WO2021139106A1 - 分群决策模型生成、分群处理方法、装置、设备及介质 - Google Patents

分群决策模型生成、分群处理方法、装置、设备及介质 Download PDF

Info

Publication number
WO2021139106A1
WO2021139106A1 PCT/CN2020/098829 CN2020098829W WO2021139106A1 WO 2021139106 A1 WO2021139106 A1 WO 2021139106A1 CN 2020098829 W CN2020098829 W CN 2020098829W WO 2021139106 A1 WO2021139106 A1 WO 2021139106A1
Authority
WO
WIPO (PCT)
Prior art keywords
individual
sample data
group
individual sample
grouping
Prior art date
Application number
PCT/CN2020/098829
Other languages
English (en)
French (fr)
Inventor
徐卓扬
孙行智
赵惟
左磊
蒋雪涵
胡岗
Original Assignee
平安科技（深圳）有限公司
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 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Publication of WO2021139106A1 publication Critical patent/WO2021139106A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • 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
    • G06Q30/00—Commerce
    • G06Q30/06—Buying, selling or leasing transactions
    • G06Q30/0601—Electronic shopping [e-shopping]
    • G06Q30/0631—Item recommendations
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H20/00—ICT specially adapted for therapies or health-improving plans, e.g. for handling prescriptions, for steering therapy or for monitoring patient compliance
    • G16H20/30—ICT specially adapted for therapies or health-improving plans, e.g. for handling prescriptions, for steering therapy or for monitoring patient compliance relating to physical therapies or activities, e.g. physiotherapy, acupressure or exercising
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H20/00—ICT specially adapted for therapies or health-improving plans, e.g. for handling prescriptions, for steering therapy or for monitoring patient compliance
    • G16H20/60—ICT specially adapted for therapies or health-improving plans, e.g. for handling prescriptions, for steering therapy or for monitoring patient compliance relating to nutrition control, e.g. diets

