CN116362371A - Knowledge graph-based purchasing prediction system and method - Google Patents

Knowledge graph-based purchasing prediction system and method Download PDF

Info

Publication number
CN116362371A
CN116362371A CN202211672151.2A CN202211672151A CN116362371A CN 116362371 A CN116362371 A CN 116362371A CN 202211672151 A CN202211672151 A CN 202211672151A CN 116362371 A CN116362371 A CN 116362371A
Authority
CN
China
Prior art keywords
purchase
entity
information
purchasing
representation
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202211672151.2A
Other languages
Chinese (zh)
Inventor
高睿
霍胜军
杨熙鑫
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Qingdao Mengdou Network Technology Co ltd
Original Assignee
Qingdao Mengdou Network Technology Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Qingdao Mengdou Network Technology Co ltd filed Critical Qingdao Mengdou Network Technology Co ltd
Priority to CN202211672151.2A priority Critical patent/CN116362371A/en
Publication of CN116362371A publication Critical patent/CN116362371A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/02—Knowledge representation; Symbolic representation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/08—Logistics, e.g. warehousing, loading or distribution; Inventory or stock management
    • G06Q10/087—Inventory or stock management, e.g. order filling, procurement or balancing against orders
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30—Computing systems specially adapted for manufacturing

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • Economics (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Strategic Management (AREA)
  • Human Resources & Organizations (AREA)
  • Development Economics (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • Operations Research (AREA)
  • General Business, Economics & Management (AREA)
  • Marketing (AREA)
  • Finance (AREA)
  • Accounting & Taxation (AREA)
  • Game Theory and Decision Science (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a knowledge graph-based purchase prediction system and a knowledge graph-based purchase prediction method, wherein a purchase-purchase quantity interaction matrix Y and a knowledge graph G are given, a recommendation task is to predict the probability of the quantity of enterprise purchase items, the aim is to predict the relevant quantity of enterprise purchases by learning a prediction function through the knowledge graph and a randomization system, the system uses a random feature propagation framework, processes error information in purchase information through a random discarding strategy to predict the quantity of purchases, comprises a random missing component, a knowledge embedding propagation layer and a prediction layer, provides purchase list prediction for enterprises in a purchase stage, and generates an actual purchase prediction list for the enterprises by combining data such as production orders in Enterprise Resource Planning (ERP) in the enterprise production process and relevant data (such as order data, social prediction and the like) obtained by enterprise market departments and relevant information (such as purchase orders, actual purchases, stores and the like) in enterprise stores.

