Information processing device, information processing method, and information processing program
Patent Information
- Application Number
- JP2022028573
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2022-02-25
- Publication Date
- 2025-02-20
- Estimated Expiration
- 2042-02-25
Smart Images
Figure 0007633958000001 
Figure 0007633958000002 
Figure 0007633958000003
Abstract
Description
[Technical field]
[0001] The present invention relates to an information processing device, an information processing method, and an information processing program. [Background technology]
[0002] Recently, in various fields, it has become common for computers to use learning data to machine-learn mathematical models. In machine learning, the coefficients (parameters) of the mathematical model are optimized. When a user inputs certain explanatory variables into a machine-learned mathematical model, the mathematical model outputs the objective variable expected by the user. Here, "the objective variable expected by the user" means "the user can empirically evaluate as correct."
[0003] If there is a bias in the learning data as a sample obtained from a population due to artificial, social, temporal, or environmental constraints, the mathematical model will not output the objective variable expected by the user, even if the machine learning is performed technically correctly. Therefore, it becomes necessary to correct the coefficients of the machine-learned mathematical model. The information processing device of Patent Document 1 describes that by correcting the values of the coefficients multiplied by the explanatory variables of the machine-learned mathematical model, the mathematical model outputs an objective variable that is less strange to the user. [Prior art documents] [Patent documents]
[0004] [Patent Document 1] JP 2020-9502 A Summary of the Invention [Problem to be solved by the invention]
[0005] However, to correct the coefficient values of a mathematical model that has already been machine-learned, the user must select the coefficients to be corrected from among many coefficients, and then determine the values of the coefficients through trial and error. Instead of correcting the coefficients after the fact, one method is to replace the training data itself with unbiased data from the population and perform machine learning again. However, collecting new training data with the same format (variables) from the population also requires a lot of time and effort. Furthermore, the newly collected training data may also have other types of bias. Therefore, an object of the present invention is to easily correct biased learning data collected from a population, and to use the corrected learning data to machine-learn a highly accurate mathematical model. [Means for solving the problem]
[0006] The information processing device of the present invention includes a simultaneous distribution creation unit that creates simultaneous distribution information having the same variables as statistical information from training data, a weight calculation unit that calculates a degree to which a sample of the created simultaneous distribution information is biased compared to the statistical information as a population as a weight for each combination of a plurality of the variables, a prediction model creation unit that corrects the training data based on the calculated weight and learns a prediction model using the corrected training data, and an output processing unit that outputs a prediction result for the prediction data using the prediction model. the statistical information is obtained from a population including facilities where samples of the learning data were collected, the prediction model creation unit increases the number of samples of the learning data by copying the samples of the learning data based on the weights, the prediction model creation unit learns a first weak learner using the learning data before correction and learns a second weak learner using the learning data after correction, and creates an ensemble model having the first weak learner and the second weak learner; It is characterized by: Other means will be described in the description of the embodiment of the invention. Effect of the Invention
[0007] According to the present invention, biased learning data collected from a population can be easily corrected, and a highly accurate mathematical model can be machine-learned using the corrected learning data. [Brief description of the drawings]
[0008] [Figure 1] FIG. 2 is a diagram illustrating a configuration of an information processing device. [Diagram 2]1 is an example of learning data. [Diagram 3] 1 is an example of statistical information. [Figure 4] 13 is an example of joint distribution information. [Diagram 5] 11 is an example of weight information. [Figure 6] 13 is a flowchart of a joint distribution creation process. [Figure 7] 13 is a flowchart of a weight calculation process; [Figure 8] 13 is a flowchart of a prediction model creation process. [Figure 9] 13 is a flowchart of an output process procedure. [Figure 10] 1 is an example of a display screen. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0009] Hereinafter, an embodiment of the present invention (referred to as the "present embodiment") will be described in detail with reference to the drawings, etc. This embodiment is an example in which learning data collected from a single medical facility and statistical information published by the country, local government, etc. are used to correct the learning data. However, the present invention is generally applicable to fields other than medicine.
