Intelligent auxiliary decision-making model training method based on expert knowledge and data fusion

Abstract

The invention belongs to the technical field of artificial intelligence, discloses an expert knowledge and data fused intelligent aid decision-making model training method, and aims to solve the problem that expert knowledge and data are difficult to synergistically utilize by an intelligent model, the expert knowledge is converted into a feature and strategy bound decision-making sequence, so that the expert knowledge becomes a data sample; a way is provided for utilizing expert knowledge; a business target state is adopted to evaluate the strategy effect, the expert sample weight and the parameter learning rate are dynamically adjusted according to the model training process, the expert knowledge utilization effect is improved, and meanwhile, new expert samples are dynamically injected in a model evaluation mode; a sample fusion matching mode is adopted to eliminate redundancy of expert samples and data samples, the model training process is accelerated, a sample cluster with a small number of samples is identified through a density clustering mode, the number of similar samples is increased by taking the sample cluster as a starting point, and the robustness of an auxiliary decision-making model is enhanced through continuous improvement of sample richness.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a training method for an intelligent auxiliary decision-making model that integrates expert knowledge and data. Background Technology

[0002] The continuous development of artificial intelligence technology has prompted various industries to implement digital transformation and upgrading, making the construction of intelligent models crucial for intelligent empowerment. In complex business processes, how to improve the effectiveness of business decision-making based on the capabilities of intelligent models is a problem that intelligent empowerment must solve. Traditional intelligent models derive their capabilities from data, that is, fitting patterns into massive amounts of data and then solidifying and enforcing model capabilities under multiple parameter combinations by adjusting model parameters. However, many industries have accumulated years of operational experience based on expert knowledge. Unlike the diversity and richness of data, expert knowledge relies entirely on the experiential cognition and judgment of industry experts, representing a higher dimension of data knowledge. Existing methods for training intelligent models using massive amounts of data lack integration with expert knowledge, which to some extent limits the training speed and effectiveness of intelligent models. Current technologies for integrating data and expert knowledge suffer from high barriers to accessing expert knowledge and a large workload in rule-based transformation, making it difficult to flexibly utilize expert knowledge and data, and hindering the flexible training of decision-making models.

[0003] With the continuous development of large-scale artificial intelligence models, more and more industries are beginning to use intelligent models as auxiliary tools to support the implementation of various business operations. However, to apply intelligent models in specialized fields, it is necessary to explore methods for integrating domain expert knowledge and data to support the training of auxiliary decision-making models for specific domains. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a training method for an intelligent auxiliary decision-making model that integrates expert knowledge and data. By appropriately introducing expert knowledge, the intelligent auxiliary decision-making model can simultaneously learn from expert knowledge and data patterns during training, thereby continuously improving and enhancing the robustness of the auxiliary decision-making model.

[0005] The present invention discloses a training method for an intelligent auxiliary decision-making model that integrates expert knowledge and data, comprising the following steps: Step 1: Transform existing business expert knowledge into trainable expert samples, construct a query graph, and describe each expert sample in graph form to obtain the expert sample set. ; Step 2: Transform business data without expert knowledge application into trainable non-expert samples, construct a target graph, and describe each non-expert sample in graph form to obtain a non-expert sample set. ; Step 3: Evaluate the expert strategies for the expert samples and the non-expert strategies for the non-expert samples, and based on the evaluation results... The number of expert samples in the data was adjusted. Step 4: Based on the results obtained in Step 3, perform the following steps again: and The samples in the model are evaluated, and the evaluation results are used as the rewards for the corresponding samples, thereby constructing the intelligent auxiliary decision-making model. sample; Step 5: Traverse The expert samples are used to perform subgraph matching between each query graph and the target graph to obtain the training sample set. ; Step 6: Extract the training sample set The learning rate of the extracted samples is calculated, and the policy network of the intelligent assisted decision-making model is trained using the experience-first replay method. Step 7: Select a test sample set, evaluate the performance of the intelligent decision-making assistance model on the test sample, construct a new expert sample based on the difference between the output of the intelligent decision-making assistance model and the expert selection results, and integrate it into the system. In this process, the auxiliary decision-making model can be corrected under the guidance of experts; Step 8, for The samples in the training sample are clustered, and the operable samples are extracted as the starting point for exploration. The training sample is expanded by randomly applying strategies to multiple starting points for exploration. Step 9: Repeat steps 1 to 8 to enable the intelligent decision support model to continuously learn based on dynamically updated expert knowledge and data, providing accurate decision support for business operations.