Description

Knowledge graph-based purchasing prediction system and method
Technical Field
The invention belongs to the field of data analysis mining, and particularly relates to a purchasing prediction system and method based on a knowledge graph.
Background
The present era is in an information explosion era, a large amount of information is updated at any time, such as our online browsing records, bank account balance records and the like, the data contains social development changes, online browsing records are increased, the internet is in vigorous development, bank account transfer records also contain economic development changes, and if we can reasonably utilize effective information, valuable energy contained in the information is extracted, so that the life of our is greatly facilitated, the social progress is promoted, and a valuable wealth is created for human society.
In recent years, network data updating also enters the era of rapid development, the development of small and medium-sized enterprises is greatly influenced, and the enterprises cannot possess the same male funds as large enterprises to cope with the problems of backlog of products, too slow information updating and the like. The medium and small enterprises are urgently required to extract the change trend of market economy and the change trend of product supply and demand relationship from a large amount of information data.
Under the environment that the data is updated quickly, a recommendation algorithm starts to be developed vigorously, and a basic recommendation algorithm can recommend commodities according to the history of a user, so that the time for the user to screen the commodities required by the user from massive data is greatly reduced. Only the recommendation algorithm can find the data relation from massive historical data, thereby playing a prediction effect on the development of future data and reducing the problems of warehouse backlog, overlarge purchasing quantity and market demand deviation of small and medium enterprises as much as possible.
In the purchasing system, when a user wants to determine a purchasing plan according to experience or related data, a long time is often required to determine purchasing ability, warehouse pressure accumulating ability, social demand and the like of the enterprise. The massive data can enable users to search the relationship among the data, and search the social development trend contained in the data according to the relationship among the data, but the difficulty of predicting information is increased to a certain extent due to the messy data.
In recent years, knowledge maps for recommendation systems have attracted considerable attention because these methods can capture structured information by associating items with their attributes, rather than using only interaction data between users and items. The knowledge graph is connected with the user and the article through different relations to obtain a candidate list which can be interpreted by the target user, the path between the target user and the recommended article is taken as an interpretation basis, but in the practical application of enterprises, false information and partial outdated information generated during enterprise competition are inevitably recorded, the accuracy of a prediction algorithm is reduced due to the data,
in a recommendation system, cold start and data sparseness can affect the accuracy of recommendation, even cause excessive fitting and cause misprediction, so that it is particularly important to ensure the accuracy of prediction when the data sparseness is achieved.
Disclosure of Invention
In order to solve the problems in the prior art, the purchasing prediction system and method based on the knowledge graph provided by the invention predict the relevant quantity of enterprise purchasing by learning a prediction function through the knowledge graph and the randomization system.
In order to achieve the above purpose, the present invention provides the following technical solutions:
in one aspect, the present invention provides a knowledge graph based purchase prediction system that uses a random feature propagation framework to predict purchase quantity by processing error information in purchase information through a random discard strategy, comprising a random loss component, a knowledge embedding propagation layer, and a prediction layer, wherein,
the random losing component is used for randomly discarding the element information to generate a disturbance feature matrix so as to prevent enterprise purchasing information from being influenced by false information generated by enterprise competition and outdated information generated by social market change;
the knowledge embedding propagation layer captures the structural information of the purchasing entity from the neighborhood thereof in the propagation process, learns potential high-order interests through repeated propagation processes, and deeply digs social market demand data and explores social market demand change data;
and the prediction layer is used for enhancing the inner product of the plan embedding and the purchase quantity embedding of the purchasing entity in the feature matrix and predicting the demand probability of the purchasing entity on the purchase quantity.
Optionally, the random missing component is givenA knowledge graph G, the original entity characteristic matrix E R is obtained by an initializer |ε|*d Epsilon is the set of all entities, |epsilon| is the number of entities, d is the dimension of the entity's feature, the discard policy is designed to generate a perturbed feature matrix, in order not to destroy the impact of the structural information, all elements of the partial entity feature full vector are randomly set to 0, some of the entity's feature representation is randomly discarded in the discard node, but the final structural and feature information is not damaged and incomplete, the erroneous information in the procurement information is randomly discarded, but warehouse inventory, market plan procurement, and actual procurement information are not affected.