[0010] (Population and Sample) The population in this embodiment includes citizens, local residents, etc. residing in a country, a local government, etc. The specimen (sample) in this embodiment is a subject (subject) in a specific medical facility belonging to a country, a local government, or the like. The specimen does not necessarily accurately reflect the characteristics of the population. Each medical facility may inevitably have characteristics such as a large number of elderly people, a small number of men, etc. Due to such characteristics, a deviation occurs between the characteristics of the population and the characteristics of the sample. This deviation is also called "bias."
[0011] (Configuration of information processing device, etc.) 1 is a diagram for explaining the configuration of an information processing device 1. The information processing device 1 is a general computer, and includes a central control device 11, an input device 12 such as a mouse or a keyboard, an output device 13 such as a display, a main memory device 14, an auxiliary memory device 15, and a communication device 16. These are connected to each other via a bus. The auxiliary memory device 15 stores joint distribution information 31, weight information 32, and a prediction model 33. The prediction model 33 includes a first weak learner 34 and a second weak learner 35.
[0012] As is generally known, a "weak learner" is a mathematical model that has a relatively low prediction accuracy in an individual state. A mathematical model that is composed of multiple weak learners and outputs the result of majority voting of the outputs of the individual weak learners is called an ensemble model. The prediction accuracy of an ensemble model is higher than that of individual weak learners. The learning model 33 of this embodiment is such an ensemble model.
[0013] The joint distribution creation unit 21, the weight calculation unit 22, the prediction model creation unit 23, and the output processing unit 22 in the main memory device 14 are programs. The central control device 11 reads these programs from the auxiliary memory device 15 to the main memory device 14, thereby realizing the functions of each program (described in detail later).
[0014] The information processing device 1 is connected to a database 2 via a wired or wireless network 3. The database 2 stores learning data 41, prediction data 42, and statistical information 43 (described in detail below). The database 2 may be divided into, for example, a database of a country, a local government, etc. that stores the statistical information 43, a database of a specific medical facility that stores the learning data 41, and a database of other medical facilities that stores the prediction data 43 (described in detail below).
[0015] (Learning data) 2 is an example of the learning data 41. The learning data 41 is collected from a specific single medical facility. In the learning data 41, an event occurrence flag is stored in an event occurrence column 102, an age is stored in an age column 103, and a gender is stored in a gender column 104 in association with a subject ID stored in a subject ID column 101.
[0016] The subject ID in the subject ID column 101 is an identifier that uniquely identifies a subject. A subject is a subject of a specific examination by a single medical facility. The event occurrence flag in the event occurrence column 102 is either "0" or "1." "1" indicates that the test revealed that the subject is suffering from a particular disease. "0" indicates that the test revealed that the subject is not suffering from a particular disease. In other words, the "event" indicates that a disease occurs. The age in the age column 103 is the age of the subject. The gender in the gender column 104 indicates the gender of the subject.
[0017] (statistics) FIG. 3 is an example of the statistical information 43. The statistical information 43 is created from administrative documents (not shown) created by the country, local government, etc. The administrative documents assumed in this embodiment are public documents that correlate attributes of citizens or local residents (variables such as event occurrence, age, sex, occupation, address, number of vaccinations received, etc.). While the sample of the learning data 41 is a "part," the population of the administrative documents is the "whole" that includes that part and is much broader than that part. In addition, the attributes in the administrative documents often include attributes (variables) other than the "event occurrence," "age," and "sex" in the learning data 41.
[0018] In the statistical information 43, age is stored in the age column 112, gender is stored in the gender column 113, and the number of statistical information examples is stored in the number of statistical information examples column 114, in association with the event occurrence flag stored in the event occurrence column 111.
[0019] The event occurrence flags in the event occurrence column 111 are the same as the event occurrence flags in FIG. The age in the age column 112 is the age of the citizen or other person covered by the administrative documents. The gender in the gender column 113 is the gender of the citizen or other person covered by the administrative document. The number of statistical information examples in the statistical information example number column 114 is the number of citizens (population) corresponding to the combination of the event occurrence, age, and gender. The combination of the event occurrence, age, and gender matches columns 102 to 104 of the learning data 41 (FIG. 2).
