JPWO2016121053A1 - Computer system and graphical model management method - Google Patents

Abstract

A computer system including a computing device and a computer having a memory, and a probability that a node to be inferred has a predetermined state value by accepting input of evidence data and a graphical model construction unit that constructs a graphical model using learning data An inference unit that calculates an inference result that is a distribution of values, and the inference unit includes an impact calculation unit that calculates an impact evaluation value, and a probability value differential amount calculation unit that calculates a probability value differential amount The probability value derivative calculation unit extracts a node having a dependency relationship with the inference target node, selects a processing target node from the extracted nodes, obtains a probability table of the processing target node, A set of one or more nodes that have a dependency relationship with the target node is extracted as a family node, and the inference target node, the processing target node, and the family node are combined. Was calculated joint probability distribution of the joint probability variables, calculates the probability value differential quantity using the probability table of the joint probability distribution and processing node.

Description

  The present invention relates to a probability inference method in a graphical model.

  Graphical models such as Bayesian networks are often used in causality analysis and prediction model construction.

  The graphical model is composed of nodes that are random variables and edges that connect the nodes. An edge indicates a dependency between random variables. Of the two nodes having a dependency relationship, a node that affects one node is called a parent node, and a node that is affected by one node is called a child node. Furthermore, a set of one node and a parent node of the node is referred to as a family node of the node.

  Each node is given a conditional probability table. The conditional probability table represents a distribution of conditional probability values of a certain node for each state value vector of a parent node of the certain node. Here, the conditional probability distribution means a distribution of conditional probability values of state values that a node can take.

  In the graphical model, when a value is input to a predetermined random variable, the probability value of the random variable to be predicted can be obtained. Thereby, for example, future trends and the like can be predicted from past data.

  In this specification, the random variable to be predicted is called the inference target, and the distribution of probability values (probability distribution) using the conditional probability value is called inference, and the probability distribution of the inference target is inferred. Call the result. In this specification, the confidence interval of each probability value in the probability distribution of the inference object is referred to as a confidence interval.

Tim Van Allen, Ajit Singh, Russell Greiner, Peter Hooper, “Quantifying the uncertainty of a belief net response: Bayesian error-bars for belief net inference”, Artificial Intelligence 172 (2008) 483-513

  Since the graphical model is generated by learning using a finite number of data, it has an index of statistical reliability. It is important to quantify the statistical uncertainty of the graphical model, based on this, evaluate the reliability of the prediction using the graphical model, and reconstruct the graphical model.

  As described in Non-Patent Document 1, the influence of the probability value in the probability table on the inference target is given as shown in the following equation (1). Equation (1) corresponds to Equation (8) in Non-Patent Document 1.

  Here, the probability value differential amount represents the change amount of the probability value of the inference result with respect to the minute change of the conditional probability value. As shown in Expression (1), the influence evaluation value is given as a function of the probability value differential amount.

  Here, Non-Patent Document 1 describes that the probability value differential amount is defined as in the following equation (2), and specifically, the following equation (3) may be calculated. Equation (2) corresponds to Equation (6) in Non-Patent Literature 1, and Equation (3) corresponds to Equation (14) in Non-Patent Literature 1.

  In the case of the computer method using the intermediate result of the conventional variable elimination method as shown in Equation (3), the computational cost of the variable elimination method, which is a strict inference method, increases as the scale of the graphical model increases. There is a problem that you can not.

  The present invention provides an apparatus and a method capable of calculating an impact evaluation value while suppressing calculation cost even in a large-scale graphical model.

  A typical example of the invention disclosed in the present application is as follows. That is, a computer system including an arithmetic unit that executes a program and one or more computers having a memory that stores the program, and manages learning data including records including a plurality of items corresponding to random variables A learning data storage unit, and using the learning data, a node corresponding to the random variable, an edge indicating a dependency relationship between the nodes, and a distribution of probability values for each state value of the random variable corresponding to the node A graphical model constructing unit that constructs a graphical model including a probability table indicating the structure information of the graphical model constructed by the graphical model constructing unit, and the probability of each of a plurality of nodes included in the graphical model A model information storage unit for managing the table, and a compound variable serving as the random variable. An inference unit that accepts input of evidence data that is composed of records containing items and stores values in at least one item, and calculates inference results that are distributions of probability values that cause the inference target node to have a predetermined state value And the inference unit calculates an influence evaluation value indicating an influence evaluation value indicating a degree of influence of a probability value of a node having a dependency relationship with the inference target node on the probability value of the inference target node. And a probability value differential amount that is a change amount of the probability value of the inference target node with respect to a minute change in the probability value of the node having a dependency relationship with the inference target node. A probability value differential amount calculating unit, wherein the probability value differential amount calculating unit refers to structure information of the graphical model managed by the model information storage unit, and the inference Extracting a node having a dependency relationship with an elephant node, selecting a node to be processed from the extracted nodes, and obtaining the probability table of the node to be processed managed by the model information storage unit , Referring to the structural information of the graphical model managed by the model information storage unit, extracting a set of one or more nodes having a dependency relationship with the processing target node as a family node, and the inference target node , Using the probability table of the processing target node and the family node, calculating a joint probability distribution of a joint probability variable combining the inference target node, the processing target node, and the family node, The probability value differential amount is calculated using a joint probability distribution and a probability table of the processing target node.

  According to the present invention, the probability value differential amount is calculated using the joint probability distribution. Therefore, since the impact evaluation value can be calculated using the approximate reasoning method, it is possible to calculate the impact evaluation value while suppressing the calculation cost even in a large-scale graphical model. Problems, configurations, and effects other than those described above will become apparent from the description of the following examples.

6 is a flowchart illustrating an example of processing executed by an inference unit of the computer according to the first embodiment. 1 is a block diagram illustrating an example of a configuration of a computer system according to a first embodiment. It is explanatory drawing which shows an example of the learning data stored in the database of Example 1. 3 is an explanatory diagram illustrating an example of a graphical model according to Embodiment 1. FIG. 6 is an explanatory diagram illustrating an example of structure information of a graphical model according to Embodiment 1. FIG. 6 is an explanatory diagram illustrating an example of structure information of a graphical model according to Embodiment 1. FIG. It is explanatory drawing which shows an example of the conditional probability table of Example 1. FIG. It is explanatory drawing which shows an example of the evidence data of Example 1. FIG. It is explanatory drawing which shows an example of the inference result management information of Example 1. It is explanatory drawing which shows an example of the influence management information of Example 1. FIG. It is explanatory drawing which shows an example of the influence management information of Example 1. FIG. 6 is a flowchart illustrating an example of processing executed by a graphical model construction unit of the computer according to the first embodiment. FIG. 6 is an explanatory diagram illustrating an example of a user interface displayed by the display unit according to the first embodiment. 10 is a flowchart for explaining a graphical model reconstruction process executed by the computer according to the second embodiment. 12 is a flowchart for explaining a graphical model reconstruction process executed by a computer according to a third embodiment. 15 is a flowchart for explaining a graphical model reconstructing process executed by the computer according to the fourth embodiment.

  In the first embodiment, the computer 200 constructs a causal relationship and a transition relationship between a plurality of random variables as a graphical model based on the input learning data. Further, the computer 200 calculates the probability distribution (inference result) of the inference object using the approximate inference method based on the evidence data input to the constructed graphical model. At this time, the computer 200 calculates an impact evaluation value, and also calculates a confidence interval of the conditional probability value of the inference result using the impact evaluation value.

  When predicting future trends, etc. from past data (sample data) using a graphical model, if the confidence interval of the probability value in the inference result can be shown as well as the inference result, the reliability of the inference result and It leads to improvement of persuasive power.

  For example, when the phenomenon that a coin is thrown and the coin becomes the front side or the back side is considered, the true probability that the coin becomes the front side or the back side is 50%. However, if the number of coins to be thrown is a finite number of times, for example, 10 times, 5 times out of 10 times does not necessarily become the surface. Here, the number of times the coin is thrown is called the number of samples.

  Therefore, when the number of samples is a finite value, the probability that the coin becomes the surface is a value deviated from 50%. In the trial described above, the probability that the coin will be on the surface is 60%. As the number of coins to be thrown increases, the fluctuation of the obtained result becomes smaller. For example, the probability value obtained by throwing a coin 10 times may be less reliable than the probability value obtained by throwing a coin 100 times. I understand.

  Since the probability value obtained from the finite number of samples as described above is an estimated value of the true probability value, it is necessary to evaluate the reliability of the probability value. The reliability evaluation value for evaluating the reliability of the probability value is given as a function of the number of samples.

  When a graphical model is constructed by machine learning using learning data, a conditional probability table can be given to each node. The conditional probability value in the conditional probability table of each node is more reliable as the number of learning data, that is, the number of samples is larger, as in the above-described trial.

  Therefore, since the reliability of the inference result using the graphical model also depends on the number of samples, the more accurate inference result can be obtained as the number of samples increases. That is, the confidence interval of the inference result is reduced. From this, the reliability of the inference result can also be expressed as a function of the number of samples.

  With such a connection, the reliability of the conditional probability value due to the finiteness of the learning data is finally reflected in the reliability of the inference result and expressed in the form of a confidence interval.

  Here, as will be described later, the learning data is a set of records including a plurality of items corresponding to the random variables, and values are stored in at least one item. The graphical model is a graph that treats each random variable as a node and expresses a dependency relationship between nodes as an edge. In the graphical model, a probability table is given to the node.

  There are several types of graphical models, such as Bayesian networks and Markov networks. In this embodiment, a Bayesian network will be described as an example. In a Bayesian network, the dependency between random variables is expressed as a conditional probability, and a probability table indicating the distribution of probability values of nodes is a conditional probability table. Also, in this embodiment, some or all of items (probability variables) that do not have a value among items corresponding to random variables of evidence data described later are treated as inference targets.

  The approximate reasoning method is an approximate method for calculating the probability distribution of the reasoning object. As a representative approximate reasoning method, the Loopy belief propagation method or the Gibbs sampling method is known.

  FIG. 2 is a block diagram illustrating an example of the configuration of the computer system according to the first embodiment.

  The computer system includes a computer 200 and a database 206.

  The computer 200 constructs a graphical model, and calculates an inference result and a confidence interval of a probability value in the inference result using the graphical model. The computer 200 of this embodiment includes an arithmetic device 201, a memory 202, a storage medium 203, an input device 204, and an output device 205, which are connected to each other via an internal bus or the like.

  The arithmetic device 201 is an arithmetic device that executes a program stored in the memory 202, and includes, for example, a CPU and a GPU. In the following description, when processing and functions are described using a functional unit as a subject, it indicates that a program for realizing the functional unit is being executed by the arithmetic device 201. The memory 202 stores a program executed by the arithmetic device 201 and information used by the program. The memory 202 may be either a volatile memory or a non-volatile memory.

  The storage medium 203 stores programs and the like that realize various functions of the computer 200. In this embodiment, the arithmetic unit 201 reads a program from the storage medium 203, loads the read program onto the memory 202, and executes the loaded program. The programs stored in the storage medium 203 of this embodiment will be described later.

  Note that the program stored in the storage medium 203 can be obtained from a removable medium such as a CD-ROM and a flash memory, or a distribution server connected via a network. When acquiring a program from a removable medium, the computer 200 includes an interface connected to the removable medium.

  The input device 204 is a device for inputting various information to the computer 200, and includes, for example, a keyboard, a mouse, a touch panel, and the like. The output device 205 is a device that outputs a processing result executed by the computer 200, and includes, for example, a display.

  The database 206 stores various data managed by the computer 200. In this embodiment, it is assumed that the database 206 is constructed using a storage system (not shown). The storage system includes a controller, an external interface, and a plurality of storage media. The storage system can configure RAID using a plurality of storage media. The storage system can also provide a plurality of logical storage areas using a RAID volume.

  The database 206 includes a learning data storage unit 241, a model information storage unit 242, an evidence data storage unit 243, an inference result storage unit 244, and an influence degree storage unit 245.

  The learning data storage unit 241 stores learning data 300 used when a graphical model is constructed. Details of the learning data 300 will be described with reference to FIG. The model information storage unit 242 stores structure information 500 and 510 indicating the structure of the graphical model, and a conditional probability table 600. Details of the structure information 500 and 510 will be described with reference to FIGS. 5A and 5B, and details of the conditional probability table 600 will be described with reference to FIG.

  The evidence data storage unit 243 stores the evidence data 700. Here, the evidence data 700 indicates information relating to health such as test values and image data acquired from patients undergoing health examinations, etc., and information relating to medical treatment such as diagnostic information, treatments, and prescription drugs by doctors. Details of the evidence data 700 will be described with reference to FIG. The inference result storage unit 244 stores the inference result management information 800. Details of the inference result management information 800 will be described with reference to FIG. The influence degree storage unit 245 stores influence degree management information 900. Details of the influence management information 900 will be described with reference to FIG.

  Here, the program stored in the storage medium 203 will be described.

  The storage medium 203 stores programs for realizing the graphical model construction unit 210, the inference unit 220, and the display unit 230.

  The graphical model construction unit 210 constructs a graphical model and generates various types of information related to the constructed graphical model. The graphical model construction unit 210 is composed of a plurality of modules. The graphical model construction unit 210 of this embodiment includes a model structure learning unit 211, a sample number calculation unit 212, and a probability table calculation unit 213.

  The model structure learning unit 211 uses the learning data 300 to construct a graphical model (Bayesian network). The model structure learning unit 211 stores the structure information 500 and 510 of the graphical model constructed in the database 206 via the model information storage unit 242. Here, it is assumed that the model structure learning unit 211 constructs a graphical model using an existing structure learning algorithm of the Bayesian network. As an existing structure learning algorithm of the Bayesian network, for example, there is a Hill Climbing method.

  The sample number calculation unit 212 calculates the number of records that meet a predetermined condition from the records included in the learning data 300 as the number of samples. The sample number calculation unit 212 stores the calculated sample number in the influence management information 900 via the influence storage unit 245.

  The probability table calculation unit 213 calculates a conditional probability table 600 for each node of the constructed graphical model. The probability table calculation unit 213 stores the conditional probability table 600 via the model information storage unit 242.

  The inference unit 220 receives an input of an inference object, and calculates a distribution (inference result) of the conditional probability value of the input inference object. In addition, the inference unit 220 of this embodiment calculates the magnitude of the influence of a certain conditional probability value on the probability value of the inference result as an impact evaluation value, and the probability value of the inference result based on the calculated impact evaluation value Compute confidence intervals for. The inference unit 220 includes a plurality of modules. The inference unit 220 of this embodiment includes an inference result calculation unit 221, a probability value differential amount calculation unit 222, an influence degree calculation unit 223, and a confidence interval calculation unit 224.

  The inference result calculation unit 221 calculates a probability distribution of the inference target using a strict inference method such as a variable elimination method or an approximate inference method such as Loop Belief Propagation.

  The probability value differential amount calculation unit 222 calculates the probability value differential amount necessary for calculating the impact evaluation value. Here, the probability value differential amount is a change amount of the probability value of the inference result with respect to a minute change amount of the conditional probability value, and is expressed as Expression (2).

  In this embodiment, the probability value differential amount is given by the following equation (4). When calculating the joint probability distribution included in Equation (4) using strict inference, the result matches the calculation result of the probability value differential amount described in Non-Patent Document 1. A specific method for calculating the impact evaluation value and the probability value differential amount will be described later.

Here, the derivation of Expression (4) will be described. In the following description, an inference object is T, the random variable corresponding to the child nodes of the conditional probability value is A, the random variable corresponding to the parent node B _i. Note that i is an integer from 1 to m. Also, a set of possible values of T and R (T), a set of possible values of A and R (A), a set of possible values of B _i and R (B _i). At this time, the following expression (5) holds.

  By using the Bayes' theorem as shown in the following equations (6) and (7), the equation (5) becomes the following equation (8).

Here, t ₁ is an arbitrary value of R (T), a _k is an arbitrary value of R (A), b _{i, j} is an arbitrary value of R (B _i ), and the following equation (9) is calculated. .

  At this time, since there is no term including the following formula (11) other than the term shown in the following formula (10) on the right side, the formula (9) becomes as shown in the formula (12).

  When the right side of equation (12) is transformed using Bayes' theorem, the following equation (13) is obtained. Therefore, Expression (12) can be expressed as shown in Expression (14). That is, it agrees with Expression (4). The above is the description of the derivation method of Equation (4).

  The influence degree calculation unit 223 calculates an influence evaluation value using the probability value differential amount. The confidence interval calculation unit 224 calculates a confidence interval of the probability value of the inference result based on the impact evaluation value and the number of samples.

  The display unit 230 displays various information to be output to the user via the output device 205 or the like. The display unit 230 includes a plurality of modules. The display unit 230 of this embodiment includes a model display unit 231 and an inference result display unit 232.

  The model display unit 231 generates display data for displaying the constructed graphical model, and outputs information related to the graphical model via the output device 205 based on the generated display data.

  The inference result display unit 232 generates display data for displaying the calculated inference result, and outputs information on the inference result via the output device 205 based on the generated display data.

  FIG. 3 is an explanatory diagram illustrating an example of the learning data 300 stored in the database 206 according to the first embodiment.

  The learning data 300 includes records composed of a plurality of columns corresponding to identification information and random variables. The record of this embodiment includes a patient ID 301, a BMI value 302, a blood pressure value 303, a blood glucose level 304, a heart disease 305, and a diabetes 306.

  The patient ID 301 is patient identification information. The BMI value 302, the blood pressure value 303, and the blood glucose level 304 are the BMI value, blood pressure value, and blood glucose level of the patient. Heart disease 305 and diabetes 306 are information indicating whether or not the patient falls into heart disease and diabetes. Patient ID 301 is identification information of the record, and BMI value 302, blood pressure value 303, blood glucose level 304, heart disease 305, and diabetes 306 are random variables.

  When the patient has heart disease or diabetes, “Yes” is stored in the heart disease 305 or diabetes 306, and when the patient does not have heart disease or diabetes, the heart disease 305 or diabetes 306 has “ “No” is stored.

  The record in the first line from the top of FIG. 3 shows that the patient ID 301 is “K0001”, the BMI value is “32”, the blood pressure value is “90”, and the blood glucose level is “5”. Indicates that it does not fall under any of the above.

  Note that the BMI value 302, blood pressure value 303, blood glucose level 304, heart disease 305, and diabetes 306 do not necessarily have to store values. In this case, information indicating that data is missing is stored in the column. The information indicating the data loss may be any of a numerical value, a character, and a Boolean value.

  Next, the graphical model and structure information 500 and 510 corresponding to the learning data 300 shown in FIG. 3 will be described.

  FIG. 4 is an explanatory diagram illustrating an example of the graphical model 400 according to the first embodiment.

  The graphical model 400 includes a plurality of nodes 410 and an edge 420 that connects the nodes 410. Each node 410 of the graphical model 400 shown in FIG. 4 corresponds to the BMI value 302, blood pressure value 303, blood glucose level 304, heart disease 305, and diabetes 306 of the learning data 300.

  Since a Bayesian network is assumed in the present embodiment, there is a direction at the edge 420 connecting the nodes 410. The node 410 corresponding to the start point of the edge 420 is called a parent node, and the node 410 corresponding to the end point of the edge 420 is called a child node. For example, the “diabetes” node 410 is a child node of the “blood glucose level” node, and the “blood glucose level” node is a parent node for the “diabetes” node. Each node 410 is given a conditional probability table. In a Bayesian network, the probability distribution of child nodes depends on the probability value of the parent node.

  5A and 5B are explanatory diagrams illustrating examples of the structure information 500 and 510 of the graphical model 400 according to the first embodiment. The structure information of the graphical model 400 of this embodiment includes information on the node 410 and information on the edge 420.

  FIG. 5A shows structural information 500 regarding the node 410 in the graphical model 400. The structure information 500 includes one record for one node 410, and the record includes a node ID 501 and an item name 502.

  The node ID 501 is identification information for uniquely identifying the node 410. The item name 502 is identification information of a random variable corresponding to the node 410. The item name 502 corresponds to the item name of the learning data 300.

  FIG. 5B shows structural information 510 regarding the edge 420 in the graphical model 400. The structure information 510 includes one record for one edge 420, and the record includes an edge ID 511, a parent node 512, and a child node 513.

  The edge ID 511 is identification information for uniquely identifying the edge 420. The parent node 512 is identification information of the node 410 corresponding to the parent node. The child node 513 is identification information of the node 410 corresponding to the child node. The same information as the item name 502 is stored in the parent node 512 and the child node 513. Note that the same information as the node ID 501 may be stored in the parent node 512 and the child node 513.

  FIG. 6 is an explanatory diagram illustrating an example of the conditional probability table 600 according to the first embodiment.

  The conditional probability table 600 stores a conditional probability value at which a child node takes an arbitrary state value with respect to the state value of the parent node. The conditional probability table 600 is given to the node 410 corresponding to the child node.

  The conditional probability table 600 includes a parent node 601, a child node 602, a conditional probability 603, and the number of samples 604.

  The parent node 601 is a state value of the parent node. When there are a plurality of parent nodes, the parent node 601 has as many columns as the number of parent nodes. Child node 602 is a state value of the child node. The conditional probability 603 is a probability value that takes the state value set in the child node 602 with respect to the state value set in the parent node 601. The sample number 604 is the number of records that match the state value of the parent node 601 among the records included in the learning data 300.

  Note that the number of samples may be managed in a table different from the conditional probability table 600.

  In the case of the top record in FIG. 6, when the blood pressure value is “90” and the blood glucose level is “5”, the probability of heart disease is “9%”. Further, in the case of the top record in FIG. 6, the number of records whose blood pressure value and blood glucose level state values as parent nodes are “90” and “5”, respectively, is “1563”. The conditional probability 603 can be obtained from the number of records that match the state value of the parent node 601 included in the learning data 300 and the number of records that match the state values of the parent node 601 and the child node 602.

  In general, the greater the value of the sample number 604, the higher the reliability indicated by the conditional probability 603.

  FIG. 7 is an explanatory diagram of an example of the evidence data 700 according to the first embodiment.

  The evidence data 700 according to the first embodiment has the same configuration as the records included in the learning data 300. Specifically, the evidence data 700 includes a patient ID 701, a BMI value 702, a blood pressure value 703, a blood sugar level 704, a heart disease 705, and diabetes 706. In addition, the symbol which shows that data is missing is stored in the value of the item which is not acquired from a patient.

  FIG. 8 is an explanatory diagram illustrating an example of the inference result management information 800 according to the first embodiment.

  The inference result management information 800 stores a conditional probability value at which the inference target node 410 calculated using the graphical model 400 and the evidence data 700 becomes an arbitrary state value. The inference result management information 800 according to the first embodiment includes evidence 801, an inference object 802, a conditional probability 803, and a confidence interval 804.

  The evidence 801 is a state value of the node 410 having a dependency relationship with the node 410 related to the inference target 802 in the evidence data 700. The inference target 802 is a state value of the node 410 that is the inference target.

  In the inference result management information 800, as many records as the state values that can be taken by the inference target node 410 are generated for one record included in the evidence data 700. In this embodiment, since the node 410 corresponding to “heart disease” takes a state value of “Yes” or “No”, the inference result management information 800 includes one record of the evidence data 700. , Two records are stored. For example, when the number of records in the evidence data 700 is “M”, the number of records in the inference result management information 800 is “2M”.

  The conditional probability 803 is a conditional probability value to be inferred. A set of conditional probabilities 803 of the two records of the inference result management information 800 corresponding to one record of the evidence data 700 is an inference result in one record of the evidence data 700.

  The confidence interval 804 is a confidence interval for evaluating the reliability of the conditional probability value. In this embodiment, it is assumed that the confidence interval 804 stores the value of the 95% confidence interval.

  The top record shown in FIG. 8 shows that the probability that a patient with a BMI value of “21”, a blood pressure value of “90”, and a blood glucose level of “5” will have heart disease is 95% with a probability of 5%. 9%.

  9A and 9B are explanatory diagrams illustrating an example of the impact management information 900 according to the first embodiment. Since the number of records of the impact management information 900 is large, the impact management information 900 is shown in two parts of FIGS. 9A and 9B.

  The influence degree management information 900 manages an influence evaluation value indicating the degree of influence of the conditional probability 603 on the inference result. The influence management information 900 includes a parent node 901, a child node 902, a conditional probability 903, a sample number 904, an inference target 905, a joint probability 906, a probability value derivative 907, and an influence evaluation value 908.

  The parent node 901, the child node 902, the conditional probability 903, and the number of samples 904 are the same as the parent node 601, the child node 602, the conditional probability 603, and the number of samples 604.

  Note that one piece of influence management information 900 exists for the conditional probability table 600 of the node 410 having a dependency relationship with the inference target. FIG. 9 shows the influence level management information 900 corresponding to the blood glucose level node 410. In addition, as many records as the number of state values that can be taken as an inference target are generated in the influence management information 900 for one conditional probability value. Since the node 410 corresponding to “heart disease” takes two state values “Yes” or “No”, the impact management information 900 includes two records for one record of the conditional probability table 600. Is stored. For example, when the number of records in the conditional probability table 600 is “N”, the number of records in the impact management information 900 is “2N”.

  The inference object 905 is a state value of the inference object. The joint probability 906 is a joint probability value of related random variables in a combination of state values corresponding to the record. The probability value derivative 907 is a probability value derivative of the conditional probability 903. The influence evaluation value 908 is the degree of influence that the conditional probability 903 has on the probability value of the inference result.

  For example, the first record of FIG. 9 shows the influence of the conditional probability value “9%” that causes heart disease on the probability value of the inference result when the BMI value is “20” and the blood glucose level is “5”. The magnitude, that is, the influence evaluation value is “0.2”.

  The value stored in the impact evaluation value 908 is given as a value of a function of the amount of change in the probability value of the inference result for a minute change in the conditional probability 603. It is assumed that a function for calculating the impact evaluation value is set in the inference unit 220 in advance. Also, the value stored in the confidence interval 804 of the inference result management information 800 is calculated based on the number of samples and the impact evaluation value, as will be described later.

  Next, processing executed by the computer 200 will be described. First, the construction process of the graphical model 400 will be described with reference to FIG. FIG. 10 is a flowchart illustrating an example of processing executed by the graphical model construction unit 210 of the computer 200 according to the first embodiment.

  The computer 200 starts processing described below when the construction of the graphical model 400 is instructed.

  The computer 200 accepts input of learning data 300 (step S1001). For example, the learning data 300 may be input using the input device 204 or the like.

  At this time, the computer 200 stores the learning data 300 input via the learning data storage unit 241 of the database 206 in the database 206. When learning data 300 in various data formats is input, the learning data storage unit 241 stores the converted learning data 300 after converting the input data into the format shown in FIG. Also good.

  Next, the computer 200 executes a discretization process for the learning data 300 (step S1002). Specifically, the model structure learning unit 211 of the graphical model construction unit 210 discretizes state values for items in which the state value stored in the item has a continuous value among the items of the record of the learning data 300. For example, it is discretized so that only an integer is handled as the state value of the blood sugar level. In this case, the number after the decimal point is rounded off, rounded down or rounded up. The granularity of discretization can be set arbitrarily.

  Next, the computer 200 executes a restriction condition setting process used to construct the graphical model 400 (step S1003). For example, the model structure learning unit 211 of the graphical model construction unit 210 receives a restriction condition input using the input device 204 or the like, and stores the restriction condition in the memory 202.

  Here, as the limiting condition, a dependency relationship between the nodes 410 can be considered. For example, information such as “there is no edge between the first node and the second node” and “there is an edge between the third node and the fourth node” is input as the restriction information.

  Next, the computer 200 executes a model structure learning process using the learning data 300 (step S1004). Specifically, the model structure learning unit 211 of the graphical model construction unit 210 generates the structure information 500 of the node 410 and the structure information 510 of the edge 420 based on the learning data 300 and the constraint conditions, thereby the graphical model 400. Build up. The Hill Climbing method is known as a Bayesian network structure learning algorithm. In this embodiment, any learning algorithm may be used.

  Next, the computer 200 executes data collation processing (step S1005). Specifically, when the sample number calculation unit 212 of the graphical model construction unit 210 sets a certain node 410 as a child node from among records included in the learning data 300, the child node state value and the parent node Extract records that match the status value combination.

  Next, the computer 200 executes a sample number calculation process (step S1006). Specifically, the number-of-samples calculation unit 212 of the graphical model construction unit 210 calculates the number of records extracted in the data matching process as the number of samples, and the calculated number of samples is the state value of the child node and the parent node. The data is temporarily stored in the memory 202 in association with the state value.

  Next, the computer 200 executes a calculation process of the conditional probability table 600 (step S1007). Specifically, the following processing is executed.

  The probability table calculation unit 213 of the graphical model construction unit 210 selects the node 410 to be processed, identifies the parent node for the selected node 410 based on the structure information 500 and 510, and creates a child in the conditional probability table 600. Records are generated for the number of combinations of the state values of the node and the parent node. Further, the probability table calculation unit 213 stores the number of samples calculated in step S1006 in the number of samples 604 of the generated record.

  Further, the probability table calculation unit 213 calculates, as a conditional probability value, a ratio of records in which the state value of the child node 602 is a predetermined value among the records included in the learning data 300. Further, the probability table calculation unit 213 stores the calculated conditional probability value in the conditional probability 603 of a predetermined record in the conditional probability table 600.

  Through the above processing, the structure information 500 and 510 as shown in FIGS. 5A and 5B and the conditional probability table 600 as shown in FIG. 6 are generated. That is, a graphical model 400 as shown in FIG. 4 is constructed.

  Next, a process for calculating the inference result will be described with reference to FIG. FIG. 1 is a flowchart illustrating an example of processing executed by the inference unit 220 of the computer 200 according to the first embodiment.

  When instructing unit 220 is instructed to start processing, inference unit 220 starts processing described below. At this time, the evidence data 700 is input to the inference unit 220. In addition, since the calculation method of the inference result using the evidence data 700 is a well-known method, description is abbreviate | omitted. Here, it is assumed that the inference result calculation unit 221 calculates an inference result using a strict inference method such as a variable elimination method, or an approximate inference method such as Loopy Belief Propagation.

  The inference unit 220 selects an inference target (step S101). Specifically, the probability value differential amount calculation unit 222 of the inference unit 220 receives the identification information of the node 410 from the user or the like via the input device 204, and sets the node 410 as an inference target.

  Next, the inference unit 220 selects the conditional probability table 600 to be processed (step S102). Specifically, the following processing is executed.

  The probability value differential amount calculation unit 222 of the inference unit 220 refers to the structure information 500 and 510 and follows the graphical model 400 along the edge 420 to the root node using the inference object as a leaf node, thereby depending on the inference object. A plurality of related nodes 410 are extracted. The probability value differential amount calculation unit 222 reads the conditional probability table 600 corresponding to each of the extracted nodes 410 from the database 206 and stores it in the memory 202.

  The probability value differential amount calculation unit 222 selects one processing target node 410 from among the plurality of extracted nodes 410. For example, a method of selecting a parent node when the inference target is a child node and then selecting the parent node in that order can be considered. Note that this embodiment does not depend on the method of selecting the node 410 to be processed. Hereinafter, the selected node 410 is also referred to as a selection node 410.

  The probability value differential amount calculation unit 222 acquires the conditional probability table 600 corresponding to the selected node 410 from the plurality of conditional probability tables 600 stored in the memory 202. Further, the probability value differential amount calculation unit 222 generates the degree of influence management information 900 as shown in FIGS. 9A and 9B based on the acquired conditional probability table 600. Specifically, a record is generated for each state value that can be inferred for a combination of state values of a parent node and a child node. When the inference target is “heart disease”, since there are two state values, two records are generated in the influence management information 900 for one record in the conditional probability table 600.

  The probability value derivative calculation unit 222 adds the parent node 601, child node 602, and the selected conditional probability table 600 to the parent node 901, child node 902, conditional probability 903, and number of samples 904 of the generated record. The values of the conditional probability 603 and the number of samples 604 are stored, and the state values that the inference object can take are stored in the inference object. At this time, no value is stored in the joint probability 906, the probability value differential amount 907, and the impact evaluation value 908. The above is the description of the processing in step S102.

  Next, the inference unit 220 extracts a related random variable (step S103). Specifically, the probability value differential amount calculation unit 222 of the inference unit 220 refers to the structure information 500 and 510 and uses the selected node 410 as a child node, and the parent node 410 of the selected node that is dependent on the selected node 410. Is extracted as a family node 410. Further, the inference unit 220 extracts the inference object, the selected node 410, and the extracted family node 410 as related random variables.

  Next, the inference unit 220 calculates a joint probability distribution of related random variables (step S104). Specifically, the probability value differential amount calculation unit 222 of the inference unit 220 uses an inference method such as Loop Belief Propagation using the conditional probability table 600 of the inference object, the selected node 410, and the extracted family node 410. Based on the above, a joint probability distribution of related random variables is calculated. At this time, the probability value derivative calculation unit 222 calculates a joint probability value to the joint probability 906 of the record in which the combination of the state value of the parent node, the child node, and the inference target 905 matches based on the calculated joint probability distribution. Store.

  Next, the inference unit 220 calculates a probability value differential amount using the selected conditional probability table 600 and the calculated joint probability distribution (step S105). Specifically, the following processing is executed.

  The probability value derivative calculation unit 222 of the inference unit 220 refers to the state values of the parent node 601 and the child node 602 in the selected conditional probability table 600, and combines any state value in the calculated joint probability distribution. Select records that match The probability value derivative calculation unit 222 reads the values of the conditional probability 903 and the joint probability 906 of the retrieved record.

  The probability value derivative calculation unit 222 calculates the probability value derivative by substituting the values of the conditional probability 903 and the joint probability 906 into the equation (4). The probability value derivative calculation unit 222 stores the calculated probability value derivative in the probability value derivative 907 of the selected record. The above is the description of the processing in step S105.

  Next, using the probability value differential amount, the inference unit 220 calculates the degree of influence that the conditional probability value of the selection node 410 has on the probability value of the inference result as an influence evaluation value (step S106).

  In general, the impact evaluation value is given as a function of the probability value differential amount, and can be set as an arbitrary function according to various conditions. In the present embodiment, the definition is such that the probability value derivative itself becomes the influence evaluation value. Note that a plurality of probability value differential amount functions may be averaged using a weighted function, and the function may be used as an impact evaluation value.

  In step S106, the probability value derivative calculation unit 222 of the inference unit 220 calculates an influence evaluation value by substituting the value of the probability value derivative 907 of the predetermined record into a preset function, and the influence of the record The evaluation value calculated in the evaluation value 908 is stored.

  Next, the inference unit 220 determines whether or not the processing has been completed for all the read conditional probability tables 600 (step S107).

  When it is determined that the processing has not been completed for all the read conditional probability tables 600, the inference unit 220 returns to step S102 and executes the same processing.

  When it is determined that the processing has been completed for all the read conditional probability tables 600, the inference unit 220 determines the confidence interval of the probability value of the inference result based on the number of samples 904 and the impact evaluation value 908. Calculate (step S108). For example, the confidence interval can be calculated using a known method such as the following equation (15). Note that the square of the variance is given by the following equation (16).

  Here, the processing of FIG. 1 is illustrated by taking as an example a serial Bayesian network of node D, node C, the parent node of node D, node B, the parent node of node C, and node A, the parent node of node B. The general flow will be described. At this time, conditional probability table 600 of node A is a distribution of conditional probability P (A), conditional probability table 600 of node B is a distribution of conditional probability P (B | A), conditional probability table of node C 600 is given as a distribution of conditional probability P (C | B), and the conditional probability table 600 of node D is given as a distribution of conditional probability P (D | C).

  In step S101, the inference unit 220 selects the node D as an inference target. In step S102, the inference unit 220 extracts the node A, the node B, and the node C as the node 410 having a dependency relationship with the inference target. In addition, the inference unit 220 selects the node C as the selection node 410.

  In step S103, the inference unit 220 extracts the node B and the node C as the family node 410, and extracts the node B, the node C, and the node D as related random variables. In step S104, the inference unit 220 calculates the distribution of the joint probability P (B, C, D).

  In step S105, the inference unit 220 sets the joint probability P (B = b, C = c, D = d) to the conditional probability P (C = C = D = d) for each combination of the state values of the nodes B, C, and D. Divide by c | B = b) to calculate the probability derivative. In step S106, the inference unit 220 calculates an impact evaluation value using the probability differential amount. As a result, the magnitude of the influence of the conditional probability value of node C on the probability value of node D can be estimated.

  In step S107, the inference unit 220 determines that processing has not been completed for all conditional probability tables 600. Therefore, the inference unit 220 returns to step S102 and selects the node B as the selection node 410. In step S103, the inference unit 220 extracts node A and node B as family nodes, and extracts node A, node B, and node D as related random variables. In step S104, the inference unit 220 calculates the distribution of the joint probability P (A, B, D).

  In step S <b> 105, the inference unit 220 determines the joint probability P (A = a ′, B = b ′, D = d ′) as the conditional probability P for each combination of the state values of the node A, node B, and node D. Divide by (B = b ′ | A = a ′) to calculate the probability differential amount. In step S106, the inference unit 220 calculates an impact evaluation value using the probability differential amount. As a result, the magnitude of the influence of the conditional probability value of node B on the probability value of node D can be estimated.

  In step S107, the inference unit 220 determines that processing has not been completed for all conditional probability tables 600. Therefore, the inference unit 220 returns to Step S102 and selects the node A as the selection node 410. In step S103, the inference unit 220 extracts the node A and the node D as related random variables since the parent node 410 does not exist because the node A is the root node 410. In step S104, the inference unit 220 calculates a distribution of the joint probability P (A, D).

  In step S105, the inference unit 220 sets the joint probability P (A = a ″, D = d ″) to the conditional probability P (A = a ″) for each combination of the state values of the node A and the node D. Divide by to calculate the probability derivative. In step S106, the inference unit 220 calculates an impact evaluation value using the probability differential amount. As a result, the magnitude of the influence of the conditional probability value of node A on the probability value of node D can be estimated.

  In step S107, the inference unit 220 determines that the processing has been completed for all conditional probability tables 600. In step S108, the inference unit 220 calculates the degree of influence of the probability value of the inference result using the calculated influence evaluation value. The above is the description of the specific flow of the processing of FIG.

  An example of information displayed by the display unit 230 will be described with reference to FIG. FIG. 11 is an explanatory diagram illustrating an example of a user interface 1100 displayed by the display unit 230 according to the first embodiment.

  The user interface 1100 includes an inference result display area 1110, a graphical model display area 1120, an impact evaluation value sort button 1130, and a sample number sort button 1140.

  The inference result display area 1110 is an area for displaying various information related to the inference result. In FIG. 11, the inference result display area 1110 displays the influence degree management information 900 and the inference result management information 800. In this case, the display unit 230 acquires the influence level management information 900 from the influence level storage unit 245, acquires the inference result management information 800 from the inference result storage unit 244, and displays each acquired information in the display area. To do.

  The graphical model display area 1120 is an area for displaying the graphical model 400 used for calculating the inference result. The display unit 230 acquires the structure information 500 of the node 410 and the structure information 510 of the edge 420 from the model information storage unit 242 and generates display data of the graphical model 400 to display the graphical model 400 in the display area.

  The impact evaluation value sort button 1130 and the sample number sort button 1140 are operation buttons for sorting the records of the impact management information 900. When the impact evaluation value sort button 1130 is operated, the display unit 230 displays the impact management information 900 in which the records are sorted in ascending order or decreasing order of the impact evaluation value 908. When the sample number sort button 1140 is operated, the display unit 230 displays the influence management information 900 in which the records are sorted in ascending order or descending order of the sample number 904.

  Note that the display format of the user interface 1100 and the information displayed are merely examples, and the present invention is not limited thereto. The display unit 230 can display various information in various display formats using information stored in the database 206.

  According to the first embodiment, the computer 200 calculates the probability value differential amount for calculating the influence evaluation value using a mathematical expression as shown in Expression (4). Since the joint probability distribution can be calculated based on the approximate reasoning method, the calculation cost can be reduced by the conventional equation (2). Therefore, the impact evaluation value can be calculated even in a large-scale graphical model. Further, the confidence interval of the probability value of the inference result can be calculated using the influence degree evaluation value. By displaying the inference result together with the confidence interval, the reliability and persuasive power of the inference result can be improved.

  In Example 2, the graphical model is reconstructed by adding reinforcement data based on the impact assessment value. Hereinafter, the second embodiment will be described focusing on differences from the first embodiment.

  The computer system of the second embodiment is the same as that of the first embodiment. The hardware configuration and software configuration of the computer 200 of the second embodiment are the same as those of the first embodiment. Further, the information stored in the database 206 of the second embodiment is different from the first embodiment in that the reinforcement data managed by the learning data storage unit 241 is newly stored. Other information is the same as in the first embodiment.

  In the second embodiment, after the inference result and the confidence interval of the probability value of the inference result are calculated, the following processing is executed. At this time, it is assumed that supplementary data information including a plurality of reinforcement data is input.

  FIG. 12 is a flowchart for explaining a graphical model reconstruction process executed by the computer 200 according to the second embodiment.

  The computer 200 selects the impact management information 900 (step S1201), and sorts the records in the conditional probability table 600 corresponding to the selected impact management information 900 in descending order of the impact evaluation value 908 (step S1202). ).

  Next, the computer 200 selects a record whose influence evaluation value 908 is greater than a predetermined threshold value as a conditional probability value to be reinforced (step S1203).

  Next, based on the combination of the state values of the node 410 corresponding to the selected conditional probability value, the computer 200 performs collation with the record included in the supplement data and extracts a usable record (step S1204). . For example, the computer 200 collates supplementary data using the state values of the plurality of nodes 410 as state value vectors.

  Next, the computer 200 updates the learning data 300 by adding the record extracted from the reinforcement data to the learning data 300 (step S1205).

  The computer 200 determines whether or not the processing has been completed for all the impact management information 900 (step S1206).

  When it is determined that the processing has not been completed for all the impact management information 900, the computer 200 returns to step S1201 and executes the same processing.

  When it is determined that the processing has been completed for all the influence management information 900, the computer 200 executes the construction process of the graphical model 400 using the updated learning data 300 (step S1207). The construction process of the graphical model 400 is the same process as FIG.

  It is assumed that the processing from step S1201 to step S1206 is executed by the inference unit 220. Note that the graphical model construction unit 210 or other functional units may execute.

  According to the second embodiment, the confidence interval of the inference result can be reduced by reconstructing the graphical model 400 based on the influence evaluation value. That is, the graphical model 400 that can calculate a highly reliable inference result can be constructed.

  In the third embodiment, the graphical model 400 is made compact by re-evaluating the dependency of any node 410 on the inference result using the impact evaluation value. Hereinafter, the third embodiment will be described focusing on differences from the first embodiment.

  The computer system of the third embodiment is the same as that of the first embodiment. The hardware configuration and software configuration of the computer 200 of the third embodiment are the same as those of the first embodiment. The information stored in the database 206 of the third embodiment is the same as that of the first embodiment.

  In the third embodiment, after the inference result and the confidence interval of the probability value of the inference result are calculated, the following processing is executed.

  FIG. 13 is a flowchart for explaining the graphical model reconstruction process executed by the computer 200 according to the third embodiment.

  The computer 200 calculates the total impact evaluation value of the corresponding node 410 using each impact management information 900 (step S1301).

  For example, the computer 200 calculates the weighted average of the impact evaluation values 908 of all the records of the selected impact management information 900 as the total impact evaluation value of the node 410. Since the influence evaluation value of the conditional probability value exists for each state value to be inferred, the calculated total influence evaluation value of the node 410 also exists for each state value to be inferred. By calculating the weighted average of the influence evaluation values 908 for all the state values of the inference object for one node 410, it can be estimated as the influence evaluation value of the node 410 for the inference object.

  Note that the total impact evaluation value of the node 410 described above is an example, and other calculation methods may be used.

  Next, the computer 200 selects the important node 410 from among the plurality of nodes 410 having a dependency relationship with the inference target based on the calculated total impact evaluation value of the node 410 (step S1302).

  For example, the computer 200 selects, as the important node 410, the node 410 whose total influence evaluation value of the node 410 is larger than the threshold based on a preset threshold. The threshold value can be set based on the computer resource amount of the computer 200, the calculation time, and the like.

  The computer 200 executes the construction process of the graphical model 400 again (step S1303). The processing in step S1303 is almost the same as that in FIG. 10, except that only the selected important node 410 is input as information of the node 410 having the dependency in step S1003. Other processes are the same as those in FIG.

  It is assumed that the inference unit 220 executes the processes in steps S1301 and S1302. Note that the graphical model construction unit 210 or other functional units may execute.

  It is also possible to make the graphical model compact by repeatedly executing the processing described above.

  According to the third embodiment, the graphical model can be made compact based on the influence evaluation value.

  In Example 4, the graphical model 400 is reconstructed by correcting the discretization granularity of the learning data using the impact evaluation value and the number of samples. Here, the discretization granularity indicates the width of a continuous numerical value.

  For example, for continuous real numbers, discretization that handles only integer values has a greater granularity than discretization that handles numbers up to the first decimal point. When more specific medical data is considered, the blood pressure value is a continuous value, and the value is distributed over a wide range from “60” to “200”. As described in the first embodiment, when the graphical model is constructed, the state value of the node 410 may need to be finite. At this time, a method of discretizing the blood glucose level into four values of “80 or less”, “80 to 100”, “100 to 120”, and “120 or more” can be considered.

  The width of the discretization affects the prediction performance of the graphical model. That is, in the case of discretization, the inference result is an approximate value of the original continuous value. Therefore, if the discretization width is reduced, the error due to discretization is reduced. On the other hand, if the discretization width is small, the number of learning data that matches the state value of the node decreases. Therefore, the reliability of the conditional probability value in the conditional probability table 600 is lowered, and as a result, the reliability and accuracy of the inference result are lowered. Therefore, it is necessary to set an appropriate discretization width.

  Hereinafter, the fourth embodiment will be described focusing on differences from the first embodiment.

  The computer system of the fourth embodiment is the same as that of the first embodiment. The hardware configuration and software configuration of the computer 200 of the fourth embodiment are the same as those of the first embodiment. The information stored in the database 206 of the fourth embodiment is the same as that of the first embodiment.

  In the fourth embodiment, after the inference result and the confidence interval of the probability value of the inference result are calculated, the following processing is executed.

  FIG. 14 is a flowchart for explaining a graphical model reconstruction process executed by the computer 200 according to the fourth embodiment.

  The computer 200 selects one node 410 whose state value is discretized from the nodes 410 in which the influence degree management information 900 exists (step S1401). For example, when the graphical model 400 is constructed, it is conceivable to add a flag or the like to the node 410 whose state values are discretized in advance.

  Next, the computer 200 determines the discretization granularity of the selected node 410 based on the influence management information 900 (step S1402). For example, the following four methods can be considered as a method for determining the granularity of discretization.

  (1) The discretization granularity is determined so that the value of the number of samples 904 is the same for all records or the error becomes small.

  (2) The granularity of discretization is determined so that the product or weighted average value of the influence evaluation value 908 and the number of samples 904 of each record is the same or the error is reduced.

  (3) The granularity of discretization is determined so that the number of records whose influence evaluation value 908 is equal to or greater than a predetermined threshold is half the number of records whose influence evaluation value 908 is smaller than the predetermined threshold.

  (4) The discretization granularity is determined by classifying the state value width into a predetermined number based on the influence evaluation value 908.

  The discretization granularity determination method described above is merely an example, and any method using the influence management information 900 may be used.

  Next, the computer 200 determines whether or not the processing has been completed for all the nodes 410 whose state values are discretized (step S1403).

  When it is determined that the processing has not been completed for all the nodes 410 whose state values are discretized, the computer 200 returns to step S1401 and executes the same processing.

  When it is determined that the processing has been completed for all the nodes 410 whose state values are discretized, the computer 200 executes the construction process of the graphical model 400 again (step S1404). At this time, in step S1002, the discretization process of the learning data is executed based on the determined discretization granularity. Other processes are the same as those in FIG.

  It is assumed that the processing from step S1401 to step S1403 is executed by the inference unit 220. Note that the graphical model construction unit 210 or other functional units may execute.

  An appropriate discretization granularity can be determined by repeatedly executing the processing described above.

  According to the fourth embodiment, by adjusting the discretization granularity based on the influence evaluation value, the number of samples, and the like, it is possible to construct the graphical model 400 that can calculate a highly reliable inference result.

  In addition, this invention is not limited to an above-described Example, Various modifications are included. Further, for example, the above-described embodiments are described in detail for easy understanding of the present invention, and are not necessarily limited to those provided with all the described configurations. Further, a part of the configuration of each embodiment can be added to, deleted from, or replaced with another configuration.

  Each of the above-described configurations, functions, processing units, processing means, and the like may be realized by hardware by designing a part or all of them with, for example, an integrated circuit. The present invention can also be realized by software program codes that implement the functions of the embodiments. In this case, a storage medium in which the program code is recorded is provided to the computer, and a CPU included in the computer reads the program code stored in the storage medium. In this case, the program code itself read from the storage medium realizes the functions of the above-described embodiments, and the program code itself and the storage medium storing it constitute the present invention. As a storage medium for supplying such a program code, for example, a flexible disk, a CD-ROM, a DVD-ROM, a hard disk, an SSD (Solid State Drive), an optical disk, a magneto-optical disk, a CD-R, a magnetic tape, A non-volatile memory card, ROM, or the like is used.

  Further, the program code for realizing the functions described in the present embodiment can be implemented by a wide range of programs or script languages such as assembler, C / C ++, perl, Shell, PHP, Java, and the like.

