JP6463961B2 - Information processing apparatus, information processing method, and program - Google Patents

Description

  Embodiments described herein relate generally to an information processing apparatus, an information processing method, and a program.

  Stratification means using variables (also referred to as data items or parameters) that exist in common within a set of data, and classifying the data into different tiers (also referred to as groups) based on the difference in values. It is. By classifying data with different properties in advance and building a model in each group, it is possible to expect the effect of obtaining a high-quality model. Here, the quality of the model evaluates the quality of the model based on data such as KL (Kullback-Leibler) information amount, likelihood, AIC (Akaike's Information Criterion), decision coefficient, and independence between variables. There are various indicators. Note that the index is not limited to any one of these as long as it can be obtained by calculation processing. The variable used by stratification is hereinafter referred to as “stratified variable”. Further, what defines which layer the data is classified into when the value of which layer variable matches what (or in which range) is referred to as “layer specification”.

  For example, consider handling human data. The variable representing the character of a person can be gender, age, height, weight, location, etc. An example of a stratified specification is: “Gender is a stratified variable and if the value matches“ male ”, it is classified as“ male ”, and if the value matches“ female ”, it is classified as“ female ”. ”Can be considered. In addition, with height as a stratified variable, a stratified specification that divides the value by 5 cm increments based on 0 cm, a “high stature (180 cm or higher) layer” with a wider section width with the same stratified variable and “content height ( Various variations are possible, such as a layered specification that divides the value into “160 to 179 cm) layer” and “low stature (less than 160 cm) layer”. In addition, there may be a “high stature male” class defined by combining gender and height as stratification variables and combining the categories.

  Use stratification specifications to build a high quality model unless you know in advance which stratification variables affect the quality of the model, or if you want to see if a particular variable affects the quality of the model. It is difficult to make a decision by repeating trial and evaluation. Although it depends on the data, the stratification method is not uniquely determined as in the above example. In particular, when there are multiple candidates for stratification variables or when there are multiple classification specifications, the number of stratification specifications increases so that the worker cannot manage them. It becomes difficult to decide.

  With reference to FIG. 1, it will be described that there are many possible stratified specifications for data. FIG. 1 is a diagram for explaining a data segmentation method. The circles in FIG. 1 represent data, and these circles are plotted on an xy two-dimensional plane. As shown in FIG. 1, there are a and b as the sorting method in the x-axis direction, and c as the sorting method in the y-axis direction. The classification of a may be divided into a maximum of three layers of data, and the granularity may be coarsened to two layers by considering a superset combining two of them. Layers, that is, layers may not be separated. On the other hand, the division of b is classified into a maximum of four layers, but two of them can be combined to make two layers, or all can be combined into one layer. Similarly, c is a maximum of 3 layers, but it is also possible to change the particle size to 2 layers or 1 layer. However, a and b cannot be employed simultaneously. At this time, the number of possible layered specifications is as follows: x-axis layered specification number x y-axis layered specification number = (a layered specification number + b layered specification number) × c layered specification Since number = (3 + 4) × 3 = 21, there are 21 ways.

  Thus, enumerating all applicable stratification specifications will multiply each variable by the number of stratifications that can be applied. Then, when the number of stratification variables and the stratification method increase, so-called combination explosion easily occurs, and there is a problem that it becomes difficult to select a better stratification specification.

Japanese Patent No. 5085016 JP 2013-174951 A

  A problem to be solved by an embodiment of the present invention is to provide an information processing apparatus, an information processing method, and a program that enable selection of a suitable stratified specification.

  According to one embodiment, the information processing apparatus receives data and a stratified specification as input, and performs a plurality of subdivisions to change the stratified specification so that the stratification becomes finer. For each of the different specifications, the data is classified according to the subdivided stratified specifications, and the stratified evaluation values that are the stratified evaluation values are determined based on the classified stratified data. Based on the above, the subdivided stratified specification is selected, and the selected stratified specification is output.

It is a figure for demonstrating the classification method of data. It is a figure which shows the structure of the information processing apparatus 1 in 1st Embodiment. 3 is an example of data stored in a data storage unit 11. 3 is an example of data representing a stratified specification stored in a stratified specification storage unit 12; 3 is an example of data representing a stratified relationship stored in a relationship storage unit 13; It is an example of the data memorize | stored in the evaluation value memory | storage part 14 according to a layer. It is an example of the interface which selects the evaluation value according to a stratification and the subdivision execution condition. It is an example of the selection item which can be selected as a stratified evaluation value. It is an example of a selection item that can be selected as a subdivision execution condition. 4 is a diagram illustrating a configuration example of a display screen of the display device 40. FIG. It is a flowchart which shows an example of the whole process in 1st Embodiment. It is a figure which shows an example of the layered specification whose layered specification ID is 0 previously recorded on the layered specification memory | storage part 12. FIG. It is a figure which shows an example of the layered specification whose layered specification ID recorded in the layered specification memory | storage part 12 is 1. FIG. It is a figure which shows an example of the layered specification whose layered specification ID recorded in the layered specification memory | storage part 12 is 2. FIG. It is a figure which shows an example of the layered specification whose layered specification ID recorded in the layered specification memory | storage part 12 is 3. FIG. It is a figure which shows an example of the layered specification whose layered specification ID recorded in the layered specification memory | storage part 12 is 4. FIG. It is a figure which shows an example of the layered specification whose layered specification ID recorded in the layered specification memory | storage part 12 is 5. FIG. It is a figure which shows an example of the regression coefficient for every explanatory variable. It is a figure which shows an example of the stratified evaluation value for every stratified specification. It is a figure which shows an example of the layered specification whose layered specification ID recorded in the layered specification memory | storage part 12 is 6. FIG. It is a figure which shows the structure of the information processing apparatus 1b in 2nd Embodiment. It is an example of the interface which receives the location of ontology. It is a figure which shows the ontology which arranged the concept required in order to express a person and its characteristic. FIG. 24 is an example of a layered specification obtained by subdividing the layered specification of FIG. 16 using the ontology shown in FIG. 23. 24 is an example of the layered specification obtained by subdividing the layered specification shown in FIG. 24 using the ontology shown in FIG.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. The information processing apparatus 1 according to the present embodiment reflects the knowledge obtained by the stratified trial in the stratified design to be tried next, and tries the stratified by random or brute force, so that suitable stratified Discover in a few steps. More specifically, the information processing apparatus 1 performs the stratification using the stratification specification obtained by performing the subdivision to change the stratification specification obtained at a certain time point so that the stratification becomes finer. The process of trying and adopting a suitable one among them is repeated until the evaluation value for each layer, which is the evaluation value for each layer, is not improved. Thereby, the information processing apparatus 1 can output a stratified specification that improves the quality of a model constructed from input data.

  Subsequently, the configuration of the information processing apparatus 1 according to the present embodiment will be described with reference to FIG. FIG. 2 is a diagram illustrating a configuration of the information processing apparatus 1 according to the first embodiment. As illustrated in FIG. 2, the information processing apparatus 1 includes a data storage unit 11, a layered specification storage unit 12, a relationship storage unit 13, a layered evaluation value storage unit 14, a control unit 20, and a reception unit 31.

  The data storage unit 11 records data used in model construction. The data storage unit 11 has an electronic data recording function such as a database. In the present embodiment, it is assumed that data shown in FIG. 3 is stored in the data storage unit 11 prior to model construction. FIG. 3 is an example of data stored in the data storage unit 11. This data is about the bedridden age, the bedridden age, the age at which the person was surveyed in the past, and the survey results (specifically, gender, number of steps in daily life, degree of paralysis, The degree of contracture) is recorded.

  In the present embodiment, as an example, the control unit 20 constructs a regression model capable of predicting the age at which bedridden occurs, with the bedridden occurrence age as an objective variable and the rest as explanatory variables. Among the variables shown in FIG. 3, “sex”, “paralysis”, and “contracture” are variables that take categorical values instead of continuous values. In the present embodiment, in order to handle data that is not a numerical value in this way, a concept called a dummy variable is introduced. For example, the control unit 20 performs a process of replacing males with numbers such as 0 and females with 1. .

  Here, an example of constructing a linear regression model that applies to a data record stored in the data storage unit 11 will be described. The linear regression model is a model that expresses a relationship between variables (also referred to as data items or parameters) constituting data by the following equation (1).

このとき、ｙは目的変数、ｘ_ｉは説明変数である。α_ｉは変数ｘ_ｉにかかる係数であり、βは定数項、εは誤差項を表す。モデルの構築とは、データが当てはまるよう係数α_ｉや定数項βを決定することである。

At this time, y is an objective variable and x _i is an explanatory variable. α _i is a coefficient for the variable x _i , β is a constant term, and ε is an error term. Model construction means determining the coefficient α _i and the constant term β so that the data fits.

  The stratified specification storage unit 12 records a stratified specification used by a data stratified unit 21 described later of the control unit 20. The stratified specification storage unit 12 has an electronic data recording function such as a database. FIG. 4 is an example of data representing the stratified specification stored in the stratified specification storage unit 12. In the case of the present embodiment, as shown in FIG. 4, the stratified specification is stored as a set of four sets of “stratified specification ID, explanatory variable, value range, and layer”. The “layer-specific specification ID” is an identifier for grouping the units of processing performed by the data layer-dividing unit 21 for each layer, and one layer in the set of four groups having a common layer-specific specification ID. Treat as a separate specification. By using this layer-specific specification ID, it is possible to follow the detailed process of each layer. “Explanatory variable” is a variable in the model that can be stratified, “value range” is a range of values for determining the tier at the time of stratification, and “layer” distributes data corresponding to the value range Is a layer.

  The first line excluding the header in FIG. 4 represents the meaning that a person whose survey age is 0 to 40 years old is classified into the “0 to 40” layer. The “value range” in FIG. 4 is a variable that takes a quantitative value as a variable recorded in a numerical range, and a variable that stores an enumeration of values by commas is a variable that takes a category value.

  In addition, all the values in the row where “layer” is NA are associated with one layer. In other words, this means that no stratification is performed on that variable. Therefore, in the stratified specification of “stratified specification ID” 99 illustrated in FIG. 4, it means that stratification of 6 divisions is performed by multiplying three layers of age at the time of investigation and two layers of gender.

  In the case of the present embodiment, it is assumed that a layer-specific specification (FIG. 12) in which all layers are NA (that is, no variable is performed for any variable) is recorded in the layer-specific specification storage unit 12 in advance.

  The relationship storage unit 13 stores a subdivided relationship indicating how the subdivided specifications are subdivided. The relation storage unit 13 has an electronic data recording function such as a database. FIG. 5 is an example of data representing the stratified relationship stored in the relationship storage unit 13. In this embodiment, as shown in FIG. 5, the subdivision relationship between the stratified specifications is generated by subdividing the subdivision source ID that is the stratified specification ID for identifying the stratified specification that is the subdivision source. In addition, it is recorded as a pair with a segmentation destination ID that is a layered specification ID for identifying the layered specification.

  The stratified evaluation value storage unit 14 records the stratified evaluation value for each stratified specification. The stratified evaluation value storage unit 14 has an electronic data recording function such as a database. FIG. 6 is an example of data stored in the stratified evaluation value storage unit 14. In the present embodiment, as shown in FIG. 6, the layer-by-layer evaluation value storage unit 14 records a set of a layer-by-layer specification ID and a layer-by-layer evaluation value of the layer-by-layer specification.

  The control unit 20 acquires the data and the stratified specification, and performs a plurality of subdivisions for changing the stratified specification so that the stratification is further subdivided. The data is classified according to the stratified specification, and the stratified evaluation value that is the evaluation value for each stratification is determined based on the classified stratified data, and the subdivided layer is determined based on the plurality of determined stratified evaluation values. Select another specification and output the selected stratified specification. The determination process of the evaluation value for each layer will be described. For example, the control unit 20 constructs a model for each layer, determines a model evaluation value that is an evaluation value of the model for each constructed model, and determines a plurality of determined models. A stratified evaluation value is determined based on the evaluation value. Here, the model evaluation value is, for example, a determination coefficient that evaluates how well the model is applied to the data, and the stratified evaluation value is an average value of the determination coefficient.

  For example, the control unit 20 according to the present embodiment repeats the subdivision, the determination of the stratified evaluation value, and the selection of the stratified specification after the subdivision. Output the selected substratification specifications.

  The subdivision process in the control unit 20 includes the following two processes. One of the subdivisions is to add another variable stratification for each current layer, and the other subdivision is to change the classification of the values of the stratified variable included in the current layer. It is to be.

  In this embodiment, when subdividing the stratified specification, the control unit 20 adds a layer separately for all explanatory variables that are not stratified variables for each current layer, for example, for each current layer. Among the layers of the stratified variables included, all the layers that can be changed so that the value classification becomes finer are changed so that the value classification becomes finer separately.

  Here, the control unit 20 includes a data stratification unit 21, a model construction unit 22, a stratification evaluation unit 23, a segmentation determination unit 24, a segmentation unit 25, a display control unit 26, and an output unit 27.

  The data stratification unit 21 stratifies data according to the stratification specification. Specifically, for example, the data stratification unit 21 reads the data from the data storage unit 11, stratifies the data according to the stratification specification stored in the stratification specification storage unit 12, and sets the stratified data set. To the model construction unit 22.

  The model construction unit 22 generates a model for each layer that is classified by using data included in the layer. Specifically, for example, the model construction unit 22 constructs a model for each layer from the layered data received from the data layer separation unit 21, and outputs the model and data to the layer evaluation unit 23.

  It should be noted that which of the variables constituting the data is used as a target variable to build a model and which model to use is given to the model building unit 22 in advance. In the case of the present embodiment, for example, the model construction unit 22 constructs a linear regression model with the age of occurrence as the objective variable and the survey age, sex, number of steps in daily life, paralysis, and contracture as explanatory variables.

  The stratified evaluation unit 23 determines a model evaluation value that is an evaluation value of a plurality of statistical models constructed by the model construction unit 22, and uses the plurality of model evaluation values to determine a plurality of stratified evaluation values. . Specifically, for example, the stratified evaluation unit 23 uses the model calculated by the model construction unit 22 and the data used at that time to calculate a model evaluation value for each model by a separately specified evaluation method, and model evaluation Summarize values in stratified units. In addition, the stratified evaluation unit 23 calculates the stratified evaluation value by a separately designated evaluation method, and outputs the stratified evaluation value obtained as a result to the subdivision determining unit 24. The stratified evaluation unit 23 stores the stratified evaluation value obtained as a result in the stratified evaluation value storage unit 14.

  As described above, the model evaluation value is an index for evaluating the quality of the model, such as the amount of KL information, likelihood, AIC, determination coefficient, and independence between variables. The stratified evaluation value is an index for comprehensively evaluating a model generated by the stratification. In the present embodiment, as an example, the determination coefficient is specified as the model evaluation value, and the average value of the determination coefficients is specified as the stratified evaluation value.

  The coefficient of determination is a representative index for evaluating the goodness of fit of a model to data. In the case of a linear regression model, when the determination coefficient is 1, it means that the regression model can express the relationship between data without error. As the coefficient of determination approaches 0, the regression model does not apply to the data.

  The average value of the determination coefficient is a value obtained by collecting the determination coefficients of each model for each layer and averaging them. For example, when the determination coefficients of the linear regression model created using the stratified data divided into three layers are 0.6, 0.7, and 0.5, respectively, this average value is 0. 6 (= (0.6 + 0.7 + 0.5) / 3) is the evaluation value for the entire layer.

  The stratified evaluation value is not limited to the average value of the determination coefficient, and the maximum value or the minimum value may be used among the model evaluation values of the model generated from a certain stratified data. In addition, among the model evaluation values of models generated from certain stratified data, the average value of the difference compared to the model evaluation value when not stratified (difference average with no stratification), the model whose evaluation value has increased Or the ratio of models whose model evaluation value has increased with respect to the total number of models included in the stratified result (rising model rate) may be used.

  In the present embodiment, as an example, the accepting unit 31 accepts an instruction for specifying a model evaluation value calculation method and an instruction for specifying a stratified evaluation value calculation method from the outside of the information processing apparatus 1. The stratified evaluation unit 23 acquires a model evaluation value calculation method and a stratified evaluation value calculation method from the reception unit 31. The calculation method of the stratified evaluation value may be specified by the user using, for example, the interface shown in FIG. FIG. 7 is an example of an interface for selecting a stratified evaluation value and a subdivision execution condition. FIG. 8 is an example of selection items that can be selected as the stratified evaluation value.

  In the interface of FIG. 7, one model evaluation value can be selected from the selection items shown in FIG. 8 in the list box. Here, it is assumed that an average value of the determination coefficient is designated as the evaluation value for each layer.

  The subdivision determination unit 24 determines whether or not to subdivide further stratified specifications using these stratified evaluation values. Specifically, for example, the subdivision determining unit 24 summarizes the stratified evaluation values received from the stratified evaluation unit 23 by the unit of the stratified specification that is the source of the subdivision, and sets separately specified conditions. Use to determine whether to further subdivide the stratification. Then, when the subdivision determining unit 24 subdivides as a result of the determination, the subdivision determining unit 24 outputs the stratified specification to the subdividing unit 25.

  In the case of the present embodiment, as an example, the reception unit 31 receives an instruction for specifying a subdivision execution condition from the outside of the information processing apparatus 1. The subdivision determination unit 24 acquires the subdivision execution condition from the reception unit 31. A user may specify execution conditions for subdivision via an interface as shown in FIG. In the interface of FIG. 7, one subdivision execution condition can be selected from the selection items shown in FIG. 9 in the list box. FIG. 9 is an example of selection items that can be selected as execution conditions for segmentation. Here, only the maximum value is specified as the subdivision execution condition. In other words, the subdivision execution condition is such that only the largest evaluation value (in this case, the average coefficient of determination) of the stratified specifications obtained by subdividing a stratified specification is further subdivided. Is specified.

  In addition, as an execution condition of subdivision, the difference is compared with the evaluation value when the evaluation value by layer is 0.6 or more and the evaluation value by layer is not stratified up to the top three cases in descending order of evaluation value by layer. It may be 0.1 or more, and the difference between the evaluation values by layer is 0.1 or more compared to the evaluation value of the segmentation source. However, it is a common condition that the evaluation value for each layer changes in a more desirable direction. For example, in the case of the average value of the determination coefficient, it is determined that it is not necessary to subdivide if it does not change.

  When it is determined that further stratification specifications are further subdivided, the subdivision unit 25 refers to the above data and adds another variable stratification to the current stratification and / or the current stratification. Subdivide the layers by changing the value categories of the included variables. Specifically, for example, the subdivision unit 25 subdivides the layered specifications recorded in the layered specification storage unit 12 designated by the model construction unit 22 into subdivided specifications so that the classification is fine. Generated and recorded in the layered specification storage unit 12. The subdivision unit 25 records the relationship between the subdivision source and the subdivision destination in the relationship storage unit 13. Further, the subdivision unit 25 transmits the subdivided stratification specification to the data stratification unit 21 and instructs to perform the data stratification.

  As described above, there are two types of subdivision: one is to add a variable that is not used as a stratification variable to the stratification variable, and the other is the value of a variable that is used as a stratification variable. It is to make the division of. The subdivision unit 25 performs these two kinds of details, and generates subdivided specifications for all cases.

  Hereinafter, first, a case where a variable that is not used as a stratification variable is added to the stratification variable will be described. A variable that is not used as a stratified variable is obtained by finding a row whose layer is NA as described in the explanation of FIG. The definition of the layer is determined with reference to the data stored in the data storage unit 11.

  In this embodiment, when a new stratified specification is provided, all possible patterns are transmitted to the stratified specification storage unit 12 by adding one variable that is not used as a stratified variable to the stratified variable. At that time, the division is divided into two, and when the variable takes a quantitative value, the average value of the variable held by the data storage unit 11 at that time is set as a threshold value, and the layer becomes less than the layer. And add the definition to divide into two layers. For example, in the case of subdivision in which “daily life steps” is added as a stratified variable based on the stratified specification shown in FIG. 4, the average value 5000 (= (4000 + 12000 + 2000 + 2000) / 4) is a threshold value for classifying the variable. It becomes.

  On the other hand, if the variable takes a category value, add a definition that stratifies only that type. For example, in the case of the “paralysis” variable, the subdividing unit 25 records the stratified specifications that are changed so as to stratify “none”, “mild”, and “severe”.

  Next, a description will be given of a case where the value classification of variables used as stratified variables is made fine. When based on the stratified specification shown in FIG. 4, the classification of the value of “sex”, which is a variable that takes a category value, is provided for each of all values. In other words, since it is already a minimum unit, it is performed for “survey age”, which is a variable for taking a quantitative value. The variable used as the stratified variable is obtained by finding a row whose layer is not NA as described in the explanation of FIG. In the present embodiment, as an example, when the value range of a stratified variable that takes a quantitative value is made fine, a layer is added to redefine the value range so that the range is divided into two by the median value. For example, for the stratified specification in the state of FIG. 4, the value range related to the age at the time of the survey is divided into two, so that before the change, “0-40”, “41-65”, “66-” are divided into three. What was stratified is a six-divided stratification of “0-20”, “21-40”, “41-53”, “54-65”, “66-83” and “84-100” .

  The display control unit 26 displays the subdivision relationship of the stratified specifications generated in the processing process on the display device 40 in a tree structure. For example, the display control unit 26 includes a layered specification stored in the layered specification storage unit 12, a relationship storage unit 13, and a layered evaluation value storage unit 14, a subdivision relationship between the layered specifications, and a model included in each layer. And an interface that outputs the model evaluation value received from the model construction unit 22 and the model evaluation value received from the layer evaluation unit 23 to the outside of the apparatus.

  The display control unit 26 displays the models constructed and evaluated along the subdivision performed by the subdivision unit 25 and the relationship between the models on the display device 40. FIG. 10 is a diagram illustrating a configuration example of a display screen of the display device 40. Tables T2, T3, and T4 to which the paralysis stratification variables are added are shown, and tables T5 and T6 to which the contracture stratification variables are added are shown with respect to the table T1 of the display screen of FIG. ing. Tables T1 to T6 display information about the model such as the coefficients of the respective explanatory variables and constant terms.

  Note that this is a tree structure that represents the relationship between models generated by subdividing the stratified specifications, and does not represent the subdivided relationship between the stratified specifications.

  The output unit 27 outputs an arbitrary stratified specification displayed on the display device 40. For example, the output unit 27 receives designation of a model displayed on the display device 40 and outputs the model to the outside of the information processing apparatus 1. In the screen configuration example illustrated in FIG. 10, an “output” button displayed beside each model is an interface for specifying a model for the output unit 27, and the user can press the button to select the model. It is output outside the information processing apparatus 1. As an output, the table shown in FIG. 4 may be output as a file in CSV format.

  The operation of the information processing apparatus 1 having the above configuration will be described using FIG. 11 with reference to FIGS. FIG. 11 is a flowchart illustrating an example of overall processing in the first embodiment. FIG. 12 is a diagram illustrating an example of a layered specification whose layered specification ID is 0 recorded in advance in the layered specification storage unit 12. FIG. 13 is a diagram illustrating an example of a layered specification having a layered specification ID 1 recorded in the layered specification storage unit 12. FIG. 14 is a diagram illustrating an example of a layered specification with a layered specification ID 2 recorded in the layered specification storage unit 12. FIG. 15 is a diagram illustrating an example of a layered specification with a layered specification ID 3 recorded in the layered specification storage unit 12.

  FIG. 16 is a diagram illustrating an example of a layered specification having a layered specification ID 4 recorded in the layered specification storage unit 12. FIG. 17 is a diagram illustrating an example of a layered specification having a layered specification ID 5 recorded in the layered specification storage unit 12. FIG. 18 is a diagram illustrating an example of a regression coefficient for each explanatory variable. FIG. 19 is a diagram illustrating an example of a stratified evaluation value for each stratified specification. FIG. 20 is a diagram illustrating an example of a layered specification having a layered specification ID 6 recorded in the layered specification storage unit 12.

  (Step S101) First, via the interface shown in FIG. 7, the stratified evaluation unit 23 accepts the evaluation method of the stratified evaluation value from the accepting unit 31, and the subdivision determining unit 24 accepts the subdivision execution condition. 31. Here, it is assumed that “the average value of the determination coefficient” is selected as the evaluation value for each layer, and “only the maximum value” is selected as the subdivision execution condition.

  (Step S102) Next, the data stratification unit 21 stratifies the data stored in the data storage unit 11 in accordance with the stratification specification stored in the stratification specification storage unit 12, and models each layer. Data is output to the construction unit 22.

  For the first time, a layered specification (see FIG. 12) with a layered specification ID of 0 recorded in advance in the layered specification storage unit 12 is used. Since this stratified specification is a specification in which the entire group is the same group, the data registered in the data storage unit 11 is passed to the model building unit 22 as one group of data without being stratified as it is.

  (Step S103) Next, the model construction unit 22 constructs a linear regression model for each data group received from the data stratification unit 21. At this time, in the model, the coefficients and constant terms of each explanatory variable are expressed in a working memory (not shown) in the control unit 20 in the form shown in FIG.

  (Step S104) Next, the stratified evaluation unit 23 calculates a determination coefficient, which is an evaluation method received by the reception unit 31, from the linear regression model constructed by the model construction unit 22 and its data. At this time, the determination coefficient is assumed to be 0.5.

  (Step S105) Next, the stratification evaluation unit 23 summarizes the model evaluation values calculated in step S104 for each stratification, and calculates the stratification evaluation value based on the designated method. In the case of this embodiment, the average value of the determination coefficient is obtained. In the first case (that is, when there is no stratification), since the model is unique, 0.5, which is the determination coefficient obtained in step S104, is the stratified evaluation value.

  (Step S106) Next, the subdivision determination unit 24 determines execution and cancellation of subdivision of the stratified specification based on the designated subdivision execution condition. For example, the subdivision determination unit 24 determines execution of subdivision of the stratified specification when the specified subdivision execution condition is satisfied. In the first case (that is, when there is no stratification), since there is one stratified specification to be compared, this is the maximum value, execution of subdivision is determined, and the process proceeds to step S107.

  If the specified subdivision execution condition is not satisfied, it is determined that the subdivision of the stratified specification is to be stopped, the processing relating to the stratified specification is finished, and the processing proceeds to step S109.

  (Step S107) The subdivision unit 25 selects one of the variables not currently used as a stratification variable for the stratification specification in FIG. 12, subdivides the stratification specification, and outputs the result as a layer. It records in the separate specification storage unit 12 and the relationship storage unit 13. More specifically, since there are five variables that are not stratified, the stratified specifications of FIGS. 13 to 17 that are subdivided by adding each to the stratified variables are recorded. The stratification method is as described above.

  (Step S108) In parallel with step S107, the subdivision unit 25 selects one of the variables that take the quantitative value already used as the stratification variable for the stratification specification in FIG. In addition, it tries to generate a stratified specification with finer division of values. However, with respect to the stratified specification shown in FIG. 12, no stratified variable exists, so a new stratified specification is not recorded.

  Next, in step S102, the data stratification unit 21 converts the data stored in the data storage unit 11 into the stratified specification stored in the stratified specification storage unit 12 newly generated by subdivision. The data is stratified along the line, and the data is sent to the model construction unit 22 in units of the layer (group). At this time, since the stratified specifications storage unit 12 stores the stratified specifications of FIG. 13 to FIG. 17, stratification based on each is performed, and data is sent to the model building unit 22 in units of groups. It should be noted that an empty set group having no data record belonging to a layer is not sent.

  In step S103, the model construction unit 22 constructs a linear regression model for each group using the data received from the data stratification unit 21.

  Next, in step S104, the stratified evaluation unit 23 calculates a coefficient of determination, which is a designated model evaluation method, from the linear regression model constructed by the model construction unit 22 and its data. At this time, it is assumed that the determination coefficient has a result as shown in the model evaluation value (determination coefficient) in FIG.

  Next, in step S105, the stratification evaluation unit 23 summarizes the model evaluation values calculated in step S104 for each stratification, and calculates the stratification evaluation value based on the designated method. In the case of the present embodiment, it is assumed that the average value of the determination coefficient is obtained for each stratified specification, and the result is as shown in the stratified evaluation value (average value of the deciding coefficient) in FIG.

  Next, in step S <b> 106, the subdivision determination unit 24 determines execution or cancellation of subdivision of the stratified specification based on the designated subdivision execution condition. In the present embodiment, only the evaluation value with the largest stratification value is the target of segmentation, and therefore, with reference to FIG. 19, the stratification evaluation value is the largest (in FIG. 19, the stratification evaluation value is 0.67). ), A stratified specification (see FIG. 16) with a stratified specification ID of 4 using paralysis by stratification is sent to the subdividing unit 25.

  Next, in step S107, the subdivision unit 25 adds one of the variables that are not currently used as a stratification variable to the stratification specification with respect to the stratification specification of FIG. Is recorded in the layered specification storage unit 12. The subdivision unit 25 records the subdivision relationship in the relationship storage unit 13.

  For example, when the survey age is added to the stratification variable, the stratification specification is as shown in FIG. At this time, the layer was “survey age = 40-62 and paralysis = none”, “survey age = 63-85, and paralysis = none”, “survey age = 40-62, and paralysis = mild”, ... and so on, it is divided into 6 layers by multiplying 2 categories of age at the time of survey and 3 categories of paralysis. Further, the subdivision unit 25 records the subdivision relationship in the relationship storage unit 13 with the subclassification specification having the subclassification specification ID of 4 and the subdivision specification having the subclassification specification having the subclassification specification ID of 6. To do.

  In the same manner, the flow of subdivision of stratified specifications, model construction, model evaluation, and stratified evaluation will not be subdivided, or there will be no stratified specifications that satisfy the subdivision execution conditions. Repeat until the process.

  (Step S109) Specifically, in step S109, the subdividing unit 25 determines whether or not there is a layered specification to be subdivided.

  (Step S110) When it is determined in step S109 that there is no segmentation specification to be subdivided, that is, when it is determined that there is no further segmentation for all the segmentation specifications, the display control unit updates the display screen, The model construction based on the stratification process is finished.

  In the present embodiment, it is assumed that the quality improvement beyond the stratification regarding the paralysis is not obtained when any other stratification is added, and as a result, the stratification regarding the paralysis is finally selected.

  Through the above process, the stratification relating to “paralysis” is the stratification that best fits the regression model, and the model when the stratification is performed is obtained.

  As described above, in the information processing apparatus 1 according to the first embodiment, the control unit 20 acquires a data and a stratified specification, and performs a plurality of subdivisions for changing the stratified specification so that the stratified specification becomes finer. And classifying the data according to the subdivided stratified specification for each subdivided stratified specification, and determining a stratified evaluation value that is a stratified evaluation value based on the stratified data after classification Based on the plurality of stratified evaluation values, the subdivided stratified specification is selected, and the selected stratified specification is output.

  According to the first embodiment, on average, the quality can be reduced with a smaller number of steps than the method of trying the stratified specifications at random and the method of trying several stratified specifications without taking advantage of the knowledge obtained in the middle. A high model can be obtained.

(Second Embodiment)
Next, the second embodiment will be described. In the first embodiment, when subdividing, the addition of variables to the stratified variables is tried for all variables that are not stratified variables. In contrast, in the second embodiment, an ontology that is a concept hierarchy that represents a concept appearing in the target world and a relationship between these concepts is searched using a predetermined method, and an element is used. Try subdividing.

  First, the information processing apparatus 1b according to the second embodiment will be described with reference to FIG. FIG. 21 is a diagram illustrating a configuration of the information processing device 1b according to the second embodiment. Elements common to those in FIG. 2 are denoted by the same reference numerals, and detailed description thereof is omitted. In the configuration of the information processing apparatus 1b in the first embodiment, the control unit 20 is changed to the control unit 20b and the reception unit 31 is changed to the reception unit 31b with respect to the configuration of the information processing apparatus 1 in the first embodiment. It has been done. Specifically, in the control unit 20b of the second embodiment, the subdivision unit 25 is changed to the subdivision unit 25b as compared with the control unit 20 of the first embodiment.

  The control unit 20b uses the elements acquired by searching for the ontology, which is a concept hierarchy representing the concept appearing in the world targeted by the information processing apparatus 1b, and the relationship between these concepts, using a predetermined method. Try the above subdivision. Here, this element belongs to, for example, another variable belonging to the same concept as the variable used in the current layer and / or other concepts related to the concept to which the variable used in the current layer belongs. A variable or a segment of a variable used for each current layer is a segment that is further subdivided.

  The reception unit 31b has the same function as the reception unit 31 of the first embodiment, but further has the following functions. For example, the reception unit 31 b receives the location of the ontology specified by the user of the information processing apparatus 1 and outputs the location to the subdivision unit 25.

  FIG. 22 is an example of an interface that accepts the location of an ontology. For example, the location of the ontology may be received as URL text using the interface shown in FIG. The receiving unit 31b may receive data directly in a file format or a stream format.

  Here, the ontology is a concept hierarchy representing concepts (also referred to as classes) appearing in the world targeted by the information processing apparatus 1 and relationships (also referred to as properties) between the concepts. An ontology can be expressed as a graph network with classes as nodes and properties as links. A property represents a characteristic of a certain class by linking with another class. Of the classes connected by the property, one is the class that is the owner of the property, and the other is the class that is the value constraint on the range of classes (values) that can be connected. Here, among the relationships between classes, a relationship representing a classification called an is-a relationship is stretched between a higher class and a lower class, and is distinguished from a property as a special one.

  For example, FIG. 23 is a diagram showing an ontology in which humans and concepts necessary for expressing their characteristics are arranged. In FIG. 23, squares represent classes and arrows represent properties. For example, the characteristics of the “person” class are expressed by an “age” property, a “health state” property, and the like. For example, in the ontology of FIG. 23, the “age” property refers to “aged”, “middle age”, and “young” as value constraints, which means that an instance of a certain “person” class is “aged”. , “Middle-age” and “young” classes (and their subclasses) can be referred to via the “age” property. Similarly, an instance of the “health” class means that instances of the “present” and “none” classes (and their subclasses) can be referred to via the “paralysis” and “contract” properties. ing. The “presence / absence value” class is classified into two classes, “present” and “none”, and the “present” class is connected in an is-a relationship so as to be classified into two classes of “severe” and “mild”. . “Age” is classified into three classes “elderly”, “middle-aged”, and “young”, and “elderly” class is classified into classes such as “80s” and “70s”. ing.

  For the ontology recording method and the ontology processing method, the ontology is expressed using, for example, the ontology description language OWL (Web Ontology Language) established by the World Wide Web Consortium (W3C), and Jena (URL: http: //Jena.sourceforge.net/) and The OWL API (URL: http://owlapi.sourceforge.net/) can be used to manage and process the content.

  When adding a new stratification variable to the stratification specification recorded in the stratification specification storage section 12, the subdivision section 25b does not change the stratification specification for all the variables that can be added. Referring to the property of, stratified variable candidates are extracted only within the range of variables corresponding to other properties possessed by a class having a property corresponding to the stratified variable determined at a certain time.

  A specific example will be described below in the case of subdividing by adding another stratification variable from the stratification specification having “paralysis” shown in FIG. 16 as a stratification variable. In the ontology shown in FIG. 23, a class having the “paralysis” property is “health state”. The only property of “health state” other than paralysis is “contracture”. Therefore, for the stratified specification shown in FIG. 16, the subdividing unit 25 generates a stratified specification having two variables of “paralysis” and “contracture” as stratified variables, and records them in the stratified specification storage unit 12. To do.

  At this time, the values of the “contracture” variable are “none”, “mild”, and “severe” according to FIG. 4, but according to the ontology shown in FIG. “None” is classified, and “Yes” is classified into “mild” and “severe”. 24 is an example of the layered specification obtained by subdividing the layered specification of FIG. 16 using the ontology shown in FIG. FIG. 25 is an example of the layered specification obtained by subdividing the layered specification shown in FIG. 24 using the ontology shown in FIG.

  The subdivision unit 25b searches the ontology in FIG. 23, and the values of the property “contract” corresponding to the “contract” variable are first classified into “present” and “none”, and these are used as layers. Define segmentation. Since the upper class of “mild” and “severe” is “yes”, as shown in FIG. 24, the subdividing unit 25b takes the layer of “mild, severe” as the value range of the “yes” layer. Generate another specification. This is different from the stratified specification that divides all possible values of “contracture” into different layers as shown in FIG. 17 in the first embodiment.

  Then, when subdividing the stratified specification of FIG. 24, since it is obtained from the ontology that “Yes” can be classified into “mild” and “severe”, the subdividing unit 25b performs the stratified specification shown in FIG. The layered specifications shown in FIG. 25 are generated in which the “contract” is “present” and the “contract” is “severe” and “mild”.

  As described above, in the information processing device 1b according to the second embodiment, the control unit 20b displays the ontology that is a concept hierarchy representing the concept that appears in the world targeted by the information processing device 1b and the relationship between these concepts. The subdivision is tried using elements obtained by searching in a predetermined manner. As a result, a layered specification in which layers are added within the range of properties of the same concept in the ontology is obtained. For this reason, a stratified specification reflecting the concept hierarchy defined by the ontology is obtained. An ontology is a group of meanings originally and describes concepts that exist in the target world. Therefore, using an ontology to add stratified specifications from a range where the number of search steps on the network is close (abstractly speaking, strongly related) is more than introducing a distant concept into stratification suddenly. Will develop the stratification applied at a certain point in the near range, and asymptotically subdivide the stratification specification.

  Note that an information processing system including a plurality of devices may process each processing of the information processing device 1 or 1b of each embodiment in a distributed manner by the plurality of devices.

  Further, a program for executing each process of the information processing apparatus 1 or 1b of each embodiment is recorded on a computer-readable recording medium, the program recorded on the recording medium is read into a computer system, and the processor By executing, the above-described various processes related to the information processing apparatus 1 or 1b of each embodiment may be performed.

  Here, the “computer system” may include an OS and hardware such as peripheral devices. Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used. The “computer-readable recording medium” means a flexible disk, a magneto-optical disk, a ROM, a writable nonvolatile memory such as a flash memory, a portable medium such as a CD-ROM, a hard disk built in a computer system, etc. This is a storage device.

  The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, and what is called a difference file (difference program) may be sufficient.

  As described above, the present invention is not limited to the above-described embodiments as they are, and can be embodied by modifying the components without departing from the scope of the invention in the implementation stage. In addition, various inventions can be formed by appropriately combining a plurality of components disclosed in the embodiment. For example, you may delete some components from all the components shown by embodiment. Furthermore, the constituent elements over different embodiments may be appropriately combined.

1, 1b Information processing device 11 Data storage unit 12 Stratified specification storage unit 13 Relational storage unit 14 Stratified evaluation value storage unit 20 Control unit 21 Data stratification unit 22 Model construction unit 23 Stratification evaluation unit 24 Segmentation determination unit 25 , 25b Subdivision unit 26 Display control unit 27 Output unit 31, 31b Reception unit

Claims (16)

Data as input and stratification specifications for a plurality of variables, and the subdivision of changing the layer-specific data as further the stratification finer plurality trial, the subdivided for each stratification specifications after subdivision The data is classified according to the later stratified specifications, and the stratified evaluation value, which is the valuation value for each stratification, is determined based on the classified stratified data. After the subdivision based on the determined stratified evaluation values An information processing apparatus that selects a stratified specification and outputs the selected stratified specification. The model is constructed for each layer, the model evaluation value that is the evaluation value of the model is determined for each constructed model, and the evaluation value for each layer is determined based on the plurality of determined model evaluation values. The information processing apparatus described. The model evaluation value is a coefficient of determination that evaluates the goodness of fit of the model to the data,
The information processing apparatus according to claim 2, wherein the stratified evaluation value is an average value of the determination coefficients. When the subdivision, determination of the evaluation value by stratification, and selection of the stratified specification after subdivision are repeated, and the execution condition of the subdivision is not satisfied, the stratified specification after the subdivision selected last is output The information processing apparatus according to any one of claims 1 to 3. The information processing apparatus according to any one of claims 1 to 4, wherein the subdivision is adding another variable stratification according to the current stratification. The information processing apparatus according to any one of claims 1 to 4, wherein the subdivision is a change so that a classification of a value of a variable included in each current layer becomes finer. When trying to subdivide the stratified specifications, add the layers separately for all explanatory variables that are not stratified variables for each current tier, and among the stratified variable layers included in the current tier The information processing apparatus according to any one of claims 1 to 6, wherein all the layers that can be changed so that the value classification becomes fine are changed so that the value classification becomes fine separately. 6. The subdivision is tried using an element obtained by searching an acquired ontology, which is a concept hierarchy representing a concept targeted by the information processing apparatus and the relationship between the concepts, using a predetermined method. The information processing apparatus described. 9. The element is a variable belonging to a concept related to a concept to which another variable included in a concept to which a variable used in each current layer belongs and / or a variable used in a current layer is belonging. The information processing apparatus described in 1. The information processing apparatus according to claim 8, wherein the element is a classification obtained by further subdividing a variable classification used for each current layer. A data stratification section for stratifying the data according to stratification specifications;
For each layer that is stratified, a model building unit that generates a statistical model using data included in the layer; and
A model evaluation value that is an evaluation value of a plurality of statistical models constructed by the model construction unit; and a stratified evaluation unit that determines the plurality of stratification evaluation values using the plurality of model evaluation values;
Using the stratified evaluation value, a subdivision determining unit that determines whether or not to subdivide further stratified specifications;
If it is decided to subdivide the further stratification specification, referring to the data, addition of another variable to the current stratification and / or variables included in the current stratification A subdivision unit for executing the change of the value category as the subdivision;
The information processing apparatus according to claim 1, further comprising: The information processing apparatus according to any one of claims 1 to 11, further comprising a reception unit that receives an instruction for specifying a calculation method of the stratified evaluation value from outside the information processing apparatus. The information processing apparatus according to claim 4, further comprising: a reception unit that receives an instruction for specifying the execution condition of the subdivision from the outside of the information processing apparatus. A storage unit for storing a subdivision relationship indicating how subdivision is made between stratified specifications;
A display control unit for displaying the subdivision relationship of the stratified specification generated in the process in a tree structure on a display device;
An output unit for outputting an arbitrary stratified specification displayed on the display device;
The information processing apparatus according to claim 1, further comprising: The computer inputs the data and the stratified specifications regarding a plurality of variables, and performs a plurality of subdivisions for changing the stratified specifications so that the stratification becomes finer. The data is classified according to the subdivided stratified specifications, the stratified evaluation value that is the stratified evaluation value is determined based on the categorized stratified data, and the subdivision is performed based on the determined stratified evaluation values. An information processing method for selecting a classified stratified specification and outputting the selected stratified specification. On the computer,
Using data and stratified specifications for multiple variables as input, try multiple types of subdivision to change the stratified specifications so that the stratification becomes finer, and subdivide each of the subdivided stratified specifications The data is classified according to the later stratified specifications, and the stratified evaluation value, which is the valuation value for each stratification, is determined based on the classified stratified data. After the subdivision based on the determined stratified evaluation values A program for selecting the stratified specifications and outputting the selected stratified specifications.

Images