[0006] Furthermore, in step 1, existing business expert knowledge is described as being composed of object characteristics. Expert strategy and the next feature after the expert strategy takes effect Expert samples formed by combination ,in, and The feature descriptions required for expert knowledge to take effect. The actual strategy corresponding to expert knowledge; the set of expert samples is defined as... , The set is defined as ; Set of expert samples Each sample in the graph is described as a query graph. Query Image The nodes in the graph represent object features, and the edges in the graph represent expert strategies and feature transfer directions, thus realizing the graph representation of expert knowledge.

[0007] Furthermore, in step 2, the data without the application of expert knowledge is described as being composed of object characteristics. Non-expert strategies and the next feature after the strategy takes effect Non-expert samples formed by combination ,in, and The feature description required for non-expert knowledge to take effect. The actual strategy corresponding to non-expert knowledge; the set of non-expert samples is defined as... , The set is defined as ; non-expert sample set Described as target graph Target image The nodes in the graph represent object features, and the edges in the graph represent non-expert strategies and feature transfer directions, thus realizing the graph representation of samples without expert knowledge.

[0008] Furthermore, in step 3, the expert sample and the non-expert sample are evaluated separately, and the ideal business indicator is described as follows: Assess the sample status based on sample characteristics. , , , Corresponding business metrics , , , Through with The relative distance between the features is used to assess the quality of the features at any given time, thereby evaluating the strategy. and for: (1), (2), in, Indicates the distance between features. and This is an adjustable positive constant; when the strategy works well, and Will be compared and Closer Thus, the strategy evaluation value is positive, otherwise it is negative.

[0009] Furthermore, on The expert sample was copied m times; Suppose the model is specific to the features of a particular sample. Corresponding features The output strategy is The next feature obtained is Then set the value of m to: (3), in, These are adjustable positive numbers; This is a floor operation, meaning that as the model policy performance increases with training, the value of m decreases until the model policy performance is greater than or equal to the expert policy, at which point the value of m is 0, thus achieving adaptive adjustment of the number of expert samples.

[0010] Furthermore, in step 5, the acquired expert sample set is... With non-expert sample set ,use and These represent the sample sets before and after processing, respectively. In the initial stage, and deposited at the same time In, and will deposit In the middle, traversal Query graph corresponding to a single sample Perform the following operations: 1) To and Subgraph matching is performed, and a similarity threshold for subgraph matching is constructed based on the similarity calculation between nodes and edges. When the calculated similarity is greater than or equal to the preset subgraph matching similarity threshold, extract... The corresponding subgraph Based on subgraph Reconstruct the sample from the nodes and edges. If the similarity is less than the matching threshold, then directly... The corresponding expert samples were copied multiple times. deposit middle; 2) For successfully matched subgraphs With query graph The corresponding sample With sample Strategies were evaluated separately. and The effect; like Superior Then In Delete, and The corresponding expert samples were copied multiple times. deposit In this process, the final training sample set is obtained. .

[0011] Furthermore, step 6 specifically includes: The input and output dimensions of the intelligent auxiliary decision-making model are determined based on the characteristics of the training samples and the number of available policies, and the probability output of each policy is realized through the policy network. Extraction using a priority experience replay method The training samples are used to construct a sample set. And the parameters of the intelligent auxiliary decision-making model are optimized using gradient descent. For the sample set extracted during model training Calculate the proportion of expert samples among them. Then, for this training round, based on the current model's base learning rate... Adjust the learning rate based on as follows: (4), The learning rate is an adjustable normal number, meaning that the higher the proportion of expert samples, the greater the learning rate, thus allowing for a more comprehensive fit to the expert samples.

[0012] Furthermore, step 8 specifically includes: A density-based clustering algorithm is used to... The training samples were processed, and the samples in the normal cluster and the noise cluster were labeled separately, and the noise samples were extracted separately. Select operable samples from the noisy samples as one of the starting points for sample exploration; set a distance threshold for the noisy samples, traverse the operable samples in the new samples, and when the distance between the new operable sample and the noisy sample meets the threshold, add the new operable sample to the exploration starting point sample set. by Starting with a set of noisy samples, one of the strategies permitted by expert knowledge is randomly applied to obtain new training samples, which are then constructed as follows: Added in the form of middle.

[0013] The beneficial effects of this invention are as follows: 1) The method described in this invention addresses the problem that expert knowledge and data are difficult to be used collaboratively by intelligent models. It transforms expert knowledge into a decision sequence that binds features and strategies, making expert knowledge a data sample and providing a way to utilize expert knowledge. 2) The business objective status evaluation strategy is adopted. The weight of expert samples and the parameter learning rate are dynamically adjusted according to the model training process, which improves the utilization of expert knowledge. At the same time, new expert samples are dynamically injected through model evaluation. 3) This invention eliminates the redundancy of expert samples and data samples by using sample fusion matching, which accelerates the model training process. It identifies sample clusters with a small number of samples by density clustering and expands the number of similar samples from these clusters. The robustness of the auxiliary decision-making model is enhanced by continuously improving the richness of the samples. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the model training module structure; Figure 2 This is the training flowchart. Detailed Implementation

[0015] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0016] like Figure 1 and Figure 2 As shown, the intelligent auxiliary decision-making model training method based on the fusion of expert knowledge and data according to the present invention includes the following steps: Step 1: Transform existing business expert knowledge into trainable expert samples, construct a query graph, and describe each expert sample in graph form to obtain the expert sample set. ; Step 2: Transform business data without expert knowledge application into trainable non-expert samples, construct a target graph, and describe each non-expert sample in graph form to obtain a non-expert sample set. ; Step 3: Evaluate the expert strategies for the expert samples and the non-expert strategies for the non-expert samples, and based on the evaluation results... The number of expert samples in the data was adjusted. Step 4: Based on the results obtained in Step 3, perform the following steps again: and The samples in the model are evaluated, and the evaluation results are used as the rewards for the corresponding samples, thereby constructing the intelligent auxiliary decision-making model. sample; Step 5: Traverse The expert samples are used to perform subgraph matching between each query graph and the target graph to obtain the training sample set. ; Step 6: Extract the training sample set The learning rate of the extracted samples is calculated, and the policy network of the intelligent assisted decision-making model is trained using the experience-first replay method. Step 7: Select a test sample set, evaluate the performance of the intelligent decision-making assistance model on the test sample, construct a new expert sample based on the difference between the output of the intelligent decision-making assistance model and the expert selection results, and integrate it into the system. In this process, the auxiliary decision-making model can be corrected under the guidance of experts; Step 8, for The samples in the training sample are clustered, and the operable samples are extracted as the starting point for exploration. The training sample is expanded by randomly applying strategies to multiple starting points for exploration. Step 9: Repeat steps 1 to 8 to enable the intelligent decision support model to continuously learn based on dynamically updated expert knowledge and data, providing accurate decision support for business operations.

[0017] In this embodiment, the intelligent auxiliary model uses a multi-layer graph transformation network to construct the main body of the model and splices a fully connected network according to the auxiliary decision-making requirements; the input and output dimensions of the intelligent auxiliary decision-making model are determined according to the characteristics of the samples to be used and the number of strategies to be adopted.

[0018] The method described in this invention will be explained below with reference to practical examples.

[0019] In clinical medical decision-making, the complex mechanisms of disease require significant effort from medical experts, leading to inefficient treatment planning and a heavy workload for physicians. Limited by physician experience and condition, treatment plans may overlook crucial information, resulting in suboptimal treatment outcomes. Furthermore, a large concentration of high-quality medical resources in major cities hinders the accessibility of cutting-edge expert knowledge at the grassroots level. Therefore, a training method that integrates clinical expert knowledge with data can be employed to build intelligent auxiliary decision-making models.

[0020] Clinical expert prescription data, non-pharmacological treatment plans, and operational guidelines are used as expert knowledge to form a multi-dimensional health status assessment of patients. Expert strategy The patient's status after following the doctor's instructions for follow-up visits or visits. Composition of expert sample Define multiple expert samples as a set. The expert strategy set is defined as And construct a query graph for each expert sample. Transform clinically sequenced patient physical examinations, health monitoring, and other data that do not involve expert strategies into data based on patient health status. Non-expert strategies implemented by patients The patient's physical examination or monitoring status after following a non-expert strategy. Non-expert sample Define multiple patient sample sets as The set of non-expert strategies is defined as follows: .