Definitions

  • This application relates to the field of big data, and in particular to a method, device, device, and medium for grouping decision model generation and grouping processing.
  • the division of similar groups is a very important task for many companies, and the division of similar groups needs to be applied in many application scenarios, such as: food Optimization, exercise recommendation and product recommendation of similar groups, etc.
  • the inventor realizes that the current method of dividing similar groups has the following shortcomings: 1. At present, most of the group knowledge bases considered in decision-making are insufficient; 2. If multiple group knowledge bases are considered when dividing groups, but in the process of grouping There is often at least one conflicting part of group knowledge in multiple group knowledge bases, and there is currently no plan to make decisions on this conflicting part; the two points mentioned above will lead to low accuracy and comprehensiveness of group division. Low sex and unreasonable problems. Therefore, those skilled in the art urgently need to find a technical solution that can solve the above-mentioned problems.
  • a grouping decision model generation method including:
  • each of the group knowledge bases contains group knowledge associated with all the individual sample data, and the individual sample data Comprising a plurality of individual characteristics; at least two group knowledge related to the same individual sample data in the group knowledge base conflict with each other;
  • the short-term outcome model predicted by the feature contribution evaluator calls the SHAP method to evaluate the short-term contribution of the individual features in the individual sample data
  • the long-term outcome model predicted by the feature contribution evaluator calls the SHAP method to evaluate the individual sample data.
  • the long-term contribution of the individual feature after the short-term contribution and the long-term contribution are input to the preset contribution degree function of the feature contribution degree evaluator, the feature contribution degree of the individual feature is output; the feature contribution degree evaluation Established based on the SHAP method of XGBoost model;
  • the behavior variables, state variables, and reward values of the individual sample data are input into the free-stable DQN network of the preset grouping decision model to be trained for training, and the free-stable DQN network is generated according to the behavior variables and the state variables Q value, and input the reward value and the Q value of the individual sample data into the loss function of the grouping decision type to be trained to obtain the loss value of the individual sample data;
  • the grouping decision model to be trained is marked as a grouping decision model that has been trained.
  • a grouping decision model generation device including:
  • the establishment module is used to obtain individual sample data to be grouped, and establish at least two group knowledge bases using a preset grouping decision tree; each of the group knowledge bases contains group knowledge associated with all the individual sample data, so The individual sample data includes a plurality of individual characteristics; the group knowledge associated with the same individual sample data in at least two of the group knowledge bases conflict with each other;
  • the output module is used to call the SHAP method to evaluate the short-term contribution of the individual features in the individual sample data through the predictive short-term outcome model of the feature contribution evaluator, and call the SHAP method to evaluate the results of the individual feature through the predictive long-term outcome model of the feature contribution evaluator.
  • the long-term contribution of the individual feature in the individual sample data, after the short-term contribution and the long-term contribution are input to the preset contribution degree function of the feature contribution degree evaluator, the feature contribution degree of the individual feature is output;
  • the feature contribution degree evaluator is established based on the SHAP method of the XGBoost model;
  • the definition module is used to obtain the matching relationship between the individual sample data and the group knowledge associated in each of the group knowledge bases, and mark the sample population value for the individual sample data according to all the obtained matching relationships After labeling, input the feature contribution degree of all the individual characteristics of the individual sample data into the preset reward function, output the reward value of the individual sample data, and define the group knowledge value label of the individual sample data Is a behavior variable, and the individual characteristics of the individual sample data are defined as a state variable;
  • the obtaining module is used to input the behavior variables, state variables and reward values of the individual sample data into the free-stable DQN network of the preset grouping decision model to be trained for training, and obtain the free-stable DQN network according to the behavior variables and Generating the Q value from the state variable, and inputting the reward value and the Q value of the individual sample data into the loss function of the grouping decision type to be trained to obtain the loss value of the individual sample data;
  • the marking module is used to mark the grouping decision model to be trained as a grouping decision model that has been trained when it is determined that the loss value does not decrease after the individual sample data has undergone a preset training early stopping coefficient.
  • a grouping processing method including:
  • the grouping decision result is the final grouping result of the individual data.
  • a grouping processing device includes:
  • the group decision-making result acquisition module is used to acquire individual data of the group to be determined, and to acquire at least two group decision-making results corresponding to at least two group knowledge bases in the individual data;
  • one group knowledge base contains at least one group Determine a group, the grouping decision result means that the individual data belongs to one of the determined groups included in the group knowledge base;
  • one of the determined groups is associated with a group knowledge in the group knowledge base;
  • the input module is configured to, if at least two of the grouping decision results are inconsistent, input the individual data into the grouping decision model to obtain the final grouping result output by the grouping decision model; the final grouping result and the individual to which it belongs Data group value label association;
  • the final grouping result determination module is configured to determine that the grouping decision result is the final grouping result of the individual data if at least two of the grouping decision results are consistent.
  • a computer device including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, and the processor implements the above-mentioned grouping decision model generation method or grouping processing when the processor executes the computer program method.
  • a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned grouping decision model generation method or grouping processing method is realized.
  • the above-mentioned method, device, equipment, and medium for grouping decision-making model are used to process the individual sample data to be grouped, and then solve the problem that the group knowledge associated with the same individual sample data in the two group knowledge bases conflict with each other, which is finally passed Train and establish a grouping decision model to identify the individual sample data to be grouped and output the sample group value label of the individual sample data. Then the grouping decision model can accurately and efficiently determine the group to which the individual sample data to be grouped belongs, and in the model In the recognition process, the group knowledge in the group knowledge base can be fully considered, and the group to which the individual sample data output after the recognition belongs also has the advantage of high rationality.
  • the above-mentioned grouping processing method, device, equipment and medium, through the grouping decision model completed by the above training, can accurately, efficiently, reasonably and comprehensively determine the final grouping result to which the individual data belongs to the individual data of the group to be determined.
  • FIG. 1 is a schematic diagram of an application environment of a method for generating a grouping decision model in an embodiment of the present application
  • FIG. 2 is a schematic flowchart of a method for generating a grouping decision model in an embodiment of the present application
  • FIG. 3 is a schematic structural diagram of a group decision model generating device in an embodiment of the present application.
  • FIG. 4 is a schematic flowchart of a grouping processing method in an embodiment of the present application.
  • FIG. 5 is a schematic structural diagram of a grouping processing device in an embodiment of the present application.
  • Fig. 6 is a schematic diagram of a computer device in an embodiment of the present application.
  • the method for generating a grouping decision model provided by the present application can be applied in the application environment as shown in Fig. 1, in which the client communicates with the server through the network.
  • the client can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices.
  • the server can be implemented as an independent server or a server cluster composed of multiple servers.
  • a method for generating a grouping decision model is provided.
  • the method is applied to the server in FIG. 1 as an example for description, including the following steps:
  • the individual sample data to be grouped can refer to the long-term tracking data of a large number of individuals in a certain application scenario (a tracking number will be recorded after each tracking data collection is completed), for example, a large number of individuals are optimizing their diet and/or medication Tracking data in application scenarios such as management method/sports method recommendation/commodity recommendation of similar groups; and the individual sample data to be grouped have multiple individual characteristics, so multiple individual sample data can be obtained by collecting individual sample data for a preset time period Individual characteristics.
  • the identity information in individual sample data can be collected.
  • the individual characteristics including but not limited to the gender, age and weight of the individual), individual characteristics of health (including but not limited to BMI, body Quality index, body fat rate, etc.), individual characteristics of diet, individual characteristics of short-term outcome (whether the individual next BMI is within the control range) and individual characteristics of long-term outcome (whether the individual has died next time), or the scene of medication management
  • the individual characteristics of inspection indicators, individual characteristics of disease history, individual characteristics of medication history, individual characteristics of short-term efficacy and individual characteristics of long-term efficacy in individual sample data can be collected;
  • the group knowledge base contains group knowledge, which mainly indicates the knowledge of group division
  • the group knowledge in the group knowledge base is the recommended group classification method for the diet.
  • the preset group decision tree can be used to convert the various rules of the above-mentioned dietary recommendation method into the form of a decision tree, and it should be noted that each decision tree is the same individual
  • the group knowledge associated with the sample data may conflict with each other.
  • the A group knowledge base, the B group knowledge base and the amount of various diets associated with the same individual sample data conflict with each other.
  • the obtaining individual sample data to be grouped includes:
  • the types of individual sample data are different (different application scenarios), and the preset time period should also be different.
  • the preset time period can be one quarter; Because each application scenario is different, the individual characteristics in the individual sample data must also be different; specifically, the corresponding preset individual characteristic items in each application scenario can be established through the preset data atlas, and an application scenario is determined to be collected
  • various individual characteristics related to the individual sample data can be generated; through this embodiment, each individual characteristic of the individual sample data to be collected can be determined, accelerating the standardization of collection and improving the efficiency of collection .
  • the feature contribution rate evaluator is based on the XGBoost model and the SHAP method, and the feature contribution rate evaluator mainly includes a predictive short-term outcome model and a predictive long-term outcome model (the two models have inconsistent prediction goals). All models are XGBoost models.
  • the XGBoost model uses integrated ideas to serially generate multiple decision trees (where each decision tree fits the residual of the previous decision tree) during training, and uses all decision trees to obtain a prediction result after multiple rounds of decision-making.
  • both the XGBoost model for predicting short-term outcomes and the XGBoost model for predicting long-term outcomes use a large number of individual sample data extracted from S10 (except for the last tracking data, because the last tracking data generally does not have short-term and long-term outcomes. Individual characteristics).
  • the input features of the two XGBoost models are the individual features (non-outcomes) in the individual sample data that have been extracted.
  • the output of the XGBoost model predicting short-term outcome is the individual feature of the short-term outcome
  • the output of the XGBoost model predicting the long-term outcome is the individual feature of the long-term outcome .
  • the SHAP method can be called using two XGBoost models (the importance of features is defined by the SHAP value, showing how the XGBoost model finally reaches the prediction output of the prediction result, the prediction result is the short-term outcome mentioned above And long-term outcomes), the SHAP method is a classic method to measure the importance of features of machine learning models; in the above-mentioned dietary optimization application scenarios, the prediction short-term outcome model can be used to predict whether an individual’s next BMI is within the control range.
  • Short-term contributions, etc., and the predicting long-term outcome model can be used to predict whether the individual has died next time, the long-term contribution, etc.;
  • the short-term outcome is generally defined as a benign outcome, so the larger the short-term outcome, the better, while the long-term outcome is generally defined as a malignant outcome (for example, whether the individual will die next time) , Therefore, the smaller the long-term outcome result, the better, i refers to individual sample data, j refers to the number of tracking individual sample data, k refers to a certain individual characteristic of individual sample data;
  • SHAP M1 is the prediction short-term outcome model (the short-term prediction model is marked as M1) the short-term contribution of the output, and SHAP M2 is the long-term contribution of the output of the prediction short-term outcome model (the prediction long-term outcome model is marked as M2).
  • the short-term outcome model predicted by the feature contribution rate evaluator calls the SHAP method to evaluate the short-term contribution of the individual features in the individual sample data
  • the long-term outcome model predicted by the feature contribution rate evaluator calls the SHAP method to evaluate the results. Describe the long-term contribution of individual characteristics in individual sample data, including:
  • Acquiring the individual sample data corresponds to at least two grouping decision results of at least two of the group knowledge bases; one of the group knowledge bases includes at least one group, and the grouping decision result is that the individual sample data belongs to the group One said group contained in the knowledge base; one said group is related to one said group knowledge;
  • the short-term outcome model of the feature contribution evaluator is used to call the SHAP method to evaluate the short-term contribution of the individual features in the individual sample data
  • the long-term prediction of the feature contribution evaluator is The outcome model calls the SHAP method to evaluate the long-term contribution of individual characteristics in the individual sample data.
  • the number of groups contained in the group knowledge base is related to the group knowledge contained in the group knowledge base.
  • the individual sample data can be obtained by obtaining at least two grouping decision results corresponding to at least two group knowledge bases, and used to determine the grouping. Whether the decision results are consistent, for example, in the application scenario of diet optimization, the group knowledge in the knowledge base of group A recommends to the individual sample data that the vegetable diet of the group belongs to 80g-100g, while the group in the group B knowledge base Knowledge recommends to the individual sample data that the vegetable diet of the group belongs to 30g-50g. It can be seen that the results of the two grouping decisions are inconsistent and will cause conflicts. Therefore, it is necessary to perform the training steps after step S20 to solve the above-mentioned conflicting problems. It can be seen that this embodiment can effectively solve the problem of the group to which the individual sample data belongs through corresponding means.
  • the method further includes:
  • the individual sample data is deleted.
  • S30 Obtain the matching relationship between the individual sample data and the group knowledge associated in each of the group knowledge bases, and after all the obtained matching relationships are marked with a sample population value label for the individual sample data, After inputting the feature contribution degrees of all the individual characteristics of the individual sample data into the preset reward function, outputting the reward value of the individual sample data, and defining the group knowledge value label of the individual sample data as a behavior variable , Define the individual characteristics of the individual sample data as state variables;