Optionally, the knowledge embedding propagation layer generates a first-order representation of the entity according to the characteristic representation of the entity and the neighborhood entity thereof through a single knowledge graph convolution layer, explores the supply and demand relationship of the enterprise by robustly popularizing the single layer to multiple layers, and predicts the purchasing list of the enterprise according to the supply and demand relationship.
On the other hand, the invention provides a purchasing prediction method based on a knowledge graph, which comprises the following steps of:
step 1: processing the information of the purchasing entity of the enterprise, and processing for each entity i Randomly generating a binary mask co i Bernoulli (1-delta) is the probability of discarding the node, and the probability of performing discarding the node by the feature vector in the entity feature matrix is delta;
step 2: calculating the product of each row feature vector of entity purchase matrix and its mask according to purchase information and data prediction and the purchase limit required by enterprise resource planning
Figure SMS_1
Wherein->
Figure SMS_2
Is the ith row vector of matrix E, the expectation of ε is 1 (1- δ) +0.δ=1- δ, matrix +.>
Figure SMS_3
Is 1-delta multiplied by the matrix E in order to ensure that the perturbation feature matrix has the same as the original feature matrix EIt is desirable to obtain the final disturbance feature matrix by multiplying the matrix E by a factor 1/(1-delta)>
Figure SMS_4
The following is shown:
Figure SMS_5
step 3: the first order neighborhood representation of the entity with the relational weights is calculated as follows: given a purchasing entity u and a purchasing plan v, N (v) represents the set of 1-order neighbors of purchasing project plan v
Figure SMS_6
Representing entity e j And e i Relationship between them. The scoring function is used for calculating the score between the purchase and the purchase quantity, and represents the importance of the relation r to the purchase u, and the scoring function formula of the specific relation of the purchasing entity +.>
Figure SMS_7
Wherein u is e i d And r.epsilon.i d The feature vectors of the purchasing entity u and the relation r are respectively used, d is the dimension of the vector, and g is i d ×i d I is a microtransactable function;
step 4: to obtain the attention value corresponding to each type of relationship, pair
Figure SMS_8
The softmax function was applied as follows:
Figure SMS_9
n (e) is a first order neighbor of the purchasing entity set, e j And e i Is a subset of the collection of entities, the relationship between them being used
Figure SMS_10
A representation;
step 5: aggregating the purchasing entity neighborhood representation with the self-embedded representation into a vector to update the purchasing plan entity representation;
step 6: at the deepest layer, aggregating the neighborhood representation of the purchasing entity with the self representation generated in the previous information spreading step, and calculating the latest representation of the purchasing entity by using an aggregation function;
step 7: after the latest representation of the purchasing entity is calculated, the entity feature representation matrix is updated, feature embedding and transmission are carried out on each random discarding node, and the final representation of the entity contains high-order structural information in KG and reflects the importance of purchasing related information.
Step 8: in the drop node of layer s, the high-order representation of item v is defined as
Figure SMS_11
Calculating a purchase plan representation u and a purchase quantity representation +.>
Figure SMS_12
To predict probability;
step 9: in the model, a set S of enhancement feature matrices is generated, the vectors of the same entity in different enhancement matrices are consistent, and the distance between the two matrices is calculated by calculating the S-th feature matrix
Figure SMS_13
And (s-1) th feature matrix +.>
Figure SMS_14
The mean square error between the two is calculated, and the basic loss and the consistency regularization loss are combined together in each period;
step 10: and obtaining a target loss function, and ensuring the accuracy of data prediction according to the target loss function and consistency regularization.
Optionally, in step 5, the aggregator sums the two vectors before applying the nonlinear transformation, the neighborhood information is aggregated into the project layer v layer by layer, there is a first order relationship between the purchase plan and the directly adjacent purchase information, the necessity of modeling the higher order structure information and establishing a deeper network to capture the long-term interest of the user, and the single layer is extended to multiple layers to explore the higher order neighborhood aggregation information through multiple operations, so that the information factors affecting the purchase specific plan are aggregated in a higher order.
Optionally, in the step 8, a negative sampling strategy is selected for optimizing the recommendation model.
Compared with the prior art, the invention has the following beneficial effects:
1. the method solves the problem of predicting information, realizes the prediction of the purchasing information list of the enterprise, and improves the anti-interference performance of the enterprise in the modern rapid development social environment;
2. and optimizing the interpretable candidate list through a random system algorithm to obtain a final prediction list.
3. The method has better performance in the aspects of reducing excessive smoothing and predicting enterprise preference, particularly in a data sparse scene, the problem of larger prediction deviation in the data sparse scene is greatly relieved by adopting a random system algorithm, and the problems of excessive fitting and excessive smoothing are reduced to capture the higher-order and long-term supply and demand relations of enterprises, so that the recommendation accuracy is improved.
Drawings
Fig. 1 is a flowchart of a purchasing prediction method based on a knowledge graph according to an embodiment of the present invention.
Detailed Description
The following description of the technical solutions in the embodiments of the present invention will be clear and complete, and it is obvious that the described embodiments are only some embodiments of the present invention, but not all embodiments.
In this embodiment, given a purchase-purchase quantity interaction matrix Y and a knowledge graph G, the recommendation task is to predict the probability of the quantity of the purchased items of the enterprise, and the objective is to predict the relevant quantity of the purchase of the enterprise by learning a prediction function through the knowledge graph and the randomization system, where a specific prediction function formula is as follows:
y uv ＝F(u,v|θ,Y,G)
wherein y is uv And a score indicating the number v of purchased items of the predicted purchase demand u, and θ indicating the model parameter function F.
The embodiment of the application provides a purchasing prediction system based on a knowledge graph, which uses a random feature propagation framework to predict purchasing quantity by processing error information in purchasing information through a random discarding strategy, and mainly comprises three components: a random loss component, a knowledge embedding propagation layer, and a prediction layer, wherein:
1. the random discarding component generates a perturbation feature matrix by randomly discarding some of the element information to prevent enterprise procurement information from being affected by spurious information generated by enterprise competition and outdated information generated by social market changes. Giving a knowledge graph G, and obtaining an original entity characteristic matrix E R through an initializer |ε|*d . Epsilon is the set of all entities, |epsilon| is the number of entities and d is the dimension of the entity's feature. The discard strategy is designed to generate a perturbation feature matrix. In order not to destroy the influence of the structural information, all elements of the full vector of the characteristics of part of the entity are randomly set to 0, some characteristics of the entity are randomly discarded in the discarding node, but the final structural and characteristic information is not damaged or incomplete, and error information in purchase information is randomly discarded, but information such as warehouse inventory, market plan purchase, actual purchase and the like is not influenced. Because the entity information of the deletion feature is supplemented with its neighbor information according to the homogeneity assumption. Discarding the nodes also avoids the dependence on specific neighbors caused by deterministic propagation, prevents negative influence of false data on prediction due to outdated data, and improves the robustness of the model.
2. The knowledge embedding propagation layer captures the structural information of the purchasing entity from the neighborhood of the knowledge embedding propagation layer in the propagation process based on the GCN, learns potential high-order interests through repeated propagation processes, deeply digs social market demand data, explores social market demand change data and aims at making long-term targets for enterprise purchasing plans. In the knowledge graph, the enterprise purchasing entities are related by their relationships and have high-level connectivity. The neighbor set of an entity is extended by the entity associated with its information (e.g., warehouse inventory, market plan purchases, actual purchases, etc.). The first-order neighbors of an entity are defined as neighbors that are directly connected to the entity. Similarly, an n-order neighbor of an entity is a neighbor of an n-hop distance connected to the entity. Information of purchasing entity neighbors is aggregated to extend potential vectors of entities for predicting short-term needs of an enterprise purchasing inventory. And generating a first-order representation of the entity according to the characteristic representation of the entity and the neighborhood entity thereof through a single knowledge graph convolution layer. By steadily promoting the single layer to multiple layers, the long-term supply and demand relation of the enterprise is explored, the purchasing list of the enterprise is predicted according to the supply and demand relation, and high-quality long-term help is provided for the development of the enterprise.
3. And the prediction layer is used for making an inner product of the plan embedding and the purchase quantity embedding of the purchasing entity in the enhanced feature matrix and predicting the demand probability of the purchasing entity on the purchase quantity.
As shown in fig. 1, the present embodiment further provides a knowledge graph-based purchase prediction method, including the following steps:
step 1: firstly, business purchasing entity information is processed for each entity co i Randomly generating a binary mask co i Bernoulli (1-delta) is the probability of dropping a node. The probability that a feature vector in the entity feature matrix performs discarding a node is δ.
Step 2: calculating the product of each row feature vector of the entity purchase matrix and the mask thereof according to the actual purchase information and data prediction of the market part and the purchase limit required by the enterprise resource planning
Figure SMS_15
Wherein->
Figure SMS_16
Is the i-th row vector of matrix E. The expectation of e is 1· (1- δ) +0·δ=1- δ, matrix +.>
Figure SMS_17
Is 1-delta times the matrix E. To ensure that the disturbance feature matrix has the same expectations as the original feature matrix E, the final disturbance feature matrix is obtained by multiplying the matrix E by a factor of 1/(1-delta)>
Figure SMS_18
The following is shown:
Figure SMS_19
step 3: the first order neighborhood representation of the entity with the relational weights is calculated as follows: given a purchasing entity u and a purchasing plan v, N (v) represents a collection of 1-order neighbors of purchasing project plan v (entity directly connected to the project), while
Figure SMS_20
Representing entity e j And e i Relationship between them. The scoring function is used for calculating the score between the purchase and the purchase quantity, and represents the importance of the relation r to the purchase u, and the scoring function formula of the specific relation of the purchasing entity +.>
Figure SMS_21
Wherein u is e i d And r.epsilon.i d The feature vectors of the purchasing entity u and the relation r are respectively used, d is the dimension of the vector, and g is i d ×i d I is a microtransactable function;
step 4: to obtain the attention value corresponding to each type of relationship, pair
Figure SMS_22
The softmax function was applied as follows:
Figure SMS_23
n (e) is a first order neighbor of the purchasing entity set, e j And e i Is a subset of the collection of entities, the relationship between them being used
Figure SMS_24
And (3) representing. The relationships between entities in the knowledge graph are complex, resulting in a large number of neighbors between the entities. A set of fixed-size neighbors is set for each entity to achieve more efficient computation, and duplicate entities are selected as N (e) from all neighbors of an entity to keep the sizes consistent. Using multiple discard nodes, each entity randomly aggregates a portion of neighbor information each time, which facilitates modulo operationRobustness of the model. The neighborhood information for an entity is represented by a linear combination of each neighborhood representation of the entity.
Step 5: the purchasing plan entity representation is updated by aggregating the purchasing entity neighborhood representation with the self-embedded representation into a vector. The aggregator sums the two vectors before applying the nonlinear transformation, the neighborhood information is aggregated layer by layer to the project layer v, a first order relationship exists between the purchasing plan and the directly adjacent purchasing information, the necessity of modeling the higher order structural information and establishing a deeper network to capture the long-term interest of the user, and the single layer is expanded to multiple layers through multiple operations to explore the higher order neighborhood aggregation information, so that the information factors influencing the purchasing specific plan are aggregated in a higher order mode to explore the long-term social market trend and the enterprise purchasing plan trend.
Step 6: at the deepest level, the neighborhood representation of the purchasing entity is aggregated with the self-representation generated by the previous information dissemination step, and the latest representation of the purchasing entity is calculated using an aggregation function.
Step 7: after the latest representation of the purchasing entity is calculated, the entity characteristic representation matrix is updated, and characteristic embedding and propagation are carried out on each random discarding node. The final representation of the entity contains higher-order structural information in the KG and reflects the importance of the purchase related information.
Step 8: in the drop node of layer s, the high-order representation of item v is defined as
Figure SMS_25
Calculating a purchase plan representation u and a purchase quantity representation +.>
Figure SMS_26
To predict probabilities. The invention selects the negative sampling strategy to effectively optimize the recommendation model.
Step 9: in the model, a set S of enhancement feature matrices is generated, the vectors of the same entity in different enhancement matrices are consistent, and the distance between the two matrices is calculated by calculating the S-th feature matrix
Figure SMS_27
And (s-1)Feature matrix
Figure SMS_28
And the mean square error between. At each epoch, the base loss and the consistency regularization loss are combined. And finally, obtaining a target loss function, and ensuring the accuracy of data prediction according to the target loss function and consistency regularization.
The system and the method provided by the embodiment provide a purchase list prediction for enterprises in a purchase stage, and combine data such as production orders in Enterprise Resource Planning (ERP) in the enterprise production process and related data (such as order data, social prediction and the like) taken by enterprise market departments and related information (such as purchase orders, actual purchases, warehouse inventory and the like) in enterprise warehouses to generate an actual purchase prediction list for the purchase departments for the enterprises, so that the problems of backlog of the enterprise warehouses, overlarge deviation of purchase quantity and market demand and the like are alleviated.
The foregoing is only a preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art, who is within the scope of the present invention, should make equivalent substitutions or modifications according to the technical scheme of the present invention and the inventive concept thereof, and should be covered by the scope of the present invention.

