WO2010027046A1 - Information processing device, information processing method, information storage medium, and program - Google Patents

Information processing device, information processing method, information storage medium, and program Download PDF

Info

Publication number
WO2010027046A1
WO2010027046A1 PCT/JP2009/065479 JP2009065479W WO2010027046A1 WO 2010027046 A1 WO2010027046 A1 WO 2010027046A1 JP 2009065479 W JP2009065479 W JP 2009065479W WO 2010027046 A1 WO2010027046 A1 WO 2010027046A1
Authority
WO
WIPO (PCT)
Prior art keywords
phenomenon
data
value
influence
candidate
Prior art date
Application number
PCT/JP2009/065479
Other languages
French (fr)
Japanese (ja)
Inventor
達也 原田
崇 渋谷
康夫 國吉
Original Assignee
国立大学法人東京大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 国立大学法人東京大学 filed Critical 国立大学法人東京大学
Publication of WO2010027046A1 publication Critical patent/WO2010027046A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/29—Graphical models, e.g. Bayesian networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching
    • G06F2218/16—Classification; Matching by matching signal segments
    • G06F2218/20—Classification; Matching by matching signal segments by applying autoregressive analysis

Definitions

  • the present invention relates to an information processing apparatus, an information processing method, an information storage medium, and a program.
  • Non-Patent Document 1 discloses a method for calculating the above-mentioned index between phenomena whose possible values are discrete based on the amount of information given by the value of a certain phenomenon to the value of another phenomenon.
  • Non-Patent Document 2 discloses a method of extending the method disclosed in Non-Patent Document 1 so that the above-described index can be calculated between phenomena where possible values are continuous. Note that the methods disclosed in Non-Patent Document 1 and Non-Patent Document 2 are methods supported by information theory.
  • the present invention has been made in view of the above problems, and is an information processing apparatus and information processing method capable of modeling a phenomenon using a time series value of the phenomenon selected based on the strength of the influence
  • An object of the present invention is to provide an information storage medium and a program.
  • an information processing apparatus includes modeling target phenomenon data including a time series value of a modeling target phenomenon to be modeled, and candidates that affect the modeling target phenomenon.
  • a phenomenon data acquisition means for acquiring influence candidate phenomenon data including time series values of each of a plurality of influence candidate phenomena, a time series value of the influence candidate phenomenon for each of the influence candidate phenomena, and the modeling target
  • An influence index associating means for calculating an influence index indicating the strength of the influence of the influence candidate phenomenon on the modeling target phenomenon based on the relationship with the time series value of the phenomenon, and associating with the influence candidate phenomenon;
  • a phenomenon selection means for selecting at least one phenomenon from the plurality of influence candidate phenomena based on an influence index associated with each of the influence candidate phenomena; and the phenomenon selection means.
  • correspondence rule Based on the time series value of the phenomenon to be selected and the time series value of the phenomenon to be modeled, a correspondence rule between the value of the phenomenon selected by the phenomenon selecting means and the value of the modeled phenomenon And correspondence rule data generating means for generating corresponding rule data to be shown.
  • the information processing method includes modeling target phenomenon data including a time series value of a modeling target phenomenon to be modeled, and a plurality of influence candidates that are candidates for influencing the modeling target phenomenon.
  • the information storage medium includes modeling target phenomenon data including a time series value of a modeling target phenomenon to be modeled, and a plurality of influence candidates that are candidates for influencing the modeling target phenomenon.
  • Effect candidate phenomenon data including time series values of each phenomenon
  • phenomenon data acquisition means for acquiring, for each of the effect candidate phenomena, the time series value of the effect candidate phenomenon, the time series value of the modeling target phenomenon, Based on the relationship, an influence index indicating the strength of the influence of the candidate effect phenomenon on the modeling target phenomenon is calculated, and an influence index associating means for associating with the influence candidate phenomenon is associated with each influence candidate phenomenon.
  • a phenomenon selection means for selecting at least one phenomenon from the plurality of influence candidate phenomena based on an influence index; a timeline of the phenomenon selected by the phenomenon selection means; Correspondence that generates correspondence rule data indicating a correspondence rule between the value of the phenomenon selected by the phenomenon selection means and the value of the phenomenon to be modeled based on the value and the time series value of the phenomenon to be modeled
  • a computer-readable information storage medium storing a program characterized by causing a computer to function as rule data generation means.
  • the program according to the present invention includes modeling target phenomenon data including time series values of a modeling target phenomenon to be modeled, and a plurality of influence candidate phenomena that are candidates for affecting the modeling target phenomenon.
  • the effect candidate phenomenon data including the time series value of the phenomenon candidate data acquisition means for obtaining the relationship between the time series value of the candidate effect phenomenon and the time series value of the phenomenon to be modeled
  • the influence index indicating the strength of the influence of the influence candidate phenomenon on the modeling target phenomenon
  • a phenomenon selection means for selecting at least one phenomenon from among the plurality of influence candidate phenomena, and a time series of the phenomena selected by the phenomenon selection means
  • a correspondence rule for generating correspondence rule data indicating a correspondence rule between the value of the phenomenon selected by the phenomenon selection unit and the value of the phenomenon to be modeled based on the time series value of the phenomenon to be modeled
  • a computer is made to function as data generation means.
  • the influence index indicating the strength of the influence of the influence candidate phenomenon on the modeling target phenomenon is calculated. .
  • at least one influence candidate phenomenon is selected from a plurality of influence candidate phenomena based on the influence index.
  • correspondence rule data indicating a correspondence rule between the value of the selected phenomenon and the value of the modeled phenomenon is obtained.
  • the influence index association means is based on a relationship between a time series value of the influence candidate phenomenon that is qualitative data and a time series value of the modeling target phenomenon that is quantitative data.
  • the influence index is calculated. In this way, it is possible to evaluate the strength of the influence of the phenomenon that is qualitative data on the phenomenon that is quantitative data.
  • the influence index associating means has a relationship between a time series value of the influence candidate phenomenon that is quantitative data and a time series value of the modeling target phenomenon that is qualitative data. Based on this, the influence index is calculated. In this way, it is possible to evaluate the strength of the influence of the phenomenon that is quantitative data on the phenomenon that is qualitative data.
  • the value of the candidate effect phenomenon is based on the relationship between the time series value of the candidate effect phenomenon and the time series value of the modeling target phenomenon.
  • Basic correspondence rule data indicating a correspondence rule with the value of the modeling target phenomenon is generated, and further includes basic correspondence rule data associating means for associating with the influence candidate phenomenon, and the correspondence rule data generation means is the phenomenon selection means.
  • the correspondence rule data is generated based on the basic correspondence rule data associated with the phenomenon selected by the above. In this way, when the correspondence rule data is generated, the basic correspondence rule data indicating the correspondence rule between the value of each selected phenomenon and the value of the modeling target phenomenon can be used.
  • the apparatus further includes a predicted value calculation unit that calculates a predicted value of the modeling target phenomenon at a given time in accordance with the corresponding rule indicated by the corresponding rule data. In this way, the value of the phenomenon to be modeled can be predicted.
  • calculation rule data storage means for storing a plurality of calculation rule data indicating rules for calculating the influence index, a time series value of the influence candidate phenomenon, and a time series of the modeling target phenomenon
  • calculation rule data selection means for selecting at least one calculation rule data from a plurality of the calculation rule data based on the relationship between the values, and the influence index associating means includes the calculation rule data selection
  • the influence index is calculated according to a rule indicated by calculation rule data selected by the means, and is associated with the influence candidate phenomenon. In this way, when calculating the influence index, a calculation rule according to the relationship between the time series value of the influence candidate phenomenon and the time series value of the modeling target phenomenon can be used.
  • each of the influence candidate phenomena is associated with phenomenon type data indicating whether the value of the influence candidate phenomenon is qualitative data or quantitative data
  • the calculation rule data selection means includes Calculation rule data may be selected based on phenomenon type data associated with an influence candidate phenomenon. In this way, when calculating the influence index, a calculation rule according to whether the value of the influence candidate phenomenon is qualitative data or quantitative data can be used.
  • the correspondence rule data generation unit generates correspondence rule data indicating a coefficient of a predetermined function. In this way, since it is not necessary to determine the function when generating the correspondence rule data, the generation of the correspondence rule data is facilitated.
  • the influence index indicates a causal influence strength that the influence candidate phenomenon has on the prediction target phenomenon. In this way, it is possible to predict the value of the phenomenon to be predicted using the time series value of the phenomenon selected based on the strength of the causal influence received.
  • the image processing device further includes image output means for outputting an image based on a selection result by the phenomenon selection means. In this way, the user can obtain the selection result by the phenomenon selection means as an image.
  • the information processing apparatus 10 includes a control unit 12, a storage unit 14, and a user interface (UI) unit 16, as illustrated in the hardware configuration diagram of FIG.
  • the control unit 12, the storage unit 14, and the UI unit 16 are connected via a bus 18.
  • the control unit 12 is a program control device such as a CPU, and operates according to a program installed in the information processing apparatus 10.
  • the storage unit 14 is a storage element such as a RAM or a hard disk.
  • the storage unit 14 stores a program executed by the control unit 12.
  • the storage unit 14 also operates as a work memory for the control unit 12.
  • the UI unit 16 includes a display, a microphone, a speaker, a button, and the like, and outputs the content of the operation performed by the user and the voice input by the user to the control unit 12. In addition, the UI unit 16 displays and outputs information according to an instruction input from the control unit 12.
  • FIG. 2 is a functional block diagram illustrating an example of functions realized by the information processing apparatus 10 according to the present embodiment.
  • the information processing apparatus 10 functionally includes a phenomenon data storage unit 20, a phenomenon data acquisition unit 22, an association unit 24, a calculation rule data storage unit 26, and calculation rule data. It functions as including a selection unit 28, a phenomenon selection unit 30, a corresponding rule data generation unit 32, a predicted value calculation unit 34, and an image output unit 36.
  • the phenomenon data storage unit 20 and the calculation rule data storage unit 26 are realized mainly by the storage unit 14.
  • the phenomenon data acquisition unit 22, the association unit 24, the calculation rule data selection unit 28, the phenomenon selection unit 30, the corresponding rule data generation unit 32, the predicted value calculation unit 34, the image output unit 36, 12 is mainly realized.
  • These elements are realized by executing a program installed in the information processing apparatus 10 that is a computer by the control unit 12 such as a CPU included in the information processing apparatus 10.
  • the program is supplied to the information processing apparatus 10 via a computer-readable information transmission medium such as a CD-ROM or DVD-ROM, or via a communication network such as the Internet.
  • the phenomenon data storage unit 20 stores phenomenon data including time series values of a plurality of phenomena.
  • the phenomenon data includes a plurality of phenomenon unit data 38 (see FIG. 3).
  • FIG. 3 is a diagram illustrating an example of the data structure of the phenomenon unit data 38.
  • the time series value of the phenomenon is not limited to the time series value related to the natural phenomenon such as the observed temperature, the measured distance, the rotation speed of the motor, the dosage, etc. It may be a time series value related to various phenomena or a time series value related to psychological phenomena.
  • the time series value of the phenomenon is not limited to the quantitative data as described above, for example, whether or not the event is held, time zone, day of the week, weather, room number, presence or absence of symptoms, proper noun, switch state, Or other qualitative data.
  • the phenomenon unit data 38 is a phenomenon ID 40 that is an identifier of a phenomenon, a phenomenon name data 42 that indicates the name of the phenomenon, and a phenomenon value indicated by the phenomenon unit data 38 is qualitative data.
  • Phenomenon type data 44 indicating whether the data is quantitative data or time series value data 46 indicating the time series value of the phenomenon.
  • FIG. 4 is a diagram illustrating an example of the data structure of the time series value data 46. As illustrated in FIG. 4, the time-series value data 46 includes a plurality of combinations of time data 48 indicating a time point and value data 50 indicating a value (for example, an observed value or a measured value) corresponding to the time point. .
  • the phenomenon type data 44 is a flag.
  • the phenomenon type data 44 takes a value of “1” and indicates the phenomenon indicated by the phenomenon unit data 38.
  • the value is quantitative data, the value is “0”.
  • the phenomenon unit data 38 indicates the value of the phenomenon associated with each of a plurality of time points.
  • the value of the phenomenon may be a value that can be handled as a random variable that varies depending on the time.
  • the value indicated by the value data 50 is qualitative data, the value indicated by the value data 50 is, for example, the value of a dummy variable.
  • the control unit 12 may acquire an operation signal via the UI unit 16 such as a mouse or a keyboard, generate phenomenon data based on the operation signal, and output the phenomenon data to the phenomenon data storage unit 20. Good.
  • the phenomenon data acquisition unit 22 acquires phenomenon data.
  • the phenomenon data acquisition unit 22 acquires, for example, phenomenon data stored in the phenomenon data storage unit 20.
  • the control unit 12 may generate phenomenon data corresponding to an operation signal acquired via a mouse, a keyboard, or the like, and the phenomenon data acquisition unit 22 may acquire the generated phenomenon data.
  • the phenomenon indicated by each phenomenon unit data 38 stored in the phenomenon data storage unit 20 is a phenomenon to be modeled (modeling target phenomenon) (more specifically, for example, given It is also a phenomenon (prediction target phenomenon) for which a prediction value is calculated at the time point, and is also a phenomenon (influence candidate phenomenon) that is a candidate for affecting the modeling target phenomenon.
  • the phenomenon data acquisition unit 22 includes a time series value of a modeling target phenomenon to be modeled and a time series value of each of a plurality of influence candidate phenomena that are candidates for affecting the modeling target phenomenon. The phenomenon data will be acquired.
  • An influence index 52 indicating (for example, the intensity of causal influence) is calculated and associated with the influence candidate phenomenon (see FIG. 5).
  • the associating unit 24 for each candidate effect phenomenon based on the relationship between the time series value of the effect candidate phenomenon and the time series value of the modeling target phenomenon, the value of the effect candidate phenomenon and the model Basic correspondence rule data indicating a correspondence rule (specifically, a relational expression (model expression) such as a function indicating a relation between a value of an influence candidate phenomenon and a value of a modeling target phenomenon), for example 54 is generated and associated with the influence candidate phenomenon (see FIG. 5).
  • the associating unit 24 may estimate a correspondence rule (for example, a relational expression such as a function, a coefficient or a parameter of the relational expression, etc.) in the process of calculating the influence index 52.
  • the associating unit 24 generates, for example, influence index data 56 (see FIG. 5).
  • FIG. 5 is a diagram illustrating an example of the data structure of the influence index data 56.
  • the influence index data 56 includes an influence candidate phenomenon ID 58 corresponding to the phenomenon ID 40 of the influence candidate phenomenon, a modeling target phenomenon ID 60 corresponding to the phenomenon ID 40 of the modeling target phenomenon, the above-described influence index 52, The basic correspondence rule data 54 described above is included.
  • the associating unit 24 associates the influence index 52 with the combination of the influence candidate phenomenon and the modeling target phenomenon.
  • the associating unit 24 associates the basic correspondence rule data 54 with the combination of the influence candidate phenomenon and the modeling target phenomenon.
  • the associating unit 24 when the number of phenomenon unit data 38 is n, the associating unit 24 generates, for each phenomenon unit data 38, influence index data 56 corresponding to other (n ⁇ 1) phenomena. To do. That is, the associating unit 24 generates n (n ⁇ 1) pieces of influence index data 56.
  • the calculation rule data storage unit 26 stores a plurality of calculation rule data indicating rules for calculating the influence index 52.
  • the calculation rule data selection unit 28 selects at least one calculation rule data from among a plurality of calculation rule data based on the relationship between the time series value of the influence candidate phenomenon and the time series value of the modeling target phenomenon. .
  • the calculation rule data selection unit 28 quantifies, for example, whether the influence candidate phenomenon corresponds to qualitative data or quantitative data, and whether the modeling target phenomenon corresponds to qualitative data. Calculation rule data is selected based on whether it corresponds to data.
  • the calculation rule data selection unit 28 uses, for example, the value of the phenomenon type data 44 included in the phenomenon unit data 38 corresponding to the influence candidate phenomenon and the phenomenon unit data 38 corresponding to the modeling target phenomenon. Calculation rule data is selected based on a combination with the value of the included phenomenon type data 44.
  • the associating unit 24 calculates the influence index 52 according to the rule indicated by the calculation rule data selected by the calculation rule data selecting unit 28. That is, the influence index 52 includes the value of the phenomenon type data 44 included in the phenomenon unit data 38 corresponding to the influence candidate phenomenon, the value of the phenomenon type data 44 included in the phenomenon unit data 38 corresponding to the modeling target phenomenon, Are associated with the combination.
  • the associating unit 24 may generate the basic correspondence rule data 54 based on the calculation rule data. Specifically, for example, when the calculation rule data indicates a predetermined function type, the associating unit 24 may estimate the coefficient of the function. Then, the associating unit 24 may generate basic correspondence rule data 54 indicating a function for which a coefficient is determined.
  • the calculation rule data includes the value of the phenomenon type data 44 included in the phenomenon unit data 38 corresponding to the influence candidate phenomenon and the phenomenon included in the phenomenon unit data 38 corresponding to the modeled phenomenon. It is associated with the combination of the value of the type data 44.
  • the time series value of the influence candidate phenomenon is Y
  • the time series value of the modeling target phenomenon is X.
  • Transfer Entropy which is an index based on information theory, disclosed in Non-Patent Document 1, is used as an influence index 52 when an influence candidate phenomenon and a modeling target phenomenon correspond to qualitative data.
  • p represents the transition probability.
  • This equation then calculates the Kullback-Leibler information amount between p (x t
  • the above equation can be expressed as the following equation using entropy (average information amount) H (X).
  • X t , X t-1 (k) , Y t-1 (l) included in Equation 2 are x t , x t-1 (k) , y t-1 each composed of the value data 50. It means a set of (l) .
  • the symbol which encloses the cross mark in a circle means a direct product.
  • the associating unit 24 specifically includes, for example, p (x t
  • the basic correspondence rule data 54 indicating the model expression of the transition probability model is generated.
  • the time series value of the influence candidate phenomenon is Y
  • the time series value of the modeling target phenomenon is X.
  • k is an embedding dimension and a ⁇ R k is an autoregressive coefficient vector.
  • ⁇ t (x) is an error according to a normal distribution with an average of 0 and a variance ⁇ x 2 and is independent of the value of x t ⁇ 1 (k) .
  • the maximum likelihood estimation value of the autoregressive coefficient vector a and the variance ⁇ x 2 is estimated by the least square method.
  • the likelihood L and log likelihood l of the autoregressive model is a function of a and ⁇ x 2 And expressed as the following equation.
  • Equation 3 each term on the right side is expressed by the following equation based on Equation 3.
  • Equation 7 Solving the equation shown in Equation 7 gives the following equation.
  • the log-likelihood becomes a function of only the autoregressive coefficient vector a and is expressed as the following expression.
  • the maximum likelihood estimate of ⁇ x 2 may be minimized.
  • the least squares method can be implemented using, for example, a pseudo inverse matrix.
  • the embedding dimension k can be selected by a method using an Akaike information criterion (AIC).
  • Equation 3 the conditional probability distribution of x t is expressed by the following equation.
  • the variance-covariance matrix of this multidimensional normal distribution is expressed by the following equation.
  • This formula is a specific example of a calculation formula of the influence index 52 (T Y ⁇ X ) when the influence candidate phenomenon and the modeling target phenomenon correspond to the quantitative data.
  • TY ⁇ X calculated by the following equation. You may use as an influence parameter
  • the value of the influence index 52 is based on the variance of the error in the regression model (in this embodiment, the variance ratio). Yes.
  • the associating unit 24 specifically, for example, the coefficient is estimated by, for example, the least square method, etc.
  • Basic correspondence rule data 54 indicating the model expression of the mixed regression model of Expression 16 is generated.
  • the time series value of the influence candidate phenomenon is S
  • the time series value of the modeling target phenomenon is X.
  • Transfer Entropy is expressed by the following equation.
  • b (s t-1 (l) ) is a coefficient vector that switches for each state of the qualitative data s t-1 (l) , and x t-1 (k) is changed to s t-1 (l)
  • each s t-1 (l) is estimated by the least square method. This is because the autoregressive coefficient vector changes with the change of s t-1 (l) if the value of the influence candidate phenomenon has some influence on the value of the phenomenon to be modeled.
  • the regression coefficient vector is based on the concept that it will not change and will be constant.
  • This equation is a formula for calculating the influence index 52 (T S ⁇ X ) when the value of the influence candidate phenomenon is qualitative data (for example, symbol data) and the value of the modeling target phenomenon is quantitative data.
  • the influence index 52 is an index expressing how much the accuracy is improved when the time series value X is divided for each state of s t ⁇ 1 (l) by the average number of bits.
  • T S ⁇ X calculated by the following equation is used. You may use as an influence parameter
  • the value of the influence candidate phenomenon is qualitative data (for example, symbol data) and the value of the phenomenon to be modeled is quantitative data
  • the value of the influence index 52 is an error variance in the regression model. (In this embodiment, based on the dispersion ratio).
  • the associating unit 24 specifically includes, for example, the coefficient The basic correspondence rule data 54 indicating the model expression of the Switching autoregressive model of the above-described Expression 23, which is estimated by the least square method or the like, is generated.
  • the time series value of the influence candidate phenomenon is X
  • the time series value of the modeling target phenomenon is S.
  • Transfer Entropy is expressed by the following equation.
  • Transfer Entropy is as follows:
  • time series value of the qualitative data corresponding to the phenomenon to be modeled is s t , s t-1 (l)
  • time series value x t-1 Assuming that k) follows a normal distribution, the following relationship is established.
  • This formula is a specific example of a formula for calculating the influence index 52 (T X ⁇ S ) when the influence candidate phenomenon corresponds to quantitative data and the modeling target phenomenon corresponds to qualitative data (for example, symbol data). It becomes.
  • T X ⁇ S calculated by the following equation is used. You may use as an influence parameter
  • the associating unit 24 specifically specifies, for example, p (s t
  • the calculation formula shown above is a specific example of the calculation formula for the influence index 52.
  • the calculation formula of the influence index 52 may be any expression that calculates the influence index 52 that can mutually evaluate the strength of the influence by comparing the values, and is not limited to the above calculation formula.
  • the phenomenon selection unit 30 selects at least one phenomenon from a plurality of influence candidate phenomena based on the influence index 52.
  • the phenomenon selection unit 30 specifically includes, for example, an influence candidate phenomenon ID 58 and a modeling target phenomenon ID 60 included in the influence index data 56 including the influence index 52 larger than a predetermined value. Select a combination.
  • the influence indexes 52 that can be compared with each other are calculated based on various calculation formulas. Further, in the present embodiment, regardless of whether the value of the influence candidate phenomenon is qualitative data or quantitative data, and whether the value of the modeling target phenomenon is qualitative data or quantitative data, A unified influence index 52 is calculated. Therefore, it is possible to uniformly evaluate the influence between phenomena without being conscious of the distinction between qualitative data and quantitative data.
  • the influence index 52 obtained in the present embodiment is a value supported by information theory. Further, in the present embodiment, even when continuous quantitative data is handled, it can be handled without performing quantization.
  • the correspondence rule data generation unit 32 determines the value and model of the phenomenon selected by the phenomenon selection unit 30 based on the time series value of the phenomenon selected by the phenomenon selection unit 30 and the time series value of the phenomenon to be modeled. Correspondence rule data indicating a correspondence rule with the value of the phenomenon to be normalized is generated.
  • the correspondence rule data generation unit 32 may generate correspondence rule data indicating coefficients, parameters, and the like in a predetermined function.
  • the correspondence rule data generation unit 32 specifically includes, for example, correspondence rule data indicating a model expression represented by a linear combination of model expressions of the mixed regression model, and a model expression of the transition probability model described above.
  • the correspondence rule data generation unit 32 may generate the correspondence rule data based on the basic correspondence rule data 54 associated with the phenomenon selected by the phenomenon selection unit 30. Specifically, for example, correspondence rule data indicating a relational expression (for example, a function) obtained by combining (for example, linear combination) a relational expression (for example, a function) indicated by the basic correspondence rule data 54 may be generated. .
  • Corresponding rule data indicates, for example, a relational expression (prediction formula) indicating the relationship between the value of the phenomenon selected by the phenomenon selection unit 30 and the value of the modeling target phenomenon.
  • the predicted value calculation unit 34 calculates the predicted value of the modeling target phenomenon based on the time series value of the phenomenon selected by the phenomenon selection unit 30 and the time series value of the modeling target phenomenon.
  • the predicted value calculation unit 34 calculates the predicted value of the phenomenon to be modeled based on the corresponding rule (for example, prediction formula) indicated by the corresponding rule data generated by the corresponding rule data generating unit 32. Also good. More specifically, the predicted value calculation unit 34 predicts a value of x t + 1 from x t (k) and y t (l) based on, for example, a prediction formula indicated by the corresponding rule data.
  • prediction time data indicating a time point to be predicted is stored in the storage unit 14 in advance. Then, for example, the predicted value calculation unit 34 acquires the predicted time point data, and calculates the predicted value at the predicted time point indicated by the predicted time point data.
  • the formula used for calculating the influence index 52 can be used as it is for the prediction of the phenomenon to be modeled.
  • the predicted value calculation unit 34 may calculate a predicted value of quantitative data or a predicted value of qualitative data.
  • the image output unit 36 outputs (displays output) an image based on the selection result by the phenomenon selection unit 30 to the UI unit 16 such as a display, for example.
  • the image output unit 36 based on, for example, the calculated predicted value, the relational expression (prediction formula) indicated by the corresponding rule data, the influence index data 56 selected by the phenomenon selection unit 30, and the like 6 is generated and output to the UI unit 16 such as a display.
  • a phenomenon image 64 showing a phenomenon corresponding to the phenomenon unit data 38, an arrow image 66 shown between phenomena having a strong influence, and a point to be predicted are displayed.
  • the result display screen 62 may include an image showing a relational expression (prediction formula).
  • the associating unit 24 checks whether or not the influence index data 56 has been generated for all ordered combinations of the phenomenon unit data 38 included in the phenomenon data stored in the phenomenon data storage unit 20 (S101). If it has been generated (S101: Y), the processing of this processing example is terminated.
  • the associating unit 24 selects an ordered combination of the phenomenon unit data 38 for which the influence index data 56 has not yet been generated, and one of them is the phenomenon unit data corresponding to the influence candidate phenomenon. 38, and the other is the phenomenon unit data 38 corresponding to the phenomenon to be modeled (S102).
  • the calculation rule data selection unit 28 acquires the value of the phenomenon type data 44 included in the phenomenon unit data 38 corresponding to the influence candidate phenomenon and the phenomenon unit data 38 corresponding to the modeling target phenomenon (S103). At this time, specifically, for example, as described above, the calculation rule data selection unit 28 acquires a value of “1” when the phenomenon unit data 38 corresponds to the qualitative data. If the unit data 38 corresponds to quantitative data, a value of “0” is acquired.
  • the calculation rule data selection unit 28 sets the value of the phenomenon type data 44 included in the phenomenon unit data 38 corresponding to the influence candidate phenomenon and the phenomenon type data 44 included in the phenomenon unit data 38 corresponding to the modeling target phenomenon.
  • One calculation rule data is selected from a plurality of calculation rule data stored in the calculation rule data storage unit 26 based on a combination of values (in this embodiment, there are four patterns). (S104).
  • the associating unit 24 indicates the time series value indicated by the value data 50 included in the phenomenon unit data 38 corresponding to the influence candidate phenomenon. Then, an influence index 52 based on the time series value indicated by the value data 50 included in the phenomenon unit data 38 corresponding to the phenomenon to be modeled is calculated (S105).
  • the association unit 24 calculates the calculation rule (calculation formula) indicated by the calculation rule data selected in the process exemplified in S104, and the time series value indicated by the value data 50 included in the phenomenon unit data 38 corresponding to the influence candidate phenomenon
  • the basic correspondence rule data 54 is generated based on the time-series values indicated by the value data 50 included in the phenomenon unit data 38 corresponding to the phenomenon to be modeled (S106).
  • the associating unit 24 selects the effect candidate phenomenon ID 58 corresponding to the value of the phenomenon ID 40 included in the phenomenon unit data 38 corresponding to the influence candidate phenomenon, and the value of the phenomenon ID 40 included in the phenomenon unit data 38 corresponding to the modeled phenomenon.
  • the impact index data 56 including the modeling target phenomenon ID 60 corresponding to, the impact index 52 calculated in the process exemplified in S105, and the basic correspondence rule data 54 generated in the process exemplified in S106 is generated (S107).
  • the present embodiment it is possible to predict the value of a phenomenon to be predicted using the time series value of the phenomenon selected based on the strength of the influence. Specifically, for example, it is possible to perform future prediction of a phenomenon to be modeled with high efficiency and high prediction ability by using time series values of a phenomenon that is strongly influenced by causal effects.
  • the present embodiment can be applied in various fields such as marketing, psychology, economics, physiology, medicine, and engineering.
  • the present embodiment can also be applied to estimation of neural activity, estimation of relevance between human behavior and robot behavior and external sensors, and the like. Therefore, it is expected that this embodiment will be useful for production of intelligent systems that provide services to humans and intelligent robots that behave in the same manner as humans depending on the situation.
  • the present invention is not limited to the above embodiment.
  • the calculation formula indicated by the above calculation rule data is derived on the assumption of quantitative data that is continuous value data, but the present invention is applied to quantitative data that is discrete value data. But of course.
  • the phenomenon selection unit 30 selects the influence index data 56 based on a linear regression model or a mixed regression model.
  • the influence index data 56 may be selected using a kernel method or nonlinear feature extraction.
  • the influence index data 56 may be selected using PEV (Polynomial Embedding Vector) for the time series value having nonlinearity. That is, the present invention is applicable to both time series values assumed to be linear and time series values assumed to be non-linear.
  • some of the phenomenon data storage unit 20 and the calculation rule data storage unit 26 are provided in another computer outside the information processing apparatus 10 and communicate with the information processing apparatus 10 via a communication unit included in the information processing apparatus 10.
  • the present invention may be applied to a distributed information processing system configured as described above.
  • the information processing apparatus 10 may be configured by a single casing or a plurality of casings.

Abstract

Provided is an information processing device which can model a phenomenon by using a time-series value of the phenomenon selected in accordance with the intensity of the affects received.  A phenomenon data acquisition unit (22) acquires prediction object phenomenon data containing a time-series value of the phenomenon to be modeled and affect candidate phenomenon data containing a time-series value of each of the affect candidate phenomena.  For each of the affect candidate phenomena, a correlation unit (24) calculates an affect index in accordance with the relationship between the time-series value of the affect candidate phenomenon and the time-series value of the phenomenon to be modeled and correlates the index with the affect candidate phenomenon.  A phenomenon selection unit (30) selects at least one phenomenon in accordance with the affect index.  A correspondence rule data generation unit (32) generates correspondence rule data indicating a rule of correspondence between the value of the phenomenon selected by the phenomenon selection unit (30) and the value of the phenomenon to be modeled in accordance with the time-series value of the phenomenon selected by the phenomenon selection unit (30) and the time-series value of the phenomenon to be modeled.

Description

情報処理装置、情報処理方法、情報記憶媒体及びプログラムInformation processing apparatus, information processing method, information storage medium, and program
　本発明は、情報処理装置、情報処理方法、情報記憶媒体及びプログラムに関する。 The present invention relates to an information processing apparatus, an information processing method, an information storage medium, and a program.
　ある現象が他の現象へ与える影響の強さを示す指標を算出する方法が知られている。非特許文献１には、とりうる値が離散的である現象間について、ある現象の値が他の現象の値に与えている情報量に基づいて上述の指標を算出する方法が開示されている。非特許文献２には、とりうる値が連続的である現象間について、上述の指標を算出できるよう非特許文献１に開示されている方法を拡張する方法が開示されている。なお、非特許文献１や非特許文献２に開示されている方法は、情報理論に裏付けられた方法である。 There is a known method for calculating an index indicating the strength of the influence of a certain phenomenon on other phenomena. Non-Patent Document 1 discloses a method for calculating the above-mentioned index between phenomena whose possible values are discrete based on the amount of information given by the value of a certain phenomenon to the value of another phenomenon. . Non-Patent Document 2 discloses a method of extending the method disclosed in Non-Patent Document 1 so that the above-described index can be calculated between phenomena where possible values are continuous. Note that the methods disclosed in Non-Patent Document 1 and Non-Patent Document 2 are methods supported by information theory.
　上述の従来の方法などによれば、ある現象が他の現象へ与える影響の強さを示す指標を算出することができる。そして、この指標を用いて、例えば、現象間の因果関係などの、現象間の関係を発見することができる。しかしながら、従来の方法では、ある現象と、その現象が影響を強く与える現象との間の具体的な関係やダイナミクスを特定することができなかった。そのため、従来の方法を現象の値のモデル化に活かすことは困難であった。 According to the conventional method described above, it is possible to calculate an index indicating the strength of the influence of a certain phenomenon on other phenomena. Then, using this index, a relationship between phenomena such as a causal relationship between phenomena can be found. However, in the conventional method, a specific relationship and dynamics between a certain phenomenon and a phenomenon that the phenomenon strongly influences cannot be specified. For this reason, it has been difficult to apply the conventional method to modeling the value of a phenomenon.
　本発明は上記課題に鑑みてなされたものであって、受ける影響の強さに基づいて選択される現象の時系列値を用いて現象のモデル化を行うことができる情報処理装置、情報処理方法、情報記憶媒体及びプログラムを提供することを目的とする。 The present invention has been made in view of the above problems, and is an information processing apparatus and information processing method capable of modeling a phenomenon using a time series value of the phenomenon selected based on the strength of the influence An object of the present invention is to provide an information storage medium and a program.
　上記課題を解決するために、本発明に係る情報処理装置は、モデル化の対象となるモデル化対象現象の時系列値を含むモデル化対象現象データと、前記モデル化対象現象に影響を与える候補となる複数の影響候補現象それぞれの時系列値を含む影響候補現象データと、を取得する現象データ取得手段と、前記各影響候補現象について、当該影響候補現象の時系列値と、前記モデル化対象現象の時系列値と、の関係に基づいて、当該影響候補現象が前記モデル化対象現象に与える影響の強さを示す影響指標を算出して、当該影響候補現象に関連付ける影響指標関連付け手段と、前記各影響候補現象に関連付けられる影響指標に基づいて、前記複数の影響候補現象のうちから少なくとも１つの現象を選択する現象選択手段と、前記現象選択手段により選択される現象の時系列値と、前記モデル化対象現象の時系列値と、に基づいて、前記現象選択手段により選択される現象の値と前記モデル化対象現象の値との対応規則を示す対応規則データを生成する対応規則データ生成手段と、を含むことを特徴とする。 In order to solve the above problem, an information processing apparatus according to the present invention includes modeling target phenomenon data including a time series value of a modeling target phenomenon to be modeled, and candidates that affect the modeling target phenomenon. A phenomenon data acquisition means for acquiring influence candidate phenomenon data including time series values of each of a plurality of influence candidate phenomena, a time series value of the influence candidate phenomenon for each of the influence candidate phenomena, and the modeling target An influence index associating means for calculating an influence index indicating the strength of the influence of the influence candidate phenomenon on the modeling target phenomenon based on the relationship with the time series value of the phenomenon, and associating with the influence candidate phenomenon; A phenomenon selection means for selecting at least one phenomenon from the plurality of influence candidate phenomena based on an influence index associated with each of the influence candidate phenomena; and the phenomenon selection means. Based on the time series value of the phenomenon to be selected and the time series value of the phenomenon to be modeled, a correspondence rule between the value of the phenomenon selected by the phenomenon selecting means and the value of the modeled phenomenon And correspondence rule data generating means for generating corresponding rule data to be shown.
　また、本発明に係る情報処理方法は、モデル化の対象となるモデル化対象現象の時系列値を含むモデル化対象現象データと、前記モデル化対象現象に影響を与える候補となる複数の影響候補現象それぞれの時系列値を含む影響候補現象データと、を取得する現象データ取得ステップと、前記各影響候補現象について、当該影響候補現象の時系列値と、前記モデル化対象現象の時系列値と、の関係に基づいて、当該影響候補現象が前記モデル化対象現象に与える影響の強さを示す影響指標を算出して、当該影響候補現象に関連付ける影響指標関連付けステップと、前記各影響候補現象に関連付けられる影響指標に基づいて、前記複数の影響候補現象のうちから少なくとも１つの現象を選択する現象選択ステップと、前記現象選択手段により選択される現象の時系列値と、前記モデル化対象現象の時系列値と、に基づいて、前記現象選択手段により選択される現象の値と前記モデル化対象現象の値との対応規則を示す対応規則データを生成する対応規則データ生成ステップと、を含むことを特徴とする。 In addition, the information processing method according to the present invention includes modeling target phenomenon data including a time series value of a modeling target phenomenon to be modeled, and a plurality of influence candidates that are candidates for influencing the modeling target phenomenon. A phenomenon data acquisition step of acquiring influence candidate phenomenon data including a time series value of each phenomenon; for each of the influence candidate phenomena, a time series value of the influence candidate phenomenon and a time series value of the modeling target phenomenon; Based on the relationship, the influence index indicating the strength of the influence of the candidate influence phenomenon on the modeling target phenomenon is calculated, and the influence index associating step for associating with the influence candidate phenomenon, A phenomenon selection step of selecting at least one phenomenon from the plurality of influence candidate phenomena based on the associated influence index; and the phenomenon selection means. A correspondence rule indicating a correspondence rule between the value of the phenomenon selected by the phenomenon selecting unit and the value of the modeled phenomenon based on the time series value of the phenomenon to be modeled and the time series value of the modeled phenomenon And a corresponding rule data generation step for generating rule data.
　また、本発明に係る情報記憶媒体は、モデル化の対象となるモデル化対象現象の時系列値を含むモデル化対象現象データと、前記モデル化対象現象に影響を与える候補となる複数の影響候補現象それぞれの時系列値を含む影響候補現象データと、を取得する現象データ取得手段、前記各影響候補現象について、当該影響候補現象の時系列値と、前記モデル化対象現象の時系列値と、の関係に基づいて、当該影響候補現象が前記モデル化対象現象に与える影響の強さを示す影響指標を算出して、当該影響候補現象に関連付ける影響指標関連付け手段、前記各影響候補現象に関連付けられる影響指標に基づいて、前記複数の影響候補現象のうちから少なくとも１つの現象を選択する現象選択手段、前記現象選択手段により選択される現象の時系列値と、前記モデル化対象現象の時系列値と、に基づいて、前記現象選択手段により選択される現象の値と前記モデル化対象現象の値との対応規則を示す対応規則データを生成する対応規則データ生成手段、としてコンピュータを機能させることを特徴とするプログラムが記憶されたコンピュータ読み取り可能な情報記憶媒体である。 The information storage medium according to the present invention includes modeling target phenomenon data including a time series value of a modeling target phenomenon to be modeled, and a plurality of influence candidates that are candidates for influencing the modeling target phenomenon. Effect candidate phenomenon data including time series values of each phenomenon, phenomenon data acquisition means for acquiring, for each of the effect candidate phenomena, the time series value of the effect candidate phenomenon, the time series value of the modeling target phenomenon, Based on the relationship, an influence index indicating the strength of the influence of the candidate effect phenomenon on the modeling target phenomenon is calculated, and an influence index associating means for associating with the influence candidate phenomenon is associated with each influence candidate phenomenon. A phenomenon selection means for selecting at least one phenomenon from the plurality of influence candidate phenomena based on an influence index; a timeline of the phenomenon selected by the phenomenon selection means; Correspondence that generates correspondence rule data indicating a correspondence rule between the value of the phenomenon selected by the phenomenon selection means and the value of the phenomenon to be modeled based on the value and the time series value of the phenomenon to be modeled A computer-readable information storage medium storing a program characterized by causing a computer to function as rule data generation means.
　また、本発明に係るプログラムは、モデル化の対象となるモデル化対象現象の時系列値を含むモデル化対象現象データと、前記モデル化対象現象に影響を与える候補となる複数の影響候補現象それぞれの時系列値を含む影響候補現象データと、を取得する現象データ取得手段、前記各影響候補現象について、当該影響候補現象の時系列値と、前記モデル化対象現象の時系列値と、の関係に基づいて、当該影響候補現象が前記モデル化対象現象に与える影響の強さを示す影響指標を算出して、当該影響候補現象に関連付ける影響指標関連付け手段、前記各影響候補現象に関連付けられる影響指標に基づいて、前記複数の影響候補現象のうちから少なくとも１つの現象を選択する現象選択手段、前記現象選択手段により選択される現象の時系列値と、前記モデル化対象現象の時系列値と、に基づいて、前記現象選択手段により選択される現象の値と前記モデル化対象現象の値との対応規則を示す対応規則データを生成する対応規則データ生成手段、としてコンピュータを機能させることを特徴とする。 The program according to the present invention includes modeling target phenomenon data including time series values of a modeling target phenomenon to be modeled, and a plurality of influence candidate phenomena that are candidates for affecting the modeling target phenomenon. The effect candidate phenomenon data including the time series value of the phenomenon candidate data acquisition means for obtaining the relationship between the time series value of the candidate effect phenomenon and the time series value of the phenomenon to be modeled Based on the above, the influence index indicating the strength of the influence of the influence candidate phenomenon on the modeling target phenomenon is calculated, and the influence index associating means associated with the influence candidate phenomenon, the influence index associated with each of the influence candidate phenomena A phenomenon selection means for selecting at least one phenomenon from among the plurality of influence candidate phenomena, and a time series of the phenomena selected by the phenomenon selection means And a correspondence rule for generating correspondence rule data indicating a correspondence rule between the value of the phenomenon selected by the phenomenon selection unit and the value of the phenomenon to be modeled based on the time series value of the phenomenon to be modeled A computer is made to function as data generation means.
　本発明では、影響候補現象の時系列値と、モデル化対象現象の時系列値と、の関係に基づいて、影響候補現象がモデル化対象現象に与える影響の強さを示す影響指標を算出する。そして、影響指標に基づいて複数の影響候補現象のうちから少なくとも１つの影響候補現象を選択する。そして、選択された現象の時系列値と、モデル化対象現象の時系列値と、に基づいて、選択された現象の値とモデル化対象現象との値との対応規則を示す対応規則データを生成する。そのため、本発明によると、受ける影響の強さに基づいて選択される現象の時系列値を用いて、対応規則データによる現象のモデル化を行うことができる。 In the present invention, based on the relationship between the time series value of the influence candidate phenomenon and the time series value of the modeling target phenomenon, the influence index indicating the strength of the influence of the influence candidate phenomenon on the modeling target phenomenon is calculated. . Then, at least one influence candidate phenomenon is selected from a plurality of influence candidate phenomena based on the influence index. Based on the time series value of the selected phenomenon and the time series value of the phenomenon to be modeled, correspondence rule data indicating a correspondence rule between the value of the selected phenomenon and the value of the modeled phenomenon is obtained. Generate. Therefore, according to the present invention, it is possible to model a phenomenon using correspondence rule data by using a time series value of a phenomenon selected based on the strength of the influence.
　本発明の一態様では、前記影響指標関連付け手段が、質的データである前記影響候補現象の時系列値と、量的データである前記モデル化対象現象の時系列値と、の関係に基づいて、前記影響指標を算出することを特徴とする。こうすれば、質的データである現象が量的データである現象に与える影響の強さを評価することができる。 In one aspect of the present invention, the influence index association means is based on a relationship between a time series value of the influence candidate phenomenon that is qualitative data and a time series value of the modeling target phenomenon that is quantitative data. The influence index is calculated. In this way, it is possible to evaluate the strength of the influence of the phenomenon that is qualitative data on the phenomenon that is quantitative data.
　また、本発明の一態様では、前記影響指標関連付け手段が、量的データである前記影響候補現象の時系列値と、質的データである前記モデル化対象現象の時系列値と、の関係に基づいて、前記影響指標を算出することを特徴とする。こうすれば、量的データである現象が質的データである現象に与える影響の強さを評価することができる。 Further, in one aspect of the present invention, the influence index associating means has a relationship between a time series value of the influence candidate phenomenon that is quantitative data and a time series value of the modeling target phenomenon that is qualitative data. Based on this, the influence index is calculated. In this way, it is possible to evaluate the strength of the influence of the phenomenon that is quantitative data on the phenomenon that is qualitative data.
　また、本発明の一態様では、前記各影響候補現象について、当該影響候補現象の時系列値と、前記モデル化対象現象の時系列値と、の関係に基づいて、前記影響候補現象の値と前記モデル化対象現象の値との対応規則を示す基礎対応規則データを生成し、前記影響候補現象に関連付ける基礎対応規則データ関連付け手段、をさらに含み、前記対応規則データ生成手段が、前記現象選択手段により選択される現象に関連付けられている前記基礎対応規則データに基づいて、前記対応規則データを生成することを特徴とする。こうすれば、対応規則データを生成する際に、選択される各現象の値とモデル化対象現象の値との対応規則を示す基礎対応規則データを用いることができる。 In one aspect of the present invention, for each candidate effect phenomenon, the value of the candidate effect phenomenon is based on the relationship between the time series value of the candidate effect phenomenon and the time series value of the modeling target phenomenon. Basic correspondence rule data indicating a correspondence rule with the value of the modeling target phenomenon is generated, and further includes basic correspondence rule data associating means for associating with the influence candidate phenomenon, and the correspondence rule data generation means is the phenomenon selection means. The correspondence rule data is generated based on the basic correspondence rule data associated with the phenomenon selected by the above. In this way, when the correspondence rule data is generated, the basic correspondence rule data indicating the correspondence rule between the value of each selected phenomenon and the value of the modeling target phenomenon can be used.
　また、本発明の一態様では、前記対応規則データが示す対応規則に従って、所与の時点における前記モデル化対象現象の予測値を算出する予測値算出手段、をさらに含むことを特徴とする。こうすれば、モデル化対象現象の値の予測を行うことができる。 Further, according to an aspect of the present invention, the apparatus further includes a predicted value calculation unit that calculates a predicted value of the modeling target phenomenon at a given time in accordance with the corresponding rule indicated by the corresponding rule data. In this way, the value of the phenomenon to be modeled can be predicted.
　また、本発明の一態様では、前記影響指標を算出する規則を示す算出規則データを複数記憶する算出規則データ記憶手段と、前記影響候補現象の時系列値と、前記モデル化対象現象の時系列値と、の関係に基づいて、複数の前記算出規則データのうちから少なくとも１つの算出規則データを選択する算出規則データ選択手段と、をさらに含み、前記影響指標関連付け手段が、前記算出規則データ選択手段により選択される算出規則データが示す規則に従って前記影響指標を算出して、前記影響候補現象に関連付けることを特徴とする。こうすれば、影響指標を算出する際に、影響候補現象の時系列値と、モデル化対象現象の時系列値と、の関係に応じた算出規則を用いることができる。 In one aspect of the present invention, calculation rule data storage means for storing a plurality of calculation rule data indicating rules for calculating the influence index, a time series value of the influence candidate phenomenon, and a time series of the modeling target phenomenon And calculation rule data selection means for selecting at least one calculation rule data from a plurality of the calculation rule data based on the relationship between the values, and the influence index associating means includes the calculation rule data selection The influence index is calculated according to a rule indicated by calculation rule data selected by the means, and is associated with the influence candidate phenomenon. In this way, when calculating the influence index, a calculation rule according to the relationship between the time series value of the influence candidate phenomenon and the time series value of the modeling target phenomenon can be used.
　この態様では、前記各影響候補現象が、当該影響候補現象の値が質的データであるか量的データであるかを示す現象種別データに関連付けられており、前記算出規則データ選択手段が、前記影響候補現象に関連付けられている現象種別データに基づいて算出規則データを選択するようにしてもよい。こうすれば、影響指標を算出する際に、影響候補現象の値が質的データであるか量的データであるかに応じた算出規則を用いることができる。 In this aspect, each of the influence candidate phenomena is associated with phenomenon type data indicating whether the value of the influence candidate phenomenon is qualitative data or quantitative data, and the calculation rule data selection means includes Calculation rule data may be selected based on phenomenon type data associated with an influence candidate phenomenon. In this way, when calculating the influence index, a calculation rule according to whether the value of the influence candidate phenomenon is qualitative data or quantitative data can be used.
　また、本発明の一態様では、前記対応規則データ生成手段が、予め定められた関数の係数を示す対応規則データを生成することを特徴とする。こうすれば、対応規則データを生成する際に、関数については決定する必要がないので、対応規則データの生成が容易となる。 Further, according to an aspect of the present invention, the correspondence rule data generation unit generates correspondence rule data indicating a coefficient of a predetermined function. In this way, since it is not necessary to determine the function when generating the correspondence rule data, the generation of the correspondence rule data is facilitated.
　また、本発明の一態様では、前記影響指標が、前記影響候補現象が前記予測対象現象に与える因果的な影響の強さを示すことを特徴とする。こうすれば、受ける因果的な影響の強さに基づいて選択される現象の時系列値を用いて予測の対象となる現象の値を予測することができる。 Further, according to an aspect of the present invention, the influence index indicates a causal influence strength that the influence candidate phenomenon has on the prediction target phenomenon. In this way, it is possible to predict the value of the phenomenon to be predicted using the time series value of the phenomenon selected based on the strength of the causal influence received.
　また、本発明の一態様では、前記現象選択手段による選択結果に基づく画像を出力する画像出力手段、をさらに含むことを特徴とする。こうすれば、ユーザが、現象選択手段による選択結果を画像として得ることができる。 Further, according to an aspect of the present invention, the image processing device further includes image output means for outputting an image based on a selection result by the phenomenon selection means. In this way, the user can obtain the selection result by the phenomenon selection means as an image.
本発明の一実施形態に係る情報処理装置のハードウェア構成の一例を示す図である。It is a figure which shows an example of the hardware constitutions of the information processing apparatus which concerns on one Embodiment of this invention. 本発明の一実施形態に係る情報処理装置の機能ブロックの一例を示す図である。It is a figure which shows an example of the functional block of the information processing apparatus which concerns on one Embodiment of this invention. 現象単位データのデータ構造の一例を示す図である。It is a figure which shows an example of the data structure of phenomenon unit data. 時系列値データのデータ構造の一例を示す図である。It is a figure which shows an example of the data structure of time series value data. 影響指標データのデータ構造の一例を示す図である。It is a figure which shows an example of the data structure of influence parameter | index data. 結果表示画面の一例を示す図である。It is a figure which shows an example of a result display screen. 本実施形態に係る情報処理装置で行われる処理の流れの一例を示すフロー図である。It is a flowchart which shows an example of the flow of the process performed with the information processing apparatus which concerns on this embodiment.
　以下、本発明の実施形態について図面に基づき詳細に説明する。 Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
　図１のハードウェア構成図に例示するように、本実施形態に係る情報処理装置１０は、制御部１２、記憶部１４、ユーザインタフェース（ＵＩ）部１６を含んで構成される。制御部１２、記憶部１４、ＵＩ部１６、は、バス１８を介して接続される。 1, the information processing apparatus 10 according to the present embodiment includes a control unit 12, a storage unit 14, and a user interface (UI) unit 16, as illustrated in the hardware configuration diagram of FIG. The control unit 12, the storage unit 14, and the UI unit 16 are connected via a bus 18.
　制御部１２はＣＰＵ等のプログラム制御デバイスであり、情報処理装置１０にインストールされるプログラムに従って動作する。 The control unit 12 is a program control device such as a CPU, and operates according to a program installed in the information processing apparatus 10.
　記憶部１４は、ＲＡＭ等の記憶素子やハードディスクなどである。記憶部１４には、制御部１２によって実行されるプログラムなどが記憶される。また、記憶部１４は、制御部１２のワークメモリとしても動作する。 The storage unit 14 is a storage element such as a RAM or a hard disk. The storage unit 14 stores a program executed by the control unit 12. The storage unit 14 also operates as a work memory for the control unit 12.
　ＵＩ部１６は、ディスプレイ、マイク、スピーカー、ボタンなどであり、利用者が行った操作の内容や、利用者が入力した音声を制御部１２に出力する。また、ＵＩ部１６は、制御部１２から入力される指示に従って情報を表示出力したり音声出力したりする。 The UI unit 16 includes a display, a microphone, a speaker, a button, and the like, and outputs the content of the operation performed by the user and the voice input by the user to the control unit 12. In addition, the UI unit 16 displays and outputs information according to an instruction input from the control unit 12.
　図２は、本実施形態に係る情報処理装置１０により実現される機能の一例を示す機能ブロック図である。図２に例示するように、本実施形態に係る情報処理装置１０は、機能的には、現象データ記憶部２０、現象データ取得部２２、関連付け部２４、算出規則データ記憶部２６、算出規則データ選択部２８、現象選択部３０、対応規則データ生成部３２、予測値算出部３４、画像出力部３６、を含むものとして機能する。これらの要素のうち、現象データ記憶部２０、算出規則データ記憶部２６、は記憶部１４を主として実現される。これらの要素のうち、現象データ取得部２２、関連付け部２４、算出規則データ選択部２８、現象選択部３０、対応規則データ生成部３２、予測値算出部３４、画像出力部３６、は、制御部１２を主として実現される。これらの要素は、コンピュータである情報処理装置１０にインストールされたプログラムを、情報処理装置１０に含まれるＣＰＵ等の制御部１２で実行することにより実現されている。なお、このプログラムは、例えば、ＣＤ－ＲＯＭ、ＤＶＤ－ＲＯＭなどのコンピュータ可読な情報伝達媒体を介して、あるいは、インターネットなどの通信ネットワークを介して情報処理装置１０に供給される。 FIG. 2 is a functional block diagram illustrating an example of functions realized by the information processing apparatus 10 according to the present embodiment. As illustrated in FIG. 2, the information processing apparatus 10 according to the present embodiment functionally includes a phenomenon data storage unit 20, a phenomenon data acquisition unit 22, an association unit 24, a calculation rule data storage unit 26, and calculation rule data. It functions as including a selection unit 28, a phenomenon selection unit 30, a corresponding rule data generation unit 32, a predicted value calculation unit 34, and an image output unit 36. Among these elements, the phenomenon data storage unit 20 and the calculation rule data storage unit 26 are realized mainly by the storage unit 14. Among these elements, the phenomenon data acquisition unit 22, the association unit 24, the calculation rule data selection unit 28, the phenomenon selection unit 30, the corresponding rule data generation unit 32, the predicted value calculation unit 34, the image output unit 36, 12 is mainly realized. These elements are realized by executing a program installed in the information processing apparatus 10 that is a computer by the control unit 12 such as a CPU included in the information processing apparatus 10. The program is supplied to the information processing apparatus 10 via a computer-readable information transmission medium such as a CD-ROM or DVD-ROM, or via a communication network such as the Internet.
　現象データ記憶部２０は、複数の現象の時系列値を含む現象データを記憶する。本実施形態では、現象データには、現象単位データ３８が複数含まれている（図３参照）。図３は、現象単位データ３８のデータ構造の一例を示す図である。ここで、現象の時系列値は、観測される気温や測定される距離やモータの回転数や投薬量などといった、自然現象に関する時系列値に限定されず、例えば、株価や売り上げなどといった経済的な現象に関する時系列値や、心理的な現象に関する時系列値などであってもよい。また、現象の時系列値は、上述のような量的データに限定されず、例えば、催し物の開催有無や、時間帯、曜日、天気、部屋番号、症状の有無、固有名詞、スイッチの状態、などの質的データであってもよい。 The phenomenon data storage unit 20 stores phenomenon data including time series values of a plurality of phenomena. In the present embodiment, the phenomenon data includes a plurality of phenomenon unit data 38 (see FIG. 3). FIG. 3 is a diagram illustrating an example of the data structure of the phenomenon unit data 38. Here, the time series value of the phenomenon is not limited to the time series value related to the natural phenomenon such as the observed temperature, the measured distance, the rotation speed of the motor, the dosage, etc. It may be a time series value related to various phenomena or a time series value related to psychological phenomena. In addition, the time series value of the phenomenon is not limited to the quantitative data as described above, for example, whether or not the event is held, time zone, day of the week, weather, room number, presence or absence of symptoms, proper noun, switch state, Or other qualitative data.
　現象単位データ３８は、本実施形態では、図３に例示するように、現象の識別子である現象ＩＤ４０、現象の名称を示す現象名称データ４２、現象単位データ３８が示す現象の値が質的データであるか量的データであるかを示す現象種別データ４４、現象の時系列値を示す時系列値データ４６、を含む。図４は、時系列値データ４６のデータ構造の一例を示す図である。図４に例示するように、時系列値データ４６は、時点を示す時点データ４８と、その時点に対応する値（例えば、観測値や測定値）を示す値データ５０と、の組合せを複数含む。本実施形態では、例えば、現象種別データ４４は、フラグであり、現象単位データ３８が示す現象の値が質的データである場合は「１」の値をとり、現象単位データ３８が示す現象の値が量的データである場合は「０」の値をとる。 In this embodiment, as shown in FIG. 3, the phenomenon unit data 38 is a phenomenon ID 40 that is an identifier of a phenomenon, a phenomenon name data 42 that indicates the name of the phenomenon, and a phenomenon value indicated by the phenomenon unit data 38 is qualitative data. Phenomenon type data 44 indicating whether the data is quantitative data or time series value data 46 indicating the time series value of the phenomenon. FIG. 4 is a diagram illustrating an example of the data structure of the time series value data 46. As illustrated in FIG. 4, the time-series value data 46 includes a plurality of combinations of time data 48 indicating a time point and value data 50 indicating a value (for example, an observed value or a measured value) corresponding to the time point. . In the present embodiment, for example, the phenomenon type data 44 is a flag. When the value of the phenomenon indicated by the phenomenon unit data 38 is qualitative data, the phenomenon type data 44 takes a value of “1” and indicates the phenomenon indicated by the phenomenon unit data 38. When the value is quantitative data, the value is “0”.
　このように、現象単位データ３８は、複数の時点それぞれに対応付けられる、現象の値を示している。そして、現象の値は、時点によって変動する確率変数として扱える値であってもよい。値データ５０が示す値が質的データである場合は、値データ５０が示す値は、例えば、ダミー変数の値などである。 Thus, the phenomenon unit data 38 indicates the value of the phenomenon associated with each of a plurality of time points. The value of the phenomenon may be a value that can be handled as a random variable that varies depending on the time. When the value indicated by the value data 50 is qualitative data, the value indicated by the value data 50 is, for example, the value of a dummy variable.
　なお、制御部１２が、マウスやキーボード等のＵＩ部１６を介して操作信号を取得して、その操作信号に基づいて現象データを生成して、現象データ記憶部２０に出力するようにしてもよい。 The control unit 12 may acquire an operation signal via the UI unit 16 such as a mouse or a keyboard, generate phenomenon data based on the operation signal, and output the phenomenon data to the phenomenon data storage unit 20. Good.
　現象データ取得部２２は、現象データを取得する。本実施形態では、現象データ取得部２２は、例えば、現象データ記憶部２０に記憶されている現象データを取得する。なお、制御部１２が、マウスやキーボード等を介して取得する操作信号に対応する現象データを生成して、現象データ取得部２２が、生成された現象データを取得するようにしてもよい。 The phenomenon data acquisition unit 22 acquires phenomenon data. In the present embodiment, the phenomenon data acquisition unit 22 acquires, for example, phenomenon data stored in the phenomenon data storage unit 20. Note that the control unit 12 may generate phenomenon data corresponding to an operation signal acquired via a mouse, a keyboard, or the like, and the phenomenon data acquisition unit 22 may acquire the generated phenomenon data.
　本実施形態では、現象データ記憶部２０に記憶されている各現象単位データ３８が示す現象は、モデル化の対象となる現象（モデル化対象現象）（より具体的には、例えば、所与の時点における予測値を算出する対象となる現象（予測対象現象））でもあり、モデル化対象現象に影響を与える候補となる現象（影響候補現象）でもある。すなわち、現象データ取得部２２は、モデル化の対象となるモデル化対象現象の時系列値と、モデル化対象現象に影響を与える候補となる複数の影響候補現象それぞれの時系列値と、を含む現象データを取得することとなる。 In the present embodiment, the phenomenon indicated by each phenomenon unit data 38 stored in the phenomenon data storage unit 20 is a phenomenon to be modeled (modeling target phenomenon) (more specifically, for example, given It is also a phenomenon (prediction target phenomenon) for which a prediction value is calculated at the time point, and is also a phenomenon (influence candidate phenomenon) that is a candidate for affecting the modeling target phenomenon. That is, the phenomenon data acquisition unit 22 includes a time series value of a modeling target phenomenon to be modeled and a time series value of each of a plurality of influence candidate phenomena that are candidates for affecting the modeling target phenomenon. The phenomenon data will be acquired.
　関連付け部２４は、各影響候補現象について、影響候補現象の時系列値と、モデル化対象現象の時系列値と、の関係に基づいて、影響候補現象がモデル化対象現象に与える影響の強さ（例えば、因果的な影響の強さ）を示す影響指標５２を算出して、影響候補現象に関連付ける（図５参照）。また、本実施形態では、関連付け部２４は、各影響候補現象について、影響候補現象の時系列値と、モデル化対象現象の時系列値と、の関係に基づいて、影響候補現象の値とモデル化対象現象の値との対応規則（具体的には、例えば、影響候補現象の値とモデル化対象現象の値との関係を示す関数などの関係式（モデル式））を示す基礎対応規則データ５４を生成して、影響候補現象に関連付ける（図５参照）。関連付け部２４は、例えば、影響指標５２の算出の過程で対応規則（例えば、関数などの関係式や、関係式の係数やパラメータなど）を推定するようにしてもよい。 The associating unit 24, for each influence candidate phenomenon, based on the relationship between the time series value of the influence candidate phenomenon and the time series value of the modeling target phenomenon, the strength of the influence of the influence candidate phenomenon on the modeling target phenomenon An influence index 52 indicating (for example, the intensity of causal influence) is calculated and associated with the influence candidate phenomenon (see FIG. 5). Further, in the present embodiment, the associating unit 24 for each candidate effect phenomenon, based on the relationship between the time series value of the effect candidate phenomenon and the time series value of the modeling target phenomenon, the value of the effect candidate phenomenon and the model Basic correspondence rule data indicating a correspondence rule (specifically, a relational expression (model expression) such as a function indicating a relation between a value of an influence candidate phenomenon and a value of a modeling target phenomenon), for example 54 is generated and associated with the influence candidate phenomenon (see FIG. 5). For example, the associating unit 24 may estimate a correspondence rule (for example, a relational expression such as a function, a coefficient or a parameter of the relational expression, etc.) in the process of calculating the influence index 52.
　関連付け部２４は、本実施形態では、例えば、影響指標データ５６を生成する（図５参照）。図５は、影響指標データ５６のデータ構造の一例を示す図である。図５に例示するように、影響指標データ５６は、影響候補現象の現象ＩＤ４０に対応する影響候補現象ＩＤ５８、モデル化対象現象の現象ＩＤ４０に対応するモデル化対象現象ＩＤ６０、上述の影響指標５２、上述の基礎対応規則データ５４、を含む。 In the present embodiment, the associating unit 24 generates, for example, influence index data 56 (see FIG. 5). FIG. 5 is a diagram illustrating an example of the data structure of the influence index data 56. As illustrated in FIG. 5, the influence index data 56 includes an influence candidate phenomenon ID 58 corresponding to the phenomenon ID 40 of the influence candidate phenomenon, a modeling target phenomenon ID 60 corresponding to the phenomenon ID 40 of the modeling target phenomenon, the above-described influence index 52, The basic correspondence rule data 54 described above is included.
　このようにして、本実施形態では、例えば、関連付け部２４が、影響候補現象とモデル化対象現象との組合せに影響指標５２を関連付ける。また、本実施形態では、例えば、関連付け部２４が、影響候補現象とモデル化対象現象との組合せに基礎対応規則データ５４を関連付ける。 Thus, in the present embodiment, for example, the associating unit 24 associates the influence index 52 with the combination of the influence candidate phenomenon and the modeling target phenomenon. In the present embodiment, for example, the associating unit 24 associates the basic correspondence rule data 54 with the combination of the influence candidate phenomenon and the modeling target phenomenon.
　本実施形態では、関連付け部２４は、現象単位データ３８の数がｎである場合には、各現象単位データ３８について、他の（ｎ－１）個の現象に対応する影響指標データ５６を生成する。すなわち、関連付け部２４は、ｎ（ｎ－１）個の影響指標データ５６を生成する。 In this embodiment, when the number of phenomenon unit data 38 is n, the associating unit 24 generates, for each phenomenon unit data 38, influence index data 56 corresponding to other (n−1) phenomena. To do. That is, the associating unit 24 generates n (n−1) pieces of influence index data 56.
　算出規則データ記憶部２６は、影響指標５２を算出する規則を示す算出規則データを複数記憶する。算出規則データ選択部２８は、影響候補現象の時系列値と、モデル化対象現象の時系列値、との関係に基づいて、複数の算出規則データのうちから少なくとも１つの算出規則データを選択する。算出規則データ選択部２８は、本実施形態では、例えば、影響候補現象が質的データに対応するか量的データに対応するか、及び、モデル化対象現象が質的データに対応するか量的データに対応するか、に基づいて、算出規則データを選択する。すなわち、算出規則データ選択部２８は、本実施形態では、例えば、影響候補現象に対応する現象単位データ３８に含まれる現象種別データ４４の値と、モデル化対象現象に対応する現象単位データ３８に含まれる現象種別データ４４の値と、の組合せに基づいて、算出規則データを選択する。 The calculation rule data storage unit 26 stores a plurality of calculation rule data indicating rules for calculating the influence index 52. The calculation rule data selection unit 28 selects at least one calculation rule data from among a plurality of calculation rule data based on the relationship between the time series value of the influence candidate phenomenon and the time series value of the modeling target phenomenon. . In this embodiment, the calculation rule data selection unit 28 quantifies, for example, whether the influence candidate phenomenon corresponds to qualitative data or quantitative data, and whether the modeling target phenomenon corresponds to qualitative data. Calculation rule data is selected based on whether it corresponds to data. That is, in this embodiment, the calculation rule data selection unit 28 uses, for example, the value of the phenomenon type data 44 included in the phenomenon unit data 38 corresponding to the influence candidate phenomenon and the phenomenon unit data 38 corresponding to the modeling target phenomenon. Calculation rule data is selected based on a combination with the value of the included phenomenon type data 44.
　そして、本実施形態では、関連付け部２４は、算出規則データ選択部２８により選択される算出規則データが示す規則に従って影響指標５２を算出する。すなわち、影響指標５２は、影響候補現象に対応する現象単位データ３８に含まれる現象種別データ４４の値と、モデル化対象現象に対応する現象単位データ３８に含まれる現象種別データ４４の値と、の組合せに対応付けられる。なお、関連付け部２４は、算出規則データに基づいて、基礎対応規則データ５４を生成しても構わない。具体的には、例えば、算出規則データが、予め定められた関数の型を示している場合に、関連付け部２４が、その関数の係数を推定してもよい。そして関連付け部２４が、係数が定められた関数を示す基礎対応規則データ５４を生成するようにしてもよい。 In this embodiment, the associating unit 24 calculates the influence index 52 according to the rule indicated by the calculation rule data selected by the calculation rule data selecting unit 28. That is, the influence index 52 includes the value of the phenomenon type data 44 included in the phenomenon unit data 38 corresponding to the influence candidate phenomenon, the value of the phenomenon type data 44 included in the phenomenon unit data 38 corresponding to the modeling target phenomenon, Are associated with the combination. Note that the associating unit 24 may generate the basic correspondence rule data 54 based on the calculation rule data. Specifically, for example, when the calculation rule data indicates a predetermined function type, the associating unit 24 may estimate the coefficient of the function. Then, the associating unit 24 may generate basic correspondence rule data 54 indicating a function for which a coefficient is determined.
　ここで、複数の算出規則データそれぞれが示す、影響指標５２を算出する規則（算出式）の具体例について説明する。上述のとおり、本実施形態では、算出規則データは、影響候補現象に対応する現象単位データ３８に含まれる現象種別データ４４の値と、モデル化対象現象に対応する現象単位データ３８に含まれる現象種別データ４４の値と、の組合せに対応付けられている。 Here, a specific example of a rule (calculation formula) for calculating the influence index 52 indicated by each of a plurality of calculation rule data will be described. As described above, in the present embodiment, the calculation rule data includes the value of the phenomenon type data 44 included in the phenomenon unit data 38 corresponding to the influence candidate phenomenon and the phenomenon included in the phenomenon unit data 38 corresponding to the modeled phenomenon. It is associated with the combination of the value of the type data 44.
　以下に示す算出式において、現象単位データ３８に含まれる値データ５０の組合せを時系列値Ｘ＝（ｘ１，ｘ２，・・・，ｘＮ）や、時系列値Ｓ＝（ｓ１，ｓ２，・・・，ｓＮ）や、時系列値Ｙ＝（ｙ１，ｙ２，・・・，ｙＮ）などのように表現する。そして、時系列値に基づく特徴量である遅延埋め込みベクトルをｘｔ （ｋ）＝（ｘｔ，ｘｔ－１，・・・，ｘｔ－ｋ＋１）Ｔや、ｓｔ （ｌ）＝（ｓｔ，ｓｔ－１，・・・，ｓｔ－ｌ＋１）Ｔや、ｙｔ （ｌ）＝（ｙｔ，ｙｔ－１，・・・，ｙｔ－ｌ＋１）Ｔなどと定義する。なお、ｋやｌの値は、埋め込み次元を表す。 In the following calculation formula, combinations of value data 50 included in the phenomenon unit data 38 are time series values X = (x 1 , x 2 ,..., X N ) or time series values S = (s 1 , s 2 ,..., s N ) and time series values Y = (y 1 , y 2 ,..., y N ). Then, when a feature based on sequence value delay embedding vector x t (k) = (x t, x t-1, ···, x t-k + 1) T or, s t (l) = ( s t, s t-1, ··· , s t-l + 1) T and, y t (l) = ( y t, y t-1, ···, is defined as such as y t-l + 1) T . Note that the values of k and l represent embedding dimensions.
　まず、影響候補現象の値及びモデル化対象現象の値が質的データである場合の影響指標５２の算出式の具体例について説明する。なお、この具体例では、影響候補現象の時系列値をＹとし、モデル化対象現象の時系列値をＸとする。 First, a specific example of the calculation formula of the influence index 52 when the value of the influence candidate phenomenon and the value of the modeling target phenomenon are qualitative data will be described. In this specific example, the time series value of the influence candidate phenomenon is Y, and the time series value of the modeling target phenomenon is X.
　本実施形態では、非特許文献１に開示されている、情報理論に基づく指標であるＴｒａｎｓｆｅｒ　Ｅｎｔｒｏｐｙを影響候補現象及びモデル化対象現象が質的データに対応する際の影響指標５２として用いる。 In this embodiment, Transfer Entropy, which is an index based on information theory, disclosed in Non-Patent Document 1, is used as an influence index 52 when an influence candidate phenomenon and a modeling target phenomenon correspond to qualitative data.
　すなわち、影響候補現象及びモデル化対象現象が質的データに対応する際の影響指標５２（ＴＹ→Ｘ）の算出式の一具体例は、次式の通りである。 That is, a specific example of the calculation formula of the influence index 52 (T Y → X ) when the influence candidate phenomenon and the modeling target phenomenon correspond to the qualitative data is as follows.
Figure JPOXMLDOC01-appb-M000001
Figure JPOXMLDOC01-appb-M000001
　ここで、ｐは遷移確率を表している。そして、この式はｐ（ｘｔ｜ｘｔ－１ （ｋ），ｙｔ－１ （ｌ））とｐ（ｘｔ｜ｘｔ－１ （ｋ））の間のＫｕｌｌｂａｃｋ－Ｌｅｉｂｌｅｒ情報量を計算するものである。すなわち、この式は、Ｙの情報を加えることでＸの遷移確率がどれほど変化するかを表現したものである。また、ＸとＹを入れ替えることにより、ＴＹ→Ｘの値が変動するので、情報の流れを区別することが可能である。また、上述の式は、エントロピー（平均情報量）Ｈ（Ｘ）を用いて次式のように表すことができる。 Here, p represents the transition probability. This equation then calculates the Kullback-Leibler information amount between p (x t | x t-1 (k) , y t-1 (l) ) and p (x t | x t-1 (k) ). To do. That is, this expression expresses how much the transition probability of X changes when Y information is added. Further, by changing X and Y, the value of TY → X varies, so that the information flow can be distinguished. The above equation can be expressed as the following equation using entropy (average information amount) H (X).
Figure JPOXMLDOC01-appb-M000002
Figure JPOXMLDOC01-appb-M000002
　なお、数式２に含まれる、Ｘｔ，Ｘｔ－１ （ｋ），Ｙｔ－１ （ｌ）はそれぞれ値データ５０から構成されたｘｔ，ｘｔ－１ （ｋ），ｙｔ－１ （ｌ）の集合を意味する。また、バツ印を丸で囲んだ記号は直積を意味する。 X t , X t-1 (k) , Y t-1 (l) included in Equation 2 are x t , x t-1 (k) , y t-1 each composed of the value data 50. It means a set of (l) . Moreover, the symbol which encloses the cross mark in a circle means a direct product.
　そして、影響候補現象の値及びモデル化対象現象の値が質的データである場合は、関連付け部２４は、具体的には、例えば、ｐ（ｘｔ｜ｘｔ－１ （ｋ），ｙｔ－１ （ｌ））の遷移確率モデルのモデル式を示す基礎対応規則データ５４を生成する。 Then, when the value of the influence candidate phenomenon and the value of the modeling target phenomenon are qualitative data, the associating unit 24 specifically includes, for example, p (x t | x t−1 (k) , y t -1 (l) ) The basic correspondence rule data 54 indicating the model expression of the transition probability model is generated.
　次に、影響候補現象の値及びモデル化対象現象の値が量的データである場合の影響指標５２の算出式の具体例について説明する。なお、この具体例でも、影響候補現象の時系列値をＹとし、モデル化対象現象の時系列値をＸとする。 Next, a specific example of the calculation formula of the influence index 52 when the value of the influence candidate phenomenon and the value of the modeling target phenomenon are quantitative data will be described. Also in this specific example, the time series value of the influence candidate phenomenon is Y, and the time series value of the modeling target phenomenon is X.
　ここでは、次式で表される自己回帰モデルを導入する。 Here, an autoregressive model represented by the following formula is introduced.
Figure JPOXMLDOC01-appb-M000003
Figure JPOXMLDOC01-appb-M000003
　なお、ｋは埋め込み次元であり、ａ∈Ｒｋは自己回帰係数ベクトルである。また、εｔ （ｘ）は平均０、分散σｘ ２の正規分布に従う誤差で、ｘｔ－１ （ｋ）の値と独立である。 Note that k is an embedding dimension and aεR k is an autoregressive coefficient vector. Also, ε t (x) is an error according to a normal distribution with an average of 0 and a variance σ x 2 and is independent of the value of x t−1 (k) .
　ここで、自己回帰係数ベクトルａと分散のσｘ ２の最尤推定値が最小二乗法により推定されることを説明する。 Here, it will be described that the maximum likelihood estimation value of the autoregressive coefficient vector a and the variance σ x 2 is estimated by the least square method.
　ｋが与えられていると仮定して、値｛ｘｔ，ｘｔ－１ （ｋ）｝が与えられたとき、自己回帰モデルの尤度Ｌ及び対数尤度ｌはａとσｘ ２の関数となり、次式のように表される。 Assuming that k is given, given the value {x t , x t−1 (k) }, the likelihood L and log likelihood l of the autoregressive model is a function of a and σ x 2 And expressed as the following equation.
Figure JPOXMLDOC01-appb-M000004
Figure JPOXMLDOC01-appb-M000004
　ここで、右辺の各項は、数式３に基づき、次式のようになる。 Here, each term on the right side is expressed by the following equation based on Equation 3.
Figure JPOXMLDOC01-appb-M000005
Figure JPOXMLDOC01-appb-M000005
　そのため、Ｎ個の値が得られたとすると、対数尤度は次式のように表される。 Therefore, assuming that N values are obtained, the log likelihood is expressed as follows.
Figure JPOXMLDOC01-appb-M000006
Figure JPOXMLDOC01-appb-M000006
　このｌ（ａ，σｘ ２）を最大とするａ及びσｘ ２を求めることにより、それぞれの最尤推定値が求められる。任意に定められた自己回帰係数ベクトルａに対して、上述の対数尤度式を最大とする分散σｘ ２を求めるためには、次式の方程式を解く必要がある。 The l (a, σ x 2) by determining the a and sigma x 2 to maximize, the respective maximum likelihood estimates obtained. In order to obtain the variance σ x 2 that maximizes the above-mentioned log-likelihood formula for an arbitrarily determined autoregressive coefficient vector a, it is necessary to solve the following equation.
Figure JPOXMLDOC01-appb-M000007
Figure JPOXMLDOC01-appb-M000007
　数式７に示す方程式を解くと、次式が得られる。 Solving the equation shown in Equation 7 gives the following equation.
Figure JPOXMLDOC01-appb-M000008
Figure JPOXMLDOC01-appb-M000008
　これを、数式６の対数尤度式に代入すると、対数尤度は自己回帰係数ベクトルａだけの関数となり、次式のように表される。 If this is substituted into the log-likelihood expression of Equation 6, the log-likelihood becomes a function of only the autoregressive coefficient vector a and is expressed as the following expression.
Figure JPOXMLDOC01-appb-M000009
Figure JPOXMLDOC01-appb-M000009
　このことから、対数尤度を最大とするためにはσｘ ２の最尤推定値を最小とすればよいことがわかる。以上のようにして、自己回帰モデルの自己回帰係数ベクトル及び分散の最尤推定値は最小二乗法によって求められることが示された。最小二乗法は、例えば、疑似逆行列を用いて実装することができる。なお、データのモデリングにおいて、例えば、赤池情報量基準（ＡＩＣ）を用いる方法により、埋め込み次元ｋを選択することができる。 From this, it can be seen that in order to maximize the log likelihood, the maximum likelihood estimate of σ x 2 may be minimized. As described above, it was shown that the autoregressive coefficient vector of the autoregressive model and the maximum likelihood estimate of the variance can be obtained by the least square method. The least squares method can be implemented using, for example, a pseudo inverse matrix. In data modeling, for example, the embedding dimension k can be selected by a method using an Akaike information criterion (AIC).
　時系列値Ｘ＝（ｘ１，ｘ２，・・・，ｘＮ）が自己回帰モデルに従うと仮定した場合、その同時確率分布は多次元正規分布となることが知られている。ここで、ｘｔ－１ （ｋ）の確率分布ｐ（ｘｔ－１ （ｋ））は次式で表される。 When it is assumed that the time series value X = (x 1 , x 2 ,..., X N ) follows an autoregressive model, it is known that the joint probability distribution is a multidimensional normal distribution. Here, x t-1 probability distribution p of the (k) (x t-1 (k)) is expressed by the following equation.
Figure JPOXMLDOC01-appb-M000010
Figure JPOXMLDOC01-appb-M000010
　また、数式３に示した自己回帰モデルを仮定すると、ｘｔの条件付き確率分布は、次式で表される。 Further, assuming an autoregressive model shown in Equation 3, the conditional probability distribution of x t is expressed by the following equation.
Figure JPOXMLDOC01-appb-M000011
Figure JPOXMLDOC01-appb-M000011
　以上の２つの式から、ｘｔとｘｔ－１ （ｋ）の直積の同時確率分布を計算すると、次式に示されるような多次元正規分布となる。 When the simultaneous probability distribution of the direct product of x t and x t−1 (k) is calculated from the above two expressions, a multidimensional normal distribution as shown in the following expression is obtained.
Figure JPOXMLDOC01-appb-M000012
Figure JPOXMLDOC01-appb-M000012
　この多次元正規分布の分散共分散行列は、次式で表される。 The variance-covariance matrix of this multidimensional normal distribution is expressed by the following equation.
Figure JPOXMLDOC01-appb-M000013
Figure JPOXMLDOC01-appb-M000013
　そして、この分散共分散行列の行列式は、次式で表される。 And the determinant of the variance-covariance matrix is expressed by the following equation.
Figure JPOXMLDOC01-appb-M000014
Figure JPOXMLDOC01-appb-M000014
　すなわち、次式が成立する。 That is, the following equation holds.
Figure JPOXMLDOC01-appb-M000015
Figure JPOXMLDOC01-appb-M000015
　そして、ｘｔとｘｔ－１ （ｋ）とｙｔ－１ （ｌ）の直積の同時確率分布について、次式の混合回帰モデルを仮定する。ここでは、εｔ （ｘ｜ｙ）の分散をσｘ｜ｙ ２とする。 Then, a mixed regression model of the following equation is assumed for the simultaneous probability distribution of the direct product of x t , x t−1 (k) and y t−1 (l) . Here, the variance of ε t (x | y) is σ x | y 2 .
Figure JPOXMLDOC01-appb-M000016
Figure JPOXMLDOC01-appb-M000016
　すると、数式１５と同様に、次式の関係が成立する。 Then, as in Equation 15, the relationship of the following equation is established.
Figure JPOXMLDOC01-appb-M000017
Figure JPOXMLDOC01-appb-M000017
　ここで、上述のＴｒａｎｓｆｅｒ　Ｅｎｔｒｏｐｙを拡張したＣｏｎｔｉｎｏｕｓ　Ｔｒａｎｓｆｅｒ　Ｅｎｔｒｏｐｙについて説明する。 Here, the Continuous Transfer Entropy, which is an extension of the above-mentioned Transfer Entropy, will be described.
　ｘｔ－１ （ｋ）が正規分布であると仮定した際のエントロピーは次式で表される。 The entropy when x t-1 (k) is assumed to be a normal distribution is expressed by the following equation.
Figure JPOXMLDOC01-appb-M000018
Figure JPOXMLDOC01-appb-M000018
　この仮定をｘｔ、及び、ｙｔ－１ （ｌ）にも適用し、数式２に代入すると、次式が導かれる。 Applying this assumption to x t and y t−1 (l) and substituting them into Equation 2, the following equation is derived.
Figure JPOXMLDOC01-appb-M000019
Figure JPOXMLDOC01-appb-M000019
　この値が、非特許文献２に開示されているＣｏｎｔｉｎｏｕｓ　Ｔｒａｎｆｅｒ　Ｅｎｔｒｏｐｙに対応する。 This value corresponds to the Continuous Transfer Entropy disclosed in Non-Patent Document 2.
　そして、数式１５及び数式１７を数式１９に代入すると、次式が導出される。 Then, substituting Formula 15 and Formula 17 into Formula 19, the following formula is derived.
Figure JPOXMLDOC01-appb-M000020
Figure JPOXMLDOC01-appb-M000020
　この式が、影響候補現象及びモデル化対象現象が量的データに対応する際の影響指標５２（ＴＹ→Ｘ）の算出式の一具体例となる。 This formula is a specific example of a calculation formula of the influence index 52 (T Y → X ) when the influence candidate phenomenon and the modeling target phenomenon correspond to the quantitative data.
　なお、誤差の分散が０になり影響指標５２が無限大に発散することを防ぐため、微小数δ（例えば、δ＝１０－３００）を用いて、次式で算出されるＴＹ→Ｘを影響候補現象及びモデル化対象現象が量的データに対応する際の影響指標５２として用いても構わない。 In order to prevent the error variance from becoming zero and the influence index 52 to diverge infinitely, using a small number δ (for example, δ = 10 −300 ), TY → X calculated by the following equation is used. You may use as an influence parameter | index 52 when an influence candidate phenomenon and a modeling object phenomenon respond | correspond to quantitative data.
Figure JPOXMLDOC01-appb-M000021
Figure JPOXMLDOC01-appb-M000021
　このように、影響候補現象の値及びモデル化対象現象の値が量的データである場合は、影響指標５２の値は、回帰モデルにおける誤差の分散（本実施形態では、分散比）に基づいている。 As described above, when the value of the influence candidate phenomenon and the value of the modeling target phenomenon are quantitative data, the value of the influence index 52 is based on the variance of the error in the regression model (in this embodiment, the variance ratio). Yes.
　そして、影響候補現象の値及びモデル化対象現象の値が量的データである場合は、関連付け部２４は、具体的には、例えば、係数が例えば最小二乗法などにより推定されている、上述の数式１６の混合回帰モデルのモデル式を示す基礎対応規則データ５４を生成する。 When the value of the influence candidate phenomenon and the value of the modeling target phenomenon are quantitative data, the associating unit 24 specifically, for example, the coefficient is estimated by, for example, the least square method, etc. Basic correspondence rule data 54 indicating the model expression of the mixed regression model of Expression 16 is generated.
　次に、影響候補現象の値が質的データ（例えば、シンボルデータ）であり、モデル化対象現象の値が量的データである場合の影響指標５２の算出式の具体例について説明する。なお、この具体例では、影響候補現象の時系列値をＳとし、モデル化対象現象の時系列値をＸとする。 Next, a specific example of the calculation formula of the influence index 52 when the value of the influence candidate phenomenon is qualitative data (for example, symbol data) and the value of the modeling target phenomenon is quantitative data will be described. In this specific example, the time series value of the influence candidate phenomenon is S, and the time series value of the modeling target phenomenon is X.
　この場合の、Ｔｒａｎｆｅｒ　Ｅｎｔｒｏｐｙは次式のように表される。 In this case, Transfer Entropy is expressed by the following equation.
Figure JPOXMLDOC01-appb-M000022
Figure JPOXMLDOC01-appb-M000022
　このとき、ｐ（ｘｔ，ｘｔ－１ （ｋ）｜ｓｔ－１ （ｌ））及びｐ（ｘｔ｜ｘｔ－１ （ｋ），ｓｔ－１ （ｌ））は、ｓｔ－１ （ｌ）を所与とした条件付き確率分布である。この条件付き確率分布に対して、次式のＳｗｉｔｃｈｉｎｇ自己回帰モデルを導入する。 At this time, p (x t , x t-1 (k) | s t-1 (l) ) and p (x t | x t-1 (k) , s t-1 (l) ) are expressed as s t −1 is a conditional probability distribution given (l) . For this conditional probability distribution, the following switching autoregressive model is introduced.
Figure JPOXMLDOC01-appb-M000023
Figure JPOXMLDOC01-appb-M000023
　ここで、ｂ（ｓｔ－１ （ｌ））は、質的データｓｔ－１ （ｌ）の状態ごとに切り替わる係数ベクトルであり、ｘｔ－１ （ｋ）をｓｔ－１ （ｌ）の状態ごとに切り分ける処理を実行した後、各々のｓｔ－１ （ｌ）に対して最小二乗法により推定される。これは、影響候補現象の値がモデル化対象現象の値に何らかの影響を与えていればｓｔ－１ （ｌ）の変化に伴って自己回帰係数ベクトルが変化し、影響をあたえていなければ自己回帰係数ベクトルは変化せず一定であるであろうというコンセプトに基づいている。 Here, b (s t-1 (l) ) is a coefficient vector that switches for each state of the qualitative data s t-1 (l) , and x t-1 (k) is changed to s t-1 (l) After performing the process of carving for each state, each s t-1 (l) is estimated by the least square method. This is because the autoregressive coefficient vector changes with the change of s t-1 (l) if the value of the influence candidate phenomenon has some influence on the value of the phenomenon to be modeled. The regression coefficient vector is based on the concept that it will not change and will be constant.
　このような自己回帰モデルについてもｓｔ－１ （ｌ）の状態ごとに切り分ければ、同時確率分布は正規分布に従う、よって、数式１９に倣って、次式のようにＴｒａｎｆｅｒ　Ｅｎｔｒｏｐｙは表される。 If such an autoregressive model is also carved for each state of s t-1 (l) , the joint probability distribution follows a normal distribution. Therefore, according to Equation 19, Transfer Entropy is expressed as the following equation: .
Figure JPOXMLDOC01-appb-M000024
Figure JPOXMLDOC01-appb-M000024
　また、係数ベクトルｂ（ｓｔ－１ （ｌ））に対応する誤差εｔ （ｘ｜ｙ）の分散をσｘ｜ｙ ２（ｓｔ－１ （ｌ））とすると、次式の関係が成立する。 If the variance of the error ε t (x | y) corresponding to the coefficient vector b (s t-1 (l) ) is σ x | y 2 (s t-1 (l) ), the relationship of To establish.
Figure JPOXMLDOC01-appb-M000025
Figure JPOXMLDOC01-appb-M000025
　従って、数式１５及び数式２５を数式２４に代入して下記の式が導出される。 Therefore, the following formula is derived by substituting Formula 15 and Formula 25 into Formula 24.
Figure JPOXMLDOC01-appb-M000026
Figure JPOXMLDOC01-appb-M000026
　この式が、影響候補現象の値が質的データ（例えば、シンボルデータ）であり、モデル化対象現象の値が量的データである場合の影響指標５２（ＴＳ→Ｘ）の算出式の一具体例となる。この影響指標５２は、時系列値Ｘをｓｔ－１ （ｌ）の状態ごとに切り分けたときにどれほど精度が向上するかを、平均ビット数で表現した指標となっている。 This equation is a formula for calculating the influence index 52 (T S → X ) when the value of the influence candidate phenomenon is qualitative data (for example, symbol data) and the value of the modeling target phenomenon is quantitative data. This is a specific example. The influence index 52 is an index expressing how much the accuracy is improved when the time series value X is divided for each state of s t−1 (l) by the average number of bits.
　なお、誤差の分散が０になり影響指標５２が無限大に発散することを防ぐため、微小数δ（例えば、δ＝１０－３００）を用いて、次式で算出されるＴＳ→Ｘを影響候補現象及びモデル化対象現象が量的データに対応する際の影響指標５２として用いても構わない。 In order to prevent the error variance from becoming 0 and the influence index 52 to diverge infinitely, using a small number δ (for example, δ = 10 −300 ), T S → X calculated by the following equation is used. You may use as an influence parameter | index 52 when an influence candidate phenomenon and a modeling object phenomenon respond | correspond to quantitative data.
Figure JPOXMLDOC01-appb-M000027
Figure JPOXMLDOC01-appb-M000027
　このように、影響候補現象の値が質的データ（例えば、シンボルデータ）であり、モデル化対象現象の値が量的データである場合は、影響指標５２の値は、回帰モデルにおける誤差の分散（本実施形態では、分散比）に基づいている。 As described above, when the value of the influence candidate phenomenon is qualitative data (for example, symbol data) and the value of the phenomenon to be modeled is quantitative data, the value of the influence index 52 is an error variance in the regression model. (In this embodiment, based on the dispersion ratio).
　そして、影響候補現象の値が質的データ（例えば、シンボルデータ）であり、モデル化対象現象の値が量的データである場合は、関連付け部２４は、具体的には、例えば、係数が例えば最小二乗法などにより推定されている、上述の数式２３のＳｗｉｔｃｈｉｎｇ自己回帰モデルのモデル式を示す基礎対応規則データ５４を生成する。 When the value of the influence candidate phenomenon is qualitative data (for example, symbol data) and the value of the modeling target phenomenon is quantitative data, the associating unit 24 specifically includes, for example, the coefficient The basic correspondence rule data 54 indicating the model expression of the Switching autoregressive model of the above-described Expression 23, which is estimated by the least square method or the like, is generated.
　次に、影響候補現象の値が量的データであり、モデル化対象現象の値が質的データ（例えば、シンボルデータ）である場合の影響指標５２の算出式の具体例について説明する。なお、この具体例では、影響候補現象の時系列値をＸとし、モデル化対象現象の時系列値をＳとする。 Next, a specific example of the calculation formula of the influence index 52 when the value of the influence candidate phenomenon is quantitative data and the value of the modeling target phenomenon is qualitative data (for example, symbol data) will be described. In this specific example, the time series value of the influence candidate phenomenon is X, and the time series value of the modeling target phenomenon is S.
　この場合の、Ｔｒａｎｆｅｒ　Ｅｎｔｒｏｐｙは次式のように表される。 In this case, Transfer Entropy is expressed by the following equation.
Figure JPOXMLDOC01-appb-M000028
Figure JPOXMLDOC01-appb-M000028
　そして、ベイズの定理により、次式の関係が成立する。 And by the Bayes' theorem, the following relationship is established.
Figure JPOXMLDOC01-appb-M000029
Figure JPOXMLDOC01-appb-M000029
　この式を用いると、Ｔｒａｎｆｅｒ　Ｅｎｔｒｏｐｙは次式のようになる。 Using this equation, Transfer Entropy is as follows:
Figure JPOXMLDOC01-appb-M000030
Figure JPOXMLDOC01-appb-M000030
　この式から、ｓｔ－１ （ｌ）のみの状態ごとにｘｔ－１ （ｋ）を切り分けた場合の確率分布と、ｓｔとｓｔ－１ （ｌ）の両方を考慮したｓｔとｓｔ－１ （ｌ）の直積の状態ごとにｘｔ－１ （ｋ）を切り分けた場合の確率分布との比較を行えばよいということがわかる。 From this equation, the probability distribution when carved x t-1 (k) for each state of s t-1 (l) only, and s t and s t-1 s t to both the consideration of (l) It can be seen that the comparison with the probability distribution when x t-1 (k) is carved for each state of the direct product of s t-1 (l) is sufficient.
　モデル化対象現象に対応する質的データの時系列値がｓｔ－１ （ｌ）であるときの、影響候補現象に対応する量的データの時系列値ｘｔ－１ （ｋ）が正規分布に従うと仮定すると、数式１８と同様に、次式の関係が成立する。 When the time series value of the qualitative data corresponding to the phenomenon to be modeled is s t-1 (l) , the time series value x t-1 (k) of the quantitative data corresponding to the influence candidate phenomenon is normally distributed. As in Equation 18, the relationship of the following equation is established.
Figure JPOXMLDOC01-appb-M000031
Figure JPOXMLDOC01-appb-M000031
　また、モデル化対象現象に対応する質的データの時系列値がｓｔ，ｓｔ－１ （ｌ）であるときの、影響候補現象に対応する量的データの時系列値ｘｔ－１ （ｋ）が正規分布に従うと仮定すると、次式の関係が成立する。 Further, when the time series value of the qualitative data corresponding to the phenomenon to be modeled is s t , s t-1 (l) , the time series value x t-1 ( Assuming that k) follows a normal distribution, the following relationship is established.
Figure JPOXMLDOC01-appb-M000032
Figure JPOXMLDOC01-appb-M000032
　そして、数式３１及び数式３２を数式３０に代入して次式が導き出される。 Then, the following equation is derived by substituting Equation 31 and Equation 32 into Equation 30.
Figure JPOXMLDOC01-appb-M000033
Figure JPOXMLDOC01-appb-M000033
　この式が、影響候補現象が量的データに対応し、モデル化対象現象が質的データ（例えば、シンボルデータ）に対応する際の影響指標５２（ＴＸ→Ｓ）の算出式の一具体例となる。 This formula is a specific example of a formula for calculating the influence index 52 (T X → S ) when the influence candidate phenomenon corresponds to quantitative data and the modeling target phenomenon corresponds to qualitative data (for example, symbol data). It becomes.
　なお、誤差の分散が０になり影響指標５２が無限大に発散することを防ぐため、微小数δ（例えば、δ＝１０－３００）を用いて、次式で算出されるＴＸ→Ｓを影響候補現象及びモデル化対象現象が量的データに対応する際の影響指標５２として用いても構わない。 In order to prevent the error variance from becoming zero and the influence index 52 to diverge infinitely, using a small number δ (for example, δ = 10 −300 ), T X → S calculated by the following equation is used. You may use as an influence parameter | index 52 when an influence candidate phenomenon and a modeling object phenomenon respond | correspond to quantitative data.
Figure JPOXMLDOC01-appb-M000034
Figure JPOXMLDOC01-appb-M000034
　そして、影響候補現象が量的データに対応し、モデル化対象現象が質的データ（例えば、シンボルデータ）に対応する場合は、関連付け部２４は、具体的には、例えば、ｐ（ｓｔ｜ｓｔ－１ （ｌ），ｘｔ－１ （ｋ））の遷移確率モデルのモデル式を示す基礎対応規則データ５４を生成する。 When the influence candidate phenomenon corresponds to quantitative data and the modeling target phenomenon corresponds to qualitative data (for example, symbol data), the associating unit 24 specifically specifies, for example, p (s t | Basic correspondence rule data 54 indicating the model expression of the transition probability model of s t-1 (l) , x t-1 (k) ) is generated.
　なお、上述に示した算出式は、影響指標５２の算出式の一具体例である。影響指標５２の算出式は、値を比較することによって影響の強さを互いに評価できる影響指標５２を算出する式であればよく、上述の算出式には限定されない。 The calculation formula shown above is a specific example of the calculation formula for the influence index 52. The calculation formula of the influence index 52 may be any expression that calculates the influence index 52 that can mutually evaluate the strength of the influence by comparing the values, and is not limited to the above calculation formula.
　現象選択部３０は、影響指標５２に基づいて、複数の影響候補現象のうちから少なくとも１つの現象を選択する。現象選択部３０は、本実施形態では、具体的には、例えば、予め定められた値よりも大きな影響指標５２が含まれる影響指標データ５６に含まれる、影響候補現象ＩＤ５８及びモデル化対象現象ＩＤ６０の組合せを選択する。 The phenomenon selection unit 30 selects at least one phenomenon from a plurality of influence candidate phenomena based on the influence index 52. In the present embodiment, the phenomenon selection unit 30 specifically includes, for example, an influence candidate phenomenon ID 58 and a modeling target phenomenon ID 60 included in the influence index data 56 including the influence index 52 larger than a predetermined value. Select a combination.
　本実施形態では、様々な算出式に基づいて互いに比較可能な影響指標５２が算出される。また、本実施形態では、影響候補現象の値が質的データであるか量的データであるか、モデル化対象現象の値が質的データであるか量的データであるか、に関わらず、統一的な影響指標５２が算出される。そのため、質的データ、量的データの区別を意識することなく、現象間の影響の評価を統一的に行うことができる。また、本実施形態で得られる影響指標５２は、情報理論に裏付けられた値となっている。また、本実施形態では、連続的な量的データを扱う場合も、量子化を行うことなく取り扱うことができる。 In the present embodiment, the influence indexes 52 that can be compared with each other are calculated based on various calculation formulas. Further, in the present embodiment, regardless of whether the value of the influence candidate phenomenon is qualitative data or quantitative data, and whether the value of the modeling target phenomenon is qualitative data or quantitative data, A unified influence index 52 is calculated. Therefore, it is possible to uniformly evaluate the influence between phenomena without being conscious of the distinction between qualitative data and quantitative data. The influence index 52 obtained in the present embodiment is a value supported by information theory. Further, in the present embodiment, even when continuous quantitative data is handled, it can be handled without performing quantization.
　対応規則データ生成部３２は、現象選択部３０により選択される現象の時系列値と、モデル化対象現象の時系列値と、に基づいて、現象選択部３０により選択される現象の値とモデル化対象現象の値との対応規則を示す対応規則データを生成する。対応規則データ生成部３２は、予め定められた関数における係数やパラメータなどを示す対応規則データを生成してもよい。本実施形態では、対応規則データ生成部３２は、具体的には、例えば、混合回帰モデルのモデル式の線形結合で表されるモデル式を示す対応規則データや、上述の遷移確率モデルのモデル式を示す対応規則データなどを生成する。 The correspondence rule data generation unit 32 determines the value and model of the phenomenon selected by the phenomenon selection unit 30 based on the time series value of the phenomenon selected by the phenomenon selection unit 30 and the time series value of the phenomenon to be modeled. Correspondence rule data indicating a correspondence rule with the value of the phenomenon to be normalized is generated. The correspondence rule data generation unit 32 may generate correspondence rule data indicating coefficients, parameters, and the like in a predetermined function. In the present embodiment, the correspondence rule data generation unit 32 specifically includes, for example, correspondence rule data indicating a model expression represented by a linear combination of model expressions of the mixed regression model, and a model expression of the transition probability model described above. Correspondence rule data indicating
　また、対応規則データ生成部３２は、現象選択部３０により選択される現象に関連付けられている基礎対応規則データ５４に基づいて、対応規則データを生成するようにしてもよい。具体的には、例えば、基礎対応規則データ５４が示す関係式（例えば、関数）を結合（例えば、線形結合）した関係式（例えば、関数）を示す対応規則データを生成するようにしてもよい。 Further, the correspondence rule data generation unit 32 may generate the correspondence rule data based on the basic correspondence rule data 54 associated with the phenomenon selected by the phenomenon selection unit 30. Specifically, for example, correspondence rule data indicating a relational expression (for example, a function) obtained by combining (for example, linear combination) a relational expression (for example, a function) indicated by the basic correspondence rule data 54 may be generated. .
　対応規則データは、例えば、現象選択部３０により選択される現象の値とモデル化対象現象の値との関係を示す関係式（予測式）を示している。 Corresponding rule data indicates, for example, a relational expression (prediction formula) indicating the relationship between the value of the phenomenon selected by the phenomenon selection unit 30 and the value of the modeling target phenomenon.
　予測値算出部３４は、現象選択部３０により選択される現象の時系列値と、モデル化対象現象の時系列値と、に基づいて、モデル化対象現象の予測値を算出する。ここで、予測値算出部３４が、対応規則データ生成部３２により生成される対応規則データが示す対応規則（例えば、予測式）に基づいて、モデル化対象現象の予測値を算出するようにしてもよい。予測値算出部３４は、より具体的には、例えば、対応規則データが示す予測式などに基づいて、ｘｔ （ｋ）と、ｙｔ （ｌ）とから、ｘｔ＋１の値を予測する。本実施形態では、例えば、予め予測の対象となる時点を示す予測時点データが記憶部１４に記憶されている。そして、予測値算出部３４は、例えば、この予測時点データを取得して、予測時点データが示す予測時点における予測値を算出する。 The predicted value calculation unit 34 calculates the predicted value of the modeling target phenomenon based on the time series value of the phenomenon selected by the phenomenon selection unit 30 and the time series value of the modeling target phenomenon. Here, the predicted value calculation unit 34 calculates the predicted value of the phenomenon to be modeled based on the corresponding rule (for example, prediction formula) indicated by the corresponding rule data generated by the corresponding rule data generating unit 32. Also good. More specifically, the predicted value calculation unit 34 predicts a value of x t + 1 from x t (k) and y t (l) based on, for example, a prediction formula indicated by the corresponding rule data. In the present embodiment, for example, prediction time data indicating a time point to be predicted is stored in the storage unit 14 in advance. Then, for example, the predicted value calculation unit 34 acquires the predicted time point data, and calculates the predicted value at the predicted time point indicated by the predicted time point data.
　このように、本実施形態では、影響指標５２の算出に用いた式をそのままモデル化対象現象の予測に活用することができる。なお、予測値算出部３４は、量的データの予測値を算出しても、質的データの予測値を算出してもよい。 Thus, in the present embodiment, the formula used for calculating the influence index 52 can be used as it is for the prediction of the phenomenon to be modeled. Note that the predicted value calculation unit 34 may calculate a predicted value of quantitative data or a predicted value of qualitative data.
　画像出力部３６は、現象選択部３０による選択結果に基づく画像を、例えば、ディスプレイなどのＵＩ部１６に出力（表示出力）する。画像出力部３６は、具体的には、例えば、算出される予測値や、対応規則データが示す関係式（予測式）や、現象選択部３０が選択する影響指標データ５６などに基づいて、図６に例示する結果表示画面６２を生成して、ディスプレイなどのＵＩ部１６に出力する。図６に例示する結果表示画面６２には、現象単位データ３８に対応する現象を示す現象画像６４と、強い影響を与えている現象間に示されている矢印画像６６と、予測対象となる時点を示す予測時点画像６８と、他の現象から強く影響を受け、予測の対象となる現象（モデル化対象現象）の予測時点における予測値を示す予測値画像７０と、が表示されている。なお、結果表示画面６２には、関係式（予測式）を示す画像が含まれていてもよい。 The image output unit 36 outputs (displays output) an image based on the selection result by the phenomenon selection unit 30 to the UI unit 16 such as a display, for example. Specifically, the image output unit 36, based on, for example, the calculated predicted value, the relational expression (prediction formula) indicated by the corresponding rule data, the influence index data 56 selected by the phenomenon selection unit 30, and the like 6 is generated and output to the UI unit 16 such as a display. In the result display screen 62 illustrated in FIG. 6, a phenomenon image 64 showing a phenomenon corresponding to the phenomenon unit data 38, an arrow image 66 shown between phenomena having a strong influence, and a point to be predicted are displayed. And a predicted value image 70 indicating a predicted value at a predicted time of a phenomenon (modeling target phenomenon) that is strongly influenced by other phenomena and that is a target of prediction is displayed. The result display screen 62 may include an image showing a relational expression (prediction formula).
　ここで、本実施形態に係る情報処理装置１０における、影響指標データ５６を生成する処理の一例を、図７に例示するフロー図を参照しながら説明する。 Here, an example of processing for generating the influence index data 56 in the information processing apparatus 10 according to the present embodiment will be described with reference to a flowchart illustrated in FIG.
　まず、関連付け部２４が、現象データ記憶部２０に記憶されている現象データに含まれる現象単位データ３８のすべての順序付き組合せについて影響指標データ５６が生成されたか否かを確認する（Ｓ１０１）。生成された場合は（Ｓ１０１：Ｙ）、本処理例の処理を終了する。 First, the associating unit 24 checks whether or not the influence index data 56 has been generated for all ordered combinations of the phenomenon unit data 38 included in the phenomenon data stored in the phenomenon data storage unit 20 (S101). If it has been generated (S101: Y), the processing of this processing example is terminated.
　生成されていない場合は（Ｓ１０１：Ｎ）、関連付け部２４は、まだ影響指標データ５６が生成されていない現象単位データ３８の順序付き組合せを選択し、一方を影響候補現象に対応する現象単位データ３８とし、他方をモデル化対象現象に対応する現象単位データ３８とする（Ｓ１０２）。 If not generated (S101: N), the associating unit 24 selects an ordered combination of the phenomenon unit data 38 for which the influence index data 56 has not yet been generated, and one of them is the phenomenon unit data corresponding to the influence candidate phenomenon. 38, and the other is the phenomenon unit data 38 corresponding to the phenomenon to be modeled (S102).
　そして、算出規則データ選択部２８が、影響候補現象に対応する現象単位データ３８とモデル化対象現象に対応する現象単位データ３８に含まれる現象種別データ４４の値を取得する（Ｓ１０３）。このとき、具体的には、例えば、上述のように、算出規則データ選択部２８は、この現象単位データ３８が質的データに対応している場合は「１」の値を取得し、この現象単位データ３８が量的データに対応している場合は「０」の値を取得する。 Then, the calculation rule data selection unit 28 acquires the value of the phenomenon type data 44 included in the phenomenon unit data 38 corresponding to the influence candidate phenomenon and the phenomenon unit data 38 corresponding to the modeling target phenomenon (S103). At this time, specifically, for example, as described above, the calculation rule data selection unit 28 acquires a value of “1” when the phenomenon unit data 38 corresponds to the qualitative data. If the unit data 38 corresponds to quantitative data, a value of “0” is acquired.
　そして、算出規則データ選択部２８が、影響候補現象に対応する現象単位データ３８に含まれる現象種別データ４４の値と、モデル化対象現象に対応する現象単位データ３８に含まれる現象種別データ４４の値と、の組合せ（本実施形態では、４通りのパターンがある。）に基づいて、算出規則データ記憶部２６に記憶されている複数の算出規則データのうちから、１つの算出規則データを選択する（Ｓ１０４）。 Then, the calculation rule data selection unit 28 sets the value of the phenomenon type data 44 included in the phenomenon unit data 38 corresponding to the influence candidate phenomenon and the phenomenon type data 44 included in the phenomenon unit data 38 corresponding to the modeling target phenomenon. One calculation rule data is selected from a plurality of calculation rule data stored in the calculation rule data storage unit 26 based on a combination of values (in this embodiment, there are four patterns). (S104).
　そして、関連付け部２４が、Ｓ１０４に例示する処理で選択された算出規則データが示す算出規則（算出式）に従って、影響候補現象に対応する現象単位データ３８に含まれる値データ５０が示す時系列値と、モデル化対象現象に対応する現象単位データ３８に含まれる値データ５０が示す時系列値と、に基づく影響指標５２を算出する（Ｓ１０５）。 Then, according to the calculation rule (calculation formula) indicated by the calculation rule data selected by the process exemplified in S104, the associating unit 24 indicates the time series value indicated by the value data 50 included in the phenomenon unit data 38 corresponding to the influence candidate phenomenon. Then, an influence index 52 based on the time series value indicated by the value data 50 included in the phenomenon unit data 38 corresponding to the phenomenon to be modeled is calculated (S105).
　そして、関連付け部２４が、Ｓ１０４に例示する処理で選択された算出規則データが示す算出規則（算出式）と、影響候補現象に対応する現象単位データ３８に含まれる値データ５０が示す時系列値と、モデル化対象現象に対応する現象単位データ３８に含まれる値データ５０が示す時系列値と、に基づいて、基礎対応規則データ５４を生成する（Ｓ１０６）。 Then, the association unit 24 calculates the calculation rule (calculation formula) indicated by the calculation rule data selected in the process exemplified in S104, and the time series value indicated by the value data 50 included in the phenomenon unit data 38 corresponding to the influence candidate phenomenon The basic correspondence rule data 54 is generated based on the time-series values indicated by the value data 50 included in the phenomenon unit data 38 corresponding to the phenomenon to be modeled (S106).
　そして、関連付け部２４が、影響候補現象に対応する現象単位データ３８に含まれる現象ＩＤ４０の値に対応する影響候補現象ＩＤ５８、モデル化対象現象に対応する現象単位データ３８に含まれる現象ＩＤ４０の値に対応するモデル化対象現象ＩＤ６０、Ｓ１０５に例示する処理で算出された影響指標５２、Ｓ１０６に例示する処理で生成された基礎対応規則データ５４、を含む影響指標データ５６を生成する（Ｓ１０７）。 Then, the associating unit 24 selects the effect candidate phenomenon ID 58 corresponding to the value of the phenomenon ID 40 included in the phenomenon unit data 38 corresponding to the influence candidate phenomenon, and the value of the phenomenon ID 40 included in the phenomenon unit data 38 corresponding to the modeled phenomenon. The impact index data 56 including the modeling target phenomenon ID 60 corresponding to, the impact index 52 calculated in the process exemplified in S105, and the basic correspondence rule data 54 generated in the process exemplified in S106 is generated (S107).
　そして、再度Ｓ１０１に例示する処理を実行する。 Then, the process exemplified in S101 is executed again.
　本実施形態によれば、受ける影響の強さに基づいて選択される現象の時系列値を用いて予測の対象となる現象の値を予測することができる。具体的には、例えば、因果的な影響を強く受ける現象の時系列値を用いて、効率的で予測能力の高い、モデル化対象現象の将来予測を行うことができる。 According to the present embodiment, it is possible to predict the value of a phenomenon to be predicted using the time series value of the phenomenon selected based on the strength of the influence. Specifically, for example, it is possible to perform future prediction of a phenomenon to be modeled with high efficiency and high prediction ability by using time series values of a phenomenon that is strongly influenced by causal effects.
　また、本実施形態は、例えば、マーケティング、心理学、経済学、生理学、医学、工学などの様々な分野において適用することができる。そして、本実施形態は、神経活動の推定や、人間の行動やロボットの行動と外部のセンサとの関連性の推定などにも適用することができる。そのため、人間へサービスを提供する知的システムや状況に応じて人間と同じ振る舞いをする知的ロボットの制作などに、本実施形態を役立てることが期待される。 Also, the present embodiment can be applied in various fields such as marketing, psychology, economics, physiology, medicine, and engineering. The present embodiment can also be applied to estimation of neural activity, estimation of relevance between human behavior and robot behavior and external sensors, and the like. Therefore, it is expected that this embodiment will be useful for production of intelligent systems that provide services to humans and intelligent robots that behave in the same manner as humans depending on the situation.
　なお、本発明は上記実施形態に限定されるものではない。 The present invention is not limited to the above embodiment.
　例えば、上記実施形態では、上述の算出規則データが示す算出式は連続値データである量的データを前提として、導出しているが、本発明を離散値データである量的データに適用しても、もちろん構わない。 For example, in the above embodiment, the calculation formula indicated by the above calculation rule data is derived on the assumption of quantitative data that is continuous value data, but the present invention is applied to quantitative data that is discrete value data. But of course.
　また、上記実施形態では、現象選択部３０は、線形モデルの自己回帰モデルや混合回帰モデルに基づいて、影響指標データ５６の選択を行っているが、非線形性を持つ時系列値に対しては、例えば、カーネル法や非線形な特徴抽出を用いて、影響指標データ５６の選択を行っても構わない。また、非線形性を持つ時系列値に対して、ＰＥＶ（Ｐｏｌｙｎｏｍｉａｌ　Ｅｍｂｅｄｄｉｎｇ　Ｖｅｃｔｏｒ）を用いて、影響指標データ５６の選択を行っても構わない。すなわち、本発明は、線形性が仮定される時系列値にも非線形性が仮定される時系列値にも適用可能である。 In the above embodiment, the phenomenon selection unit 30 selects the influence index data 56 based on a linear regression model or a mixed regression model. However, for a time series value having nonlinearity, For example, the influence index data 56 may be selected using a kernel method or nonlinear feature extraction. Further, the influence index data 56 may be selected using PEV (Polynomial Embedding Vector) for the time series value having nonlinearity. That is, the present invention is applicable to both time series values assumed to be linear and time series values assumed to be non-linear.
　また、現象データ記憶部２０、算出規則データ記憶部２６のうちのいくつかを情報処理装置１０外の別のコンピュータに設け、情報処理装置１０と、情報処理装置１０が備える通信部を介して通信する構成とした分散型情報処理システムに本発明を適用してもよい。 Further, some of the phenomenon data storage unit 20 and the calculation rule data storage unit 26 are provided in another computer outside the information processing apparatus 10 and communicate with the information processing apparatus 10 via a communication unit included in the information processing apparatus 10. The present invention may be applied to a distributed information processing system configured as described above.
　また、情報処理装置１０は、一つの筐体により構成されていても、複数の筐体により構成されていてもよい。 Further, the information processing apparatus 10 may be configured by a single casing or a plurality of casings.

Claims (13)

  1. 　モデル化の対象となるモデル化対象現象の時系列値を含むモデル化対象現象データと、前記モデル化対象現象に影響を与える候補となる複数の影響候補現象それぞれの時系列値を含む影響候補現象データと、を取得する現象データ取得手段と、
    　前記各影響候補現象について、当該影響候補現象の時系列値と、前記モデル化対象現象の時系列値と、の関係に基づいて、当該影響候補現象が前記モデル化対象現象に与える影響の強さを示す影響指標を算出して、当該影響候補現象に関連付ける影響指標関連付け手段と、
    　前記各影響候補現象に関連付けられる影響指標に基づいて、前記複数の影響候補現象のうちから少なくとも１つの現象を選択する現象選択手段と、
    　前記現象選択手段により選択される現象の時系列値と、前記モデル化対象現象の時系列値と、に基づいて、前記現象選択手段により選択される現象の値と前記モデル化対象現象の値との対応規則を示す対応規則データを生成する対応規則データ生成手段と、
    　を含むことを特徴とする情報処理装置。     Modeling target phenomenon data including a time series value of a modeling target phenomenon to be modeled, and an influence candidate phenomenon including time series values of each of a plurality of candidate effecting phenomena that are candidates for affecting the modeling target phenomenon A phenomenon data acquisition means for acquiring data;
    For each candidate effect phenomenon, the strength of the influence of the candidate effect phenomenon on the modeling target phenomenon based on the relationship between the time series value of the candidate effect phenomenon and the time series value of the modeling target phenomenon An impact index associating means for calculating an impact index indicating
    A phenomenon selecting means for selecting at least one phenomenon from the plurality of influence candidate phenomena based on an influence index associated with each of the influence candidate phenomena;
    Based on the time series value of the phenomenon selected by the phenomenon selecting means and the time series value of the modeling target phenomenon, the value of the phenomenon selected by the phenomenon selecting means and the value of the modeling target phenomenon Correspondence rule data generating means for generating correspondence rule data indicating the correspondence rules of
    An information processing apparatus comprising:
  2. 　前記影響指標関連付け手段が、質的データである前記影響候補現象の時系列値と、量的データである前記モデル化対象現象の時系列値と、の関係に基づいて、前記影響指標を算出する、
    　ことを特徴とする請求項１に記載の情報処理装置。     The influence index association means calculates the influence index based on the relationship between the time series value of the candidate effect phenomenon that is qualitative data and the time series value of the modeled phenomenon that is quantitative data. ,
    The information processing apparatus according to claim 1.
  3. 　前記影響指標関連付け手段が、量的データである前記影響候補現象の時系列値と、質的データである前記モデル化対象現象の時系列値と、の関係に基づいて、前記影響指標を算出する、
    　ことを特徴とする請求項１に記載の情報処理装置。     The influence index associating means calculates the influence index based on a relationship between a time series value of the candidate effect phenomenon that is quantitative data and a time series value of the modeling target phenomenon that is qualitative data. ,
    The information processing apparatus according to claim 1.
  4. 　前記各影響候補現象について、当該影響候補現象の時系列値と、前記モデル化対象現象の時系列値と、の関係に基づいて、前記影響候補現象の値と前記モデル化対象現象の値との対応規則を示す基礎対応規則データを生成し、前記影響候補現象に関連付ける基礎対応規則データ関連付け手段、をさらに含み、
    　前記対応規則データ生成手段が、前記現象選択手段により選択される現象に関連付けられている前記基礎対応規則データに基づいて、前記対応規則データを生成する、
    　ことを特徴とする請求項１に記載の情報処理装置。     For each candidate effect phenomenon, the value of the candidate effect phenomenon and the value of the modeling target phenomenon are based on the relationship between the time series value of the candidate effect phenomenon and the time series value of the modeling target phenomenon. A basic correspondence rule data associating means for generating basic correspondence rule data indicating a correspondence rule and associating it with the candidate effect phenomenon;
    The correspondence rule data generation means generates the correspondence rule data based on the basic correspondence rule data associated with the phenomenon selected by the phenomenon selection means;
    The information processing apparatus according to claim 1.
  5. 　前記対応規則データが示す対応規則に従って、所与の時点における前記モデル化対象現象の予測値を算出する予測値算出手段、をさらに含む、
    　ことを特徴とする請求項１に記載の情報処理装置。     A predicted value calculating means for calculating a predicted value of the modeling target phenomenon at a given time point according to the corresponding rule indicated by the corresponding rule data;
    The information processing apparatus according to claim 1.
  6. 　前記影響指標を算出する規則を示す算出規則データを複数記憶する算出規則データ記憶手段と、
    　前記影響候補現象の時系列値と、前記モデル化対象現象の時系列値と、の関係に基づいて、複数の前記算出規則データのうちから少なくとも１つの算出規則データを選択する算出規則データ選択手段と、をさらに含み、
    　前記影響指標関連付け手段が、前記算出規則データ選択手段により選択される算出規則データが示す規則に従って前記影響指標を算出して、前記影響候補現象に関連付ける、
    　ことを特徴とする請求項１に記載の情報処理装置。     Calculation rule data storage means for storing a plurality of calculation rule data indicating rules for calculating the influence index;
    Calculation rule data selection means for selecting at least one calculation rule data from the plurality of calculation rule data based on the relationship between the time series value of the candidate effect phenomenon and the time series value of the phenomenon to be modeled And further including
    The influence index associating means calculates the influence index according to a rule indicated by calculation rule data selected by the calculation rule data selection means, and associates it with the candidate effect phenomenon;
    The information processing apparatus according to claim 1.
  7. 　前記各影響候補現象が、当該影響候補現象の値が質的データであるか量的データであるかを示す現象種別データに関連付けられており、
    　前記算出規則データ選択手段が、前記影響候補現象に関連付けられている現象種別データに基づいて算出規則データを選択する、
    　ことを特徴とする請求項６に記載の情報処理装置。     Each influence candidate phenomenon is associated with phenomenon type data indicating whether the value of the influence candidate phenomenon is qualitative data or quantitative data,
    The calculation rule data selection means selects calculation rule data based on the phenomenon type data associated with the influence candidate phenomenon;
    The information processing apparatus according to claim 6.
  8. 　前記対応規則データ生成手段が、予め定められた関数の係数を示す対応規則データを生成する、
    　ことを特徴とする請求項１に記載の情報処理装置。     The correspondence rule data generating means generates correspondence rule data indicating a coefficient of a predetermined function;
    The information processing apparatus according to claim 1.
  9. 　前記影響指標が、前記影響候補現象が前記予測対象現象に与える因果的な影響の強さを示す、
    　ことを特徴とする請求項１に記載の情報処理装置。     The influence index indicates the strength of causal influence of the influence candidate phenomenon on the prediction target phenomenon;
    The information processing apparatus according to claim 1.
  10. 　前記現象選択手段による選択結果に基づく画像を出力する画像出力手段、をさらに含む、
    　ことを特徴とする請求項１に記載の情報処理装置。     Image output means for outputting an image based on the selection result by the phenomenon selection means,
    The information processing apparatus according to claim 1.
  11. 　モデル化の対象となるモデル化対象現象の時系列値を含むモデル化対象現象データと、前記モデル化対象現象に影響を与える候補となる複数の影響候補現象それぞれの時系列値を含む影響候補現象データと、を取得する現象データ取得ステップと、
    　前記各影響候補現象について、当該影響候補現象の時系列値と、前記モデル化対象現象の時系列値と、の関係に基づいて、当該影響候補現象が前記モデル化対象現象に与える影響の強さを示す影響指標を算出して、当該影響候補現象に関連付ける影響指標関連付けステップと、
    　前記各影響候補現象に関連付けられる影響指標に基づいて、前記複数の影響候補現象のうちから少なくとも１つの現象を選択する現象選択ステップと、
    　前記現象選択手段により選択される現象の時系列値と、前記モデル化対象現象の時系列値と、に基づいて、前記現象選択手段により選択される現象の値と前記モデル化対象現象の値との対応規則を示す対応規則データを生成する対応規則データ生成ステップと、
    　を含むことを特徴とする情報処理方法。     Modeling target phenomenon data including a time series value of a modeling target phenomenon to be modeled, and an influence candidate phenomenon including time series values of each of a plurality of candidate effecting phenomena that are candidates for affecting the modeling target phenomenon A phenomenon data acquisition step for acquiring data, and
    For each candidate effect phenomenon, the strength of the influence of the candidate effect phenomenon on the modeling target phenomenon based on the relationship between the time series value of the candidate effect phenomenon and the time series value of the modeling target phenomenon An impact index associating step for calculating an impact index indicating
    A phenomenon selection step of selecting at least one phenomenon from the plurality of influence candidate phenomena based on an influence index associated with each of the influence candidate phenomena;
    Based on the time series value of the phenomenon selected by the phenomenon selecting means and the time series value of the modeling target phenomenon, the value of the phenomenon selected by the phenomenon selecting means and the value of the modeling target phenomenon A correspondence rule data generation step for generating correspondence rule data indicating the correspondence rules of
    An information processing method comprising:
  12. 　モデル化の対象となるモデル化対象現象の時系列値を含むモデル化対象現象データと、前記モデル化対象現象に影響を与える候補となる複数の影響候補現象それぞれの時系列値を含む影響候補現象データと、を取得する現象データ取得手段、
    　前記各影響候補現象について、当該影響候補現象の時系列値と、前記モデル化対象現象の時系列値と、の関係に基づいて、当該影響候補現象が前記モデル化対象現象に与える影響の強さを示す影響指標を算出して、当該影響候補現象に関連付ける影響指標関連付け手段、
    　前記各影響候補現象に関連付けられる影響指標に基づいて、前記複数の影響候補現象のうちから少なくとも１つの現象を選択する現象選択手段、
    　前記現象選択手段により選択される現象の時系列値と、前記モデル化対象現象の時系列値と、に基づいて、前記現象選択手段により選択される現象の値と前記モデル化対象現象の値との対応規則を示す対応規則データを生成する対応規則データ生成手段、
    　としてコンピュータを機能させることを特徴とするプログラムが記憶されたコンピュータ読み取り可能な情報記憶媒体。     Modeling target phenomenon data including a time series value of a modeling target phenomenon to be modeled, and an influence candidate phenomenon including time series values of each of a plurality of candidate effecting phenomena that are candidates for affecting the modeling target phenomenon Data and phenomenon data acquisition means for acquiring,
    For each candidate effect phenomenon, the strength of the influence of the candidate effect phenomenon on the modeling target phenomenon based on the relationship between the time series value of the candidate effect phenomenon and the time series value of the modeling target phenomenon An impact index associating means for calculating an impact index indicating
    A phenomenon selecting means for selecting at least one phenomenon from the plurality of influence candidate phenomena based on an influence index associated with each of the influence candidate phenomena;
    Based on the time series value of the phenomenon selected by the phenomenon selecting means and the time series value of the modeling target phenomenon, the value of the phenomenon selected by the phenomenon selecting means and the value of the modeling target phenomenon Correspondence rule data generating means for generating correspondence rule data indicating correspondence rules of
    A computer-readable information storage medium storing a program characterized by causing a computer to function as:
  13. 　モデル化の対象となるモデル化対象現象の時系列値を含むモデル化対象現象データと、前記モデル化対象現象に影響を与える候補となる複数の影響候補現象それぞれの時系列値を含む影響候補現象データと、を取得する現象データ取得手段、
    　前記各影響候補現象について、当該影響候補現象の時系列値と、前記モデル化対象現象の時系列値と、の関係に基づいて、当該影響候補現象が前記モデル化対象現象に与える影響の強さを示す影響指標を算出して、当該影響候補現象に関連付ける影響指標関連付け手段、
    　前記各影響候補現象に関連付けられる影響指標に基づいて、前記複数の影響候補現象のうちから少なくとも１つの現象を選択する現象選択手段、
    　前記現象選択手段により選択される現象の時系列値と、前記モデル化対象現象の時系列値と、に基づいて、前記現象選択手段により選択される現象の値と前記モデル化対象現象の値との対応規則を示す対応規則データを生成する対応規則データ生成手段、
    　としてコンピュータを機能させることを特徴とするプログラム。     Modeling target phenomenon data including a time series value of a modeling target phenomenon to be modeled, and an influence candidate phenomenon including time series values of each of a plurality of candidate effecting phenomena that are candidates for affecting the modeling target phenomenon Data and phenomenon data acquisition means for acquiring,
    For each candidate effect phenomenon, the strength of the influence of the candidate effect phenomenon on the modeling target phenomenon based on the relationship between the time series value of the candidate effect phenomenon and the time series value of the modeling target phenomenon An impact index associating means for calculating an impact index indicating
    A phenomenon selecting means for selecting at least one phenomenon from the plurality of influence candidate phenomena based on an influence index associated with each of the influence candidate phenomena;
    Based on the time series value of the phenomenon selected by the phenomenon selecting means and the time series value of the modeling target phenomenon, the value of the phenomenon selected by the phenomenon selecting means and the value of the modeling target phenomenon Correspondence rule data generating means for generating correspondence rule data indicating correspondence rules of
    A program characterized by causing a computer to function.
PCT/JP2009/065479 2008-09-04 2009-09-04 Information processing device, information processing method, information storage medium, and program WO2010027046A1 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2008-227169 2008-09-04
JP2008227169A JP2010061450A (en) 2008-09-04 2008-09-04 Information processing device, information processing method and program

Publications (1)

Publication Number Publication Date
WO2010027046A1 true WO2010027046A1 (en) 2010-03-11

Family

ID=41797212

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2009/065479 WO2010027046A1 (en) 2008-09-04 2009-09-04 Information processing device, information processing method, information storage medium, and program

Country Status (2)

Country Link
JP (1) JP2010061450A (en)
WO (1) WO2010027046A1 (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
TWI672565B (en) * 2017-03-23 2019-09-21 荷蘭商Ａｓｍｌ荷蘭公司 Methods of modelling systems or performing predictive maintenance of systems, such as lithographic systems and associated lithographic systems

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9979738B2 (en) 2012-01-23 2018-05-22 Hrl Laboratories, Llc System and method to detect attacks on mobile wireless networks based on motif analysis
US10091218B2 (en) 2012-01-23 2018-10-02 Hrl Laboratories, Llc System and method to detect attacks on mobile wireless networks based on network controllability analysis
US10460008B2 (en) 2013-03-14 2019-10-29 Hrl Laboratories, Llc Predicting system trajectories toward critical transitions
CN105122004B (en) * 2013-03-14 2017-11-28 赫尔实验室有限公司 The system trajectory of pre- direction finding critical transitions
US20160218931A1 (en) * 2013-09-30 2016-07-28 Nec Corporation System for determining inter-element influence, device for determining inter-element influence, method for determining inter-element influence, and recording medium
JP7040065B2 (en) * 2018-01-31 2022-03-23 富士通株式会社 Detection program, detection method and detection device

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH06266993A (en) * 1993-03-11 1994-09-22 Hitachi Ltd Highway traffic controller
JP2006079298A (en) * 2004-09-09 2006-03-23 Toyota Motor Corp Sales volume forecast system
JP2007207101A (en) * 2006-02-03 2007-08-16 Infocom Corp Graph generation method, graph generation program, and data mining system

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH06266993A (en) * 1993-03-11 1994-09-22 Hitachi Ltd Highway traffic controller
JP2006079298A (en) * 2004-09-09 2006-03-23 Toyota Motor Corp Sales volume forecast system
JP2007207101A (en) * 2006-02-03 2007-08-16 Infocom Corp Graph generation method, graph generation program, and data mining system

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
J. TIKKA ET AL.: "Learning Linear Dependency Trees from Multivariate Time-series Data", PROCEEDINGS OF THE ICDM2004 WORKSHOP ON TEMPORAL DATA MINING, 2004 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
TWI672565B (en) * 2017-03-23 2019-09-21 荷蘭商Ａｓｍｌ荷蘭公司 Methods of modelling systems or performing predictive maintenance of systems, such as lithographic systems and associated lithographic systems
TWI706232B (en) * 2017-03-23 2020-10-01 荷蘭商Ａｓｍｌ荷蘭公司 Methods of modelling systems or performing predictive maintenance of systems, such as lithographic systems and associated lithographic systems
US11461675B2 (en) 2017-03-23 2022-10-04 Asml Netherlands B.V. Methods of modelling systems or performing predictive maintenance of systems, such as lithographic systems and associated lithographic systems

Also Published As

Publication number Publication date
JP2010061450A (en) 2010-03-18

Similar Documents

Publication Publication Date Title
Mondal et al. Approximate Bayesian computation with deep learning supports a third archaic introgression in Asia and Oceania
WO2010027046A1 (en) Information processing device, information processing method, information storage medium, and program
Höhle et al. Bayesian nowcasting during the STEC O104: H4 outbreak in Germany, 2011
Bebbington et al. The discrete additive Weibull distribution: A bathtub-shaped hazard for discontinuous failure data
Phaladiganon et al. Bootstrap-based T 2 multivariate control charts
Paiva et al. A data-driven model to describe and forecast the dynamics of COVID-19 transmission
Bélisle et al. Evaluation of machine learning interpolation techniques for prediction of physical properties
Teng et al. Persistence and extinction for a class of stochastic SIS epidemic models with nonlinear incidence rate
Chao et al. Expected Shannon entropy and Shannon differentiation between subpopulations for neutral genes under the finite island model
Schimit et al. Disease spreading in complex networks: A numerical study with Principal Component Analysis
JP2022529178A (en) Features of artificial intelligence recommended models Processing methods, devices, electronic devices, and computer programs
Rahman et al. Accuracy comparison of ARIMA and XGBoost forecasting models in predicting the incidence of COVID-19 in Bangladesh
Zhang et al. Inference of gene regulatory networks using pseudo-time series data
Bell et al. A distribution-free multivariate phase I location control chart for subgrouped data from elliptical distributions
Triacca et al. Forecasting the number of confirmed new cases of COVID-19 in Italy for the period from 19 May to 2 June 2020
Bien et al. Non-convex global minimization and false discovery rate control for the TREX
Xiao et al. Inferring gene regulatory networks using conditional regulation pattern to guide candidate genes
Zhou et al. Estimating time to the most recent common ancestor (TMRCA): comparison and application of eight methods
JP6636282B2 (en) A method, apparatus, and program for calculating an interaction parameter between a polymer and a solvent.
Burger et al. Neural networks for self-adjusting mutation rate estimation when the recombination rate is unknown
Bruneau et al. A clustering package for nucleotide sequences using Laplacian Eigenmaps and Gaussian Mixture Model
CN107342857B (en) Grouping method and device
Bagirov et al. Estimation of a regression function by maxima of minima of linear functions
Fierro et al. Statistical inference on a stochastic epidemic model
Papadantonakis et al. CoMuS: simulating coalescent histories and polymorphic data from multiple species

Legal Events

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

Ref document number: 09811571

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 09811571

Country of ref document: EP

Kind code of ref document: A1