  • step S10 is located only establishes a group knowledge base associated with individual sample data
  • step S20 filters out individual sample data with consistent grouping decision results of group knowledge in each group knowledge base
  • this embodiment is In order to determine the direct matching relationship between the individual sample data and the associated group knowledge base, that is, to determine the matching relationship between the individual sample data and the group knowledge in the group knowledge base from the individual characteristics of the individual sample data, for example, in the dietary optimization
  • the individual characteristic of the diet in the individual sample data is that the daily vegetable diet is 40g
  • the group knowledge in the A group knowledge base recommends the group’s vegetable diet to the individual sample data to be 80g-100g
  • the group knowledge in the database recommends to the individual sample data that the vegetable diet of the group belongs to 30g-50g.
  • the individual sample data does not match the group knowledge in the A group knowledge base but matches the group knowledge in the A group knowledge base. All the matching relationships obtained are used to mark the sample group value label (0,1) for the individual sample data, and the similar principle can be used (mark 1 if it conforms to the group knowledge in the group knowledge base, and mark it if it does not conform to the group knowledge in the group knowledge base. 0)
  • the sample population value labels (mathematically coordinate values) in other situations can be obtained.
  • the Q-learning-based reinforcement learning method uses a Q-value table to store each state and the Q value of each action under this state state.
  • many dimensions in the state are continuous. It is very difficult to use a table to store the Q value of each action of each state, and free-stable DQN (using a four-layer DQN network, the input layer and state have the same dimensions, and the two dimensions in the middle are respectively The hidden layers of 32 and 64, the output layer has the same dimensions as the action, and the four layers mentioned above are all fully connected.) Fusion neural network and Q-learning, so the neural network is used to generate the Q value to solve the state
  • the dimension is a continuous problem; the formula for generating the Q value is And the loss function is among them, ⁇ j is the state variable at step j (j mentioned here represents the jth step of an episode of individual sample data in reinforcement learning, which can be understood as the number of tracking individual sample data mentioned in S20 above), Is the behavior variable at step j, ⁇ is the network parameter, r
  • the use of the grouping decision-making loss function to be trained can reduce the subjectivity of the reward value defined above, and adding a function similar to the regularization term to the loss function can be Reduce the problem of parameter fluctuations in reinforcement learning, thereby improving the stability of the group decision-making to be trained.
  • the loss value in the grouping decision model to be trained does not fluctuate significantly and does not decrease, which can indicate that the grouping decision model to be trained tends to converge. It can indicate that the group decision model to be trained has been trained as a group decision model.
  • the above provides a method for generating a grouping decision model.
  • the individual sample data to be grouped is processed for data and then the problem that the group knowledge associated with the same individual sample data in the two group knowledge bases conflict with each other is solved. That is, finally, a grouping decision model is established through training to identify the individual sample data to be grouped and output the sample group value label of the individual sample data, and then the grouping decision model can accurately and efficiently determine the group to which the individual sample data to be grouped belongs And in the process of model recognition, the group knowledge in the group knowledge base can be considered more comprehensively, and the group to which the individual sample data output after the recognition belongs also has the advantage of high rationality.
  • a group decision model generation device is provided, and the group decision model generation device corresponds to the group decision model generation method in the above-mentioned embodiment in a one-to-one correspondence.
  • the group decision model generation device includes an establishment module 11, an output module 12, a definition module 13, an acquisition module 14 and a marking module 15.
  • the detailed description of each functional module is as follows:
  • the establishment module 11 is used to obtain individual sample data to be grouped, and use a preset grouping decision tree to establish at least two group knowledge bases; each of the group knowledge bases contains group knowledge associated with all the individual sample data, The individual sample data includes a plurality of individual characteristics; the group knowledge associated with the same individual sample data in at least two of the group knowledge bases conflict with each other;
  • the output module 12 is configured to call the SHAP method to evaluate the short-term contribution of individual features in the individual sample data through the predictive short-term outcome model of the feature contribution evaluator, and call the SHAP method to evaluate the long-term outcome model of the feature contribution evaluator.
  • the long-term contribution of the individual feature in the individual sample data, after the short-term contribution and the long-term contribution are input to the preset contribution degree function of the feature contribution degree evaluator, the feature contribution degree of the individual feature is output;
  • the feature contribution degree evaluator is established based on the SHAP method of the XGBoost model;
  • the definition module 13 is used to obtain the matching relationship between the individual sample data and the group knowledge associated in each of the group knowledge bases, and mark the sample group for the individual sample data according to all the obtained matching relationships
  • the feature contribution degrees of all the individual characteristics of the individual sample data are input into the preset reward function, the reward value of the individual sample data is output, and the group knowledge value of the individual sample data is labeled
  • the behavior variable and define the individual characteristics of the individual sample data as a state variable
  • the obtaining module 14 is used to input the behavior variables, state variables, and reward values of the individual sample data into the free-stable DQN network of the preset grouping decision model to be trained for training, and obtain the free-stable DQN network according to the behavior variables Generating a Q value with the state variable, and inputting the reward value and the Q value of the individual sample data into the loss function of the grouping decision type to be trained to obtain the loss value of the individual sample data;
  • the marking module 15 is configured to mark the grouping decision model to be trained as a grouping decision model that has been trained when it is determined that the loss value does not decrease after the individual sample data has undergone a preset training early stopping coefficient.
  • the establishment module includes:
  • the collection sub-module is used to obtain all the individual sample data for a preset time period, and use the individual sample data to query at least one preset individual characteristic item in the preset data atlas, and collect it with the preset individual characteristic item The individual characteristics.
  • the output module includes:
  • the obtaining submodule is used to obtain at least two grouping decision results of the individual sample data corresponding to at least two of the group knowledge bases; one of the group knowledge bases contains at least one group, and the grouping decision result is the individual
  • the sample data belongs to one of the groups contained in the group knowledge base; one of the groups is associated with one of the group knowledge;
  • the evaluation sub-module is configured to, if at least two of the grouping decision results are inconsistent, call the SHAP method through the predictive short-term outcome model of the feature contribution degree evaluator to evaluate the short-term contribution of the individual feature in the individual sample data, and use the feature contribution
  • the predictive long-term outcome model of the degree evaluator invokes the SHAP method to evaluate the long-term contribution of individual characteristics in the individual sample data.
  • the output module further includes:
  • the deleting sub-module is configured to delete the individual sample data if at least two of the grouping decision results are consistent.
  • each module in the above-mentioned grouping decision model generating device can be implemented in whole or in part by software, hardware, and a combination thereof.
  • the above-mentioned modules may be embedded in the form of hardware or independent of the processor in the computer equipment, or may be stored in the memory of the computer equipment in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.
  • the evaluation data processing method provided by this application can be applied in the application environment as shown in Fig. 1, in which the client communicates with the server through the network.
  • the client can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices.
  • the server can be implemented as an independent server or a server cluster composed of multiple servers.
  • a grouping processing method is provided. Taking the method applied to the server in FIG. 1 as an example for description, the method includes the following steps:
  • S50 Obtain individual data of the group to be determined, and obtain at least two grouping decision results in the individual data that correspond to at least two group knowledge bases one-to-one; one group knowledge base includes at least one determined group, and the grouping
  • the decision result means that the individual data belongs to one of the certain groups contained in the group knowledge base; one of the certain groups is associated with a group knowledge in the group knowledge base; understandably, the above mentioned
  • the identified groups include but are not limited to diet groups, sports groups, and similar product recommendation groups.
  • step S60 If at least two of the grouping decision results are inconsistent, input the individual data to the grouping decision model to obtain a final grouping result output by the grouping decision model; the final grouping result and the group of individual data to which it belongs Value label association; understandably, the grouping decision model after training is used to identify the individual sample data and output the group value label of the individual sample data, and determine the group to which the group value label belongs, for example, if the obtained group value label is still the above
  • the (0, 1) mentioned in step S30 means that the group to which the individual sample data belongs is the corresponding certain group in the knowledge base of group B. If the obtained group value label is (1, 0), it means that the individual sample data belongs to The group of is the corresponding definite group in the knowledge base of group A.
  • the confirmed definite group can further recommend suitable three-meal diet, exercise recommendation and group-like product recommendation to the user.
  • the clustering decision model can be optimized during the use of the model, and the reward values, behavior variables, and state variables associated with all individual characteristics in the next new individual data of the individual data can be directly input to the clustering After the decision-making model, the group decision-making model is optimized to adapt to the new individual sample data;
  • the grouping decision model is used to identify the individual data
  • the group value label of the individual data output by the model is obtained, and the final grouping result of the individual data is determined through the group value label.
  • the principle is consistent with the above-mentioned step S50; when the two grouping decision results are consistent, the individual sample data can belong to the determined group included in the grouping decision result at the same time.
  • the above provides a grouping processing method, and the final grouping result to which the individual data belongs can be accurately, efficiently, reasonably and comprehensively determined by the grouping decision model completed by the above training to determine the individual data of the group to be determined.
  • a grouping processing device corresponds to the evaluation data processing method in the above-mentioned embodiment in a one-to-one correspondence.
  • the grouping processing device includes a grouping decision result acquisition module 21, an input module 22 and a final grouping result determination module 23.
  • the detailed description of each functional module is as follows:
  • the group decision-making result acquisition module 21 is configured to acquire individual data of the group to be determined, and obtain at least two group decision-making results corresponding to at least two group knowledge bases in the individual data; one group knowledge base contains at least A certain group, the group decision result means that the individual data belongs to one of the certain groups included in the group knowledge base; one of the certain groups is associated with a group knowledge in one of the group knowledge bases;
  • the input module 22 is configured to, if at least two of the grouping decision results are inconsistent, input the individual data to the grouping decision model to obtain the final grouping result output by the grouping decision model; the final grouping result and the grouping decision to which it belongs Group value label association of individual data;
  • the final grouping result determination module 23 is configured to determine that the grouping decision result is the final grouping result of the individual data if at least two of the grouping decision results are consistent.
  • Each module in the above-mentioned grouping processing device may be implemented in whole or in part by software, hardware, and a combination thereof.
  • the above-mentioned modules may be embedded in the form of hardware or independent of the processor in the computer equipment, or may be stored in the memory of the computer equipment in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.
  • a computer device is provided.
  • the computer device may be a server, and its internal structure diagram may be as shown in FIG. 6.
  • the computer equipment includes a processor, a memory, a network interface, and a database connected through a system bus. Among them, the processor of the computer device is used to provide calculation and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium and an internal memory.
  • the non-volatile storage medium stores an operating system, a computer program, and a database.
  • the internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium.
  • the database of the computer equipment is used to store the data involved in the grouping decision model generation method.
  • the network interface of the computer device is used to communicate with an external terminal through a network connection.
  • the computer program is executed by the processor to realize a grouping decision model generation method, or the computer program is executed by the processor to realize a grouping processing method.
  • a computer device including a memory, a processor, and a computer program stored in the memory and running on the processor.
  • the processor executes the computer program to implement the grouping decision model generation method in the above embodiment Or when the processor executes the computer program to implement the steps of the grouping processing method in the above-mentioned embodiment.
  • a computer-readable storage medium is provided with a computer program stored thereon.
  • the computer-readable storage medium may be non-volatile or volatile.
  • Non-volatile memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • Volatile memory may include random access memory (RAM) or external cache memory.
  • RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Channel (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
  • the blockchain referred to in this application is a new application mode of computer technology such as distributed data storage, point-to-point transmission, consensus mechanism, and encryption algorithm.
  • Blockchain essentially a decentralized database, is a series of data blocks associated with cryptographic methods. Each data block contains a batch of network transaction information for verification. The validity of the information (anti-counterfeiting) and generation of the next block.
  • the blockchain can include the underlying platform of the blockchain, the platform product service layer, and the application service layer.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Business, Economics & Management (AREA)
  • Databases & Information Systems (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Medical Informatics (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Epidemiology (AREA)
  • Primary Health Care (AREA)
  • Public Health (AREA)
  • Accounting & Taxation (AREA)
  • Biomedical Technology (AREA)
  • Finance (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Economics (AREA)
  • Development Economics (AREA)
  • Marketing (AREA)
  • Physical Education & Sports Medicine (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Business, Economics & Management (AREA)
  • Strategic Management (AREA)
  • Nutrition Science (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种分群决策模型生成、分群处理方法、装置、设备及介质。所述方法包括：获取个体样本数据中个体特征的短期贡献和长期贡献，将短期贡献和长期贡献输入至预设的贡献度函数后，输出特征贡献度；将特征贡献度输入至预设reward函数后，输出reward值，并将个体样本数据的群体知识值标签定义为行为变量，将个体样本数据的个体特征定义为状态变量；将行为变量、状态变量和reward值输入至待训练预设分群决策模型进行训练，获取生成的Q值，并将reward值和Q值输入至损失函数后获取损失值；在判定损失值不再下降时，将待训练分群决策模型标记为训练完成的分群决策模型。还涉及区块链技术，所述个体样本数据可存储于区块链节点中。