[0020] As is clear from the above, statistical information 43 is compiled after ignoring all attributes of the original administrative data other than those that appear in the learning data.
[0021] (Joint distribution information) 4 is an example of the simultaneous distribution information 31. In the simultaneous distribution information 31, ages are stored in an age column 122, genders are stored in a gender column 123, and the number of examples of learning data is stored in a number-of-examples-of-learning-data column 124 in association with the event occurrence flags stored in an event occurrence column 121.
[0022] The event occurrence flags in the event occurrence column 121 are the same as the event occurrence flags in FIG. The ages in the age column 122 are the same as those in FIG. The gender in the gender column 123 is the same as that in FIG. The number of training data examples in the training data example number column 124 is the number of subjects (samples) in the training data 41 (FIG. 2) that correspond to the combination of the event occurrence, age, and sex.
[0023] The attributes of the simultaneous distribution information 31 (combination of simultaneous event occurrences, age, and gender) completely match the attributes of the statistical information 43 in Fig. 3. The values of each attribute of the simultaneous distribution information 31 (binary values of "0" or "1", age range, binary values of "male" or "female") also completely match the values of each attribute of the statistical information 43.
[0024] Comparing the number of examples in the statistical information 43 (Fig. 3) and the number of examples in the learning data in the joint distribution information 31 (Fig. 4), for example, reveals the following: These are biases in the learning data 41 (sample) collected from specific medical facilities with respect to the general public (population).
[0025] In learning data 41, the ratio of men is high. In learning data 41, the proportion of event occurrence flags set to “1” is high for both men and women. In learning data 41, the proportion of females whose event occurrence flags are “1” is significantly higher than the proportion of males whose event occurrence flags are “1.” No such tendency is observed in the population. In learning data 41, the proportion of young females whose event occurrence flags are “1” is high. In learning data 41, the proportion of elderly males whose event occurrence flags are “1” is high.
[0026] (Weight information) 5 is an example of the weight information 32. In the weight information 32, ages are stored in an age column 132, genders are stored in a gender column 133, and weights are stored in a weight column 134 in association with the event occurrence flags stored in the event occurrence column 131.
[0027] The event occurrence flags in the event occurrence column 121 are the same as the event occurrence flags in FIG. The ages in the age column 132 are the same as those in FIG. The gender in the gender column 133 is the same as that in FIG. The weights in the weight column 134 are the result of dividing the number of examples of the statistical information (column 114 in FIG. 3) by the number of examples of the training data (column 124 in FIG. 4). In other words, the weights are the degree to which the samples of the joint distribution information are biased compared to the statistical information as a population.
[0028] (Meaning of weight) When the weights of each record (row) in Figure 5 are viewed vertically, if all the weights of each record are the same value, the learning data 41 is not biased relative to the statistical information 43, and it can be said that a particular medical facility is an ideal representative of the population, such as the general public. However, in many cases, these values have variation. This variation indicates the bias characteristics of the learning data 41. For example, the weight of record 135 is "5.0", which is significantly larger than the other records. The weight of record 6 is "1.0", which is significantly smaller than the other records. This indicates the following:
[0029] In a particular medical facility where learning data 41 was collected, among samples whose age is “20 to 29” and whose gender is “female”, those whose event occurrence flag is “0” are under-represented and those whose event occurrence flag is “1” are over-represented. Therefore, if a mathematical model is trained using the learning data41 collected from this medical facility, the machine-learned mathematical model will output the erroneous conclusion that “women aged 20 to 29 are more likely to suffer from a certain disease.” To prevent such an error, among the records of the training data 41, those having the attributes of record 135 are copied to increase the number of samples, while those having the attributes of record 136 are left as they are without being copied, and new training data with different contents is created. Furthermore, the mathematical model is retrained using new learning data.
[0030] The processing procedures of this embodiment are described below. There are four processing procedures, which are a joint distribution creation processing procedure, a weight calculation processing procedure, a prediction model creation processing procedure, and an output processing procedure, and are executed in this order.
