Knowledge enhancement method and system based on graph event model
By converting quantitative and qualitative background knowledge into soft constraints, the problem of graph event model learning dependence on data and inaccurate knowledge is solved, and efficient model learning and accurate prediction in an environment of insufficient data are achieved.
Patent Information
- Application Number
- CN202510340371.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-21
- Publication Date
- 2025-07-08
AI Technical Summary
Existing graph event model learning methods rely too much on complete data and over-reliance on inaccurate knowledge, resulting in high training costs and degraded model prediction performance.
By collecting quantitative and qualitative background knowledge in the target field, converting it into soft constraints, and defining incompatible functions to add them to the likelihood function, forming a regularized objective function, and optimizing the graph event model.
In the case of insufficient data, flexible integration of prior knowledge can be enhanced to enhance model learning ability, avoid the negative impact of hard logic rules, and improve the fit and accuracy of the model.
Smart Images
Figure CN120278255A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of knowledge enhancement, and more specifically, to a knowledge enhancement method and system based on a graph event model. Background Art
[0002] The graph event model is a type of probabilistic graph model used to represent the dynamic relationships between event streams and can be used to model event sequences. It requires a large amount of data to learn an accurate temporal dependence structure. Large-scale pre-trained language models encode rich world knowledge through unsupervised learning, providing new ideas for knowledge enhancement. However, in many practical applications, only limited data is available, making it very difficult to rely solely on data to learn an accurate graph event model. Therefore, studying how to use relevant domain knowledge to assist graph event model learning in the case of insufficient data has become an important research topic. Applying the structural knowledge and parameter knowledge in the knowledge base to graph event model learning can make up for the limitation of insufficient data and push the model to a higher performance limit. The combination of these two technologies provides an effective approach for knowledge-driven artificial intelligence. The knowledge in the knowledge base is of high quality and wide range, which can significantly improve the learning ability of the graph event model for practical applications. The combination of the two can achieve deep collaboration between knowledge and data, and establish a new type of intelligent system with knowledge enhancement while ensuring quality. Existing methods mainly rely on prior knowledge to define the parameter space of the event model, which requires users to provide a large amount of quantitative knowledge. However, in practice, users are more likely to obtain qualitative knowledge, such as "event A depends on event B". Therefore, how to use simple qualitative knowledge to enhance graph event model learning is particularly crucial.
[0003] The prior art has proposed a new type of method for graph event models, called the piecewise constant conditional intensity model. This model assumes that the conditional incidence rate function of each event type takes different constant values according to the most recent history of its parent nodes. Therefore, it can represent the temporal dependence relationship between events. Then the prior solution gives corresponding model parameter and structure learning methods. For parameter learning, a Bayesian estimation framework is proposed, assuming that the prior distribution of the parameters is a gamma distribution, calculating the posterior distribution through data and obtaining the point estimate of the parameters. For structure learning, a Bayesian scoring function is defined, and the local optimal graph structure is found through forward-backward greedy search. Finally, the effectiveness of the proposed model is verified on synthetic and real data. In summary, the key innovation of this prior art is to propose a new type of temporal dependence graph event model and give parameter estimation and structure learning solutions, which can better model the temporal relationships in event sequences.
[0004] However, this existing technology also has drawbacks. Pure data-driven learning methods need to collect a large amount of training sets to avoid overfitting problems, resulting in too high training costs. Such methods attempt to learn complex graph event models only from data without using any domain prior knowledge. However, learning an accurate time-dependent structure requires a large amount of event data, usually far exceeding the amount of independent and identically distributed data required by categorical Bayesian networks. Therefore, such methods need to consume a large amount of resources to collect massive training data. However, it is very difficult to obtain large-scale event data in many practical applications, making the training cost of pure data-driven methods too high to meet the requirements.
[0005] The existing technology also proposes a new class of methods for time-domain logic-driven graph event models. First, a time-domain logic language is designed, which can represent the temporal dependence relationships between events, such as increasing the probability of event B occurring after event A occurs. Then, these logical rules are mapped into parameter constraints in the new graph event model, that is, a legal parameter space that satisfies the logical rules is generated. On this basis, the maximum likelihood estimation problem is defined to learn the parameters of the graph event model that satisfy the prior time-domain logical rules. Finally, this method demonstrates the representation ability of some time-domain logical rules and verifies the effectiveness of modeling using logical rules on multiple data sets. Generally speaking, the innovation of this existing technology lies in proposing the idea of driving graph event model learning with time-domain logic and explicitly injecting prior structural knowledge through predefined logical rules.
[0006] However, this existing technology also has drawbacks. The hard logic rule method is prone to encoding incorrect prior knowledge, resulting in a high error rate. Such methods simply encode qualitative knowledge as hard logic rules, such as first-order logical expressions. However, background knowledge is usually uncertain and incomplete. If inaccurate knowledge is wrongly hard-coded, it is very easy to cause a large deviation in the learning results, ultimately leading to a decline in the model prediction performance and a high error rate. This hard-coding method also lacks the flexibility to handle uncertain knowledge. Generally speaking, the hard logic rule method relies too much on incomplete prior knowledge and is prone to keeping the error rate high. Summary of the Invention
[0007] The present invention aims to provide a knowledge enhancement method and system based on a graph event model to solve the technical problems that the existing graph event model learning methods rely too much on complete data and overly rely on inaccurate knowledge.
[0008] The technical solution adopted by the present invention to solve its technical problems is: a knowledge enhancement method based on a graph event model, including the following steps:
[0009] Collect background knowledge of the target domain; wherein, the background knowledge includes quantitative knowledge and qualitative knowledge;
[0010] Convert the background knowledge into soft constraint conditions through knowledge encoding;
[0011] Define an incompatibility function by evaluating the degree of incompatibility between the soft constraints and the graph event model;
[0012] Add the incompatibility function to the likelihood function of the graph event model to form a regularized objective function;
[0013] Obtain an enhanced graph event model by optimizing the regularized objective function.
[0014] Preferably, the collection of background knowledge related to the target domain includes:
[0015] Collect an event dataset for the target domain; the event dataset includes the dependencies and parameter distributions of different types of events;
[0016] Obtain background knowledge through pre-trained language for the dependencies and parameter distributions of different types of events; the background knowledge includes qualitative knowledge of dependencies and quantitative knowledge of parameter distributions.
[0017] Preferably, the event dataset is represented as:
[0018]
[0019] where the time label D k represents the k-th event stream; K represents the total number of event streams; t represents time; represents the specific occurrence time of the i-th event in the k-th stream.
[0020] Preferably, the conversion of the background knowledge into soft constraints through knowledge encoding includes:
[0021] Convert the qualitative knowledge of dependencies and the quantitative knowledge of parameter distributions into soft constraints through knowledge encoding; specifically:
[0022] Map the qualitative knowledge to process dependency constraints and map the quantitative knowledge to parameter inequality constraints.
[0023] Preferably, the definition of the incompatibility function by evaluating the degree of incompatibility between the soft constraints and the graph event model includes:
[0024] Utilize the degree of parameter constraint incompatibility of the graph event model to obtain an incompatibility function for evaluating parameter constraints;
[0025]
[0026] Utilize the degree of structural constraint incompatibility of the graph event model to obtain an incompatibility function for evaluating structural constraints;
[0027]
[0028] where f(·) + represents max(0, f(·)), I X represents a set of inequality statements for node X; Λ X represents a set of conditional intensity parameters for each node X; P X represents a set of process dependence or independence constraints associated with node X; U X represents the set of parent nodes of node X; h i represents a binary function whose value depends on the process dependence or independence constraints whether it is consistent with the set of parent nodes U of node X X If the constraint is consistent with U X is consistent, then h i returns 1, otherwise returns -1.
[0029] Preferably, adding the incompatible function to the likelihood function of the graph event model to form a regularized objective function includes:[[]]
[0030] The regularized objective function is specifically:[[]]
[0031] LL' = LL - w * f(C, M)
[0032] where LL represents the log-likelihood value; w represents the weight factor; C represents the set of background knowledge; M represents the specific event model.
[0033] Preferably, the enhanced graph event model includes the dependence relationship structure and parameters between events.
[0034] Preferably, a knowledge enhancement system based on the graph event model applies the knowledge enhancement method based on the graph event model, including:[[]]
[0035] A knowledge collection module for collecting background knowledge in the target domain; among them, the background knowledge includes quantitative knowledge and qualitative knowledge;
[0036] A knowledge encoding module for converting the background knowledge into soft constraint conditions through knowledge encoding;
[0037] An evaluation module for defining an incompatible function by evaluating the incompatibility degree between the soft constraint conditions and the graph event model;
[0038] A target function construction module for adding the incompatible function to the likelihood function of the graph event model to form a regularized objective function;
[0039] A model learning module for obtaining an enhanced graph event model by optimizing the regularized objective function.
[0040] The beneficial effects of the present invention are:[[]]
[0041] Compared with the prior art, the present application provides a knowledge enhancement method and system based on a graph event model to flexibly integrate different forms of prior knowledge with data, including quantitative knowledge and qualitative knowledge, so as to enhance the learning of the graph event model in an environment with insufficient data; a function for quantifying the incompatibility degree between the prior knowledge and the model is defined; the negative impact brought by simple hard logic rules is avoided, and the data is used to correct inaccurate prior knowledge to maintain the fitness to the data. Description of the Drawings
[0042] Figure 1 is a schematic flowchart of a knowledge enhancement method based on a graph event model according to the present invention;
[0043] Figure 2 is a schematic diagram of the modules of a knowledge enhancement system based on a graph event model according to the present invention.
[0044] The drawings are only for illustrative purposes and should not be construed as a limitation of the present invention; for better illustration of the embodiments, some components in the drawings will be omitted, enlarged or reduced, and do not represent the dimensions of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted. Detailed Embodiments
[0045] The present invention will be further described in detail below with reference to the drawings and specific embodiments.
[0046] As Figure 1 shown, a knowledge enhancement method based on a graph event model of the present invention includes the following steps:
[0047] S1: Collect background knowledge of the target domain; wherein, the background knowledge includes quantitative knowledge and qualitative knowledge;
[0048] S2: Convert the background knowledge into soft constraint conditions through knowledge encoding;
[0049] S3: Define an incompatibility function by evaluating the incompatibility degree between the soft constraint conditions and the graph event model;
[0050] S4: Add the incompatibility function to the likelihood function of the graph event model to form a regularized objective function;
[0051] S5: Obtain an enhanced graph event model by optimizing the regularized objective function.
[0052] In the above solution, the present invention proposes a knowledge enhancement method, which can softly introduce domain prior knowledge to make up for the incompleteness of data and enhance the model learning ability. Whether it is a quantitative joint probability distribution prior or a simple qualitative dependence relationship, it can be used to assist model training; it solves the problem that the existing graph event model learning method is too dependent on complete data and the effect is reduced in the case of incomplete actual data. The present invention introduces knowledge in the form of soft constraint conditions, defines a function to quantify the degree of incompatibility between prior knowledge and the model, avoids the negative impact brought by simple hard logic rules, and at the same time uses data to correct inaccurate prior knowledge and maintain the fitness to data; it solves the problem that the existing method uses hard logic rules to introduce prior knowledge, which lacks flexibility and may over-rely on inaccurate knowledge.
[0053] Preferably, in step S1, the collection of background knowledge related to the target domain includes:
[0054] Collect the event dataset of the target domain; the event dataset includes the dependence relationship and parameter distribution of different types of events;
[0055] Obtain background knowledge through pre-trained language for the dependence relationship and parameter distribution of different types of events; the background knowledge includes qualitative knowledge of the dependence relationship and quantitative knowledge of the parameter distribution.
[0056] In the above solution, relevant background knowledge is obtained through a large-scale pre-trained language model, including qualitative knowledge such as event dependence relationships and quantitative knowledge such as parameter distributions. The pre-trained language model, that is, a knowledge base that encodes a large amount of world knowledge, can be used as a source for extracting structural knowledge and parameter knowledge, replacing the knowledge provided by experts, and using the knowledge provided by the pre-trained model, which is richer and of higher quality, can further enhance the learning effect of the graph event model. Finally, the knowledge of the pre-trained language model can also be used to evaluate the quality of the learned model.
[0057] Obtain the qualitative knowledge S and quantitative knowledge Q of the target domain through expert provision or other means. S can be dependence relationship knowledge, and Q can be parameter prior distribution knowledge, such as λ x|y ~Gamma(a,b).
[0058] Preferably, the event dataset is expressed as:
[0059]
[0060] Among them, the time tag D k represents the kth event stream; K represents the total number of event streams; t represents time; represents the specific occurrence time of the ith event in the kth stream.
[0061] Preferably, in step S2, the background knowledge is converted into soft constraint conditions through knowledge encoding, including:
[0062] Converting the qualitative knowledge of the dependency relationship and the quantitative knowledge of the parameter distribution into soft constraint conditions through knowledge encoding; specifically:
[0063] Mapping the qualitative knowledge to process dependency constraints and the quantitative knowledge to parameter inequality constraints.
[0064] In the above solution, the qualitative knowledge S is mapped to the process dependency constraint P, such as "Y→X". The quantitative knowledge Q is mapped to the parameter inequality constraint I, such as λ x|y ≥λ x|z .
[0065] Preferably, in step S3, defining the incompatibility function by evaluating the degree of incompatibility between the soft constraint conditions and the graph event model includes:
[0066] Utilizing the degree of parameter constraint incompatibility of the graph event model to obtain the incompatibility function for evaluating parameter constraints;
[0067]
[0068] Utilizing the degree of structural constraint incompatibility of the graph event model to obtain the incompatibility function for evaluating structural constraints;
[0069]
[0070] where f(·) + represents max(0,f(·)), I X represents a set of inequality statements for node X; Λ X represents the set of conditional intensity parameters for each node X; P X represents a set of process dependencies or independence constraints related to node X; U X represents the set of parent nodes of node X; h i represents a binary function whose value depends on the process dependency or independence constraint is consistent with the set of parent nodes U of node X X If the constraint is consistent with U X is consistent, then h i returns 1, otherwise returns -1.
[0071] Preferably, in step S4, adding the incompatibility function into the likelihood function of the graph event model to form the regularization objective function includes:
[0072] The regularization objective function is specifically:
[0073] LL' = LL - w*f(C,M)
[0074] Among them, LL represents the log-likelihood value; w represents the weight factor; C represents the background knowledge set; M represents the specific event model.
[0075] Preferably, the enhanced graph event model includes the dependency relationship structure and parameters between events.
[0076] In the above solution, this method can integrate quantitative knowledge and qualitative knowledge, enhance the learning of the graph event model in a flexible manner, and expand the knowledge source; the qualitative knowledge is added in the form of soft constraints to avoid over-reliance on inaccurate knowledge; this method can obtain good model learning effects even in an environment with insufficient data, and the model accuracy in the case of data sparsity can be significantly improved by introducing background knowledge for regularization.
[0077] Preferably, as Figure 2 shown, a knowledge enhancement system based on a graph event model, which applies the knowledge enhancement method based on a graph event model, includes:
[0078] A knowledge collection module, which is used to collect background knowledge of the target domain; among them, the background knowledge includes quantitative knowledge and qualitative knowledge;
[0079] A knowledge encoding module, which is used to convert the background knowledge into soft constraint conditions through knowledge encoding;
[0080] An evaluation module, which is used to define an incompatibility function by evaluating the incompatibility degree between the soft constraint conditions and the graph event model;
[0081] A target function construction module, which is used to add the incompatibility function to the likelihood function of the graph event model to form a regularized target function;
[0082] A model learning module, which is used to obtain an enhanced graph event model by optimizing the regularized target function.
[0083] In the above solution, a knowledge enhancement system based on a graph event model is proposed, which can integrate quantitative knowledge and qualitative knowledge and enhance the graph event model learning in a flexible manner. The key of this system is to define an incompatibility function, convert knowledge into soft constraint conditions, and add them to the target function in the form of regularization. This system gives specific method examples, including converting directed relationship knowledge into parameter inequality constraints and converting dependency relationship knowledge into graph structure constraints to enhance parameter estimation and structure learning.