Description

分群决策模型生成、分群处理方法、装置、设备及介质
本申请要求于2020年05月13日提交中国专利局、申请号为202010403130.5，发明名称为“分群决策模型生成、分群处理方法、装置、设备及介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及大数据领域，尤其涉及一种分群决策模型生成、分群处理方法、装置、设备及介质。
背景技术
目前很多公司为了实现降低成本，实现个性化推荐和提升推荐效果的目的，因此相似群体的划分是很多公司非常重要的任务，并且相似群体的划分在很多应用场景中都需应用到，例如：饮食优化、运动方式推荐和相似群体的商品推荐等。但发明人意识到目前的相似群体划分方法有以下不足：1.目前大部分决策考虑的群体知识库不足；2.若对群体划分时考虑了多个群体知识库，但在进行分群的过程中多个群体知识库中往往有至少一处群体知识的冲突的部分，且当前并不存在对这种冲突的部分进行决策的方案；上述提到的两点将导致群体划分存在准确性低、全面性低和不合理的问题。因此本领域人员亟需寻找一种技术方案能解决上述提到的问题。
技术问题
基于此，有必要针对上述技术问题，提供一种分群决策模型生成、分群处理方法、装置、设备及介质，通过该分群决策模型可精准、全面且合理确定出待分群的个体样本数据所属的群体。
技术解决方案
一种分群决策模型生成方法，包括：
获取待分群的个体样本数据，利用预设分群决策树建立至少两个群体知识库；每一个所述群体知识库中均包含与所有所述个体样本数据关联的群体知识，所述个体样本数据中包括多个个体特征；至少两个所述群体知识库中与同一个所述个体样本数据关联的群体知识相互冲突；
通过特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献，将所述短期贡献和所述长期贡献输入至所述特征贡献度评估器的预设的贡献度函数后，输出所述个体特征的特征贡献度；所述特征贡献度评估器基于XGBoost模型的SHAP方法建立；
获取所述个体样本数据与各所述群体知识库中相关联的所述群体知识之间的匹配关系，根据获取的所有所述匹配关系为所述个体样本数据标记样本群体值标签后，将所述个体样本数据的所有所述个体特征的特征贡献度输入至预设reward函数后，输出所述个体样本数据的reward值，并将所述个体样本数据的群体知识值标签定义为行为变量，将所述个体样本数据的个体特征定义为状态变量；
将所述个体样本数据的行为变量、状态变量和reward值输入至待训练预设分群决策模型的free-stable DQN网络进行训练，获取free-stable DQN网络根据所述行为变量和所述状态变量生成Q值，并将所述个体样本数据的所述reward值和所述Q值输入至所述待训练分群决策型的损失函数后获取所述个体样本数据的损失值；
在判定所述个体样本数据在经历过预设训练早停系数后所述损失值不再下降时，将所述待训练分群决策模型标记为训练完成的分群决策模型。
一种分群决策模型生成装置，包括：
建立模块，用于获取待分群的个体样本数据，利用预设分群决策树建立至少两个群体知识库；每一个所述群体知识库中均包含与所有所述个体样本数据关联的群体知识，所述个体样本数据中包括多个个体特征；至少两个所述群体知识库中与同一个所述个体样本数 据关联的群体知识相互冲突；
输出模块，用于通过特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献，将所述短期贡献和所述长期贡献输入至所述特征贡献度评估器的预设的贡献度函数后，输出所述个体特征的特征贡献度；所述特征贡献度评估器基于XGBoost模型的SHAP方法建立；
定义模块，用于获取所述个体样本数据与各所述群体知识库中相关联的所述群体知识之间的匹配关系，根据获取的所有所述匹配关系为所述个体样本数据标记样本群体值标签后，将所述个体样本数据的所有所述个体特征的特征贡献度输入至预设reward函数后，输出所述个体样本数据的reward值，并将所述个体样本数据的群体知识值标签定义为行为变量，将所述个体样本数据的个体特征定义为状态变量；
获取模块，用于将所述个体样本数据的行为变量、状态变量和reward值输入至待训练预设分群决策模型的free-stable DQN网络进行训练，获取free-stable DQN网络根据所述行为变量和所述状态变量生成Q值，并将所述个体样本数据的所述reward值和所述Q值输入至所述待训练分群决策型的损失函数后获取所述个体样本数据的损失值；
标记模块，用于在判定所述个体样本数据在经历过预设训练早停系数后所述损失值不再下降时，将所述待训练分群决策模型标记为训练完成的分群决策模型。
一种分群处理方法，包括：
获取待确定群体的个体数据，并获取所述个体数据中与至少两个群体知识库一一对应的至少两个分群决策结果；一个所述群体知识库包含至少一个确定群体，所述分群决策结果是指所述个体数据属于所述群体知识库中包含的其中一个所述确定群体；一个所述确定群体与一个所述群体知识库中的一个群体知识关联；
若至少两个所述分群决策结果不一致，则将所述个体数据输入至分群决策模型后，得到所述分群决策模型输出的最终分群结果；所述最终分群结果与其所属的个体数据的群体值标签关联；
若至少两个所述分群决策结果均一致，则确定该分群决策结果为所述个体数据的最终分群结果。
一种分群处理装置，包括：
分群决策结果获取模块，用于获取待确定群体的个体数据，并获取所述个体数据中与至少两个群体知识库一一对应的至少两个分群决策结果；一个所述群体知识库包含至少一个确定群体，所述分群决策结果是指所述个体数据属于所述群体知识库中包含的其中一个所述确定群体；一个所述确定群体与一个所述群体知识库中的一个群体知识关联；
输入模块，用于若至少两个所述分群决策结果不一致，则将所述个体数据输入至分群决策模型后，得到所述分群决策模型输出的最终分群结果；所述最终分群结果与其所属的个体数据的群体值标签关联；
最终分群结果确定模块，用于若至少两个所述分群决策结果均一致，则确定该分群决策结果为所述个体数据的最终分群结果。
一种计算机设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现上述分群决策模型生成方法或分群处理方法。
一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时实现上述分群决策模型生成方法或分群处理方法。
有益效果
上述分群决策模型生成方法、装置、设备及介质，对待分群的个体样本数据进行数据处理后来解决两个群体知识库中与同一个个体样本数据关联的群体知识均相互冲突的问题，也即最后通过训练建立一个分群决策模型来对待分群的个体样本数据进行识别后输出个体 样本数据的样本群体值标签，进而通过该分群决策模型可精准高效确定出待分群的个体样本数据所属的群体，且在模型的识别过程中，能全面考虑群体知识库中的群体知识，识别完成后输出的个体样本数据所属的群体也具备高合理性高的优点。
上述分群处理方法、装置、设备及介质，通过上述训练完成的分群决策模型对待确定群体的个体数据来精准、高效、合理和全面确定出个体数据所属的最终分群结果。
附图说明
图1是本申请一实施例中分群决策模型生成方法的一应用环境示意图；
图2是本申请一实施例中分群决策模型生成方法的一流程示意图；
图3是本申请一实施例中分群决策模型生成装置的结构示意图；
图4是本申请一实施例中分群处理方法的一流程示意图；
图5是本申请一实施例中分群处理装置的结构示意图；
图6是本申请一实施例中计算机设备的一示意图。
本发明的最佳实施方式
本申请提供的分群决策模型生成方法，可应用在如图1的应用环境中，其中，客户端通过网络与服务器进行通信。其中，客户端可以但不限于各种个人计算机、笔记本电脑、智能手机、平板电脑和便携式可穿戴设备。服务器可以用独立的服务器或者是多个服务器组成的服务器集群来实现。
在一实施例中，如图2所示，提供一种分群决策模型生成方法，以该方法应用在图1中的服务器为例进行说明，包括如下步骤：
S10，获取待分群的个体样本数据，利用预设分群决策树建立至少两个群体知识库；每一个所述群体知识库中均包含与所有所述个体样本数据关联的群体知识，所述个体样本数据中包括多个个体特征；至少两个所述群体知识库中与同一个所述个体样本数据关联的群体知识相互冲突；
可理解地，待分群的个体样本数据可指代大量个体在某个应用场景下的长期跟踪数据(每次跟踪数据收集完成后会记录一个跟踪次数)，比如大量个体在饮食方式优化、/用药管理方式/运动方式推荐/相似群体的商品推荐等应用场景下的跟踪数据；且待分群的个体样本数据都存在多个个体特征，因此可通过收集预设时间段的个体样本数据而得到多个个体特征，在饮食方式优化的应用场景下，可收集个体样本数据中的身份信息个体特征(包括但不限于个体的性别、年龄和体重等)、健康情况个体特征(包括但不限于BMI,身体质量指数、体脂率等)、饮食情况个体特征、短期结局个体特征(个体下一次BMI是否在控制范围内)和长期结局个体特征(个体下一次是否已死亡)，或者在用药管理方式的场景下，可收集个体样本数据中的检查指标个体特征、疾病史个体特征、用药史个体特征、短期疗效个体特征和长期疗效个体特征；群体知识库中包含了群体知识，主要是指示人群划分的知识，在上述饮食方式优化的应用场景，群体知识库中的群体知识为饮食方式推荐的人群划分方式，比如“老年人应多吃含钙食物，儿童应补充各种营养”就划分出了“老年人”、“儿童”两个群体。具体地，对每个群体知识库，利用预设分群决策树可将上述提到的饮食方式推荐方式的各种规则转换成决策树的形式，并且需要说明的是，各决策树与同一个个体样本数据关联的群体知识有可能相互冲突，比如A群体知识库、B群体知识库与同一个个体样本数据关联的各种饮食的用量相互冲突。
进一步地，所述获取待分群的个体样本数据，包括：
获取预设时间段的所有所述个体样本数据，并在预设数据图谱中以所述个体样本数据查询出至少一个预设个体特征项目，并以预设个体特征项目收集所述个体特征。
可理解地，个体样本数据的类型不同(应用场景不同)，预设时间段也应该不同，以上述提到的饮食方式优化的应用场景，该应用场景下，预设时间段可为一个季度；由于每个应用场景不同，因此个体样本数据中的个体特征也必须不同；具体可通过预设数据图谱建立起每个应用场景下对应的预设个体特征项目，并在确定出要收集一个应用场景下的个体 样本数据时，可产生出与该个体样本数据相关的各个个体特征；通过本实施例可确定出要收集的个体样本数据的各个个体特征，加快收集的规范性和提高收集的效率性。