[0031] (Joint distribution creation process) FIG. 6 is a flowchart of the joint distribution creation process. In step S201, the simultaneous distribution creation unit 21 of the information processing device 1 acquires the learning data 41 and the statistical information 43. Specifically, the simultaneous distribution creation unit 21 acquires the learning data 41 and the statistical information 43 in the above-mentioned "administrative data" state from the database 2. The statistical data 43 acquired here does not have a regular format as shown in FIG.
[0032] In step S202, the simultaneous distribution creating unit 21 specifies variables. Specifically, first, the simultaneous distribution creating unit 21 recognizes attributes of the learning data 41 (in the example of FIG. 2, event occurrence, age, and gender) as "variables." Second, the joint distribution creation unit 21 recognizes strata for each variable. For example, it recognizes that ages in statistical information 43 (administrative data) are stratified into "0-9", "10-19", "20-29", and so on. Third, the simultaneous distribution creation unit 21 creates statistical information 43 of the type shown in Fig. 3, which has only the variables recognized in "first" of step S202 as attributes and has a record for each stratum of each variable recognized in "second" of step S202. At this time, the simultaneous distribution creation unit 21 refers to administrative documents and calculates the number of examples of statistical information (column 114 in Fig. 3) for each event occurrence, gender, and age group.
[0033] In step S203, the simultaneous distribution creating unit 21 creates simultaneous distribution information 31. Specifically, first, the simultaneous distribution creating unit 21 creates a copy of the statistical information 43 (FIG. 3) created in “third” of step S202, rewrites the “number of examples in the statistical information” (column 114 in FIG. 3) to the “number of examples in the learning data”, and then deletes the value of the number of examples in the statistical information for each record (returns it to blank). Secondly, the simultaneous distribution creation unit 21 refers to the learning data 41 and calculates the number of examples of the learning data (column 124 in FIG. 4), thereby completing the simultaneous distribution information 31 in FIG. 4. The simultaneous distribution information 31 completed here has the same variables as the statistical information 43 (FIG. 3). Thereafter, the simultaneous distribution creation process procedure ends.
[0034] (Weight calculation procedure) FIG. 7 is a flowchart of a weight calculation process. In step S211, the weight calculation unit 22 of the information processing device 1 acquires the statistical information 43 and the simultaneous distribution information 31. Specifically, the weight calculation unit 22 acquires the statistical information 43 created in the “third” of step S202, and the simultaneous distribution information 31 completed in the “second” of step S203.
[0035] In step S212, the weight calculation unit 22 creates weight information 32. Specifically, first, the weight calculation unit 22 creates a copy of the statistical information 43 (FIG. 3) acquired in step S211, rewrites the "number of examples in statistical information" (column 114 in FIG. 3) to "weight", and then deletes the value of the number of examples in the statistical information of each record (returns it to blank). Secondly, the weight calculation unit 22 divides the number of examples of the statistical information in the statistical information 43 acquired in step S211 by the number of examples of the learning data in the joint distribution information 31 acquired in step S211, and stores the result as a "weight". The weight calculation unit 22 completes the weight information 32 in Fig. 5 by calculating a weight for each combination of the event occurrence, age group, and gender. Then, the weight calculation process procedure ends.
[0036] (Prediction model creation process) FIG. 8 is a flowchart of a prediction model processing procedure. In step S221, the prediction model creation unit 23 of the information processing device 1 acquires the learning data 41 and the weighting information 32. Specifically, the prediction model creation unit 23 acquires the learning data 41 from the database 2, and also acquires the weighting information 32 completed in “secondly” of step S212. The weighting information 31 acquired here represents the bias of the learning data 41.
[0037] As described above, for example, the weight of record 135 in FIG. 5 is “5.0”. This indicates that the number of populations in which “event occurrence” is “0”, “age” is “20-29”, and “gender” is “female” in the statistical information 43 is five times the number of such samples in the learning data 41. That is, in the medical facility where the learning data 41 is collected, the number of such samples is only “one-fifth” compared to the country, etc., meaning that such samples are “five times” rarer. Note that the number “five times” does not have an absolute meaning. “Five times” is relatively prominent compared to the weights in other records, which indicates a bias.
