible rules to form a complete scheduling scheme, and generates a document-user dual-granularity scheduling strategy set.

[0124] Among them, the document cluster strategy prototype is the initial scheduling strategy for the corresponding document feature cluster, the user cluster strategy prototype is the initial scheduling strategy for the corresponding user feature cluster, the strategy fusion optimization is the operation of integrating strategy prototypes of different granularities and eliminating conflicts, and the document-user dual-granularity scheduling strategy set is a complete scheduling strategy set that takes into account both document granularity and user granularity requirements.

[0125] In one embodiment, multi-granularity clustering analysis is performed on the heterogeneity effect tensor to obtain document cluster strategy prototypes and user cluster strategy prototypes, including:

[0126] (1) Slice the heterogeneity effect tensor according to the literature dimension to obtain the literature-environment response matrix.

[0127] For example, the scheduling terminal traverses the document dimension index of the heterogeneity effect tensor. For each document dimension index, it locates all tensor elements in the heterogeneity effect tensor associated with that index, removes the core features of the user dimension that are irrelevant to the document dimension, retains the environmental dimension features and the interactive response information between the document and the environment, arranges the retained elements in an orderly manner according to the correspondence between the document index and the environmental dimension, constructs a two-dimensional structured data form, and obtains the document-environment response matrix.

[0128] Among them, the heterogeneity effect tensor is tensor data that characterizes the differences in causal effects under the multi-dimensional interaction of literature, users and environment. The literature dimension slice is the operation of splitting the heterogeneity effect tensor along the literature dimension while retaining the environmental dimension features. The literature-environment response matrix is ​​two-dimensional matrix data that reflects the differences in causal responses of different literatures under various environmental scenarios.

[0129] (2) Perform multi-scale spectral clustering on the literature-environment response matrix to obtain the literature cluster strategy prototype.

[0130] For example, the scheduling terminal constructs a multi-scale similarity matrix based on the document-environment response matrix, sets multiple neighborhood parameters at different scales, calculates the feature similarity weights between elements in the document-environment response matrix for each set of neighborhood parameters, and integrates the similarity weights at different scales to obtain the final multi-scale similarity matrix.

[0131] The scheduling terminal performs spectral decomposition on the multi-scale similarity matrix to extract the core feature vectors of the matrix. Based on the distribution density of the core feature vectors, the document elements are clustered and classified. Document elements with high feature similarity are grouped into the same cluster. For the feature distribution of each cluster, the scheduling rules and resource allocation logic are extracted to obtain the document cluster strategy prototype.

[0132] Among them, the literature-environment response matrix is ​​a two-dimensional matrix data reflecting the differences in causal responses of different literatures under various environmental scenarios, multi-scale spectral clustering is a clustering analysis method that combines multi-scale neighborhood parameters and spectral decomposition ideas, and the literature cluster strategy prototype is a preliminary scheduling strategy for corresponding literature feature clusters.

[0133] (3) Slice the heterogeneity effect tensor according to the user dimension to generate the user-environment response matrix.

[0134] For example, the scheduling terminal traverses the user dimension index of the heterogeneity effect tensor. For each user dimension index, it locates all tensor elements in the heterogeneity effect tensor associated with that index, removes the core features of the document dimension that are irrelevant to the user dimension, retains the environmental dimension features and the interactive response information between the user and the environment, arranges the retained elements in an orderly manner according to the correspondence between the user index and the environmental dimension, constructs a two-dimensional structured data form, and generates a user-environment response matrix.

[0135] Among them, the heterogeneity effect tensor is tensor data that characterizes the differences in causal effects under the multi-dimensional interaction of literature, users and environment; the user dimension slice is the operation of splitting the heterogeneity effect tensor along the user dimension while retaining the environmental dimension features; and the user-environment response matrix is ​​two-dimensional matrix data that reflects the differences in causal responses of different users in various environmental scenarios.

[0136] (4) Perform multi-scale spectral clustering on the user-environment response matrix to obtain the user cluster strategy prototype.

[0137] For example, the scheduling terminal constructs a multi-scale similarity matrix based on the user-environment response matrix, sets multiple neighborhood parameters at different scales, calculates the feature similarity weights between each element in the user-environment response matrix for each set of neighborhood parameters, and integrates the similarity weights at different scales to form the final multi-scale similarity matrix.

[0138] The scheduling terminal performs spectral decomposition on the multi-scale similarity matrix to extract the core feature vectors of the matrix. Based on the distribution density of the core feature vectors, user elements are clustered and grouped into the same cluster. For each cluster, the scheduling rules and service supply logic are extracted to meet the demand characteristics, thus obtaining the user cluster strategy prototype.

[0139] Among them, the user-environment response matrix is ​​a two-dimensional matrix data reflecting the differences in causal responses of different users in various environmental scenarios, multi-scale spectral clustering is a clustering analysis method that combines multi-scale neighborhood parameters and spectral decomposition ideas, and the user cluster strategy prototype is a preliminary scheduling strategy for corresponding user feature clusters.

[0140] like Figure 2 As shown, the dynamic optimization of the literature-user dual-granularity scheduling strategy set yields an optimized scheduling strategy set, including:

[0141] S201: Perform digital twin environment simulation on the document-user dual-granularity scheduling strategy set to obtain a strategy execution effect map.

[0142] For example, the scheduling terminal imports the document-user dual-granularity scheduling strategy set into the digital twin environment, and simultaneously constructs a virtual simulation scenario that is completely mapped to the real library scheduling scenario. It fully restores the core scenario elements such as the distribution status of document resources, user demand characteristics, physical constraints of library space, and system operation permission restrictions, and drives the document-user dual-granularity scheduling strategy set to be fully executed in the virtual simulation scenario according to the preset logic. Throughout the process, it synchronously records various operational indicators such as the real-time status of resource allocation, user service response feedback results, dynamic fluctuations of system load, and changes in document circulation efficiency.

[0143] The scheduling terminal integrates all kinds of operational indicator data recorded throughout the process in a structured manner according to the time series dimension and the scenario classification dimension, removes redundant and invalid data and fills in key missing data, and performs visualization rendering processing on the integrated data to present the effect differences and change trends of the entire strategy execution process in an intuitive graph form, thus obtaining a strategy execution effect graph.

[0144] Among them, the digital twin environment is a virtual simulation environment that fully maps to the real library scheduling scenario; the document-user dual-granularity scheduling strategy set is a set of scheduling strategies that takes into account both document granularity resource characteristics and user granularity demand characteristics; and the strategy execution effect graph is a visual graph data that integrates strategy execution process data and effect indicators.

[0145] S202: Detect drift in real-time environmental variables and generate a policy update trigger signal.

[0146] For example, the scheduling terminal synchronously collects various environmental variables in the real library scheduling scenario through the real-time data acquisition interface, including real-time changes in document inventory, dynamic changes in user demand preferences, adjustments to library space structure, system operating parameters, and load status data. The collected real-time environmental variables are then aligned with the preset benchmark environmental variables in the digital twin environment in terms of feature dimensions and distribution characteristics, and multi-dimensional difference analysis and trend judgment are performed.

[0147] Based on the results of difference analysis and trend judgment, the scheduling terminal detects the drift of real-time environmental variables relative to the benchmark environmental variables, such as distribution shift, feature mutation, and trend deviation. When the detected drift exceeds the system's preset environmental stability threshold, it determines that the environmental state of the current real library scheduling scenario has changed significantly and cannot be ignored, and then generates a strategy update trigger signal to start the strategy optimization process.

[0148] Among them, real-time environmental variables are dynamically changing environmental data in real library scheduling scenarios, baseline environmental variables are preset stable state environmental data in digital twin environments, drift detection is an analysis method to judge the degree of difference between real-time environmental variables and baseline environmental variables, and policy update trigger signal is an instruction signal used to trigger the policy optimization process.

[0149] S203: Respond to the policy update trigger signal, perform multi-objective reinforcement learning optimization on the policy execution effect graph, and generate an optimized scheduling policy set.

[0150] For example, after receiving the policy update trigger signal, the scheduling terminal clearly defines the core optimization objectives as the level of literature resource protection, user service satisfaction, and system operation load efficiency. It loads the historical policy execution data and effect index data stored in the policy execution effect graph and uses them as training samples for the multi-objective reinforcement learning model. It then constructs a multi-objective reinforcement learning optimization model adapted to the literature-user dual-granularity scheduling scenario and sets the model's iteration rules and optimization objective weight allocation standards.

[0151] The scheduling terminal continuously iterates and trains the model through multi-objective reinforcement learning optimization model, gradually adjusting the core strategy configurations such as scheduling rule logic, resource allocation ratio parameters, user service response priority, and document circulation scheduling sequence in the document-user dual-granularity scheduling strategy set. It dynamically balances the conflict relationship between various optimization objectives, making the adjusted scheduling strategy more in line with the real-time environment of the current library scheduling scenario, and generating an optimized scheduling strategy set.