S20，通过特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献，将所述短期贡献和所述长期贡献输入至所述特征贡献度评估器的预设的贡献度函数后，输出所述个体特征的特征贡献度；所述特征贡献度评估器基于XGBoost模型的SHAP方法建立；
可理解地，特征贡献度评估器是基于XGBoost模型和SHAP方法建立的，而特征贡献度评估器中主要包括预测短期结局模型和预测长期结局模型(这两个模型预测目标不一致)，且这两个模型均为XGBoost模型。XGBoost模型利用集成思想，训练时串行地生成多棵决策树(其中每棵决策树拟合上一棵决策树的残差)，使用时利用所有决策树经过多轮决策后得到一个预测结果。在训练时，预测短期结局的XGBoost模型和预测长期结局的XGBoost模型使用的数据都为S10中抽取出的大量个体样本数据(除了最后一次跟踪数据，因为最后一次跟踪数据一般没有短期结局和长期结局个体特征)。两个XGBoost模型的输入特征都为已抽取出个体样本数据中的各个体特征(非结局)，预测短期结局的XGBoost模型输出为短期结局个体特征，预测长期结局的XGBoost模型输出为长期结局个体特征。在XGBoost模型训练完成后，使用两个XGBoost模型可调用SHAP方法(通过SHAP值的方式定义特征的重要性，显示XGBoost模型最终如何到达预测输出的预测结果，该预测结果为上述提到的短期结局和长期结局)，SHAP方法是度量机器学习模型的特征重要性的经典方法；在上述提到的饮食方式优化的应用场景下，预测短期结局模型可用于预测个体下一次BMI是否在控制范围内的短期贡献等，而预测长期结局模型可用于预测个体下一次是否已死亡的长期贡献等；特征贡献度评估器的预设的贡献度函数为Contribution(i,j,k)＝SHAP M1(i,j)[k]-α*SHAP M2(i,j)[k]，其中，α为超参数，且α需大于1，大于1是因为长期结局结果一般比短期结局结果重要，所以α用于调节长期结局结果与短期结局结果的相对重要程度，运用相差是因为短期结局结果一般定义为良性结局，因此短期结局结果越大越好，而长期结局一般定义为恶性结局(例如下一次个体是否死亡)，因此长期结局结果越小越好，i指个体样本数据，j指个体样本数据的跟踪次数，k指个体样本数据的某个个体特征；SHAP M1为预测短期结局模型(预测短期结局模型标记为M1)输出的短期贡献，SHAP M2为预测短期结局模型(预测长期结局模型标记为M2)输出的长期贡献。
进一步地，通过所述特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献，包括：
获取所述个体样本数据对应于至少两个所述群体知识库的至少两个分群决策结果；一个所述群体知识库包含至少一个群体，所述分群决策结果为所述个体样本数据属于所述群体知识库包含的一个所述群体；一个所述群体与一个所述群体知识关联；
若至少两个所述分群决策结果不一致，则通过特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献。
可理解地，群体知识库中包含的群体数量与群体知识库包含的群体知识相关，一个群体知识库中包含的群体知识越多，群体知识库中包含的群体数量就越多。
在本实施例中，提前判断个体样本数据是否需要进一步地去执行步骤S20之后的训练步骤，可通过获取个体样本数据对应于至少两个群体知识库的至少两个分群决策结果，并 用来判定分群决策结果是否为一致，比如，在饮食方式优化的应用场景下，A群体知识库中的群体知识向该个体样本数据推荐所属群体的蔬菜饮食用量为80g-100g，而B群体知识库中的群体知识向该个体样本数据推荐所属群体的蔬菜饮食用量为30g-50g，可见两个分群决策结果不一致而将引起相互冲突。因此需执行步骤S20之后的训练步骤来解决上述的相互冲突的问题。可见，本实施例可有效通过对应的手段解决个体样本数据所属群体的问题。
在一实施例中，所述获取所述个体样本数据对应于至少两个所述群体知识库的至少两个分群决策结果之后，还包括：
若至少两个所述分群决策结果均一致，则删除所述个体样本数据。
S30，获取所述个体样本数据与各所述群体知识库中相关联的所述群体知识之间的匹配关系，根据获取的所有所述匹配关系为所述个体样本数据标记样本群体值标签后，将所述个体样本数据的所有所述个体特征的特征贡献度输入至预设reward函数后，输出所述个体样本数据的reward值，并将所述个体样本数据的群体知识值标签定义为行为变量，将所述个体样本数据的个体特征定义为状态变量；
可理解地，步骤S10所在的实施例只是建立了与个体样本数据关联的群体知识库，步骤S20过滤掉了各群体知识库中群体知识的分群决策结果一致的个体样本数据，而本实施例是为了确定个体样本数据与关联的群体知识库直接的匹配关系，也即从个体样本数据中的各个个体特征来确定与群体知识库中的群体知识之间的匹配关系，比如，在饮食方式优化的应用场景下，个体样本数据中的饮食情况个体特征为一天蔬菜饮食用量为40g，而A群体知识库中的群体知识向该个体样本数据推荐所属群体的蔬菜饮食用量为80g-100g，B群体知识库中的群体知识向该个体样本数据推荐所属群体的蔬菜饮食用量为30g-50g，因此个体样本数据与A群体知识库中的群体知识不匹配而跟A群体知识库中的群体知识匹配，从而通过获取的所有匹配关系为个体样本数据标记样本群体值标签(0,1)，可通过类似的原理(符合群体知识库中的群体知识就标记1，不符合群体知识库中的群体知识就标记0)从而可得到其他情况下的样本群体值标签(数学上的坐标值)，在此需要说明的是，存在的群体知识库越多，标记样本群体值标签的坐标维度就越多；预设reward函数为reward(i,j)＝∑ kek(contribution(i,j+1,k)-contribution(i,j,k)，其中，K指个体样本数据的所有个体特征(而k是指个体样本数据的一个个体特征)，i和j同上述提到的贡献度函数一致；行为变量为强化学习中的action，状态变量为强化学习中的state，强化学习是通过“试错”的方式进行学习，在state的条件下执行某个action后，通过与环境交互获得的reward值来优化action的选择。
S40，将所述个体样本数据的行为变量、状态变量和reward值输入至待训练预设分群决策模型的free-stable DQN网络进行训练，获取free-stable DQN网络根据所述行为变量和所述状态变量生成Q值，并将所述个体样本数据的所述reward值和所述Q值输入至所述待训练分群决策型的损失函数后获取所述个体样本数据的损失值；所述损失函数包括所述reward值和所述Q值；
可理解地，基于Q-learning的强化学习方法是用一个Q值表来存储每一个状态state和在这个状态state条件下每个行为action所拥有的Q值，但state中很多维度是连续的，用表格来存储每个state的每个action的Q值是一件很困难的事，而free-stable DQN(使用一个四层DQN网络，输入层与state有相同维度，中间有两个维度分别为32、64的隐藏层，输出层与action有相同维度，上述提到的四个层之间均为全连接。)融合的神经网络和Q-learning，因此利用神经网络来生成Q值以解决state维度是连续的问题；Q值的生成公式为
Figure PCTCN2020098829-appb-000001
而损失函数为
Figure PCTCN2020098829-appb-000002
其中,
Figure PCTCN2020098829-appb-000003
φ j为step j(在此提到的j表示强化学习里个体样本数据一个episode的第j个step，可以理解为上述S20提到个体样本数据的跟踪次数)时的状态变量，
Figure PCTCN2020098829-appb-000004
为step j时的行为变量，θ为网络参数，r j为个体样本数据在step j的reward值，γ为衰减因子(表示距离当前step j越远，受当前的reward影响越小)，Q为Q值，表示预期价值，φ j+1为下一个状态变量(step j+1时的状态)，
Figure PCTCN2020098829-appb-000005
为可以得到最大Q值的行为，θ -为同θ一致，也为网络参数，θ t-1为当前训练轮次的l轮之前的参数(l为可指定的超参数)。需要说明的是，由于在对样本数据进行群体划分时，群体划分效果会在某个个体样本数据的下一次跟踪数据才会产生，因此需利用多个个体样本数据不断回放来进行学习训练，但此时reward值的定义对该传统DQN模型的训练影响还是比较敏感的，且传统DQN模型中的损失函数由于与最大的Q值有关，在接近收敛时，传统DQN模型的参数可能还在波动，因此本实施例重新定义了传统DQN模型中的损失函数，因此使用该待训练分群决策型的损失函数可减少上述定义的reward值的主观性，其中在该损失函数加入类似正则化项的功能可减少强化学习中参数波动的问题，从而可提高待训练分群决策型的稳定性。
S50，在判定所述个体样本数据在经历过预设训练早停系数后所述损失值不再下降时，将所述待训练分群决策模型标记为训练完成的分群决策模型。
可理解地，在预设训练早停系数的轮数训练后，确定待训练分群决策模型中的损失值未发生大幅度波动且不再下降就可以说明待训练分群决策模型为趋于收敛，也可说明待训练分群决策模型已训练完成为分群决策模型。
综上所述，上述提供了一种分群决策模型生成方法，对待分群的个体样本数据进行数据处理后来解决两个群体知识库中与同一个个体样本数据关联的群体知识均相互冲突的问题，也即最后通过训练建立一个分群决策模型来对待分群的个体样本数据进行识别后输出个体样本数据的样本群体值标签，进而通过该分群决策模型可精准和高效确定出待分群的个体样本数据所属的群体，且在模型的识别过程中，能较全面考虑群体知识库中的群体知识，识别完成后输出的个体样本数据所属的群体也具备高合理性高的优点。
应理解，上述实施例中各步骤的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定，而不应对本申请实施例的实施过程构成任何限定。