[0084] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, rather than limitations on the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation manners here. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the claims of the present invention.
Claims
1. A knowledge enhancement method based on a graph event model, characterized in that It includes the following steps: Collect background knowledge of the target domain; among them, the background knowledge includes quantitative knowledge and qualitative knowledge; Convert the background knowledge into soft constraint conditions through knowledge encoding; Define an incompatibility function by evaluating the degree of incompatibility between the soft constraint conditions and the graph event model; Add the incompatibility function to the likelihood function of the graph event model to form a regularized objective function; Obtain an enhanced graph event model by optimizing the regularized objective function.
2. The knowledge enhancement method based on a graph event model according to claim 1, wherein The collection of background knowledge related to the target domain includes: Collect the event dataset of the target domain; the event dataset includes the dependency relationships and parameter distributions of different types of events; Obtain background knowledge through pre-trained language for the dependency relationships and parameter distributions of different types of events; the background knowledge includes the qualitative knowledge of the dependency relationships and the quantitative knowledge of the parameter distributions.
3. The knowledge enhancement method based on a graph event model according to claim 2, wherein The event dataset is represented as: Among them, the time tag D k represents the k-th event stream; K represents the total number of event streams; t represents time; represents the specific occurrence time of the i-th event in the k-th stream.
4. The knowledge enhancement method based on a graph event model according to claim 3, wherein The conversion of the background knowledge into soft constraint conditions through knowledge encoding includes: Convert the qualitative knowledge of the dependency relationships and the quantitative knowledge of the parameter distributions into soft constraint conditions through knowledge encoding; specifically: Map the qualitative knowledge to process dependency constraints and map the quantitative knowledge to parameter inequality constraints.
5. The knowledge enhancement method based on a graph event model according to claim 4, characterized in that, The definition of the incompatibility function by evaluating the degree of incompatibility between the soft constraint conditions and the graph event model includes: Utilize the degree of incompatibility between the parameter constraints and the structural constraints of the graph event model to obtain incompatibility functions for respectively evaluating the parameter constraints and the structural constraints.
6. The knowledge enhancement method based on the graph event model according to claim 5, wherein The incompatibility function for evaluating the parameter constraints is: The incompatibility function for evaluating the structural constraints is: where f(·) + denotes max(0, f(·)), I X denotes a set of inequality statements for node X; Λ X denotes the set of conditional intensity parameters for each node X; P X denotes a set of process dependencies or independence constraints associated with node X; U X denotes the set of parent nodes of node X; h i denotes a binary function whose value depends on the process dependencies or independence constraints whether it is consistent with the set of parent nodes U of node X X If the constraint is consistent with U X is consistent, then h i returns 1, otherwise returns -1.
7. The knowledge enhancement method based on the graph event model according to claim 6, characterized in that The addition of the incompatibility function to the likelihood function of the graph event model to form a regularized objective function includes: The regularized objective function is specifically: LL' = LL - w * f(C, M) Where, LL represents the log-likelihood value; w represents the weight factor; C represents the set of background knowledge; M represents the specific event model.
8. The knowledge enhancement method based on a graph event model according to claim 7, characterized in that The enhanced graph event model includes the dependency relationship structure and parameters between events.
9. A knowledge enhancement system based on a graph event model, which applies a knowledge enhancement method based on a graph event model described in claim 1, characterized in that, It includes: A knowledge collection module for collecting background knowledge of the target domain; among them, the background knowledge includes quantitative knowledge and qualitative knowledge; A knowledge encoding module for converting the background knowledge into soft constraint conditions through knowledge encoding; An evaluation module for defining an incompatibility function by evaluating the degree of incompatibility between the soft constraint conditions and the graph event model; An objective function construction module for adding the incompatibility function to the likelihood function of the graph event model to form a regularized objective function; A model learning module for obtaining an enhanced graph event model by optimizing the regularized objective function.