[0021] Define the ideal health indicators for patients It is composed of multiple dimensions of health characteristics, including blood pressure, blood sugar, and hormone levels. The number of expert samples in the training dataset is adaptively adjusted to adapt to the model training process. When the performance of the intelligent decision-making assistance model on a particular patient sample is greater than or equal to that of a medical expert, the number of expert samples is reduced to prevent overfitting. Each patient health status feature in the sample set is processed using a policy evaluation module to construct the intelligent decision-making assistance model trained based on deep reinforcement learning. information.

[0022] For expert sample set With non-expert sample set Process it, Multiple samples are used to construct a large target graph, where nodes represent patient health status features and edges represent different strategies. The query graph is constructed from each expert sample, where nodes represent patient health status features and edges represent expert strategies. A similarity matching threshold is set based on the similarity between the patient's health status and the adopted strategy; a successful match is considered achieved when this threshold is met. Subgraph matching is then used to... and Matching and elimination processes are performed to exclude low-value non-expert strategies before model training, thus obtaining the model training sample set. .

[0023] For the model training sample set The model's learning rate is adjusted based on the sample data extracted. After training on the extracted samples, test data is selected, and the model evaluation module is used to evaluate the performance of the clinical intelligent decision-making assistance model on the patient health status test samples. New expert samples are then constructed based on the difference between the model's output strategy and the strategy actually selected by the expert, achieving intelligent model correction under the guidance of clinical experts. Clustering was performed on the samples to identify clusters with a small number of patients. Patients who could follow medical advice were selected as the starting point for the exploration, and the criteria were defined by medical experts. and The optional intervention strategies in the model can be expanded to include this small sample cluster by randomly selecting intervention strategies and conducting dynamic follow-up and re-examination, thereby improving the richness of the samples received by the model.

[0024] In routine clinical decision-making, the trained intelligent decision-making assistance model is applied, and the model is adaptively adjusted through dynamic feedback and continuous training.

[0025] The above description is merely a preferred embodiment of the present invention and is not intended to further limit the present invention. All equivalent changes made based on the description and drawings of the present invention are within the protection scope of the present invention.

Claims

1. A training method for an intelligent auxiliary decision-making model that integrates expert knowledge and data, characterized in that, Includes the following steps: Step 1: Transform existing business expert knowledge into trainable expert samples, construct a query graph, and describe each expert sample in graph form to obtain the expert sample set. ; Step 2: Transform business data without expert knowledge application into trainable non-expert samples, construct a target graph, and describe each non-expert sample in graph form to obtain a non-expert sample set. ; Step 3: Evaluate the expert strategies for the expert samples and the non-expert strategies for the non-expert samples, and based on the evaluation results... The number of expert samples in the data was adjusted. Step 4: Based on the results obtained in Step 3, perform the following steps again: and The samples in the model are evaluated, and the evaluation results are used as the rewards for the corresponding samples, thereby constructing the intelligent auxiliary decision-making model. sample; Step 5: Traverse The expert samples are used to perform subgraph matching between each query graph and the target graph to obtain the training sample set. ; Step 6: Extract the training sample set The learning rate of the extracted samples is calculated, and the policy network of the intelligent assisted decision-making model is trained using the experience-first replay method. Step 7: Select a test sample set, evaluate the performance of the intelligent decision-making assistance model on the test sample, construct a new expert sample based on the difference between the output of the intelligent decision-making assistance model and the expert selection results, and integrate it into the system. In this process, the auxiliary decision-making model can be corrected under the guidance of experts; Step 8, for The samples in the training sample are clustered, and the operable samples are extracted as the starting point for exploration. The training sample is expanded by randomly applying strategies to multiple starting points for exploration. Step 9: Repeat steps 1 to 8 to enable the intelligent decision support model to continuously learn based on dynamically updated expert knowledge and data, providing accurate decision support for business operations.

2. The training method for an intelligent auxiliary decision-making model that integrates expert knowledge and data according to claim 1, characterized in that, In step 1, existing business expert knowledge is described as being composed of object characteristics. Expert strategy and the next feature after the expert strategy takes effect Expert samples formed by combination ,in, and The feature descriptions required for expert knowledge to take effect. The practical strategies corresponding to expert knowledge; Define the set of expert samples as , The set is defined as ; expert sample set Each sample in the graph is described as a query graph. Query Image The nodes in the graph represent object features, and the edges in the graph represent expert strategies and feature transfer directions, thus realizing the graph representation of expert knowledge.

3. The training method for an intelligent auxiliary decision-making model that integrates expert knowledge and data according to claim 2, characterized in that, In step 2, the data without expert knowledge application is described as being composed of object characteristics. Non-expert strategies and the next feature after the strategy takes effect Non-expert samples formed by combination ,in, and The feature description required for non-expert knowledge to take effect. The actual strategies corresponding to non-expert knowledge; Define the set of non-expert samples as , The set is defined as ; Non-expert sample set Described as target graph Target image The nodes in the graph represent object features, and the edges in the graph represent non-expert strategies and feature transfer directions, thus realizing the graph representation of samples without expert knowledge.

4. The training method for an intelligent auxiliary decision-making model that integrates expert knowledge and data according to claim 3, characterized in that, In step 3, the expert sample and the non-expert sample are evaluated separately, and the ideal business indicator is described as follows: Assess the sample status based on sample characteristics. , , , Corresponding business metrics , , , Through with The relative distance between the features is used to assess the quality of the features at any given time, thereby evaluating the strategy. and for: （1）， （2）， in, Indicates the distance between features. and This is an adjustable positive constant; when the strategy works well, and Will be compared and Closer Thus, the strategy evaluation value is positive, otherwise it is negative.

5. The training method for an intelligent auxiliary decision-making model that integrates expert knowledge and data according to claim 4, characterized in that, right The expert sample was copied m times; Suppose the model is specific to the features of a particular sample. Corresponding features The output strategy is The next feature obtained is Then set the value of m to: （3）， in, These are adjustable positive numbers; This is a floor operation, meaning that as the model policy performance increases with training, the value of m decreases until the model policy performance is greater than or equal to the expert policy, at which point the value of m is 0, thus achieving adaptive adjustment of the number of expert samples.

6. The training method for an intelligent auxiliary decision-making model that integrates expert knowledge and data according to claim 4, characterized in that, In step 5, the acquired expert sample set is processed. With non-expert sample set ,use and These represent the sample sets before and after processing, respectively. In the initial stage, and deposited at the same time In, and will deposit In the middle, traversal Query graph corresponding to a single sample Perform the following operations: 1) To and Perform subgraph matching and calculate the similarity between nodes and edges; When the calculated similarity is greater than or equal to the preset subgraph matching similarity threshold, extract... The corresponding subgraph Based on subgraph Reconstruct the sample from the nodes and edges. ; If the similarity is less than the matching threshold, then directly... The corresponding expert samples were copied multiple times. deposit middle; 2) For successfully matched subgraphs With query graph The corresponding sample With sample Strategies were evaluated separately. and The effect; like Superior Then In Delete, and The corresponding expert samples were copied multiple times. deposit In this process, the final training sample set is obtained. .

7. The training method for an intelligent auxiliary decision-making model that integrates expert knowledge and data according to claim 6, characterized in that, Step 6 specifically involves: The input and output dimensions of the intelligent auxiliary decision-making model are determined based on the characteristics of the training samples and the number of available strategies, and the probability output of each strategy is realized through the model. Extraction using a priority experience replay method The training samples are used to construct a sample set. And the parameters of the intelligent auxiliary decision-making model are optimized using gradient descent. For the sample set extracted during model training Calculate the proportion of expert samples among them. Then, for this training round, based on the current model's base learning rate... Adjust the learning rate based on as follows: （4）， The learning rate is an adjustable normal number, meaning that the higher the proportion of expert samples, the greater the learning rate, thus allowing for a more comprehensive fit to the expert samples.

8. The training method for an intelligent auxiliary decision-making model that integrates expert knowledge and data according to claim 6, characterized in that, Step 8 specifically includes: A density-based clustering algorithm is used to... The training samples were processed, and the samples in the normal cluster and the noise cluster were labeled separately, and the noise samples were extracted separately. Screening noisy samples for real-time action is one of the starting points for sample exploration. Set a distance threshold for noise samples, traverse the operable samples in the new samples, and add the new operable sample to the exploration starting point sample set when the distance between the new sample and the noise sample is greater than or equal to the threshold. Starting with n noisy samples, one of the strategies allowed by expert knowledge is randomly applied to obtain new training samples, which are then constructed as follows: Added in the form of middle.

Images