Claims (6)

1. A knowledge graph-based purchase prediction system is characterized in that the system uses a random feature propagation framework to predict the purchase quantity by processing error information in purchase information through a random discard strategy, and comprises a random loss component, a knowledge embedding propagation layer and a prediction layer,
the random losing component is used for randomly discarding the element information to generate a disturbance feature matrix so as to prevent enterprise purchasing information from being influenced by false information generated by enterprise competition and outdated information generated by social market change;
the knowledge embedding propagation layer captures the structural information of the purchasing entity from the neighborhood thereof in the propagation process, learns potential high-order interests through repeated propagation processes, and deeply digs social market demand data and explores social market demand change data;
and the prediction layer is used for enhancing the inner product of the plan embedding and the purchase quantity embedding of the purchasing entity in the feature matrix and predicting the demand probability of the purchasing entity on the purchase quantity.
2. The knowledge-based purchase prediction system as recited in claim 1, wherein the random losing component gives a knowledge pattern G, and obtains an original entity feature matrix E R through an initializer |ε|*d Epsilon is the set of all entities, |epsilon| is the number of entities, d is the dimension of the entity's feature, the discard policy is designed to generate a perturbed feature matrix, in order not to destroy the impact of the structural information, all elements of the partial entity feature full vector are randomly set to 0, some of the entity's feature representation is randomly discarded in the discard node, but the final structural and feature information is not damaged and incomplete, the erroneous information in the procurement information is randomly discarded, but warehouse inventory, market plan procurement, and actual procurement information are not affected.
3. The knowledge-graph-based purchase prediction system of claim 1, wherein the knowledge embedding propagation layer generates a first-order representation of an entity according to the characteristic representation of itself and its neighborhood entities through a single knowledge-graph convolution layer, explores supply and demand relationships of an enterprise by robustly popularizing a single layer to multiple layers, and predicts an enterprise purchase list according to the supply and demand relationships.
4. The purchasing prediction method based on the knowledge graph is characterized by comprising the following steps of:
step 1: processing the information of the enterprise purchasing entity for each entity
Figure QLYQS_1
Randomly generating a binary mask
Figure QLYQS_2
Probability of discarding a node, entityThe probability that the feature vector in the feature matrix performs discarding the node is δ;
step 2: calculating the product of each row feature vector of entity purchase matrix and its mask according to purchase information and data prediction and the purchase limit required by enterprise resource planning
Figure QLYQS_3
Wherein->
Figure QLYQS_4
Is the ith row vector of matrix E, the expectation of ε is 1 (1- δ) +0.δ=1- δ, matrix +.>
Figure QLYQS_5
Is 1-delta multiplied by the matrix E to ensure that the disturbance feature matrix has the same expectations as the original feature matrix E, the final disturbance feature matrix is obtained by multiplying the matrix E by a factor of 1/(1-delta)>
Figure QLYQS_6
The following is shown:
Figure QLYQS_7
step 3: the first order neighborhood representation of the entity with the relational weights is calculated as follows: given a purchasing entity u and a purchasing plan v, N (v) represents the set of 1-order neighbors of purchasing project plan v
Figure QLYQS_8
Representing entity e j And e i Relationship between them. The scoring function is used for calculating the score between the purchase and the purchase quantity, and represents the importance of the relation r to the purchase u, and the scoring function formula of the specific relation of the purchasing entity +.>
Figure QLYQS_9
Wherein u is e i d And r.epsilon.i d The feature vectors of the purchasing entity u and the relation r are respectively usedD is the dimension of the vector, g: i d ×i d I is a microtransactable function;
step 4: to obtain the attention value corresponding to each type of relationship, pair
Figure QLYQS_10
The softmax function was applied as follows:
Figure QLYQS_11
n (e) is a first order neighbor of the purchasing entity set, e j And e i Is a subset of the collection of entities, the relationship between them being used
Figure QLYQS_12
A representation;
step 5: aggregating the purchasing entity neighborhood representation with the self-embedded representation into a vector to update the purchasing plan entity representation;
step 6: at the deepest layer, aggregating the neighborhood representation of the purchasing entity with the self representation generated in the previous information spreading step, and calculating the latest representation of the purchasing entity by using an aggregation function;
step 7: after the latest representation of the purchasing entity is calculated, the entity feature representation matrix is updated, feature embedding and transmission are carried out on each random discarding node, and the final representation of the entity contains high-order structural information in KG and reflects the importance of purchasing related information.
Step 8: in the drop node of layer s, the high-order representation of item v is defined as
Figure QLYQS_13
Calculating a purchase plan representation u and a purchase quantity representation +.>
Figure QLYQS_14
To predict probability;
step 9: in the model, a set S of enhancement feature matrices is generated, the vectors of the same entity in different enhancement matrices beingIn agreement, the distance between the two matrices is determined by calculating the S-th feature matrix
Figure QLYQS_15
And (s-1) th feature matrix +.>
Figure QLYQS_16
The mean square error between the two is calculated, and the basic loss and the consistency regularization loss are combined together in each period;
step 10: and obtaining a target loss function, and ensuring the accuracy of data prediction according to the target loss function and consistency regularization.
5. The knowledge-graph-based purchase prediction method according to claim 4, wherein in the step 5, the aggregator sums two vectors before applying the nonlinear transformation, the neighborhood information is aggregated to the item layer v layer by layer, a first order relationship exists between the purchase plan and the directly adjacent purchase information, the necessity of modeling the higher order structure information and establishing a deeper network to capture the long-term interest of the user, and the single layer is extended to multiple layers to explore the higher order neighborhood aggregation information through multiple operations to perform the higher order aggregation of the information factors affecting the purchase specific plan.
6. The knowledge-graph-based purchase prediction method according to claim 4, wherein in the step 8, a negative sampling strategy is selected for optimizing the recommendation model.
CN202211672151.2A 2022-12-26 2022-12-26 Knowledge graph-based purchasing prediction system and method Pending CN116362371A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211672151.2A CN116362371A (en) 2022-12-26 2022-12-26 Knowledge graph-based purchasing prediction system and method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211672151.2A CN116362371A (en) 2022-12-26 2022-12-26 Knowledge graph-based purchasing prediction system and method

Publications (1)

Publication Number Publication Date
CN116362371A true CN116362371A (en) 2023-06-30

Family

ID=86939522

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211672151.2A Pending CN116362371A (en) 2022-12-26 2022-12-26 Knowledge graph-based purchasing prediction system and method

Country Status (1)

Country Link
CN (1) CN116362371A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116720819A (en) * 2023-08-10 2023-09-08 福建省闽清双棱纸业有限公司 Impregnated paper raw material management system integrating knowledge graph and neural network

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116720819A (en) * 2023-08-10 2023-09-08 福建省闽清双棱纸业有限公司 Impregnated paper raw material management system integrating knowledge graph and neural network
CN116720819B (en) * 2023-08-10 2023-10-27 福建省闽清双棱纸业有限公司 Impregnated paper raw material management system integrating knowledge graph and neural network

Similar Documents

Publication Publication Date Title
Li et al. Dealing with uncertainty: A survey of theories and practices
Lee et al. Retracted: A hybrid artificial intelligence sales‐forecasting system in the convenience store industry
CN111127142A (en) Article recommendation method based on generalized neural attention
Fu et al. Deep reinforcement learning framework for category-based item recommendation
CN108665156B (en) Supply chain selection evaluation method based on Markov chain under block chain
Liu et al. A multi-objective model for discovering high-quality knowledge based on data quality and prior knowledge
CN116362371A (en) Knowledge graph-based purchasing prediction system and method
Lupiani et al. Evaluating case-base maintenance algorithms
Li et al. A rule-based knowledge discovery engine embedded semantic graph knowledge repository for retail business
Wan et al. Similarity-based sales forecasting using improved ConvLSTM and prophet
Miller et al. Interval type-2 fuzzy modelling and stochastic search for real-world inventory management
Zuo et al. A Data-Driven Customer Profiling Method for Offline Retailers
Ocampo et al. A sustainable manufacturing strategy from different strategic responses under uncertainty
CN115344794A (en) Scenic spot recommendation method based on knowledge map semantic embedding
Xin Application of optimized support vector machine model in tax forecasting system
Yoon et al. Methodology for Improving the Performance of Demand Forecasting Through Machine Learning
Wei et al. Modeling multilevel supplier selection problem based on weighted‐directed network and its solution
Logeswaran et al. High utility itemset mining using genetic algorithm assimilated with off policy reinforcement learning to adaptively calibrate crossover operation
Jiang et al. Nonlinear time series fuzzy regression for developing explainable consumer preferences’ models based on online comments
Murthy et al. Model for Predicting Prospective Big-Mart Sales Based on Grid Search Optimization (GSO)
Feng [Retracted] Sales Forecast of Marketing Brand Based on BP Neural Network Model
Yang Sales Prediction of Walmart Sales Based on OLS, Random Forest, and XGBoost Models
Chen et al. Semi-supervised heterogeneous graph learning with multi-level data augmentation
Guan Neural Network Model of Dynamic Prediction of Cross-Border E-Commerce Sales for Virtual Community Knowledge Sharing
CN116089722B (en) Implementation method, device, computing equipment and storage medium based on graph yield label

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination