Enterprise digital management system based on multi-source data fusion

Abstract

The invention discloses an enterprise digital management system based on multi-source data fusion, and the system comprises a data collection module which is used for constructing a multi-source data view from internal and external data sources of an enterprise; the embedding mapping module is used for generating a feature domain embedding vector set; the improved extremely deep factorization machine model module is used for feature extraction and explicit high-order interactive modeling and outputting a basic matching score; the triple generation module is used for constructing a positive and negative sample triple and calculating a business importance coefficient and an exposure probability; the Bayesian personalized sorting module is used for constructing weighted sorting loss and updating model parameters; and the online recommendation module is used for calculating matching scores of the candidate schemes and sorting and outputting a recommendation management list. According to the method, accurate and intelligent recommendation and sorting decision of candidate schemes in an enterprise management scene are realized by improving an extremely deep factorization machine model and Bayesian personalized sorting.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data management, and in particular to an enterprise digital management system based on multi-source data fusion. BACKGROUND

[0002] At present, enterprises generally deploy multiple business systems such as enterprise resource planning systems, manufacturing execution systems, warehouse management systems and customer relationship management systems in the management activities of procurement, production, inventory, logistics and task collaboration, and access external data sources such as supply service platforms, logistics service platforms, credit evaluation platforms and Internet of Things devices. These systems have obvious differences in data models, field caliber, coding rules, time granularity and update frequency, forming a large number of mutually isolated business data islands. The existing solutions mostly use point-to-point interfaces, simple data synchronization or report-level splicing to centrally display multi-source data, usually only staying at the basic summary and statistical level, lacking a unified feature expression method and systematic high-order correlation modeling capability, making it difficult to build consistent data views across systems and sources among users, organizations, business needs, candidate solutions and material inventories, and unable to mine high-order correlation relationships across sources and objects in a unified view, thereby failing to effectively support fine and real-time management and decision-making of enterprises under complex constraints.

[0003] When intelligent recommendation and decision support are carried out based on multi-source business data, the existing technology mostly relies on artificially configured business rules or single-score-based sorting methods. Although some solutions introduce collaborative filtering, factor decomposition machines or deep learning models, they often only make point value predictions for a single or a small number of data sources, lack training frameworks targeting sorting optimization, do not fully utilize the implicit feedback information of a large number of historical adoption behaviors and display behaviors in enterprise operation logs, and do not incorporate business value weights and candidate solution display exposure into unified modeling. Current technology often only performs simple sorting according to predicted scores, which cannot distinguish the importance of high-value business samples and general samples in sorting targets, nor can it distinguish between the two cases of not being adopted and not being exposed, resulting in that in key scenarios such as procurement scheme selection, production scheduling optimization, inventory replenishment decision and task allocation recommendation, the recommendation results cannot balance business value priority and real preference, and the sorting accuracy and decision rationality are insufficient, which cannot meet the actual application needs of enterprises for digital management systems in intelligent recommendation and decision support.

[0004] Therefore, how to provide an enterprise digital management system based on multi-source data fusion is a problem that those skilled in the art need to solve. SUMMARY

[0005] One objective of this invention is to propose an enterprise digital management system based on multi-source data fusion. This invention achieves intelligent recommendation and ranking optimization of candidate solutions in enterprise management decision-making scenarios by uniformly modeling multi-source business data from both internal and external sources and introducing an improved deep factorization machine model and Bayesian personalized ranking. This invention comprehensively utilizes techniques such as multi-source data view construction, field source partitioning and embedding mapping, explicit high-order interaction modeling under feature domain relationship graph constraints, positive and negative sample triple construction, adaptive interval based on business importance, and exposure-corrected ranking loss. It fully describes the entire process from business data acquisition and preprocessing, feature representation learning, basic matching score calculation, ranking training, to online recommendation output. This invention innovatively introduces embedding subspace decoupling, source-aware relationship interaction, and a ranking optimization mechanism combining business value weights and exposure probabilities, enabling effective mining of high-order relationships between multi-source heterogeneous business objects and fine-tuning of candidate solution ranking results. Compared with existing technologies, this invention can provide more accurate recommendation results that better align with business value priorities and actual usage preferences in enterprise procurement decisions, production scheduling, inventory replenishment, and task allocation scenarios, thereby improving the intelligence level and decision-making efficiency of enterprise digital management.

[0006] An enterprise digital management system based on multi-source data fusion according to an embodiment of the present invention includes the following modules:

[0007] The data acquisition module is used to acquire and preprocess data from internal and external data sources, label field categories, and build multi-source data views;

[0008] The embedding mapping module is used to divide fields according to multi-source data views and perform embedding mapping to generate a set of feature domain embedding vectors.

[0009] An improved deep factorization machine model module is used to model numerical feature extraction and explicit high-order interactions in a set of embedding vectors in the feature domain, and outputs a basic matching score.

[0010] The triple generation module is used to construct positive and negative sample triples and calculate the business importance coefficient and exposure probability based on the basic matching score and operation log.

[0011] The Bayesian personalized ranking module is used to construct a weighted ranking loss and update model parameters based on the business importance coefficient and exposure probability.

[0012] The online recommendation module is used to calculate the matching score of candidate solutions based on the updated model and sort and output the recommendation management list.

[0013] Optionally, modules can be integrated using the following methods:

[0014] Business data is obtained from internal and external data sources, preprocessed, and field categories are labeled for each data field to build a multi-source data view.

[0015] Based on the multi-source data view, the source categories of the labeled data for each field are divided and embedded mapping is performed to obtain a set of feature domain embedding vectors;

[0016] An improved deep factorization machine model is constructed to perform feature representation learning and basic score calculation. Feature extraction is performed on the numerical fields of a unified multi-source data view. Explicit higher-order interactions are performed on the feature domain embedding vectors and the interaction weights are configured according to the source category. Higher-order features are extracted and concatenated to obtain the basic matching score.

[0017] Based on enterprise management operation logs, a set of triples containing positive and negative samples is constructed. The basic matching scores of positive and negative samples are calculated separately using a deep factorization machine model. The business importance coefficient of each triple is calculated based on the positive samples, and the exposure probability is calculated based on the negative samples.

[0018] Input the set of triples into Bayesian personalized ranking, determine the adaptive interval parameter of each triple according to the business importance coefficient, combine it with the difference in matching scores of positive and negative samples to construct the ranking loss term, and multiply it with the exposure probability to form the total loss function. Update the parameters of the improved deep factorization machine model according to the total loss function to obtain the updated improved deep factorization machine model.

[0019] In enterprise management operations, feature domain embedding vectors are generated from multi-source data views and input into an updated improved deep factorization machine model to obtain matching scores for each candidate scheme, and then sorted to generate a recommended management list.

[0020] Optionally, the business data includes user and organization data, business demand data, candidate solution data, material and inventory data, and exposure and interaction logs.

[0021] Optionally, constructing a multi-source data view includes:

[0022] Establish a list and connection configuration of internal and external data sources for the enterprise, determine the data source type, access method and collection frequency, collect business data to generate raw datasets, and attach source tags and timestamps to each record, convert the unit of measurement to a unified benchmark unit, and complete data type standardization and legality verification.

[0023] Based on the enterprise's business objects, field alignment and pattern mapping are performed to map each source field to a unified field set, record primary keys are established, data cleaning and deduplication are performed, null values ​​are deleted, outliers are truncated or corrected, and duplicate records are removed by primary key.

[0024] Each field is categorized and labeled as a field category. The processed data is then written to a unified storage system, and multi-source data views are generated by dimension.

[0025] Optionally, obtaining the feature domain embedding vector set includes:

[0026] The source category of each data field in the multi-source data view is labeled and divided into internal source field set and external source field set. The categorical and identifier fields are discretely indexed and encoded, and the numerical fields are standardized according to mean and variance to form an embeddable input.

[0027] Embedding mapping is performed on the internal source field set and the external source field set using corresponding embedding matrices. The two types of embedding vectors are projected onto a vector space of the same dimension through the embedding mapping and then assembled into a feature domain embedding vector set in sequence.

[0028] Optionally, obtaining the basic matching score includes:

[0029] An improved deep factorization machine model is constructed, including an embedding subspace module, a relational interaction module, a low-order feature module, and a matching output module;

[0030] The embedding subspace module splits each feature domain embedding vector in the feature domain embedding vector set into sub-vectors according to the dimension partitioning rules, including interaction-sensitive subspace vectors and business semantic subspace vectors, and applies decoupling constraints to each subspace vector to limit the correlation between different subspaces;

[0031] The relationship interaction module constructs a feature domain relationship graph based on the field categories and business object relationships in the multi-source data view. It aligns the interaction-sensitive subspace vectors with the feature domain relationship graph according to their respective feature domains. It performs explicit feature interaction calculations only on feature domain pairs with graph edge connections in the feature domain relationship graph. It configures different interaction weights and parameter sets, and obtains high-order interaction features through multi-layer stacked explicit interactions and aggregations. At the same time, it convolves the business semantic subspace vectors to obtain semantic features.

[0032] The low-order feature module takes standardized numerical fields from multi-source data views and inputs them into the factorization machine branch to calculate low-order features;

[0033] The matching output module concatenates low-order features, high-order interaction features, and semantic features obtained from the business semantic subspace vector. It then calculates the basic matching score between the target user and each candidate solution in a given business scenario and environmental context based on the concatenation result.

[0034] Optionally, the step of calculating the matching scores of positive and negative samples using a deep factorization machine model, calculating the business importance coefficient of each triple based on the positive samples, and calculating the exposure probability based on the negative samples includes:

[0035] Based on enterprise management operation logs, behavior records are extracted from multi-source data views and categorized according to user identifiers and business scenario identifiers;

[0036] In the behavior records of each type of user and business scenario, the candidate solutions marked as true and true for display are identified as positive sample candidate solutions, and the candidate solutions marked as true for display and false for adoption are identified as negative sample candidate solutions. The user identifier, positive sample candidate solution identifier and negative sample candidate solution identifier under the same user and business scenario are combined to form a triplet, forming a triplet set.

[0037] For each triple, the corresponding user features, business scenario features, positive sample candidate solution features, and negative sample candidate solution features are input into the improved deep factorization machine model to calculate the basic matching score of positive samples and the basic matching score of negative samples, respectively.

[0038] Based on the basic matching score of positive samples and the business value associated with positive samples, the business importance coefficient of the current triple is determined through the mapping relationship. The number of times the corresponding negative sample candidate solution is displayed in the user's business scenario and the total number of times all candidate solutions are displayed are counted. The exposure probability of the current negative sample candidate solution is calculated based on the ratio of the number of displays to the total number of displays.

[0039] Optionally, the obtained updated improved deep factorization machine model includes:

[0040] Perform Bayesian personalized sorting on the triple set, and read the positive sample basic matching score, negative sample basic matching score, business importance coefficient, and exposure probability for each triple;

[0041] Based on the business importance coefficient, the adaptive interval parameter of each triple is determined by a monotonic mapping function. The adaptive interval parameter is a non-negative real number. For each triple, the basic matching score of the positive sample is calculated and the basic matching score of the negative sample is subtracted to obtain the matching score difference. The corrected difference is obtained by subtracting the corresponding adaptive interval parameter from the matching score difference.

[0042] The sorting loss term is constructed using the correction difference as input. The sorting loss term is the result of taking the negative of the correction difference, inputting it into a logic function, and then taking the logarithm.

[0043] The exposure-weighted loss term is obtained by multiplying the ranking loss term of each triplet by the corresponding exposure probability. The exposure-weighted loss terms of all triplets are summed and then superimposed with a regularization term consisting of the sum of squared parameters based on the improved deep factorization machine model and Bayesian personalized ranking to form the total loss function.

[0044] Optimize the total loss function, update the parameters of the improved deep factorization machine model and the Bayesian personalized ranking until the total loss function meets the preset convergence condition, and obtain the updated improved deep factorization machine model.

[0045] Optionally, the sorting to generate the recommendation management list includes:

[0046] During the operation of enterprise digital management, monitor the decision requests initiated by users in scenarios such as procurement decision-making, production scheduling, inventory replenishment, and task allocation. Based on the decision requests, determine the target user identifier and the target business scenario identifier, and filter the candidate solution set from multi-source data views based on the target business scenario identifier.

[0047] Extract user-related fields, scenario-related fields, and candidate solution-related fields corresponding to the target user identifier, target business scenario identifier, and candidate solution set from the multi-source data view; generate a feature domain embedding vector set; input the updated improved deep factorization machine model with the numerical fields; and calculate the matching score between the target user and each candidate solution.

[0048] The candidate solutions are sorted according to the matching score, and a recommended management list is generated in descending order of matching score. This list is then output to the user interface of the enterprise digital management system. The system also records the display position and number of times each candidate solution is displayed, as well as the user's clicks and adoption behavior for each candidate solution, thus obtaining the management results.

[0049] The beneficial effects of this invention are:

[0050] This invention significantly enhances the unified modeling and high-order association mining capabilities of heterogeneous business data from within and outside enterprises by constructing a unified multi-source data view and introducing an improved deep factorization machine model with embedded subspace decoupling, feature domain relationship graph constraints, and source-aware high-order interactions. Compared to existing technologies that isolate or simply concatenate various business data, this invention maps user, business requirement, candidate solution, and material inventory fields to unified feature domain embedding vectors. Utilizing subspace partitioning and relationship graph constraints, explicit high-order interactions are performed only on feature domain pairs with business relationships, and interaction weights are configured according to source categories. While controlling model complexity, this strengthens the representation of key cross-system and cross-source association features, improves the accuracy of basic matching scores, and enhances the model's adaptability to complex business structures, providing a stable and reliable feature foundation for intelligent recommendation and decision support.

[0051] This invention introduces a Bayesian personalized ranking system with adaptive intervals based on business importance and exposure weighting during the ranking training phase. This ensures that the recommendation results fully reflect business value priorities and actual display conditions while maintaining personalized preferences. By determining the adaptive interval parameters of the triples based on positive sample matching scores and preset business value indicators, stricter ranking interval constraints are imposed on high-value samples. Simultaneously, the ranking loss is weighted by incorporating the exposure probability of negative samples in actual business scenarios, avoiding misjudging undisplayed candidate solutions as strong negative feedback. This allows high-value and genuinely compared candidate solutions to obtain more reasonable ranking positions in critical management scenarios such as procurement decisions, production scheduling, inventory replenishment, and task allocation. This invention overcomes the technical bottlenecks of existing technologies, such as low utilization of multi-source data, disconnect between ranking objectives and business value, and inability to distinguish between unadopted and unexposed solutions. It improves the accuracy and practicality of enterprise digital management systems in intelligent recommendation and refined decision support. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is a flowchart of an enterprise digital management system based on multi-source data fusion proposed in this invention;

[0054] Figure 2 This is a structural block diagram of an enterprise digital management method based on multi-source data fusion proposed in this invention;

[0055] Figure 3 This is a functional diagram of an improved deep factorization machine model for an enterprise digital management method based on multi-source data fusion proposed in this invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0057] refer to Figure 1 A digital management system for enterprises based on multi-source data fusion includes the following modules:

[0058] The data acquisition module is used to acquire and preprocess data from internal and external data sources, label field categories, and build multi-source data views;

[0059] The embedding mapping module is used to divide fields according to multi-source data views and perform embedding mapping to generate a set of feature domain embedding vectors.

[0060] An improved deep factorization machine model module is used to model numerical feature extraction and explicit high-order interactions in a set of embedding vectors in the feature domain, and outputs a basic matching score.

[0061] The triple generation module is used to construct positive and negative sample triples and calculate the business importance coefficient and exposure probability based on the basic matching score and operation log.

[0062] The Bayesian personalized ranking module is used to construct a weighted ranking loss and update model parameters based on the business importance coefficient and exposure probability.

[0063] The online recommendation module is used to calculate the matching score of candidate solutions based on the updated model and sort and output the recommendation management list.

[0064] refer to Figure 2 and Figure 3 A digital management method for enterprises based on multi-source data fusion includes:

[0065] Business data is obtained from internal and external data sources, preprocessed, and field categories are labeled for each data field to build a multi-source data view.

[0066] Based on the multi-source data view, the source categories of the labeled data for each field are divided and embedded mapping is performed to obtain a set of feature domain embedding vectors;

[0067] An improved deep factorization machine model is constructed to perform feature representation learning and basic score calculation. Feature extraction is performed on the numerical fields of a unified multi-source data view. Explicit higher-order interactions are performed on the feature domain embedding vectors and the interaction weights are configured according to the source category. Higher-order features are extracted and concatenated to obtain the basic matching score.

[0068] Based on enterprise management operation logs, a set of triples containing positive and negative samples is constructed. The basic matching scores of positive and negative samples are calculated separately using a deep factorization machine model. The business importance coefficient of each triple is calculated based on the positive samples, and the exposure probability is calculated based on the negative samples.

[0069] Input the set of triples into Bayesian personalized ranking, determine the adaptive interval parameter of each triple according to the business importance coefficient, combine it with the difference in matching scores of positive and negative samples to construct the ranking loss term, and multiply it with the exposure probability to form the total loss function. Update the parameters of the improved deep factorization machine model according to the total loss function to obtain the updated improved deep factorization machine model.

[0070] In enterprise management operations, feature domain embedding vectors are generated from multi-source data views and input into an updated improved deep factorization machine model to obtain matching scores for each candidate scheme, and then sorted to generate a recommended management list.

[0071] In this embodiment, the business data includes user and organization data, business demand data, candidate solution data, material and inventory data, and exposure and interaction logs.

[0072] In this embodiment, constructing a multi-source data view includes:

[0073] Establish a list and connection configuration of internal and external data sources for the enterprise, determine the data source type, access method and collection frequency, collect business data to generate raw datasets, and attach source tags and timestamps to each record, convert the unit of measurement to a unified benchmark unit, and complete data type standardization and legality verification.

[0074] Based on the enterprise's business objects, field alignment and schema mapping are performed to map each source field to a unified field set, establish record primary keys, perform data cleaning and deduplication, delete null values, truncate or correct outliers, and remove duplicate records by primary key. The schema mapping specifically involves:

[0075] Based on the enterprise's business definition, a unified data model is defined, including a unified field set for user objects, business requirement objects, candidate solution objects, and material and inventory objects. Each unified field set contains the standard field name, data type, unit of measurement, and value coding standard for the corresponding business object.

[0076] For the original table structure and field definition of internal and external data sources of various enterprises, the original fields in each source are renamed, type-converted, unit-converted and encoded into target fields under a unified data model;

[0077] Each field is categorized and labeled as a field category. The processed data is then written to a unified storage system, and multi-source data views are generated by dimension.

[0078] In this embodiment, obtaining the feature domain embedding vector set includes:

[0079] The source category of each data field in the multi-source data view is labeled and divided into internal source field set and external source field set. The categorical and identifier fields are discretely indexed and encoded, and the numerical fields are standardized according to mean and variance to form an embeddable input.

[0080] Embedding mapping is performed on the internal source field set and the external source field set using corresponding embedding matrices. This embedding mapping projects the two types of embedding vectors into a vector space of the same dimension, and they are then sequentially grouped into a feature domain embedding vector set, where:

[0081] The embedding matrix is ​​a parameter matrix constructed for discrete fields, used to map the values ​​after discrete index encoding to fixed-length real number vectors. For each subcategorical field in the internal source field set, an internal embedding matrix is ​​assigned to the field. The number of rows in the internal embedding matrix is ​​equal to the number of possible values ​​of the field, and the number of columns is equal to the embedding dimension. Each row corresponds to the embedding vector of a value. The subcategorical fields or identifier fields in the external source field set are configured with corresponding external embedding matrices. The row and column structure is the same as that of the internal embedding matrix, which is used to vectorize the external source fields.

[0082] Specifically, the embedding mapping involves encoding the values ​​of categorical fields and identifier fields in the multi-source data view using discrete indexes, then using the encoded values ​​as row indices to select multiple rows of embedding vectors from the corresponding internal or external embedding matrices as the embedding representation of the current field. The internal embedding matrix is ​​used to complete the embedding mapping for the set of internal source fields, and the external embedding matrix is ​​used to complete the embedding mapping for the set of external source fields. The embedding vectors of each field are then grouped into a feature domain embedding vector set in sequence.

[0083] In this embodiment, obtaining the basic matching score includes:

[0084] An improved deep factorization machine model is constructed, comprising an embedding subspace module, a relational interaction module, a low-order feature module, and a matching output module, wherein:

[0085] The improved deep factorization machine model is constructed in a modular manner, specifically as follows:

[0086] The embedding subspace module performs subspace partitioning and decoupling on the embedding vectors of each feature domain to generate interaction-sensitive subspace vectors for explicit interaction modeling and business semantic subspace vectors for semantic modeling. The relation interaction module performs multi-layer explicit high-order feature interactions on the interaction-sensitive subspace vectors based on source and relation type under the constraint of the feature domain relation graph to obtain high-order interaction features.

[0087] The low-order feature module models the numerical fields from multi-source data views through factorization machine branches to obtain low-order features. Then, the matching output module concatenates the high-order interaction features, semantic features obtained from the business semantic subspace, and low-order features. The basic matching score between the target user and the candidate solution is calculated through the output layer to form the overall structure of the improved deep factorization machine model.

[0088] The embedding subspace module splits each feature domain embedding vector in the feature domain embedding vector set into sub-vectors according to dimensionality partitioning rules, including interaction-sensitive subspace vectors and business semantic subspace vectors. Decoupling constraints are applied to each subspace vector to limit the correlation between different subspaces, where:

[0089] The specific steps for splitting subvectors according to the dimension partitioning rules are as follows:

[0090] Each feature domain embedding vector in the feature domain embedding vector set is split according to the dimension. A uniform total dimension length is configured for the embedding vector, and several continuous dimension intervals are divided on the total dimension length. The component corresponding to the first dimension interval is determined as the interaction-sensitive subspace vector, and the component corresponding to the second dimension interval is determined as the business semantic subspace vector. By splitting each feature domain embedding vector according to the dimension, the interaction-sensitive subspace vector for explicit interaction modeling and the business semantic subspace vector for semantic modeling are obtained.

[0091] The decoupling constraint applied to each subspace vector to limit the correlation between different subspaces is specifically as follows:

[0092] To limit the correlation between different subspaces, while the embedded subspace module outputs the interaction-sensitive subspace vector and the business semantic subspace vector, decoupling constraints are applied to each subspace vector. During the training of the improved deep factorization machine model, the correlation index between the interaction-sensitive subspace vector and the business semantic subspace vector of the same feature domain is calculated.

[0093] The correlation index is the inner product and covariance between two vectors. A penalty term positively correlated with the correlation index is added to the loss function of the improved deep factorization machine model to guide the interaction-sensitive subspace vector and the business semantic subspace vector to tend to be approximately orthogonal during the optimization process. Regular constraints are applied to the same type of subspace vectors in different feature domains to suppress excessively high redundant correlations within the subspace, thereby decoupling the subspace expression and limiting the correlation between different subspaces.

[0094] The relationship interaction module constructs a feature domain relationship graph based on field categories and business object relationships in the multi-source data view. It aligns the interaction-sensitive subspace vectors with the feature domain relationship graph according to their respective feature domains. Explicit feature interaction calculations are performed only on feature domain pairs connected by graph edges in the feature domain relationship graph. Different interaction weights and parameter sets are configured, and higher-order interaction features are obtained through multi-layered stacked explicit interactions and aggregations. Simultaneously, semantic features are obtained by convolving the business semantic subspace vectors. Specifically, the process of obtaining higher-order interaction features through multi-layered stacked explicit interactions and aggregations, and simultaneously obtaining semantic features by convolving the business semantic subspace vectors, involves:

[0095] Based on the field categories and business object relationships in the multi-source data view, a feature domain relationship graph is constructed. The user feature domain, business requirement feature domain, candidate solution feature domain, material and inventory feature domain, and environmental context feature domain are used as graph nodes. Graph edges are established between nodes according to business associations and historical statistical relationships. The source category label is retained on each node, and the corresponding business relationship type is labeled on each graph edge to obtain the feature domain relationship graph used to constrain feature interaction paths.

[0096] Align the interaction-sensitive subspace vectors according to the nodes in the feature domain and feature domain relationship graph. For each node connected by a graph edge, extract the corresponding interaction-sensitive subspace vectors and calculate them according to the interaction operation method. Perform element-wise multiplication on each corresponding dimension of the two interaction-sensitive subspace vectors and sum them to generate the interaction vector. Obtain the interaction result vector of the graph edge through a non-linear activation function. For different source combinations such as internal source with internal source, internal source with external source, and external source with external source, use different sets of weight parameters and bias parameters respectively.

[0097] Aggregate the interaction result vectors generated by all graph edges connected to a node by summing them according to their dimensions. Then, obtain the node's current layer update representation through a linear transformation and a nonlinear activation function. Use the update representation of each node as the input to the next layer, and obtain a high-order interaction feature representation through multi-layer stacking.

[0098] Semantic features are obtained by performing convolution operations on the business semantic subspace vectors. The business semantic subspace vectors corresponding to each feature domain are arranged in order into a vector sequence. A one-dimensional convolution kernel is selected and slid on the sequence with a fixed stride. At each sliding position, each weight coefficient in the convolution kernel is multiplied by the corresponding business semantic subspace vector in the window according to the dimension and then summed to obtain the convolution output vector at the current position. A bias parameter is added to the convolution output vector and input into a non-linear activation function to obtain local semantic features.

[0099] After performing convolution calculations at all positions along the sequence, pooling is performed on all local semantic feature vectors by taking the maximum value of each dimension to obtain the global semantic feature representation, which is used as the semantic feature output by the relational interaction module.

[0100] The low-order feature module inputs standardized numerical fields from multi-source data views into the factorization machine branch to calculate low-order features. Specifically, the calculation of these low-order features involves:

[0101] Factorization machine modeling is performed on numerical fields from multi-source data views that have been standardized. All numerical fields of the current sample are arranged in order to form a numerical feature vector. A first-order weight parameter and a fixed-dimensional latent vector parameter are configured for each numerical field. The current numerical feature vector is multiplied by the corresponding first-order weight according to the field and summed to obtain the first-order linear combination result.

[0102] For any two different fields, the latent vector parameters are extracted respectively, multiplied element-wise along the dimension and multiplied by the corresponding numerical features, and then summed along the dimension to obtain the second-order interaction result of the current field pair. The second-order interaction results of all field pairs in the sample are accumulated to obtain the overall second-order interaction result. The first-order linear combination result is concatenated with the second-order interaction result, and through a linear transformation and a nonlinear activation function, a low-dimensional vector output is obtained to represent the low-order features of the current sample, which is used as the output result of the low-order feature module.

[0103] The matching output module concatenates low-order features, high-order interaction features, and semantic features obtained from the business semantic subspace vector. It then calculates the basic matching score between the target user and each candidate solution within a given business scenario and environmental context based on the concatenation result. Specifically, this calculation involves:

[0104] The low-order features, high-order interaction features, and semantic features obtained from the business semantic subspace vector are concatenated in sequence along the dimensions and merged into a unified feature vector. For each combination of the target user and the candidate solution, the corresponding unified feature vector is input into the output layer. In the output layer, the corresponding weight parameters are configured for each dimension of the unified feature vector, and a bias parameter is set. The values ​​of each dimension of the unified feature vector are multiplied by the corresponding weight parameters and summed. The summation result is then added to the bias parameter to obtain the linear combination value of the matching score of the user and the candidate solution in the current business scenario and environmental context.

[0105] The linear combination value is input into a monotonic nonlinear function and mapped to a real number between zero and one as the basic matching score. The basic matching score is then applied to each candidate solution in the candidate solution set to obtain the basic matching score set between the target user and each candidate solution under a given business scenario and environmental context.

[0106] In this embodiment, the step of calculating the matching scores of positive and negative samples using a deep factorization machine model, calculating the business importance coefficient of each triple based on the positive samples, and calculating the exposure probability based on the negative samples includes:

[0107] Based on enterprise management operation logs, behavior records are extracted from multi-source data views and categorized according to user identifiers and business scenario identifiers;

[0108] In the behavior records of each type of user and business scenario, the candidate solutions marked as true and true for display are identified as positive sample candidate solutions, and the candidate solutions marked as true for display and false for adoption are identified as negative sample candidate solutions. The user identifier, positive sample candidate solution identifier and negative sample candidate solution identifier under the same user and business scenario are combined to form a triplet, forming a triplet set.

[0109] For each triple, the corresponding user features, business scenario features, positive sample candidate solution features, and negative sample candidate solution features are input into the improved deep factorization machine model to calculate the basic matching score of positive samples and the basic matching score of negative samples, respectively.

[0110] Based on the basic matching score of positive samples and the business value associated with positive samples, the business importance coefficient of the current triple is determined through a mapping relationship. The number of times the corresponding negative sample candidate solution is displayed in the user's business scenario and the total number of displays for all candidate solutions are calculated. The exposure probability of the current negative sample candidate solution is then calculated based on the ratio of the number of displays to the total number of displays.

[0111] The business value includes economic value, risk level, and timeliness. The dimensions, value range, and weight coefficient of each indicator dimension are determined during configuration. The original value of each business value is converted into a dimensionless value between zero and one by normalization. The normalized values ​​are weighted and summed to obtain a comprehensive business value score. The comprehensive business value score is mapped to a real number between zero and one and used as the business importance coefficient of the current triple.

[0112] The method of calculating the exposure probability of the current negative sample candidate solution based on the ratio of the number of times it is displayed to the total number of times it is displayed is as follows: For the negative sample candidate solutions in the same triplet, the number of times it is displayed in the user's corresponding business scenario is counted, and the sum of the number of times all candidate solutions in the user's business scenario are displayed in the same statistical period is counted. The number of times the negative sample is displayed is divided by the total number of times it is displayed to obtain a ratio value between zero and one. The ratio value is used as the exposure probability of the current negative sample candidate solution.

[0113] In this embodiment, obtaining the updated improved deep factorization machine model includes:

[0114] Perform Bayesian personalized sorting on the triple set, and read the positive sample basic matching score, negative sample basic matching score, business importance coefficient, and exposure probability for each triple;

[0115] The adaptive interval parameter for each triplet is determined using a monotonic mapping function based on the business importance coefficient. The adaptive interval parameter is a non-negative real number. For each triplet, the basic matching score of the positive samples is subtracted from the basic matching score of the negative samples to obtain the matching score difference. The correction difference is then obtained by subtracting the corresponding adaptive interval parameter from the matching score difference. Specifically, determining the adaptive interval parameter for each triplet using a monotonic mapping function based on the business importance coefficient involves the following steps:

[0116] Configure two parameters: minimum interval value and maximum interval value. Both are non-negative real numbers, and the maximum interval value is not less than the minimum interval value. For each triple, obtain the business importance coefficient of the current triple, multiply the business importance coefficient by the difference between the maximum interval value and the minimum interval value, add the product to the minimum interval value, and use the sum as the adaptive interval parameter of the triple, so that the larger the business importance coefficient, the larger the corresponding adaptive interval parameter.

[0117] The sorting loss term is constructed using the correction difference as input. The sorting loss term is the result of taking the negative of the correction difference, inputting it into a logic function, and then taking the logarithm.

[0118] The exposure-weighted loss term is obtained by multiplying the ranking loss term of each triplet by its corresponding exposure probability. The exposure-weighted loss terms of all triplets are summed, and a regularization term consisting of the sum of squared parameters based on the improved deep factorization machine model and Bayesian personalized ranking is added to form the total loss function. The parameters of the improved deep factorization machine model and Bayesian personalized ranking include:

[0119] Embedding matrix parameters used to map each discrete field to a vector representation, interaction weights and bias parameters used for explicit high-order feature interaction calculation in the relational interaction module, and first-order weights and latent vector parameters used for modeling low-order features in the factorization machine branch.

[0120] Optimize the total loss function, update the parameters of the improved deep factorization machine model and the Bayesian personalized ranking until the total loss function meets the preset convergence condition, and obtain the updated improved deep factorization machine model.

[0121] In this embodiment, the sorting and generation of the recommendation management list includes:

[0122] During the operation of enterprise digital management, monitor the decision requests initiated by users in scenarios such as procurement decision-making, production scheduling, inventory replenishment, and task allocation. Based on the decision requests, determine the target user identifier and the target business scenario identifier, and filter the candidate solution set from multi-source data views based on the target business scenario identifier.

[0123] Extract user-related fields, scenario-related fields, and candidate solution-related fields corresponding to the target user identifier, target business scenario identifier, and candidate solution set from the multi-source data view; generate a feature domain embedding vector set; input the updated improved deep factorization machine model with the numerical fields; and calculate the matching score between the target user and each candidate solution.

[0124] The candidate solutions are sorted according to the matching score, and a recommended management list is generated in descending order of matching score. This list is then output to the user interface of the enterprise digital management system. The system also records the display position and number of times each candidate solution is displayed, as well as the user's clicks and adoption behavior for each candidate solution, thus obtaining the management results.

[0125] Example 1:

[0126] To verify the feasibility of this invention in practice, it was applied to an equipment manufacturing company. The company simultaneously runs multiple business systems, including an Enterprise Resource Planning (ERP) system, Manufacturing Execution System (MRP) system, Warehouse Management System (WMS) system, and a supplier collaboration platform, and is also connected to a third-party logistics platform and an IoT monitoring system for some key equipment. Due to inconsistent data formats and field definitions across these systems, purchasing and planning personnel frequently had to switch between multiple system interfaces to view information on prices, delivery dates, inventory, and equipment load when making purchasing decisions, production scheduling, and replenishment decisions. Decision-making heavily relied on personal experience, resulting in both time-consuming and unstable processes. Before the pilot program, the company frequently experienced issues such as duplicate purchases of similar materials, periodic shortages of key materials, and frequent temporary adjustments to production schedules. This led to an unbalanced inventory structure, difficulties in production plan execution, and delayed management feedback. The problems of unifying the use of multi-source heterogeneous data and the inability of sorting results to reflect business value priorities were particularly prominent.

[0127] The proposed enterprise digital management system based on multi-source data fusion is deployed on the enterprise's existing data platform. The data acquisition module integrates data from various business systems, logistics platforms, and IoT systems, including orders, work orders, material inventory, quotations, delivery dates, in-transit status, equipment status, and user and departmental data, to construct a unified multi-source data view. The embedding and mapping module performs source division and embedding mapping on fields from different sources, generating a set of feature domain embedding vectors. The improved deep factorization machine model module performs high-order interactive modeling of features such as users, business needs, candidate solutions, and inventory status under the constraint of the feature domain relationship graph, outputting basic matching scores for procurement, scheduling, and replenishment plans. The triple generation module constructs triples containing positive and negative samples based on the operation logs and calculates the business importance coefficient and exposure probability. The Bayesian personalized ranking module uses adaptive interval and exposure weighting mechanisms for ranking training. During the trial operation, when a purchasing agent or planner creates a requirement or task in the system, the online recommendation module automatically generates a ranked list of candidate solutions. Users only need to check and confirm from a limited selection, and the system simultaneously records the display and adoption behavior.

[0128] During a quarter-long trial run, focusing on the current procurement and scheduling operations of an enterprise, and comparing the implementation of this invention with previous methods such as manual experience-based decision-making, fixed-rule-based recommendation methods, and traditional factorization machine recommendation methods, the embodiments of this invention demonstrate significant improvements in multi-source data utilization and ranking decision-making effectiveness. The proportion of recommended solutions actually adopted is significantly higher than that of manual experience and rule-based methods. The priority given to high-value orders and orders related to key materials in the recommendation results is significantly enhanced. The average procurement decision-making time is significantly shortened, and alarms regarding shortages of key materials and production line downtime events caused by material shortages are significantly reduced. Simultaneously, the proportion of manual adjustments to the recommendation results by business personnel is significantly reduced, and subjective feedback indicates that the recommendation results are more aligned with actual business needs, effectively alleviating the problems of fragmented multi-source data and unreasonable decision ranking.

[0129] Table 1. Comparison of the effectiveness of different decision-making methods in procurement and scheduling scenarios.

[0130] Indicator Manual experience decision Rule engine recommendation Traditional factor decomposer Method of the present invention Adoption rate of recommended solutions (%) 55.0 64.2 72.5 83.1 High-value order adoption priority rate (%) 50.3 58.7 66.9 79.4 Average procurement decision-making time (minutes / order) 22.1 18.9 15.8 11.7 Key material shortage alarm times (times / month) 48 41 33 24 Line stop events caused by material shortage (times / month) 11 9 6 4 Order on-time or early delivery rate (%) 89.3 92.1 94.0 96.4 Inventory turnover days (days) 52.4 47.5 42.9 39.7 Manual adjustment rate of recommended results (%) 100.0 35.2 24.6 12.8

[0131] As shown in Table 1, the method of this invention significantly outperforms the three comparative methods in both the recommended solution adoption rate and the high-value order priority adoption rate, two core indicators. The adoption rate of recommendations based on human experience is only about 55%, the rule engine improves to about 64%, the traditional factorization machine further improves to about 72%, while the method of this invention reaches about 83%. Regarding the high-value order priority adoption rate, human experience is about 50%, the rule method is about 59%, the traditional factorization machine is about 67%, and the method of this invention approaches 80%. This indicates that after introducing multi-source data fusion, high-order feature interaction, and a ranking mechanism that incorporates business value weights, the system's recommendations are not only more readily adopted by business personnel, but also show a stronger preference for higher-value orders.

[0132] In terms of decision-making efficiency and operational risk, the method of this invention also demonstrates significant advantages. The average procurement decision-making time gradually decreased from approximately 22 minutes using human experience to approximately 19 minutes using rule-based methods and approximately 16 minutes using factorization machines, and the method of this invention further reduced it to approximately 12 minutes. The number of critical material shortage alarms and production line stoppages caused by shortages decreased progressively with method optimization. Under the method of this invention, the monthly average number of material shortage alarms and production line stoppages were significantly lower than those of the three comparative methods. This indicates that through unified modeling of multi-source data and fine-tuning of sorting results, the system can more promptly identify and prioritize critical materials and important orders, reducing material shortage risks and production fluctuations.

[0133] From an overall operational perspective, the method of this invention has achieved comprehensive improvements in indicators such as on-time or early delivery rate of orders, inventory turnover days, and the proportion of manual adjustments to recommendation results. The on-time or early delivery rate is slightly higher than 89% and 92% under human experience and rule-based methods, respectively, while the traditional factor decomposition machine can reach approximately 94%, and the method of this invention further improves it to over 96%. Inventory turnover days have gradually decreased from over 50 days, and the method of this invention can control it within 40 days, reflecting the optimization of inventory structure and replenishment rhythm. The proportion of manual adjustments to recommendation results has gradually decreased from all manual decisions under human experience to about one-third of that under rule-based methods and about one-quarter of that under factor decomposition machines, while the method of this invention requires only slightly more than 10% manual adjustment. This indicates that the system's recommendation results are highly consistent with business judgments, which helps reduce the workload of repeated manual comparisons and adjustments, and improves the practicality and acceptance of overall digital management.

[0134] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-source data fusion-based enterprise digital management system, characterized in that, The method comprises the following steps: a data acquisition module for obtaining and preprocessing data from internal and external data sources of an enterprise, labeling field categories, and constructing a multi-source data view; an embedding mapping module for dividing fields and performing embedding mapping according to the multi-source data view, and generating a feature domain embedding vector set; an improved extreme deep factorization machine model module for extracting numerical features and modeling explicit high-order interactions in the feature domain embedding vector set, and outputting a basic matching score; a triple generation module for constructing positive and negative sample triples and calculating a business importance coefficient and an exposure probability according to the basic matching score and a running log; a Bayesian personalized ranking module for constructing a weighted ranking loss and updating model parameters according to the business importance coefficient and the exposure probability; an online recommendation module for calculating candidate scheme matching scores according to the updated model and outputting a recommended management list in a ranked order.

2. A multi-source data fusion-based enterprise digital management method applied to the multi-source data fusion-based enterprise digital management system of claim 1, characterized in that, The method comprises the following steps: obtaining business data from internal and external data sources of an enterprise, preprocessing the data, labeling field categories for each data field, and constructing a multi-source data view; dividing each field according to the source category of the data and performing embedding mapping according to the multi-source data view to obtain a feature domain embedding vector set; constructing an improved extreme deep factorization machine model for feature representation learning and basic score calculation, extracting features for numerical fields of the unified multi-source data view, performing explicit high-order interaction on the feature domain embedding vector, configuring interaction weights according to the source category, extracting high-order features, and concatenating them to obtain a basic matching score; based on the enterprise management running log, a triple set containing positive samples and negative samples is constructed, the basic matching scores of the positive samples and the negative samples are calculated through the extreme deep factorization machine model, the business importance coefficient of each triple is calculated according to the positive samples, and the exposure probability is calculated according to the negative samples; input the triple set into the Bayesian personalized ranking, determine the adaptive interval parameter of each triple according to the business importance coefficient, combine the difference between the matching scores of the positive samples and the negative samples, construct a ranking loss term, multiply it by the exposure probability to form a total loss function, update the parameters of the improved extreme deep factorization machine model according to the total loss function, and obtain the updated improved extreme deep factorization machine model; in the enterprise management running, generate feature domain embedding vectors from the multi-source data view and input them into the updated improved extreme deep factorization machine model to obtain the matching scores of each candidate scheme and generate a recommended management list in a ranked order. 3.The enterprise digital management method based on multi-source data fusion of claim 2, characterized in that, The business data includes user and organization data, business requirement data, candidate scheme data, material and inventory data, and exposure and interaction logs.

4. The enterprise digital management method based on multi-source data fusion according to claim 2, characterized in that, The method for constructing a multi-source data view comprises the following steps: establishing a list of internal and external data sources and connection configurations, determining the data source type, access method and collection frequency, collecting business data to generate an original data set, and adding a source marker and a timestamp to each record, converting the unit of measurement to a unified reference unit, completing data type specification and legality verification; aligning fields and mapping modes according to enterprise business objects, mapping each source field to a unified field set, establishing a record primary key, performing data cleaning and deduplication processing, deleting null values, truncating or correcting abnormal values, and removing duplicate records according to the primary key; The category of each field is labeled as a field category, and the processed data is written into a unified storage, and a multi-source data view is generated according to dimensions.

5. The enterprise digital management method based on multi-source data fusion according to claim 2, characterized in that, The obtained characteristic domain embedding vector set comprises: The source categories of each data field in the multi-source data view are labeled, and are divided into an internal source field set and an external source field set, the discrete index coding is performed on the classification type and identification type fields, and the standardization processing is performed on the numerical type fields according to the mean value and variance, so as to form an embeddable input; The internal source field set and the external source field set are respectively embedded and mapped by using corresponding embedding matrices, the two types of embedding vectors are projected into a vector space of a unified dimension through embedding mapping, and are sequentially grouped into a characteristic domain embedding vector set.

6. The enterprise digital management method based on multi-source data fusion according to claim 2, characterized in that, The obtained basic matching score comprises: An improved extremely deep factorization machine model is constructed, comprising an embedding subspace module, a relationship interaction module, a low-order feature module and a matching output module; The embedding subspace module splits each characteristic domain embedding vector in the characteristic domain embedding vector set into sub-vectors according to a dimension division rule, including an interaction-sensitive subspace vector and a business semantic subspace vector, and applies decoupling constraints to limit the correlation between different subspaces; The relationship interaction module constructs a characteristic domain relationship graph according to the field categories and business object relationships in the multi-source data view, aligns the interaction-sensitive subspace vectors according to the characteristic domains and the characteristic domain relationship graph, performs explicit feature interaction calculation only on the characteristic domain pairs connected by the graph edges in the characteristic domain relationship graph, configures different interaction weights and parameter sets, obtains high-order interaction features through multi-layer stacked explicit interaction and aggregation, and obtains semantic features through convolution of the business semantic subspace vectors; The low-order feature module inputs the numerical type fields from the multi-source data view and subjected to the standardization processing into the factorization machine branch, and calculates low-order features; The matching output module splices the low-order features, the high-order interaction features and the semantic features obtained from the business semantic subspace vectors, and calculates the basic matching scores of the target user and each candidate solution in a given business scenario and environmental context through the splicing results.

7. The enterprise digital management method based on multi-source data fusion according to claim 2, characterized in that, The extremely deep factorization machine model is used to calculate the positive sample and negative sample matching scores, the business importance coefficients of each triple are calculated according to the positive samples, and the exposure probability is calculated according to the negative samples, comprising: Based on the enterprise management operation log, behavior records are extracted from the multi-source data view, and the behavior records are classified according to the user identifier and the business scenario identifier; In the behavior records of each user and business scenario, the candidate solution with the display marked as true and the adoption marked as true is determined as a positive sample candidate solution, the candidate solution with the display marked as true and the adoption marked as false is determined as a negative sample candidate solution, and the user identifier, the positive sample candidate solution identifier and the negative sample candidate solution identifier under the same user and business scenario are combined to form a triple, thereby forming a triple set; For each triple, the corresponding user features, business scenario features, positive sample candidate solution features and negative sample candidate solution features are input into the improved extremely deep factorization machine model, and the positive sample basic matching score and the negative sample basic matching score are calculated respectively. The business importance coefficient of the current triple is determined according to the positive sample basic matching score and the business value associated with the positive sample through a mapping relationship, and the display times of the corresponding negative sample candidate scheme in the user's business scene and the total display times of all candidate schemes are counted, and the exposure probability of the current negative sample candidate scheme is calculated and determined according to the proportional relationship between the display times and the total display times.

8. The enterprise digital management method based on multi-source data fusion according to claim 2, characterized in that, The updated improved extreme deep factorization machine model is obtained, comprising: The Bayesian personalized ranking of the triple set is performed, and the positive sample basic matching score, the negative sample basic matching score, the business importance coefficient and the exposure probability corresponding to each triple are read; The adaptive interval parameter of each triple is determined by a monotonic mapping function according to the business importance coefficient, and the adaptive interval parameter is a non-negative real number. The matching score difference is calculated by subtracting the negative sample basic matching score from the positive sample basic matching score for each triple, and the modified difference is obtained by subtracting the corresponding adaptive interval parameter from the matching score difference; The ranking loss term is constructed by taking the modified difference as input, and the ranking loss term is the result of taking the opposite number of the modified difference as input, then inputting a logical function and taking the logarithm again; The exposure weighted loss term is obtained by multiplying the ranking loss term of each triple by the corresponding exposure probability, the sum of the exposure weighted loss terms of all triples is calculated, and the regularization term based on the improved extreme deep factorization machine model and the Bayesian personalized ranking parameter quadratic sum is added to form a total loss function; The total loss function is optimized, and the parameters of the improved extreme deep factorization machine model and the Bayesian personalized ranking are updated until the total loss function meets the preset convergence condition, and the updated improved extreme deep factorization machine model is obtained.

9. The enterprise digital management method based on multi-source data fusion according to claim 2, characterized in that, The ranking generates a recommendation management list, comprising: Monitoring the decision request initiated by the user in the procurement decision, production scheduling, inventory replenishment and task allocation scene in the process of enterprise digital management operation, determining the target user identifier and target business scene identifier according to the decision request, and screening the candidate scheme set from the multi-source data view according to the target business scene identifier; Extracting the user related field, scene related field and candidate scheme related field corresponding to the target user identifier, target business scene identifier and candidate scheme set from the multi-source data view, generating a feature domain embedding vector set, inputting the updated improved extreme deep factorization machine model with the numerical field, and calculating the matching score of the target user and each candidate scheme; According to the matching score, each candidate scheme is ranked to generate a recommendation management list arranged in descending order of the matching score, and output to the user interface of the enterprise digital management system, and record the display position, display times and user click and adoption behavior of each candidate scheme to obtain the management result.

Images