[0038] In step S222, the prediction model creation unit 23 creates the first weak learner 34 and the second weak learner 35. Specifically, first, the prediction model creation unit 23 learns the first weak learner 34 as follows.
[0039] The prediction model creation unit 23 uses the learning data 41 as is to perform machine learning to generate a mathematical model, and sets the optimized mathematical model as the first weak learner 34. The mathematical model here is a mathematical model in which age and sex are explanatory variables and the occurrence of an event is a response variable, and may have coefficients by which each explanatory variable is multiplied. In this case, each coefficient is optimized. Furthermore, the mathematical model here may be a neural network having an input layer, multiple intermediate layers, and an output layer. In this case, a weight vector (a different concept from the weights in column 134 in Figure 5) that determines how much information from a certain node is propagated to which node in the next layer is optimized.
[0040] Second, prediction model creation unit 23 trains second weak learner 35 as follows. The prediction model creation unit 23 copies (duplicates) the records (samples) of the learning data 41 according to the weights in Fig. 5. For example, the prediction model creation unit 23 increases the number of records in the learning data 41 where the age is "20 to 29" and the gender is "female" by five times, corresponding to record 135 in Fig. 5. Similarly, for all other records in FIG. 5, the prediction model creation unit 23 increases the records (samples) of the learning data 41 corresponding to those records by a factor indicated by the weight. The prediction model creation unit 23 uses the learning data 41 after the number of records has been corrected in this manner to machine-learn the mathematical model, and sets the optimized mathematical model as the second weak learner 35. Compared to the first weak learner 34, the second weak learner 35 has higher prediction accuracy.
[0041] If the number of records in the learning data 41 after correction increases drastically, it takes time to use it to machine-learn a mathematical model. For example, assume that the weight of record 135 in FIG. 5 is "50.0" and the weight of record 136 is "20.0". In this case, the prediction model creation unit 23 does not need to copy the record of the learning data 41 corresponding to record 135 50 times and copy the record of the learning data 41 corresponding to record 136 20 times. For example, the prediction model creation unit 23 may copy the record of the learning data 41 corresponding to record 135 5 times and copy the record of the learning data 41 corresponding to record 136 2 times (limited copy).
[0042] In step S223, the prediction model creation unit 23 creates the ensemble model 33. Specifically, the prediction model creation unit 23 creates the ensemble model 33 having the first weak learner 34 and the second weak learner 35. As described above, the ensemble model is a model in which the output results of the weak learners constituting the ensemble model are adopted by "majority decision." Thereafter, the prediction model creation process procedure is terminated.
[0043] (Output processing procedure) FIG. 9 is a flowchart of the output process procedure. In step S231, the output processing unit 24 of the information processing device 1 acquires the prediction data 42. Specifically, the output processing unit 24 acquires the prediction data 42 from the database 2. The prediction data 42 has the same configuration as the learning data 41 (FIG. 2). However, the event occurrence column of the prediction data 42 is blank. In other words, the learning data 41 is "supervised learning data" (past samples) in which the event occurrence is known, whereas the prediction data 42 is a current sample in which the event occurrence is unknown. In many cases, the prediction data 42 is collected from medical facilities other than a specific medical facility.
[0044] In step S232, the output processing unit 24 displays the prediction result of the event occurrence. Specifically, first, the output processing unit 24 inputs the combination of age and gender specified by the user as prediction data to the prediction model 33, and obtains the event occurrence as the output of the prediction model 33. Here, the output processing unit 24 may input the combination of age and gender specified by the user to only the first weak learner 34, or may input it to only the second weak learner 35. Secondly, the output processing unit 24 displays a display screen 51 (described in detail later) on the output device 13. Thereafter, the output processing procedure ends.
[0045] (display screen) 10 is an example of the display screen 51. Now, the user wishes to do the following. I want to see the number of men who were free of a particular disease in the past, by age group. -I want to compare the numbers at a specific medical facility with the general population, such as the general public.