在一实施例中，提供一种分群决策模型生成装置，该分群决策模型生成装置与上述实施例中分群决策模型生成方法一一对应。如图3所示，该分群决策模型生成装置包括建立模块11、输出模块12、定义模块13、获取模块14和标记模块15。各功能模块详细说明如下：
建立模块11，用于获取待分群的个体样本数据，利用预设分群决策树建立至少两个群体知识库；每一个所述群体知识库中均包含与所有所述个体样本数据关联的群体知识，所述个体样本数据中包括多个个体特征；至少两个所述群体知识库中与同一个所述个体样本数据关联的群体知识相互冲突；
输出模块12，用于通过特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献，将所述短期贡献和所述长期贡献输入至所述特征贡献度评估器的预设的贡献度函数后，输出所述个体特征的特征贡献度；所述特征贡献度评估器基于XGBoost模型的SHAP方法建立；
定义模块13，用于获取所述个体样本数据与各所述群体知识库中相关联的所述群体知 识之间的匹配关系，根据获取的所有所述匹配关系为所述个体样本数据标记样本群体值标签后，将所述个体样本数据的所有所述个体特征的特征贡献度输入至预设reward函数后，输出所述个体样本数据的reward值，并将所述个体样本数据的群体知识值标签定义为行为变量，将所述个体样本数据的个体特征定义为状态变量；
获取模块14，用于将所述个体样本数据的行为变量、状态变量和reward值输入至待训练预设分群决策模型的free-stable DQN网络进行训练，获取free-stable DQN网络根据所述行为变量和所述状态变量生成Q值，并将所述个体样本数据的所述reward值和所述Q值输入至所述待训练分群决策型的损失函数后获取所述个体样本数据的损失值；
标记模块15，用于在判定所述个体样本数据在经历过预设训练早停系数后所述损失值不再下降时，将所述待训练分群决策模型标记为训练完成的分群决策模型。
进一步地，所述建立模块包括：
收集子模块，用于获取预设时间段的所有所述个体样本数据，并在预设数据图谱中以所述个体样本数据查询出至少一个预设个体特征项目，并以预设个体特征项目收集所述个体特征。
进一步地，所述输出模块包括：
获取子模块，用于获取所述个体样本数据对应于至少两个所述群体知识库的至少两个分群决策结果；一个所述群体知识库包含至少一个群体，所述分群决策结果为所述个体样本数据属于所述群体知识库包含的一个所述群体；一个所述群体与一个所述群体知识关联；
评估子模块，用于若至少两个所述分群决策结果不一致，则通过特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献。
进一步地，所述输出模块还包括：
删除子模块，用于若至少两个所述分群决策结果均一致，则删除所述个体样本数据。
关于分群决策模型生成装置的具体限定可以参见上文中对于分群决策模型生成方法的限定，在此不再赘述。上述分群决策模型生成装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中，也可以以软件形式存储于计算机设备中的存储器中，以便于处理器调用执行以上各个模块对应的操作。
本申请还提供的评估数据处理方法，可应用在如图1的应用环境中，其中，客户端通过网络与服务器进行通信。其中，客户端可以但不限于各种个人计算机、笔记本电脑、智能手机、平板电脑和便携式可穿戴设备。服务器可以用独立的服务器或者是多个服务器组成的服务器集群来实现。
在一实施例中，如图4所示，提供一种分群处理方法，以该方法应用在图1中的服务端为例进行说明，包括如下步骤：
S50，获取待确定群体的个体数据，并获取所述个体数据中与至少两个群体知识库一一对应的至少两个分群决策结果；一个所述群体知识库包含至少一个确定群体，所述分群决策结果是指所述个体数据属于所述群体知识库中包含的其中一个所述确定群体；一个所述确定群体与一个所述群体知识库中的一个群体知识关联；可理解地，上述提到的确定群体包括但不限于饮食群体、运动群体和相似商品推荐群体等。
S60，若至少两个所述分群决策结果不一致，则将所述个体数据输入至分群决策模型后，得到所述分群决策模型输出的最终分群结果；所述最终分群结果与其所属的个体数据的群体值标签关联；可理解地，训练完成分群决策模型用于识别个体样本数据后输出个体样本数据的群体值标签，并通过群体值标签确定所属的群体，比如，若得到的群体值标签还为上述步骤S30提到的(0，1)，则说明个体样本数据所属的群体为B群体知识库中所对应的确定群体，若得到的群体值标签为(1，0)，则说明个体样本数据所属的群体为A群体 知识库中所对应的确定群体，通过确认出来的确定群体可进一步地向用户推荐适合的三餐的饮食情况、运动方式推荐和似群体的商品推荐等。需要说明的是，在模型的使用过程中可对分群决策模型进行优化，将所述个体数据的下一次新的个体数据中的所有个体特征关联的reward值、行为变量和状态变量直接输入至分群决策模型后以对分群决策模型进行优化而适应新的个体样本数据；
S70，若至少两个所述分群决策结果均一致，则确定该分群决策结果为所述个体数据的最终分群结果。
可理解地，在至少两个分群决策结果不一致时，通过分群决策模型去识别个体数据后，获取该模型输出的个体数据的群体值标签，并通过群体值标签确定出个体数据的最终分群结果，与上述提到的步骤S50原理一致；在两个分群决策结果一致时，个体样本数据可同时属于分群决策结果所包含的确定群体。
综上所述，上述提供了一种分群处理方法，通过上述训练完成的分群决策模型对待确定群体的个体数据来精准、高效、合理和全面确定出个体数据所属的最终分群结果。
应理解，上述实施例中各步骤的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定，而不应对本申请实施例的实施过程构成任何限定。
在一实施例中，提供一种分群处理装置，该分群处理装置与上述实施例中评估数据处理方法一一对应。如图5所示，该分群处理装置包括分群决策结果获取模块21、输入模块22和最终分群结果确定模块23。各功能模块详细说明如下：
分群决策结果获取模块21，用于获取待确定群体的个体数据，并获取所述个体数据中与至少两个群体知识库一一对应的至少两个分群决策结果；一个所述群体知识库包含至少一个确定群体，所述分群决策结果是指所述个体数据属于所述群体知识库中包含的其中一个所述确定群体；一个所述确定群体与一个所述群体知识库中的一个群体知识关联；
输入模块22，用于若至少两个所述分群决策结果不一致，则将所述个体数据输入至分群决策模型后，得到所述分群决策模型输出的最终分群结果；所述最终分群结果与其所属的个体数据的群体值标签关联；
最终分群结果确定模块23，用于若至少两个所述分群决策结果均一致，则确定该分群决策结果为所述个体数据的最终分群结果。
关于分群处理装置的具体限定可以参见上文中对于评估模型生成方法的限定，在此不再赘述。上述分群处理装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中，也可以以软件形式存储于计算机设备中的存储器中，以便于处理器调用执行以上各个模块对应的操作。
在一个实施例中，提供了一种计算机设备，该计算机设备可以是服务器，其内部结构图可以如图6所示。该计算机设备包括通过系统总线连接的处理器、存储器、网络接口和数据库。其中，该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统、计算机程序和数据库。该内存储器为非易失性存储介质中的操作系统和计算机程序的运行提供环境。该计算机设备的数据库用于存储分群决策模型生成方法中涉及到的数据。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机程序被处理器执行时以实现一种分群决策模型生成方法，或者该计算机程序被处理器执行时以实现一种分群处理方法。
在一个实施例中，提供了一种计算机设备，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，处理器执行计算机程序时实现上述实施例中分群决策模型生成方法的步骤，或者处理器执行计算机程序时实现上述实施例中分群处理方法的步骤。
在一个实施例中，提供了一种计算机可读存储介质，其上存储有计算机程序，所述计算机可读存储介质可以是非易失性，也可以是易失性，计算机程序被处理器执行时实现上述实施例中分群决策模型生成方法的步骤，或者计算机程序被处理器执行时实现上述实施例中分群处理方法的步骤。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机程序来指令相关的硬件来完成，所述的计算机程序可存储于一非易失性计算机可读取存储介质中，该计算机程序在执行时，可包括如上述各方法的实施例的流程。其中，本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用，均可包括非易失性和/或易失性存储器。非易失性存储器可包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限，RAM以多种形式可得，诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
本申请所指区块链是分布式数据存储、点对点传输、共识机制、加密算法等计算机技术的新型应用模式。区块链(Blockchain)，本质上是一个去中心化的数据库，是一串使用密码学方法相关联产生的数据块，每一个数据块中包含了一批次网络交易的信息，用于验证其信息的有效性(防伪)和生成下一个区块。区块链可以包括区块链底层平台、平台产品服务层以及应用服务层等。
所属领域的技术人员可以清楚地了解到，为了描述的方便和简洁，仅以上述各功能单元、模块的划分进行举例说明，实际应用中，可以根据需要而将上述功能分配由不同的功能单元、模块完成，即将所述装置的内部结构划分成不同的功能单元或模块，以完成以上描述的全部或者部分功能。
以上所述实施例仅用以说明本申请的技术方案，而非对其限制；尽管参照前述实施例对本申请进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围，均应包含在本申请的保护范围之内。

Claims (20)

  1. 一种分群决策模型生成方法，其中，包括：
    获取待分群的个体样本数据，利用预设分群决策树建立至少两个群体知识库；每一个所述群体知识库中均包含与所有所述个体样本数据关联的群体知识，所述个体样本数据中包括多个个体特征；至少两个所述群体知识库中与同一个所述个体样本数据关联的群体知识相互冲突；
    通过特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献，将所述短期贡献和所述长期贡献输入至所述特征贡献度评估器的预设的贡献度函数后，输出所述个体特征的特征贡献度；所述特征贡献度评估器基于XGBoost模型的SHAP方法建立；
    获取所述个体样本数据与各所述群体知识库中相关联的所述群体知识之间的匹配关系，根据获取的所有所述匹配关系为所述个体样本数据标记样本群体值标签后，将所述个体样本数据的所有所述个体特征的特征贡献度输入至预设reward函数后，输出所述个体样本数据的reward值，并将所述个体样本数据的群体知识值标签定义为行为变量，将所述个体样本数据的个体特征定义为状态变量；
    将所述个体样本数据的行为变量、状态变量和reward值输入至待训练预设分群决策模型的free-stable DQN网络进行训练，获取free-stable DQN网络根据所述行为变量和所述状态变量生成Q值，并将所述个体样本数据的所述reward值和所述Q值输入至所述待训练分群决策型的损失函数后获取所述个体样本数据的损失值；
    在判定所述个体样本数据在经历过预设训练早停系数后所述损失值不再下降时，将所述待训练分群决策模型标记为训练完成的分群决策模型。
  2. 根据权利要求1所述的分群决策模型生成方法，其中，所述获取待分群的个体样本数据，包括：
    获取预设时间段的所有所述个体样本数据，并在预设数据图谱中以所述个体样本数据查询出至少一个预设个体特征项目，并以预设个体特征项目收集所述个体特征。
  3. 根据权利要求1所述的分群决策模型生成方法，其中，所述通过所述特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献，包括：
    获取所述个体样本数据对应于至少两个所述群体知识库的至少两个分群决策结果；一个所述群体知识库包含至少一个群体，所述分群决策结果为所述个体样本数据属于所述群体知识库包含的一个所述群体；一个所述群体与一个所述群体知识关联；
    若至少两个所述分群决策结果不一致，则通过特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献。
  4. 根据权利要求3所述的分群决策模型生成方法，其中，所述获取所述个体样本数据对应于至少两个所述群体知识库的至少两个分群决策结果之后，还包括：
    若至少两个所述分群决策结果均一致，则删除所述个体样本数据。
  5. 一种分群处理方法，其中，包括：
    获取待确定群体的个体数据，并获取所述个体数据中与至少两个群体知识库一一对应的至少两个分群决策结果；一个所述群体知识库包含至少一个确定群体，所述分群决策结果是指所述个体数据属于所述群体知识库中包含的其中一个所述确定群体；一个所述确定群体与一个所述群体知识库中的一个群体知识关联；
    若至少两个所述分群决策结果不一致，则将所述个体数据输入至分群决策模型后，得 到所述分群决策模型输出的最终分群结果；所述最终分群结果与其所属的个体数据的群体值标签关联；
    若至少两个所述分群决策结果均一致，则确定该分群决策结果为所述个体数据的最终分群结果。
  6. 一种分群决策模型生成装置，其中，包括如下模块：
    建立模块，用于获取待分群的个体样本数据，利用预设分群决策树建立至少两个群体知识库；每一个所述群体知识库中均包含与所有所述个体样本数据关联的群体知识，所述个体样本数据中包括多个个体特征；至少两个所述群体知识库中与同一个所述个体样本数据关联的群体知识相互冲突；
    输出模块，用于通过特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献，将所述短期贡献和所述长期贡献输入至所述特征贡献度评估器的预设的贡献度函数后，输出所述个体特征的特征贡献度；所述特征贡献度评估器基于XGBoost模型的SHAP方法建立；
    定义模块，用于获取所述个体样本数据与各所述群体知识库中相关联的所述群体知识之间的匹配关系，根据获取的所有所述匹配关系为所述个体样本数据标记样本群体值标签后，将所述个体样本数据的所有所述个体特征的特征贡献度输入至预设reward函数后，输出所述个体样本数据的reward值，并将所述个体样本数据的群体知识值标签定义为行为变量，将所述个体样本数据的个体特征定义为状态变量；
    获取模块，用于将所述个体样本数据的行为变量、状态变量和reward值输入至待训练预设分群决策模型的free-stable DQN网络进行训练，获取free-stable DQN网络根据所述行为变量和所述状态变量生成Q值，并将所述个体样本数据的所述reward值和所述Q值输入至所述待训练分群决策型的损失函数后获取所述个体样本数据的损失值；
    标记模块，用于在判定所述个体样本数据在经历过预设训练早停系数后所述损失值不再下降时，将所述待训练分群决策模型标记为训练完成的分群决策模型。
  7. 根据权利要求6所述的分群决策模型生成装置，其中，所述输出模块包括：
    获取子模块，用于获取所述个体样本数据对应于至少两个所述群体知识库的至少两个分群决策结果；一个所述群体知识库包含至少一个群体，所述分群决策结果为所述个体样本数据属于所述群体知识库包含的一个所述群体；一个所述群体与一个所述群体知识关联；
    评估子模块，用于若至少两个所述分群决策结果不一致，则通过特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献。
  8. 根据权利要求7所述的分群决策模型生成装置，其中，所述输出模块还包括：
    删除子模块，用于若至少两个所述分群决策结果均一致，则删除所述个体样本数据。
  9. 根据权利要求6所述的分群决策模型生成装置，其中，所述建立模块包括：
    收集子模块，用于获取预设时间段的所有所述个体样本数据，并在预设数据图谱中以所述个体样本数据查询出至少一个预设个体特征项目，并以预设个体特征项目收集所述个体特征。
  10. 一种分群处理装置，其中，包括如下模块：
    分群决策结果获取模块，用于获取待确定群体的个体数据，并获取所述个体数据中与至少两个群体知识库一一对应的至少两个分群决策结果；一个所述群体知识库包含至少一个确定群体，所述分群决策结果是指所述个体数据属于所述群体知识库中包含的其中一个所述确定群体；一个所述确定群体与一个所述群体知识库中的一个群体知识关联；
    输入模块，用于若至少两个所述分群决策结果不一致，则将所述个体数据输入至分群决策模型后，得到所述分群决策模型输出的最终分群结果；所述最终分群结果与其所属的 个体数据的群体值标签关联；
    最终分群结果确定模块，用于若至少两个所述分群决策结果均一致，则确定该分群决策结果为所述个体数据的最终分群结果。
  11. 一种计算机设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，其中，所述处理器执行所述计算机程序时实现分群决策模型生成方法，包括：
    获取待分群的个体样本数据，利用预设分群决策树建立至少两个群体知识库；每一个所述群体知识库中均包含与所有所述个体样本数据关联的群体知识，所述个体样本数据中包括多个个体特征；至少两个所述群体知识库中与同一个所述个体样本数据关联的群体知识相互冲突；
    通过特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献，将所述短期贡献和所述长期贡献输入至所述特征贡献度评估器的预设的贡献度函数后，输出所述个体特征的特征贡献度；所述特征贡献度评估器基于XGBoost模型的SHAP方法建立；
    获取所述个体样本数据与各所述群体知识库中相关联的所述群体知识之间的匹配关系，根据获取的所有所述匹配关系为所述个体样本数据标记样本群体值标签后，将所述个体样本数据的所有所述个体特征的特征贡献度输入至预设reward函数后，输出所述个体样本数据的reward值，并将所述个体样本数据的群体知识值标签定义为行为变量，将所述个体样本数据的个体特征定义为状态变量；
    将所述个体样本数据的行为变量、状态变量和reward值输入至待训练预设分群决策模型的free-stable DQN网络进行训练，获取free-stable DQN网络根据所述行为变量和所述状态变量生成Q值，并将所述个体样本数据的所述reward值和所述Q值输入至所述待训练分群决策型的损失函数后获取所述个体样本数据的损失值；
    在判定所述个体样本数据在经历过预设训练早停系数后所述损失值不再下降时，将所述待训练分群决策模型标记为训练完成的分群决策模型。
  12. 根据权利要求11所述的计算机设备，其中，所述获取待分群的个体样本数据，包括：
    获取预设时间段的所有所述个体样本数据，并在预设数据图谱中以所述个体样本数据查询出至少一个预设个体特征项目，并以预设个体特征项目收集所述个体特征。
  13. 根据权利要求11所述的计算机设备，其中，所述通过所述特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献，包括：
    获取所述个体样本数据对应于至少两个所述群体知识库的至少两个分群决策结果；一个所述群体知识库包含至少一个群体，所述分群决策结果为所述个体样本数据属于所述群体知识库包含的一个所述群体；一个所述群体与一个所述群体知识关联；
    若至少两个所述分群决策结果不一致，则通过特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献。
  14. 根据权利要求13所述的计算机设备，其中，所述获取所述个体样本数据对应于至少两个所述群体知识库的至少两个分群决策结果之后，还包括：
    若至少两个所述分群决策结果均一致，则删除所述个体样本数据。
  15. 一种计算机设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，其中，所述处理器执行所述计算机程序时实现分群处理方法，包括：
    获取待确定群体的个体数据，并获取所述个体数据中与至少两个群体知识库一一对应的至少两个分群决策结果；一个所述群体知识库包含至少一个确定群体，所述分群决策结果是指所述个体数据属于所述群体知识库中包含的其中一个所述确定群体；一个所述确定群体与一个所述群体知识库中的一个群体知识关联；
    若至少两个所述分群决策结果不一致，则将所述个体数据输入至分群决策模型后，得到所述分群决策模型输出的最终分群结果；所述最终分群结果与其所属的个体数据的群体值标签关联；
    若至少两个所述分群决策结果均一致，则确定该分群决策结果为所述个体数据的最终分群结果。
  16. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，其中，所述计算机程序被处理器执行时实现分群决策模型生成方法，包括：
    获取待分群的个体样本数据，利用预设分群决策树建立至少两个群体知识库；每一个所述群体知识库中均包含与所有所述个体样本数据关联的群体知识，所述个体样本数据中包括多个个体特征；至少两个所述群体知识库中与同一个所述个体样本数据关联的群体知识相互冲突；
    通过特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献，将所述短期贡献和所述长期贡献输入至所述特征贡献度评估器的预设的贡献度函数后，输出所述个体特征的特征贡献度；所述特征贡献度评估器基于XGBoost模型的SHAP方法建立；
    获取所述个体样本数据与各所述群体知识库中相关联的所述群体知识之间的匹配关系，根据获取的所有所述匹配关系为所述个体样本数据标记样本群体值标签后，将所述个体样本数据的所有所述个体特征的特征贡献度输入至预设reward函数后，输出所述个体样本数据的reward值，并将所述个体样本数据的群体知识值标签定义为行为变量，将所述个体样本数据的个体特征定义为状态变量；
    将所述个体样本数据的行为变量、状态变量和reward值输入至待训练预设分群决策模型的free-stable DQN网络进行训练，获取free-stable DQN网络根据所述行为变量和所述状态变量生成Q值，并将所述个体样本数据的所述reward值和所述Q值输入至所述待训练分群决策型的损失函数后获取所述个体样本数据的损失值；
    在判定所述个体样本数据在经历过预设训练早停系数后所述损失值不再下降时，将所述待训练分群决策模型标记为训练完成的分群决策模型。
  17. 根据权利要求16所述的计算机可读存储介质，其中，所述获取待分群的个体样本数据，包括：
    获取预设时间段的所有所述个体样本数据，并在预设数据图谱中以所述个体样本数据查询出至少一个预设个体特征项目，并以预设个体特征项目收集所述个体特征。
  18. 根据权利要求16所述的计算机可读存储介质，其中，所述通过所述特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献，包括：
    获取所述个体样本数据对应于至少两个所述群体知识库的至少两个分群决策结果；一个所述群体知识库包含至少一个群体，所述分群决策结果为所述个体样本数据属于所述群体知识库包含的一个所述群体；一个所述群体与一个所述群体知识关联；
    若至少两个所述分群决策结果不一致，则通过特征贡献度评估器的预测短期结局模型调用SHAP方法评估所述个体样本数据中个体特征的短期贡献，通过所述特征贡献度评估器的预测长期结局模型调用SHAP方法评估所述个体样本数据中个体特征的长期贡献。
  19. 根据权利要求18所述的计算机可读存储介质，其中，所述获取所述个体样本数据 对应于至少两个所述群体知识库的至少两个分群决策结果之后，还包括：
    若至少两个所述分群决策结果均一致，则删除所述个体样本数据。
  20. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，其中，所述计算机程序被处理器执行时实现分群处理方法，包括：
    获取待确定群体的个体数据，并获取所述个体数据中与至少两个群体知识库一一对应的至少两个分群决策结果；一个所述群体知识库包含至少一个确定群体，所述分群决策结果是指所述个体数据属于所述群体知识库中包含的其中一个所述确定群体；一个所述确定群体与一个所述群体知识库中的一个群体知识关联；
    若至少两个所述分群决策结果不一致，则将所述个体数据输入至分群决策模型后，得到所述分群决策模型输出的最终分群结果；所述最终分群结果与其所属的个体数据的群体值标签关联；
    若至少两个所述分群决策结果均一致，则确定该分群决策结果为所述个体数据的最终分群结果。
PCT/CN2020/098829 2020-05-13 2020-06-29 分群决策模型生成、分群处理方法、装置、设备及介质 WO2021139106A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010403130.5 2020-05-13
CN202010403130.5A CN111666494B (zh) 2020-05-13 2020-05-13 分群决策模型生成、分群处理方法、装置、设备及介质

Publications (1)

Publication Number Publication Date
WO2021139106A1 true WO2021139106A1 (zh) 2021-07-15

Family

ID=72382560

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/098829 WO2021139106A1 (zh) 2020-05-13 2020-06-29 分群决策模型生成、分群处理方法、装置、设备及介质

Country Status (2)

Country Link
CN (1) CN111666494B (zh)
WO (1) WO2021139106A1 (zh)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112115322B (zh) * 2020-09-25 2024-05-07 平安科技（深圳）有限公司 用户分群方法、装置、电子设备及存储介质
CN112331254B (zh) * 2020-11-24 2024-08-06 北京泽石科技有限公司 非易失性存储器的试错表的生成方法及装置
CN112819527B (zh) * 2021-01-29 2024-05-24 百果园技术(新加坡)有限公司 一种用户分群处理方法及装置
CN112836105B (zh) * 2021-02-05 2022-05-24 浙江工业大学 一种基于运动生理表征融合的大规模学生有氧能力分群方法
CN113763111B (zh) * 2021-02-10 2024-09-20 北京沃东天骏信息技术有限公司 物品搭配方法、装置及存储介质
CN113076486B (zh) * 2021-04-29 2023-07-25 平安科技（深圳）有限公司 药物信息推送方法、装置、计算机设备及存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109255499A (zh) * 2018-10-25 2019-01-22 阿里巴巴集团控股有限公司 投诉、投诉案件处理方法、装置及设备
CN110929752A (zh) * 2019-10-18 2020-03-27 平安科技（深圳）有限公司 基于知识驱动和数据驱动的分群方法及相关设备
US20200111565A1 (en) * 2014-10-15 2020-04-09 Brighterion, Inc. Method of personalizing, individualizing, and automating the management of healthcare fraud-waste-abuse to unique individual healthcare providers
CN111105266A (zh) * 2019-11-11 2020-05-05 中国建设银行股份有限公司 基于改进决策树的客户分群方法及装置

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120095955A1 (en) * 2008-03-06 2012-04-19 Arun Darlie Koshy Facilitating relationships and information transactions
CN109801705B (zh) * 2018-12-12 2024-07-19 平安科技（深圳）有限公司 治疗推荐方法、系统、装置及存储介质
CN110956224B (zh) * 2019-08-01 2024-03-08 平安科技（深圳）有限公司 评估模型生成、评估数据处理方法、装置、设备及介质

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200111565A1 (en) * 2014-10-15 2020-04-09 Brighterion, Inc. Method of personalizing, individualizing, and automating the management of healthcare fraud-waste-abuse to unique individual healthcare providers
CN109255499A (zh) * 2018-10-25 2019-01-22 阿里巴巴集团控股有限公司 投诉、投诉案件处理方法、装置及设备
CN110929752A (zh) * 2019-10-18 2020-03-27 平安科技（深圳）有限公司 基于知识驱动和数据驱动的分群方法及相关设备
CN111105266A (zh) * 2019-11-11 2020-05-05 中国建设银行股份有限公司 基于改进决策树的客户分群方法及装置

Also Published As

Publication number Publication date
CN111666494B (zh) 2022-08-12
CN111666494A (zh) 2020-09-15

Similar Documents

Publication Publication Date Title
WO2021139106A1 (zh) 分群决策模型生成、分群处理方法、装置、设备及介质
US11922308B2 (en) Generating neighborhood convolutions within a large network
WO2022161202A1 (zh) 多媒体资源分类模型训练方法和多媒体资源推荐方法
CN111883228B (zh) 基于知识图谱的健康信息推荐方法、装置、设备及介质
TWI612488B (zh) 用於預測商品的市場需求的計算機裝置與方法
WO2020119030A1 (zh) 用于答复问题的模型训练方法、装置、设备及存储介质
WO2020143233A1 (zh) 评分卡模型的建立方法、装置、计算机设备和存储介质
CN111506723B (zh) 问答响应方法、装置、设备及存储介质
US20230385553A1 (en) Techniques to add smart device information to machine learning for increased context
WO2021203854A1 (zh) 用户分类方法、装置、计算机设备和存储介质
CN108510402A (zh) 险种信息推荐方法、装置、计算机设备及存储介质
CN113570064A (zh) 利用复合机器学习模型来执行预测的方法及系统
CN110956224A (zh) 评估模型生成、评估数据处理方法、装置、设备及介质
US20200161000A1 (en) Method and apparatus for prediction of complications after surgery
Menakadevi et al. Robust optimization based extreme learning machine for sentiment analysis in big data
US20220414687A1 (en) Method, device, equipment and medium for determining customer tabs based on deep learning
Paigude et al. Deep Learning Model for Work-Life Balance Prediction for Working Women in IT Industry
WO2023134087A1 (zh) 问诊模板生成方法、装置、电子设备及存储介质
CN117009621A (zh) 信息搜索方法、装置、电子设备、存储介质及程序产品
Hu et al. Predictive analysis of hospital HIS system usage satisfaction based on machine learning
CN114118882B (zh) 基于组合模型的服务数据处理方法、装置、设备和介质
Llanes et al. Original Research Article Comparison and Selection of Artificial Intelligence Technology in Pre-dicting Milk Yield
CN116993218A (zh) 基于人工智能的指标分析方法、装置、设备及存储介质
CN114022270A (zh) 资产数据处理方法、相关设备及介质
CN118228178A (zh) 基于人工智能的异常数据处理方法、装置、设备及介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20912411

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20912411

Country of ref document: EP

Kind code of ref document: A1