[0152] Among them, the strategy update trigger signal is the instruction signal used to start the strategy optimization process; the strategy execution effect graph is a visualization graph data that integrates strategy execution process data and effect indicators; the multi-objective reinforcement learning optimization is a strategy iterative optimization method guided by multiple core business objectives; and the optimized scheduling strategy set is the final scheduling strategy set adapted to the current real-time environment state.

[0153] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0154] like Figure 3 As shown, this application also provides a multi-dimensional scheduling system 300 for library document resources that integrates circulation big data. This system 300 includes:

[0155] The causal graph construction module 301 is used to construct a causal graph based on multi-dimensional flow data and domain knowledge constraints, resulting in a dynamic directed acyclic graph.

[0156] The confounding adjustment module 302 is used to adjust the confounding factors of the dynamic directed acyclic graph and the target scheduling intervention strategy to generate an unbiased causal effect estimate.

[0157] The dual-granularity generation module 303 is used to generate heterogeneous strategies for unbiased causal effect estimation, resulting in a literature-user dual-granularity scheduling strategy set.

[0158] The dynamic optimization module 304 is used to dynamically optimize the document-user dual-granularity scheduling strategy set to obtain the optimized scheduling strategy set.

[0159] Specifically, the scheduling terminal includes a cause-effect graph construction module 301, a hybrid adjustment module 302, a dual-granularity generation module 303, and a dynamic optimization module 304.

[0160] The cause-effect graph construction module acquires multi-dimensional circulation data in the library field through a multi-dimensional data acquisition interface. It simultaneously introduces knowledge constraints from professional rules in library document management, industry standards for resource scheduling, practical experience in document circulation, and technical standards for smart library construction. The module performs standardized preprocessing on the acquired multi-dimensional circulation data, including data cleaning to remove abnormal data, data integration to complete missing data, and data format standardization to ensure the integrity, accuracy, and standardization of the multi-dimensional circulation data.

[0161] The causal graph construction module deeply integrates preprocessed multi-dimensional circulation data and domain knowledge constraints, sorts out the potential relationships between literature resource elements, user demand elements, environmental factor elements, and scheduling intervention elements, clarifies the direction of influence transmission and the strength of the relationship between each element, and gradually builds the initial causal graph according to the construction logic and standard framework of causal graph. The initial causal graph is then subjected to multi-dimensional logical verification to check the rationality and consistency of the association links, remove conflicting links, supplement missing associations, and adjust the direction of logical contradictions. After the construction and optimization are completed, a dynamic directed acyclic graph is obtained.

[0162] Among them, multi-dimensional circulation data refers to various types of basic data that reflect the entire process of library document circulation; domain knowledge constraints refer to the professional rules and norms system in the field of library document scheduling; causal graph is a graphical structure that represents the causal relationship of each element; and dynamic directed acyclic graph is a causal relationship graph structure with dynamic update characteristics and no closed loop.

[0163] The hybrid adjustment module acquires a dynamic directed acyclic graph and the target scheduling intervention strategy. It performs a full-link traversal scan of the dynamic directed acyclic graph to identify non-causal backdoor paths connecting the target scheduling intervention strategy nodes and the literature scheduling result variable nodes. It performs a combined traversal of the nodes along each backdoor path and performs conditional independence tests to screen out key nodes that can block backdoor paths and integrate them to obtain a blocking node set. From the blocking node set, it selects nodes that simultaneously affect the target scheduling intervention strategy and the literature scheduling result variable. After attribute verification and association confirmation, it obtains a hybrid factor set. Based on the hybrid factor set, it performs multi-dimensional hierarchical division to form a homogeneous layer. Within each homogeneous layer, it performs propensity score matching, aggregates balanced samples, and removes duplicate and invalid samples to obtain the adjustment sample group.

[0164] The confounding adjustment module selects multiple machine learning algorithms with different principles and applicable scenarios as basic estimation models based on the feature types of the confounding factor set and the sample structure of the adjustment sample group. It constructs a multi-estimation model system and clarifies the training parameters, iteration rules, and validation criteria. The adjustment sample group is divided into a training sample subset and a validation sample subset. The training sample subset is input for iterative training and parameter optimization to make the model fit the correlation between the target scheduling intervention strategy and the literature scheduling result variables. The validation sample subset is used to obtain the preliminary estimation results of the causal effect. After cross-validation, consistency check, and weighted fusion, an unbiased causal effect estimate that eliminates the interference of confounding factors is generated.

[0165] Among them, the target scheduling intervention strategy is the library document resource scheduling intervention scheme to be evaluated, the confounding factor set is the set of nodes that simultaneously affect the intervention variable and the outcome variable, the adjusted sample group is the sample set after eliminating the differences in the distribution of confounding factors, and the unbiased causal effect estimation is the accurate estimation result of the true causal effect after eliminating confounding interference.

[0166] The dual-granularity generation module extracts the unbiased causal effect of literature on users in unbiased causal effect estimation, obtaining literature feature vectors, user feature vectors, and environmental feature vectors. The feature interaction function fuses these three types of feature vectors to obtain the feature interaction result. Tensor operations are performed between the feature interaction result and the unbiased causal effect, introducing a regularization term and controlling the constraint strength through the regularization coefficient to generate a heterogeneous effect tensor. Subsequently, the heterogeneous effect tensor is sliced ​​along both the literature and user dimensions, retaining the corresponding environmental dimension features and interaction response information, constructing the literature-environment response matrix and the user-environment response matrix. Multi-scale spectral clustering is then used to obtain the document cluster strategy prototype and the user cluster strategy prototype, respectively.

[0167] The dual-granularity generation module matches the document cluster strategy prototype with the user cluster strategy prototype one by one, verifies the compatibility of scheduling rules between different strategy prototypes, eliminates conflicting scheduling rule entries, adjusts the execution priority and scope of application of compatible rules, and integrates the optimized compatible rules to form a complete scheduling scheme, thus obtaining a document-user dual-granularity scheduling strategy set.

[0168] Among them, the heterogeneity effect tensor is tensor data that characterizes the differences in causal effects under the multi-dimensional interaction of literature, users and environment features; the literature cluster strategy prototype is the prototype of the scheduling strategy for the corresponding literature feature clusters; the user cluster strategy prototype is the prototype of the scheduling strategy for the corresponding user feature clusters; and the literature-user dual-granularity scheduling strategy set is a complete set of scheduling strategies that takes into account both literature granularity and user granularity requirements.

[0169] The dynamic optimization module imports a dual-granularity scheduling strategy set for documents and users into a digital twin environment, constructing a virtual simulation scenario that fully maps to the real library scheduling scenario. It restores core scenario elements such as document resource distribution, user demand characteristics, library space constraints, and system operation limitations, driving the strategy set to execute completely in the virtual simulation scenario. It records operational data such as resource allocation status, user service response results, system load fluctuations, and document circulation efficiency throughout the process, and integrates and visualizes the data according to time series and scenario dimensions to obtain a strategy execution effect graph. At the same time, it collects real-time environmental variables in the real scenario, compares and analyzes their characteristics with benchmark environmental variables, detects environmental variable drift, and generates a strategy update trigger signal when the drift exceeds a threshold.

[0170] The dynamic optimization module responds to policy update trigger signals, with the core optimization objectives being the level of literature resource protection, user service satisfaction, and system operating load efficiency. It loads historical data from the policy execution effect graph as training samples, constructs a multi-objective reinforcement learning optimization model, and sets iteration rules and weight allocation standards. Through continuous iterative training, it adjusts the core configurations of the policy set, such as scheduling rule logic, resource allocation ratio parameters, user service response priority, and literature circulation scheduling sequence, to dynamically balance the conflicting relationships of various optimization objectives, making the policy adapt to the current real-time environment state and obtaining an optimized scheduling policy set.

[0171] Among them, the digital twin environment is a virtual simulation environment that is completely mapped to the real library scheduling scenario; the strategy execution effect graph is a visual graph data that integrates strategy execution process data and effect indicators; the strategy update trigger signal is an instruction signal used to start the strategy optimization process; and the optimized scheduling strategy set is the final scheduling strategy set adapted to the current real-time environment state.

[0172] Hybrid adjustment module 302 is also used for:

[0173] Perform backdoor path blocking analysis on dynamic directed acyclic graphs to generate a set of confounding factors;

[0174] Hierarchical matching of the confounding factor set and the target scheduling intervention strategy yields the adjusted sample group;

[0175] Multiple machine learning estimations are performed on the adjusted sample group to generate unbiased causal effect estimates.

[0176] Hybrid adjustment module 302 is also used for:

[0177] The set of confounding factors is divided into multidimensional hierarchical layers to obtain a spatiotemporally homogeneous layer. The set expression of the spatiotemporally homogeneous layer is as follows:

[0178]

[0179] in, Represents a set of spatiotemporally homogeneous layers. Indicates the first A homogeneous layer, Indicates sample The confounding factor eigenvectors Presentation layer Cluster centers Represents Mahalanobis distance, Represents the covariance matrix. Indicates the intra-layer difference threshold. Indicates an indicator function, Represents the set of confounding factors. Indicates the number of homogeneous layers. Indicates the spatiotemporal homogeneous layer number, Indicates the candidate cluster center number. Indicates the sample index. Indicates the first Candidate cluster centers of the layer;

[0180] Within each spatiotemporally homogeneous layer, the target scheduling intervention strategy is matched with a preference score to generate balanced samples within the layer.

[0181] By aggregating the intralayer equilibrium samples of all spatiotemporally homogeneous layers, an adjusted sample group is obtained.

[0182] Hybrid adjustment module 302 is also used for:

[0183] Backdoor path identification is performed on a dynamic directed acyclic graph to obtain a set of candidate paths;

[0184] Perform a conditional independence test on each path in the candidate path set to generate a blocking node set. The expression for the blocking node set is:

[0185]

[0186] in, Represents the set of blocking nodes. Represents the set of candidate paths. This represents the function for testing conditional independence. Indicates the significance threshold. Indicator of node hybridity intensity Represents the complete set of nodes in a dynamic directed acyclic graph. Indicates candidate paths, Represents a condition set. Indicates the starting point of the path. Indicates the endpoint of the path. This represents a candidate node in a dynamic directed acyclic graph.

[0187] By filtering the set of blocking nodes, nodes that simultaneously affect both the intervention variable and the outcome variable are obtained, resulting in a set of confounding factors.

[0188] The dual-granularity generation module 303 is also used for:

[0189] Feature interaction enhancement is performed on the unbiased causal effect estimate to generate a heterogeneous effect tensor. The expression for the heterogeneous effect tensor is as follows:

[0190]

[0191] in, Represents the heterogeneity effect tensor. Documents For users Unbiased causal effect Represents the feature interaction function, Represents the feature vector of a document. Represents the user feature vector. Represents the environmental feature vector. Represents the regularization coefficient. Represents the regularization term;

[0192] Multi-granularity clustering analysis was performed on the heterogeneous effect tensor to obtain the document cluster strategy prototype and the user cluster strategy prototype.

[0193] The document cluster strategy prototype and the user cluster strategy prototype are fused and optimized to generate a document-user dual-granularity scheduling strategy set.

[0194] The dual-granularity generation module 303 is also used for:

[0195] By slicing the heterogeneous effects tensor along the literature dimension, a literature-environment response matrix is ​​obtained.

[0196] Multi-scale spectral clustering of the literature-environment response matrix is ​​performed to obtain a prototype of the literature cluster strategy;

[0197] Slice the heterogeneous effects tensor along the user dimension to generate a user-environment response matrix;

[0198] Multi-scale spectral clustering of the user-environment response matrix is ​​performed to obtain a prototype of the user cluster strategy.

[0199] The dynamic optimization module 304 is also used for:

[0200] A digital twin environment simulation was performed on the document-user dual-granularity scheduling strategy set to obtain a strategy execution effect map;

[0201] Detect drift in real-time environmental variables and generate policy update trigger signals;

[0202] In response to the policy update trigger signal, multi-objective reinforcement learning is used to optimize the policy execution effect graph, generating an optimized scheduling policy set.

[0203] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0204] In one embodiment, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0205] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0206] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A multi-dimensional scheduling method for library document resources integrating circulation big data, characterized in that: The method includes: A causal graph is constructed based on multi-dimensional circulation data and domain knowledge constraints, resulting in a dynamic directed acyclic graph. The dynamic directed acyclic graph and the target scheduling intervention strategy are adjusted for confounding factors to generate unbiased causal effect estimates. Heterogeneous policy generation is performed on the unbiased causal effect estimation to obtain a literature-user dual-granularity scheduling policy set; The literature-user dual-granularity scheduling strategy set is dynamically optimized to obtain an optimized scheduling strategy set.

2. The multi-dimensional scheduling method for library document resources integrating circulation big data as described in claim 1, characterized in that, The step of adjusting the dynamic directed acyclic graph and the target scheduling intervention strategy for confounding factors to generate unbiased causal effect estimates includes: Perform backdoor path blocking analysis on the dynamic directed acyclic graph to generate a set of confounding factors; Hierarchical matching is performed on the set of confounding factors and the target scheduling intervention strategy to obtain the adjustment sample group; Multiple machine learning estimations are performed on the adjusted sample group to generate the unbiased causal effect estimate.

3. The multi-dimensional scheduling method for library document resources integrating circulation big data as described in claim 2, characterized in that, The step of performing hierarchical matching of the confounding factor set and the target scheduling intervention strategy to obtain the adjusted sample group includes: The set of heterogeneous factors is divided into multidimensional hierarchical layers to obtain a spatiotemporally homogeneous layer. The set expression of the spatiotemporally homogeneous layer is as follows: in, Represents a set of spatiotemporally homogeneous layers. Indicates the first A homogeneous layer, Indicates sample The confounding factor eigenvectors Presentation layer Cluster centers Represents Mahalanobis distance, Represents the covariance matrix. Indicates the intra-layer difference threshold. Indicates an indicator function, Represents the set of confounding factors. Indicates the number of homogeneous layers. Indicates the spatiotemporal homogeneous layer number, Indicates the candidate cluster center number. Indicates the sample index. Indicates the first Candidate cluster centers of the layer; Within each of the aforementioned spatiotemporally homogeneous layers, the target scheduling intervention strategy is matched with a preference score to generate balanced samples within the layer. The adjusted sample group is obtained by aggregating the intralayer equilibrium samples of all spatiotemporally homogeneous layers.

4. The multi-dimensional scheduling method for library document resources integrating circulation big data according to claim 2, characterized in that, The backdoor path blocking analysis performed on the dynamic directed acyclic graph generates a set of confounding factors, including: Backdoor path identification is performed on the dynamic directed acyclic graph to obtain a candidate path set; Perform a conditional independence test on each path in the candidate path set to generate a blocking node set; By filtering the set of blocking nodes, nodes that simultaneously affect both the intervention variable and the outcome variable are obtained, resulting in a set of confounding factors.

5. The multi-dimensional scheduling method for library document resources integrating circulation big data according to claim 1, characterized in that, The heterogeneous policy generation process performed on the unbiased causal effect estimation yields a literature-user dual-granularity scheduling policy set, including: The unbiased causal effect estimate is enhanced by feature interaction to generate a heterogeneous effect tensor; Multi-granularity clustering analysis was performed on the heterogeneity effect tensor to obtain the document cluster strategy prototype and the user cluster strategy prototype. The document cluster strategy prototype and the user cluster strategy prototype are fused and optimized to generate the document-user dual-granularity scheduling strategy set.

6. The multi-dimensional scheduling method for library document resources integrating circulation big data according to claim 5, characterized in that, The multi-granularity clustering analysis performed on the heterogeneous effect tensor yields document cluster strategy prototypes and user cluster strategy prototypes, including: The heterogeneity effect tensor is sliced ​​along the literature dimension to obtain the literature-environment response matrix; Multi-scale spectral clustering is performed on the literature-environment response matrix to obtain the prototype of the literature cluster strategy. The heterogeneous effect tensor is sliced ​​along the user dimension to generate a user-environment response matrix; Multi-scale spectral clustering is performed on the user-environment response matrix to obtain the user cluster strategy prototype.

7. The multi-dimensional scheduling method for library document resources integrating circulation big data according to claim 1, characterized in that, The dynamic optimization of the document-user dual-granularity scheduling strategy set yields an optimized scheduling strategy set, including: A digital twin environment simulation was performed on the document-user dual-granularity scheduling strategy set to obtain a strategy execution effect graph; Detect drift in real-time environmental variables and generate policy update trigger signals; In response to the policy update trigger signal, multi-objective reinforcement learning optimization is performed on the policy execution effect graph to generate the optimized scheduling policy set.

8. A multi-dimensional scheduling system for library document resources integrating circulation big data, characterized in that: The system includes: The causal graph construction module is used to construct causal graphs based on multi-dimensional flow data and domain knowledge constraints, resulting in a dynamic directed acyclic graph. The confounding adjustment module is used to adjust the confounding factors of the dynamic directed acyclic graph and the target scheduling intervention strategy to generate an unbiased causal effect estimate. The dual-granularity generation module is used to generate heterogeneous strategies from the unbiased causal effect estimation to obtain a literature-user dual-granularity scheduling strategy set. The dynamic optimization module is used to dynamically optimize the document-user dual-granularity scheduling strategy set to obtain an optimized scheduling strategy set.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the multi-dimensional scheduling method for library document resources that integrates circulation big data as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the multi-dimensional scheduling method for library document resources that integrates circulation big data as described in any one of claims 1 to 7.

Images