  Furthermore, by distributing the program code of the software that implements the functions of the embodiments via a network, the program code is stored in a storage means such as a hard disk or memory of a computer or a storage medium such as a CD-RW or CD-R The CPU included in the computer may read and execute the program code stored in the storage unit or the storage medium.

  In the above-described embodiments, the control lines and information lines indicate what is considered necessary for the explanation, and not all control lines and information lines on the product are necessarily shown. All the components may be connected to each other.

Claims (9)

A computer system comprising an arithmetic device for executing a program and one or more computers having a memory for storing the program,
A learning data storage unit for managing learning data composed of records including a plurality of items corresponding to random variables;
Using the learning data, it is composed of a node corresponding to the random variable, an edge indicating a dependency relationship between the nodes, and a probability table indicating a distribution of probability values for each state value of the random variable corresponding to the node. A graphical model building unit for building a graphical model;
A model information storage unit that manages structure information of the graphical model constructed by the graphical model construction unit, and the probability table of each of a plurality of nodes included in the graphical model;
It is a distribution of probability values that is composed of records including a plurality of items corresponding to the random variables, accepts input of evidence data in which values are stored in at least one item, and inferred nodes become predetermined state values. An inference section for calculating an inference result,
The reasoning part is
An influence degree calculation unit for calculating an influence evaluation value indicating a magnitude of influence of a probability value of a node having a dependency relationship with the inference target node on the probability value of the inference target node;
A probability value used to calculate the impact evaluation value, and calculates a probability value derivative that is a change amount of the probability value of the inference target node with respect to a minute change in the probability value of the node having a dependency relationship with the inference target node. A differential amount calculation unit;
Including
The probability value differential amount calculation unit includes:
With reference to the structure information of the graphical model managed by the model information storage unit, a node having a dependency relationship with the inference target node is extracted,
Select a processing target node from the extracted nodes, obtain the probability table of the processing target node managed by the model information storage unit,
With reference to the structure information of the graphical model managed by the model information storage unit, a set of one or more nodes having a dependency relationship with the processing target node is extracted as a family node,
Using the probability table of the inference target node, the processing target node, and the family node, the joint probability distribution of the joint random variable combining the inference target node, the processing target node, and the family node To calculate
The computer system, wherein the probability value derivative is calculated using the joint probability distribution and a probability table of the processing target node. The computer system according to claim 1,
The computer system characterized in that the probability value differential amount calculation unit calculates the probability value differential amount using Equation (1).

The computer system according to claim 2,
The inference unit includes a confidence interval calculation unit that calculates a confidence interval of a probability value in the inference result using a function having the influence evaluation value as a variable,
The computer system includes an inference result storage unit that manages inference result management information in which the probability value of the inference result and the confidence interval are associated with each other. The computer system according to claim 2,
An influence degree storage unit for managing influence degree management information in which the probability value of the probability table is associated with the influence evaluation value for each of the probability tables of the extracted nodes;
The learning data storage unit manages reinforcement data composed of records including a plurality of items corresponding to the random variables,
The reasoning part is
After calculating the impact evaluation value for all probability values in the probability table of the extracted node, select a probability table to be processed,
With reference to the degree of influence management information corresponding to the probability table to be processed, based on the influence evaluation value, select a probability value that requires re-learning using the reinforcement data,
Based on a combination of state values of the random variable corresponding to the selected probability value, matching with a record included in the reinforcement data,
Based on the result of the collation, the record used for the relearning is extracted from the reinforcement data,
Updating the learning data by adding the extracted records;
The computer system is characterized in that the graphical model constructing unit constructs the graphical model again using the updated learning data. The computer system according to claim 2,
An influence degree storage unit for managing influence degree management information in which the probability value of the probability table is associated with the influence evaluation value for each of the probability tables of the extracted nodes;
The reasoning part is
Based on the influence degree management information, a total influence evaluation value indicating a magnitude of influence of the extracted node on the inference target node is calculated;
Based on the value of the total impact assessment value, an important node is selected from the extracted nodes,
The computer system is characterized in that the graphical model construction unit constructs the graphical model again in consideration of only the dependency relationship of the important nodes. The computer system according to claim 2,
The graphical model building unit
Constructing the graphical model by discretizing the state value of the node, which has a continuous state value, into a predetermined granularity;
The computer system includes: a probability value of the probability table; a sample number that is the number of records of the learning data that matches a combination of state values of random variables corresponding to the probability value of the probability table; A degree-of-impact management unit that manages the degree-of-impact management information associated with each of the probability tables of the extracted nodes;
The reasoning part is
Select a node whose state value is a continuous value,
Read the influence management information corresponding to the probability value table of the node,
Based on the number of samples of the degree of influence management information and the influence evaluation value, determine the granularity of the state value of the node,
The computer system is characterized in that the graphical model constructing unit constructs the graphical model again based on the newly determined granularity of the state value of the node. In a computer system including one or more computers, a probability table indicating a node corresponding to a random variable, an edge indicating a dependency relationship between the nodes, and a distribution of probability values for each state value of the random variable corresponding to the node. A graphical model management method considering the statistical uncertainty of a graphical model composed of
The one or more computers have a computing device that executes a program, and a memory that stores the program,
The computer system includes a learning data storage unit that manages learning data composed of records including a plurality of items corresponding to random variables.
The management method of the graphical model is:
The arithmetic unit constructs the graphical model using the learning data, and stores the constructed structural information of the graphical model and the probability table of each of a plurality of nodes included in the graphical model in the memory. A first step;
Probability that the arithmetic unit is configured from a record including a plurality of items corresponding to the random variable, accepts input of evidence data in which a value is stored in at least one item, and the inference target node becomes a predetermined state value A second step of calculating an inference result that is a distribution of values;
A third step in which the arithmetic unit calculates a probability value differential amount that is a change amount of the probability value of the inference target node with respect to a minute change in the probability value of the node having a dependency relationship with the inference target node;
The arithmetic unit uses the probability value differential amount to calculate an influence evaluation value indicating a magnitude of an influence of a probability value of a node having a dependency relationship with the inference target node on the probability value of the inference target node. And a fourth step of
The third step includes
A fifth step of referring to the structure information of the graphical model stored in the memory and extracting a node having a dependency relationship with the inferred node;
A sixth step of selecting a node to be processed from the extracted nodes and obtaining the probability table of the node to be processed from the memory;
Referring to structure information of the graphical model stored in the memory, a seventh step of extracting a set of one or more nodes having a dependency relationship with the processing target node as a family node;
Using the probability table of the inference target node, the processing target node, and the family node, the joint probability distribution of the joint random variable combining the inference target node, the processing target node, and the family node An eighth step of calculating
And a ninth step of calculating the probability value differential amount using the joint probability distribution and the probability table of the processing target node. A graphical model management method according to claim 7, comprising:
In the ninth step, the probability value differential amount is calculated using Expression (2), and the graphical model management method is characterized in that:

The graphical model management method according to claim 8, comprising:
The fourth step includes
Calculating the confidence interval of the probability value in the inference result using a function having the impact evaluation value as a variable after the impact evaluation value is calculated;
And storing the inference result management information in which the probability value of the inference result is associated with the confidence interval in the memory.

Images