[0046] Therefore, the user inputs "age" in the horizontal axis column 58 of the display information column 53 of the graph display column 52, inputs "event occurrence: 0 gender: male" in the legend column 59, and presses the graph display execution button 55. The output processing unit 24 then displays lines 56 and 57 for comparison on a graph 54 with the age group on the horizontal axis and the number of samples on the vertical axis. Line 56 indicates the number of examples of statistical information, and line 57 indicates the number of examples of learning data. In FIG. 10, the scale of the vertical axis is unified to one for simplification, but the output processing unit 24 may display a scale (larger digits) for line 56 in addition to a scale (smaller digits) for line 57.
[0047] In addition, users want: -We want to predict the occurrence of events for women in the age group "30-39" from the current prediction data collected from other medical facilities. · I want to use an ensemble model33 to predict the occurrence of an event.
[0048] Therefore, the user inputs, as prediction data, "30 to 39" in the age field 63 of the prediction data field 62 in the analysis field 61, "female" in the gender field 64, and "ensemble" in the prediction model field 65, and presses the analysis execution button 66. The output processing unit 24 then displays the prediction result in the event occurrence prediction result field 67, and displays an explanation of the prediction result in the prediction result explanation field 68.
[0049] From these display examples, it can be seen that the output processing unit 24 performs the following processes. The output processing unit 24 inputs the age “30 to 39” and the gender “female” to the ensemble model 33 having the first weak learner 34 and the second weak learner 35. The first weak learner 34 outputs the event occurrence "1". The second weak learner 35 also outputs the event occurrence “1”. The ensemble model 33 voted on these outputs and output the event occurrence “1”. The “C” in the event occurrence prediction result column 67 corresponds to the output “1” after the majority vote.
[0050] Incidentally, if both the first weak learner 34 and the second weak learner 35 output an event occurrence of "0", the ensemble model 33 outputs an event occurrence of "0" as a result of majority vote. Then, the output processing unit 24 displays "A" in the event occurrence prediction result column 67. If either the first weak learner 34 or the second weak learner 35 outputs an event occurrence of "0" and the other outputs an event occurrence of "1", the ensemble model 33 outputs "tie competition" as a result of majority vote. Then, the output processing unit 24 displays "B" in the event occurrence prediction result column 67.
[0051] In the example of Fig. 10, the user has desired to use the ensemble model. However, the user may wish to use only the first weak learner, or may wish to use only the second weak learner. Ultimately, according to these desires, the output display unit 24 can output the prediction result using only the first weak learner, using only the second weak learner, or using the ensemble model based on the user's selection.
[0052] (The significance of using weights to modify training data) It is of course possible to machine-train a mathematical model using the statistical information 43 (FIG. 3) as training data from the beginning. However, it is often the case that the training data collected from a specific single medical facility is biased, which is discovered after the fact. In addition, there are cases where a mathematical model machine-trained using training data collected from a specific single medical facility already exists, and the user wishes to continue using the machine-trained mathematical model solely for prediction of samples collected from the medical facility. Furthermore, as described above, the process of simply sorting the statistical information 43, dividing the number of examples for each layer, calculating weights, and copying the training data in a limited manner is simpler than the process of performing machine learning using the statistical information 43, which has a huge amount of information.
[0053] (Effects of this embodiment) The information processing device of this embodiment has the following advantages. (1) The information processing device can correct the learning data based on the bias of the learning data relative to the population. (2) The information processing device can correct sample bias in medical facilities. (3) The information processing device can modify the learning data by the simple method of copying. (4) The information processing device can use an ensemble model with high predictive accuracy. (5) The information processor allows a user to select the use of an ensemble model. (6) The information processing device can display bias in the learning data relative to a population in a comparative manner.
[0054] The present invention is not limited to the above-described embodiment, but includes various modified examples. For example, the above-described embodiment has been described in detail to easily explain the present invention, and is not necessarily limited to those including all of the configurations described. It is also possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. It is also possible to add, delete, or replace a part of the configuration of each embodiment with another configuration.
[0055] In addition, the above-mentioned configurations, functions, processing units, processing means, etc. may be realized in part or in whole by hardware, for example, by designing them as integrated circuits. In addition, the above-mentioned configurations, functions, etc. may be realized in software by a processor interpreting and executing a program that realizes each function. Information such as the program, table, file, etc. that realizes each function can be stored in a memory, a recording device such as a hard disk or SSD (Solid State Drive), or a recording medium such as an IC card, SD card, or DVD. In addition, the control lines and information lines shown are those that are considered necessary for the explanation, and not all control lines and information lines in the product are necessarily shown. In reality, it can be considered that almost all components are connected to each other. [Explanation of symbols]
[0056] 1. Information processing device 2 Database 3. Network 11 Central control unit 12 Input Devices 13 Output Devices 14 Main memory 15 Auxiliary storage 16 Communication equipment 21 Simultaneous Distribution Creation Department 22 Weight calculation unit 23 Prediction Model Creation Department 24 Output Processing Section 31 Joint distribution information 32 Weight Information 33 Prediction model (ensemble model) 34 First Weak Learner 35 Second Weak Learner 41 Training Data 42 Prediction Data 43 Statistics 51 Display screen
Claims
1. a simultaneous distribution creating unit that creates simultaneous distribution information having the same variables as the statistical information from the learning data; a weight calculation unit that calculates a degree of bias of the created sample of the joint distribution information compared to the statistical information as a population as a weight for each combination of the multiple variables; modifying the learning data based on the calculated weights; a prediction model creation unit that learns a prediction model using the corrected learning data; an output processing unit that outputs a prediction result for prediction data using the prediction model; Equipped with The statistical information is The learning data sample is obtained from a population including a facility where the learning data sample was collected, The prediction model creation unit, increasing the number of samples of the training data by copying samples of the training data based on the weights; The prediction model creation unit, training a first weak learner using the training data before the correction; training a second weak learner using the modified training data; creating an ensemble model comprising the first weak learner and the second weak learner; An information processing device comprising:
2. The output processing unit includes: outputting the prediction result using only the first weak learner, using only the second weak learner, or using the ensemble model based on a user selection; 2. The information processing apparatus according to claim 1,
3. The output processing unit includes: Illustrating the number of samples of the statistical information and the number of samples of the learning data in a comparative manner; 3. The information processing apparatus according to claim 2,
4. The simultaneous distribution creation unit of the information processing device Creating joint distribution information having the same variables as the statistical information from the training data; The weight calculation unit of the information processing device Calculating a degree of bias of the sample of the created joint distribution information compared to the statistical information as a population as a weight for each combination of the multiple variables; The prediction model creation unit of the information processing device, modifying the learning data based on the calculated weights; Training a predictive model using the corrected training data; The output processing unit of the information processing device Outputting a prediction result for prediction data using the prediction model; The statistical information is The learning data sample is obtained from a population including a facility where the learning data sample was collected, The prediction model creation unit, increasing the number of samples of the training data by copying samples of the training data based on the weights; The prediction model creation unit, training a first weak learner using the training data before the correction; training a second weak learner using the modified training data; creating an ensemble model comprising the first weak learner and the second weak learner; An information processing method for an information processing device comprising the steps of:
5. Computer, a simultaneous distribution creating unit that creates simultaneous distribution information having the same variables as the statistical information from the learning data; a weight calculation unit that calculates a degree of bias of the created sample of the joint distribution information compared to the statistical information as a population as a weight for each combination of the multiple variables; modifying the learning data based on the calculated weights; a prediction model creation unit that learns a prediction model using the corrected learning data; an output processing unit that outputs a prediction result for prediction data using the prediction model; An information processing program for causing the device to function as described above, The statistical information is The learning data sample is obtained from a population including a facility where the learning data sample was collected, The prediction model creation unit: executing a process of increasing the number of samples of the learning data by copying samples of the learning data based on the weights; The prediction model creation unit: training a first weak learner using the training data before the correction; training a second weak learner using the modified training data; executing a process of creating an ensemble model having the first weak learner and the second weak learner; An information processing program characterized by: