JP6849084B2 - Information analyzer, information analysis method and information analysis program - Google Patents

Description

The present invention relates to an information analyzer, an information analysis method, and an information analysis program for grouping a plurality of prediction targets and analyzing information.

When trying to obtain a prediction result as a whole, instead of creating a prediction model for each individual target, a plurality of targets may be grouped and prediction may be performed for each group.

For example, Patent Document 1 describes an electric power demand adjustment system related to an energy community consisting of a large number of consumers connected to an electric power system. Specifically, Patent Document 1 describes that as the number of demands increases, that is, as the size of the consumer group increases, the demand forecast accuracy of the consumer group increases and the variance decreases. .. Further, in Patent Document 1, the size of the consumer group is almost proportional to the total amount of power demand of the consumer group, and in such a situation, it is predicted when the size of the consumer group becomes a certain value. It is stated that the difference from the actual result is minimized, and the difference increases as the number of consumers increases and decreases. The electric power demand adjustment system described in Patent Document 1 determines the size of the consumer group by utilizing the relationship between the size of the consumer group and the forecast and the actual result, and determines the electric power demand of the consumer group. Predict.

International Publication No. 2016/0843313

However, the electric power demand adjustment system described in Patent Document 1 does not consider that individual consumers have different properties from each other. Specifically, the electric power demand adjustment system described in Patent Document 1 is described above between "the size of the consumer group" and "estimation accuracy / dispersion / total amount of electric power demand / difference between forecast and actual result". It is assumed that a relationship is established. However, if the individual consumers have different properties from each other, the relationships described above will vary depending on the consumers being grouped.

That is, since the electric power demand adjustment system described in Patent Document 1 does not consider which consumers are in the same group from the viewpoint of prediction accuracy, it is said that the prediction accuracy can always be improved by grouping. hard.

Therefore, in the present invention, even when a plurality of prediction targets have different properties from each other, a desired prediction target is expressed by the sum of the plurality of prediction targets. It is an object of the present invention to provide an information analyzer, an information analysis method, and an information analysis program capable of realizing a grouping of prediction targets so as to improve the prediction accuracy of the above.

The information analyzer according to the present invention has a unit model generation unit that generates a learning model for each partial prediction target when the prediction target is represented by the sum of a plurality of partial prediction targets, and a partial prediction target using the learning model. The time axis of the estimation error calculation unit that calculates the estimation error when estimating the value of is along the time axis, and the estimation error of one learning model and the estimation error of another learning model by inputting the estimation error. Based on the similarity in the above, the set generation part that generates a set of groups by grouping the partial prediction target corresponding to one learning model and the partial prediction target corresponding to another learning model into a group, and the group include It is equipped with a partial sum model generation unit that generates a learning model with the sum of the partial prediction targets as the objective variable, and a prediction target calculation unit that calculates the value of the prediction target from the prediction values of the group calculated using the learning model. It is characterized by that.

The information analysis system according to the present invention has a unit model generation unit that generates a learning model for each partial prediction target when the prediction target is represented by the sum of a plurality of partial prediction targets, and a partial prediction target using the learning model. The time axis of the estimation error calculation unit that calculates the estimation error when estimating the value of is along the time axis, and the estimation error of one learning model and the estimation error of another learning model by inputting the estimation error. Based on the similarity in the above, by grouping the partial prediction target corresponding to one learning model and the partial prediction target corresponding to another learning model into a group, the composition of the group set is different from that of the first group. For each of the set generator that generates the set and the set of the second group, and the set of the first group and the set of the second group, the sum of the partial prediction targets included in the group is used as the objective variable. For each of the partial sum model generator that generates the training model to be used, the set of the first group, and the set of the second group, the value to be predicted is calculated from the predicted value of the group calculated using the learning model. Evaluation that evaluates the set of groups based on the difference between the value of the prediction target and the actual value of the prediction target for each of the calculation unit of the prediction target to be calculated and the set of the first group and the set of the second group. It is characterized by having a part.

In the information analysis method according to the present invention, when the prediction target is represented by the sum of a plurality of partial prediction targets, a learning model is generated for each partial prediction target, and the value of the partial prediction target is estimated using the learning model. The case estimation error is calculated along the time axis, and the estimation error is used as an input, based on the similarity between the estimation error for one learning model and the estimation error for the other learning model on the time axis. A learning model in which the sum of the partial prediction targets included in the group is used as the objective variable is generated by grouping the partial prediction targets corresponding to the learning model and the partial prediction targets corresponding to other learning models into a group. It is characterized in that the value of the prediction target is calculated from the prediction value of the group generated and calculated using the learning model.

In another information analysis method according to the present invention, when the prediction target is represented by the sum of a plurality of partial prediction targets, a learning model is generated for each partial prediction target, and the value of the partial prediction target is calculated using the learning model. The estimation error when estimated is calculated along the time axis, and the estimation error is used as an input, based on the similarity between the estimation error for one learning model and the estimation error for the other learning model on the time axis. By grouping the partial prediction target corresponding to one learning model and the partial prediction target corresponding to another learning model into a group, the composition of the group set is different from that of the first group set and the second group set. And are generated respectively, and for each of the set of the first group and the set of the second group, a learning model in which the sum of the partial prediction targets included in the group is the objective variable is generated, and the first group For each of the set and the set of the second group, the value to be predicted is calculated from the predicted value of the group calculated using the learning model, and the set of the first group and the set of the second group are each calculated. On the other hand, it is characterized in that the set of groups is evaluated based on the difference between the value of the prediction target and the actual value of the prediction target.

The information analysis program according to the present invention uses a unit model generation process and a learning model to generate a learning model for each partial prediction target when the prediction target is represented by the sum of a plurality of partial prediction targets on a computer. Estimate error calculation processing that calculates the estimation error when estimating the value of the prediction target along the time axis, the time between the estimation error for one learning model and the estimation error for the other learning model with the estimation error as input A set generation process that generates a set of groups by grouping a partial prediction target corresponding to one learning model and a partial prediction target corresponding to another learning model based on the similarity on the axis, and is included in the group. Executes the partial sum model generation process that generates a learning model with the sum of the partial prediction targets as the objective variable, and the prediction target calculation process that calculates the value of the prediction target from the prediction values of the group calculated using the learning model. It is characterized by letting it.

Another information analysis program according to the present invention uses a unit model generation process and a learning model that generate a learning model for each partial prediction target when the prediction target is represented by the sum of a plurality of partial prediction targets on a computer. Estimate error calculation processing that calculates the estimation error when the value of the partial prediction target is estimated along the time axis, and the estimation error for one learning model and the estimation error for the other learning model by inputting the estimation error By grouping the partial prediction target corresponding to one learning model and the partial prediction target corresponding to another learning model into a group based on the similarity on the time axis of, the composition of the set of the groups is different from each other. The purpose is the sum of the partial prediction targets included in the group for each of the set of the first group and the set of the second group, which is the set generation process for generating the set of the first group and the set of the second group, respectively. For each of the partial sum model generation process that generates the training model as a variable, the set of the first group and the set of the second group, the value to be predicted from the predicted value of the group calculated using the training model. For each of the prediction target calculation process for calculating and the first group set and the second group set, the group set is evaluated based on the difference between the prediction target value and the prediction target actual value. It is characterized in that the evaluation process is executed.

According to the present invention, even when a plurality of prediction targets have different properties from each other, a desired prediction target is expressed by the sum of the plurality of prediction targets. It has the technical effect of being able to realize grouping of prediction targets that improves the prediction accuracy of.

It is a block diagram which shows one Embodiment of the information analyzer according to this invention. It is explanatory drawing which shows the example of the candidate of a partial prediction target. It is explanatory drawing which shows the example of the actual data. It is explanatory drawing which shows the example of the learning model generated by the partial prediction target unit. It is explanatory drawing which shows the example of the processing which calculates the estimation error. It is explanatory drawing which shows the example of the estimation error calculated along the time axis. It is explanatory drawing which shows the example which graphed the estimation error calculated along the time axis. It is explanatory drawing which shows the example of the process of grouping a partial prediction target. It is explanatory drawing which shows the example of the process which generates the set of a group. It is explanatory drawing which shows the example of the process which generates the set of a group. It is a flowchart which shows the operation example of an information analyzer. It is a block diagram which shows the structural example of the information analysis system by this invention. It is a block diagram which shows the outline of the information analyzer according to this invention. It is a block diagram which shows the outline of the information analysis system by this invention.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of an information analyzer according to the present invention. The configuration of the information analyzer of this embodiment will be described later.

In the present invention, it is assumed that the value of the prediction target Y (prediction result) at a certain time is composed of the sum _{of a plurality of prediction targets y 1} , y ₂ , ..., Y _{n at the same time.} That is, the prediction target at a certain time is represented by the sum of a plurality of partial prediction targets at the same time. Specifically, Y = y ₁ + y ₂ + ... + y _n .

Since the desired prediction target Y is formed by adding a plurality of prediction targets y ₁ , y ₂ , ..., Y _n , in the following description, the individual prediction targets y ₁ , y ₂ , ..., Y _n is referred to as a partial prediction target. Therefore, the value of the prediction target Y at a certain time is calculated by the sum _{of the partial prediction targets y 1} , y ₂ , ..., Y _{n at the same time.} For example, when the forecast target Y is the total demand for products manufactured by the factory (total number of orders), the partial forecast target y _n is the number of product demands (number of orders) for each store.

Further, in the present invention, the partial prediction targets are grouped into a plurality of groups (Y ₁ , Y ₂ , ..., Y _N ), the prediction model is learned for each group, and the prediction value and the actual value of the prediction target are calculated.

Hereinafter, the relationship between the prediction target and the partial prediction target will be described with reference to a specific example. Below, assuming that there is a factory that manufactures a certain product and delivers it to each store, from the standpoint of the production control department of this factory, predicting future product demand (number of orders) based on past actual data. To do.

FIG. 2 is an explanatory diagram showing an example of candidates for partial prediction. In the example shown in FIG. 2, it is shown that the own factory delivers the product to the 200 stores (store code: C-001 to C-200) illustrated in FIG. Further, FIG. 3 is an explanatory diagram showing an example of the number of sales of products for each day and each store as actual data. In addition, the table illustrated in FIG. 3 includes the maximum temperature and the minimum temperature of the previous day. In the following description, the past data (actual data) actually acquired as illustrated in FIG. 3 may be referred to as evaluation data.

If such information exists, it is assumed that the number of orders for products on the next day is predicted. For example, in the general method, the total number of sales of each store on each day is set as the objective variable, and the maximum temperature, the minimum temperature, etc. as included in the table illustrated in FIG. 3 are set as the explanatory variables. It is conceivable to learn a predictive model.

However, for example, there are cases where the 200 stores include stores having significantly different properties from the others (for example, stores with a large increase / decrease in sales scale and orders). There is a high possibility that high prediction accuracy cannot be obtained even if learning is performed by grouping stores having such different properties. Therefore, it is conceivable to classify such a plurality of stores into stores having similar properties and make a prediction.

However, the sales of the target products cannot always be classified only by the location and sales floor area of each store and the sales last year as illustrated in FIG. For example, even if the stores are located at the same location, the sales tendency may differ between the stores in the building and the roadside stores, and even if the stores have the same sales floor area, depending on the neighboring environment. It is possible that the demand trends for products are different.

Therefore, in the following embodiment, a method of reducing the estimation error of the predicted value with respect to the actual value of the desired prediction target by grouping stores (partial prediction targets) having similar estimation error tendencies will be described. To do. Hereinafter, the information analyzer of the present invention will be described while appropriately using the above-mentioned specific examples.

The information analysis device 100 illustrated in FIG. 1 includes a reception unit 10, a unit model generation unit 12, an estimation error calculation unit 13, a set generation unit 14, a partial sum model generation unit 16, and a prediction target calculation unit 18. , The evaluation unit 20, the output unit 22, the control unit 24, and the storage unit 30 are provided.

The storage unit 30 stores information regarding the partial prediction target. The storage unit 30 may store a partial prediction target, for example, a store illustrated in FIG. 2. In addition, the storage unit 30 may store actual data regarding the partial prediction target as illustrated in FIG.

The actual data is used as data when the unit model generation unit 12 and the partial sum model generation unit 16 described later generate a learning model. In addition, the actual data is used as data when the evaluation unit 20 described later evaluates the prediction model. For example, of the actual data, a certain percentage of data may be used as learning data, and the remaining data may be used as evaluation data.

However, the data stored by the storage unit 30 is not limited to these data. The storage unit 30 may store, for example, the learning model generated by the unit model generation unit 12 and the partial sum model generation unit 16, which will be described later. The storage unit 30 is realized by, for example, a magnetic disk device.

The reception unit 10 receives the designation of a desired prediction target. Further, the reception unit 10 may accept the designation of a plurality of partial prediction targets constituting the desired prediction target.

The unit model generation unit 12 generates a learning model in the partial prediction target unit. Specifically, the unit model generation unit 12 generates a learning model for each of _{the partial prediction targets y 1} , y ₂ , ..., And y _n.

For example, when each store illustrated in FIG. 2 is a partial prediction target, the unit model generation unit 12 generates a learning model for each store. The unit model generation unit 12 may generate a learning model (for example, a linear multiple regression model) in which the prediction target can be predicted by the weight of the explanatory variable so that the similarity of the factors affecting the prediction target can be determined. However, the learning model generated by the unit model generation unit 12 is not limited to the linear multiple regression model. The unit model generation unit 12 may generate a learning model by any method.

FIG. 4 is an explanatory diagram showing an example of a learning model generated by the unit model generation unit 12 in the partial prediction target unit. In the example shown in FIG. 4, a store code "C-001" forecast y ₀₀₁ items from the store that are identified in the previous day's highest temperature (explanatory variable x _1), one week prior to the sales (explanatory variable x ₂₎ , And, it is shown that there is a positive correlation with the sales floor area (explanatory variable x _3). On the other hand, in the example shown in FIG. 4, a store code forecast y ₂₀₀ items from the store specified by "C-200" is one week before the sales (explanatory variable x _2), and, floor space (explanatory variable x It shows that there is a _{positive correlation with 3} ), while there is a _{negative correlation with the weekend (explanatory variable x 4).}

The estimation error calculation unit 13 calculates the estimation error when the value of the partial prediction target is estimated by using the learning model along the time axis. Specifically, the estimation error calculation unit 13 calculates the estimation error when the value of the partial prediction target is estimated by using the learning model generated by the unit model generation unit 12 along the time axis.

The estimation error calculation unit 13 similarly calculates the estimation error when the value of the partial prediction target is estimated by using the learning model generated by the partial sum model generation unit 16 described later along the time axis. .. The process using the learning model generated by the partial sum model generation unit 16 will be described later.

FIG. 5 is an explanatory diagram showing an example of processing for calculating the estimation error. For example, it is assumed that the actual data of June 2017 illustrated in FIG. 3 is used as the evaluation data, and the corresponding learning model is represented by the prediction formula illustrated in FIG.

_{First, the estimation error calculation unit 13 uses a prediction formula (y 001} = 0.1 x ₁ + 0.) Illustrating the estimated value of the sales of the store specified by the store code "C-001" which is the partial prediction target. It is calculated using _{18 x 2} + 0.06 x _{3 + 1.6).} For example, when calculating the estimated value of sales on June 2, 2017, the sales floor area of the store is 120 m ² , and the maximum temperature on the previous day is 26 degrees. Further, assuming that the number of sales one week ago was 20, the estimated value of sales _{is calculated as y ^ 001} = 0.1 × 26 + 0.18 × 10 + 0.06 × 120 + 1.6 = 13.2.

Then, the estimation error calculation unit 13 sets the error (estimation error) between 15 which is the actual value of sales on June 2, 2017 and 13.2 which is the estimated value as 15-13.2 = 1.8. calculate. The estimation error calculation unit 13 calculates the estimation error (in time series) along the time axis by similarly calculating the estimation error for the following dates. Further, the estimation error calculation unit 13 similarly calculates the estimation error along the time axis for other partial prediction targets (stores).

FIG. 6 is an explanatory diagram showing an example of an estimation error calculated along the time axis. The example shown in FIG. 6 shows the result of calculating the estimation error for each store and each day using the data illustrated in FIG. Further, FIG. 7 is an explanatory diagram showing an example in which the estimation error calculated along the time axis is graphed. In the example shown in FIG. 7, the estimation error along the time axis illustrated in FIG. 6 is represented by a line graph for each store.

The estimation error calculation unit 13 may standardize the estimation error of each partial prediction target according to the difference in sales scale. Specifically, the estimation error calculation unit 13 may standardize the estimation error based on the average and variance of the estimation error of each partial prediction target. By performing such standardization, the tendency of the estimation error along the time axis can be judged on the same scale.

The set generation unit 14 generates a set of groups by grouping the partial prediction targets into a plurality of groups. The set generation unit 14 generates a set of groups according to the control of the control unit 24 described later.

First, the set generation unit 14 inputs the estimation error calculated by the estimation error calculation unit 13. Then, the set generation unit 14 makes a partial prediction corresponding to one learning model based on the similarity on the time axis between the input estimation error of one learning model and the estimation error of another learning model. A set of groups is generated by grouping the target and the partial prediction target corresponding to other learning models into a group.

The set generation unit 14 similarly generates a set of groups based on the similarity of the estimation error by the learning model generated by the partial sum model generation unit 16 described later on the time axis. The process of using the estimation error by the learning model generated by the partial sum model generation unit 16 will be described later.

The method by which the set generation unit 14 determines the similarity of the estimation error on the time axis is arbitrary. The set generation unit 14 represents, for example, a change in the estimation error of each section in the time axis direction by a vector (estimation error vector), and the sum of the inner products with the estimation error vector by another learning model to be compared is calculated by increasing the sum. It may be judged that the larger the value, the higher the similarity. Further, the set generation unit 14 may determine that the more similar the shapes of the graphs illustrated in FIG. 7, the more similar the estimation error on the time axis is.

Further, the number of groups collected by the set generation unit 14 is also arbitrary. FIG. 8 is an explanatory diagram showing an example of processing for grouping partial prediction targets. The set generation unit 14 may sequentially group the partial prediction targets corresponding to the learning models having high estimation error similarity, and repeat the grouping while excluding the grouped partial prediction targets. Specifically, as illustrated in FIG. 8A, the set generation unit 14 generates a group G1 in which partial prediction targets having high estimation error similarity are grouped, and then a portion included in the group G1. Group G2 may be generated by grouping partial prediction targets having the next highest estimation error similarity, excluding the prediction target. For example, when each pair of learning models having similar estimation errors is extracted to generate a group, a set of 100 groups is generated from the 200 partial prediction targets illustrated in FIG.

Also. The set generation unit 14 may sequentially group the partial prediction targets corresponding to the learning models having high estimation error similarity, and repeat the grouping without excluding the grouped partial prediction targets. Specifically, as illustrated in FIG. 8B, the set generation unit 14 generates a group G1 in which partial prediction targets having high estimation error similarity are grouped, and then a portion included in the group G1. A group G3 may be generated in which partial prediction targets having the next highest estimation error similarity, including the prediction target, are grouped.

Further, the set generation unit 14 may separately generate a set of groups having a configuration different from the set of groups already generated. Specifically, the set generation unit 14 may generate a set of groups so that the number of groups constituting the set of groups is different from the number of groups constituting the set of already generated groups. For example, when there are 200 partial prediction targets illustrated in FIG. 2, the set generation unit 14 generates a set of 100 groups that summarizes the two partial prediction targets, and further summarizes the four partial prediction targets. A set of 50 groups may be generated.

The number of partial prediction targets included in each group generated by the set generation unit 14 may be the same or different.

FIG. 9 is an explanatory diagram showing an example of a process of generating a set of groups. When the estimation error illustrated in FIG. 7 is calculated, for example, the tendency of the estimation error of the store specified by the store code "C-001" and the estimation error of the store specified by the store code "C-003". It can be said that the tendency of the estimation error (degree of change) is similar to the tendency of. Therefore, the set generation unit 14 collectively generates one group by combining the store specified by the store code "C-001" and the store specified by the store code "C-003". The set generation unit 14 repeats this grouping for other stores as well.

The partial sum model generation unit 16 generates a learning model in which the sum of the partial prediction targets included in the group is the objective variable. That is, the partial sum model generation unit 16 generates a learning model for each of _{a plurality of grouped partial prediction target groups (Y 1} , Y ₂ , ..., Y _N). The method in which the partial sum model generation unit 16 generates the learning model is the same as the method in which the set generation unit 14 generates the learning model.

For example, as illustrated in FIG. 9, it is assumed that the store specified by the store code "C-001" and the store specified by the store code "C-003" are collectively generated into one group. In this case, the partial sum model generation unit 16 generates a learning model in which the sum of the sales of the products of these two stores is used as the objective variable.

The prediction target calculation unit 18 calculates the value of the prediction target from the prediction value of the group calculated using the learning model. Specifically, the prediction target calculation unit 18 calculates the prediction target value by summing the prediction values of the groups calculated using the learning model generated by the unit model generation unit 12. Further, the prediction target calculation unit 18 calculates the prediction target value by summing the prediction values of the groups calculated using the learning model generated by the partial sum model generation unit 16.

The evaluation unit 20 evaluates a set of groups based on the difference between the calculated value of the prediction target and the actual value of the prediction target. The actual value of the prediction target is stored in the storage unit 30 in advance, for example. The evaluation unit 20 may use, for example, an evaluation function for evaluating the goodness of a set of groups.

Specifically, the evaluation unit 20 evaluates a set of groups using an evaluation function designed so that the smaller the difference between the calculated value of the prediction target and the actual value of the prediction target, the better the evaluation. May be good. This is because it can be said that the smaller the error between the predicted value and the measured value, the more appropriate the grouping is. The evaluation function in this case may be designed by, for example, Equation 1 illustrated below. The smaller the value of the evaluation function W, the better the evaluation. The error of the prediction target Y exemplified in Equation 1 is the total error which is the sum of the partial prediction targets.

Evaluation function W = β · (Error of prediction target Y) (Equation 1)

Further, in addition to the above-mentioned evaluation, the evaluation unit 20 may evaluate the set of groups by using an evaluation function designed so that the smaller the number of groups constituting the set of groups, the better the evaluation. This is because the smaller the number of groups, the smaller the number of models generated, which can improve human interpretation. The evaluation function in this case may be designed by, for example, Equation 2 illustrated below.

Evaluation function W = α · Number of models + β · (Error of prediction target Y) (Equation 2)

Further, instead of evaluating the number of groups constituting the set of groups, the evaluation unit 20 evaluates better as the total number of explanatory variables having a non-zero coefficient included in each learning model corresponding to each group is smaller. A set of groups may be evaluated using an evaluation function designed to be. This is because the smaller the number of explanatory variables having a non-zero coefficient, the higher the human interpretability.

Moreover, the smaller the number of explanatory variables with nonzero coefficients, the less complex the model can be. The more complex the model, the easier it is to find evaluation data that fits the model well, which may lead to inadequate evaluation. On the other hand, the more simplistic the model is, the more the above-mentioned fear can be avoided. Therefore, it is possible to suppress the phenomenon that the accuracy deteriorates during operation even though the accuracy is good at the time of model evaluation. The evaluation function in this case may be designed by, for example, Equation 3 illustrated below.

Evaluation function W = α · (number of non-zero coefficients of each Σ model) + β · (error of prediction target Y) (Equation 3)

The control unit 24 controls the operation when the set generation unit 14 generates a set of groups. Specifically, the control unit 24 controls the set generation unit 14 so as to group individual partial prediction targets into a predetermined number of groups. FIG. 10 is an explanatory diagram showing an example of a process for generating a set of groups.

For example, when there are 200 partial prediction targets illustrated in FIG. 2, the control unit 24 may control the set generation unit 14 so as to generate a set of 100 groups in which the two partial prediction targets are put together. (Step S1 in FIG. 10).

Further, the control unit 24 may control the set generation unit 14 so that the number of groups constituting the set of groups is different from the number of groups constituting the set of already generated groups. At this time, the control unit 24 may control the set generation unit 14 so as to group the individual partial prediction targets, or the group already generated by the set generation unit 14 may be further grouped into groups. You may control to generate a set of.

A method of further grouping the groups already generated by the set generation unit 14 into groups will be described more specifically. First, the control unit 24 controls the estimation error calculation unit 13 so that the estimation error for the learning model generated by the partial sum model generation unit 16 is calculated. In response to this control, the estimation error calculation unit 13 calculates the estimation error when the value of the partial prediction target is estimated using the learning model generated by the partial sum model generation unit 16 along the time axis.

Further, when the control unit 24 controls the set generation unit 14, the control unit 24 inputs the estimation error for the learning model generated by the partial sum model generation unit 16 to the set generation unit 14. In response to this control, the set generation unit 14 generates a set of groups by inputting the estimation error of the learning model generated by the partial sum model generation unit 16.

For example, when the set generation unit 14 has already generated a set of 100 groups in which two partial prediction targets are grouped together, the control unit 24 further groups the 100 groups into groups, for example, 50. It may be controlled to generate a set of groups (step S2 in FIG. 10). Further, the control unit 24 may control the set generation unit 14 so as to generate a set of 50 groups in which the four partial prediction targets are grouped with respect to the first 200 partial prediction targets (in FIG. 10). Step S3).

In this way, the control unit 24 controls the set generation unit 14 so as to generate a set of groups having a configuration different from the set of groups already generated. When a set of groups having a plurality of configurations is generated by the control of the control unit 24, the partial sum model generation unit 16 generates a learning model in which the sum of the partial prediction targets included in each group is the objective variable. Then, the prediction target calculation unit 18 calculates the value of the prediction target from the prediction value of the group calculated using the generated learning model, and the evaluation unit 20 calculates the calculated value of the prediction target and the actual result of the prediction target. Evaluate a set of groups based on the difference from the value.

The output unit 22 outputs a set of groups according to the evaluation of the evaluation unit 20. Specifically, the output unit 22 outputs a set of better evaluation groups. At this time, the output unit 22 may output not only the set of groups but also the learning model generated for each group constituting the set.

The reception unit 10, the unit model generation unit 12, the estimation error calculation unit 13, the set generation unit 14, the partial sum model generation unit 16, the prediction target calculation unit 18, the evaluation unit 20, and the output unit 22. The control unit 24 is realized by a computer processor (for example, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), an FPGA (field-programmable gate array)) that operates according to a program (information analysis program).

For example, the program is stored in the storage unit 30, the processor reads the program, and according to the program, the reception unit 10, the unit model generation unit 12, the estimation error calculation unit 13, the set generation unit 14, and the partial sum model generation unit 16. , The prediction target calculation unit 18, the evaluation unit 20, the output unit 22, and the control unit 24 may operate. Further, the function of the information analyzer may be provided in the SAAS (Software as a Service) format.

The reception unit 10, the unit model generation unit 12, the estimation error calculation unit 13, the set generation unit 14, the partial sum model generation unit 16, the prediction target calculation unit 18, the evaluation unit 20, and the output unit 22. Each of the control units 24 may be realized by dedicated hardware. Further, a part or all of each component of each device may be realized by a general-purpose or dedicated circuitry, a processor, or a combination thereof. These may be composed of a single chip or may be composed of a plurality of chips connected via a bus. A part or all of each component of each device may be realized by a combination of the above-mentioned circuit or the like and a program.

Further, when a part or all of each component of each device is realized by a plurality of information processing devices and circuits, the plurality of information processing devices and circuits may be centrally arranged or distributed. May be done. For example, the information processing device, the circuit, and the like may be realized as a form in which each is connected via a communication network, such as a client-and-server system and a cloud computing system.

Next, the operation of the information analyzer of the present embodiment will be described. FIG. 11 is a flowchart showing an operation example of the information analyzer of the present embodiment.

The unit model generation unit 12 generates a learning model for the partial prediction target unit (step S11). The prediction target calculation unit 18 calculates the prediction target value from the prediction value of the group calculated using the learning model (step S12). Then, the evaluation unit 20 evaluates the set of groups based on the difference between the calculated value of the prediction target and the actual value of the prediction target (step S13). Here, the evaluation unit 20 evaluates the case where a learning model is generated for each partial prediction target.

The control unit 24 determines whether or not to generate a set of groups having a configuration different from the set of groups already generated (step S14). When it is determined to generate a set of groups (Yes in step S14), the control unit 24 controls the estimation error calculation unit 13 to estimate the estimation error when the value of the partial prediction target is estimated using the learning model. Calculate along the time axis (step S21). Then, the control unit 24 controls the set generation unit 14 to group the partial prediction targets into a plurality of groups and generate a set of groups (step S15). Specifically, the set generation unit 14 generates a set of groups by grouping the partial prediction targets into groups based on the similarity of the calculated estimation errors on the time axis.

The partial sum model generation unit 16 generates a learning model in which the sum of the partial prediction targets included in the generated group is the objective variable (step S16). When the learning model is generated, the prediction target calculation unit 18 calculates the value of the prediction target, and the evaluation unit 20 repeats the processes of steps S12 and S13 for evaluating the set of groups.

On the other hand, when it is determined in step S14 that the set of groups is not generated (No in step S14), the output unit 22 outputs the set of the group with the best evaluation (step S17).

As described above, in the present embodiment, the unit model generation unit 12 generates a learning model for each partial prediction target, and the estimation error calculation unit 13 estimates the value of the partial prediction target using the learning model. The estimation error is calculated along the time axis. Further, the set generation unit 14 takes the estimation error as an input and corresponds to one learning model based on the similarity on the time axis between the estimation error for one learning model and the estimation error for the other learning model. A set of groups is generated by grouping the partial prediction target and the partial prediction target corresponding to other learning models into a group. Then, the partial sum model generation unit 16 generates a learning model in which the sum of the partial prediction targets included in the group is the objective variable, and the prediction target calculation unit 18 is based on the prediction value of the group calculated using the learning model. Calculate the value to be predicted.

Therefore, even if each of the plurality of prediction targets has different properties, when the desired prediction target is expressed by the sum of the plurality of prediction targets, the prediction accuracy of the desired prediction target can be determined. It is possible to realize a grouping of forecast targets that improves.

That is, in the present embodiment, a learning model is generated by grouping partial prediction targets having a similar tendency of estimation error with respect to the time axis. By grouping partial prediction targets with similar tendencies in this way, the estimation error is reduced by the law of large numbers, and as a result, the prediction accuracy can be further improved.

Next, a modified example of the above embodiment will be described. In the above embodiment, a method has been described in which the set generation unit 14 is controlled so that the control unit 24 generates a set of groups having different configurations, so that the generated groups are further grouped together and sequentially grouped.

On the other hand, it is also effective to change the grouping granularity for evaluation. For example, it is conceivable to perform grouping by k-means many times while changing the value of k. Further, since it is expected that the learning process will take a lot of time, it is also preferable to operate a plurality of servers in parallel for evaluation.

For example, suppose there are 200 partial prediction targets. At this time, another server executes the grouping process of k = 50 and the grouping process of k = 100 in parallel, and after performing the learning process, the prediction process, and the evaluation process on the set of each group. , You may output the group with the best evaluation.

FIG. 12 is a block diagram showing a configuration example of the information analysis system according to the present invention. The information analysis system 200 illustrated in FIG. 12 includes a reception unit 10, a unit model generation unit 12, an estimation error calculation unit 13, a set generation unit 14a, a partial sum model generation unit 16a, and a prediction target calculation unit 18a. , The evaluation unit 20a, the output unit 22, and the storage unit 30 are provided. The configuration of the reception unit 10, the unit model generation unit 12, the estimation error calculation unit 13, the output unit 22, and the storage unit 30 is the same as the configuration included in the information analysis device 100.

Similar to the set generation unit 14, the set generation unit 14a of this modification also sets the partial prediction target based on the similarity on the time axis between the estimation error for one learning model and the estimation error for the other learning model. A set of groups is generated by grouping them into multiple groups. Specifically, the set generation unit 14a inputs the estimation error calculated by the estimation error calculation unit 13. Then, the set generation unit 14a groups the partial prediction target corresponding to one learning model and the partial prediction target corresponding to the other learning model into a group, so that the configuration of the set of the groups is different from that of the first group. Generate a set and a second group of sets, respectively. The method in which the set generation unit 14a generates a set of groups is the same as the method in which the set generation unit 14 in the information analyzer 100 generates a set of groups.

In the example shown in FIG. 12, only one set generation unit 14a is shown, but when processing is executed by a plurality of servers, the information analysis system 200 may include the set generation unit 14a for each server. Good. Further, the type of the set of the group generated by the set generation unit 14a is not limited to two types, and may be three or more types.

The partial sum model generation unit 16a generates a learning model for each of the set of the first group and the set of the second group, with the sum of the partial prediction targets included in the group as the objective variable. The method in which the partial sum model generation unit 16a generates the learning model is the same as the method in which the partial sum model generation unit 16 in the information analyzer 100 generates the learning model.

In the example shown in FIG. 12, only one partial sum model generation unit 16a is shown, but when processing is executed by a plurality of servers, the information analysis system 200 uses the partial sum model generation unit 16a for each server. You may have it.

The prediction target calculation unit 18a calculates the value of the prediction target from the prediction value of the group calculated by using the learning model for each of the set of the first group and the set of the second group. The method in which the prediction target calculation unit 18a calculates the value of the prediction target is the same as the method in which the prediction target calculation unit 18 in the information analyzer 100 calculates the value of the prediction target.

In the example shown in FIG. 12, only one prediction target calculation unit 18a is shown, but when processing is executed by a plurality of servers, the information analysis system 200 includes a prediction target calculation unit 18a for each server. You may.

The evaluation unit 20a evaluates the set of groups based on the difference between the calculated value of the prediction target and the actual value of the prediction target for each of the set of the first group and the set of the second group. The method in which the evaluation unit 20a evaluates the group set is the same as the method in which the evaluation unit 20 in the information analyzer 100 evaluates the group set.

In the example shown in FIG. 12, only one evaluation unit 20a is shown, but when processing is executed by a plurality of servers, the information analysis system 200 may include an evaluation unit 20a for each server.

As described above, in the present modification, the set generation unit 14a groups the partial prediction targets into a plurality of groups, so that the configuration of the group sets is different from each other (the first group set and the second group). A set of groups) is generated respectively. Therefore, in addition to the effects of the above-described embodiment, the partial prediction targets can be grouped more appropriately.

Next, the outline of the present invention will be described. FIG. 13 is a block diagram showing an outline of the information analyzer according to the present invention. In the information analyzer 80 according to the present invention, when the prediction target (for example, the prediction target Y) is represented by the sum of _{a plurality of partial prediction targets (for example, partial prediction targets y 1} , y ₂ , ..., Y _n). , The unit model generation unit 81 (for example, the unit model generation unit 12) that generates a learning model for each partial prediction target, and the estimation error when the value of the partial prediction target is estimated using the training model along the time axis. With the estimation error calculation unit 82 (for example, estimation error calculation unit 13) to be calculated and the estimation error as input, the similarity between the estimation error for one learning model and the estimation error for another learning model on the time axis Based on this, a set generation unit 83 (for example, a set generation unit 14) that generates a set of groups by grouping a partial prediction target corresponding to one learning model and a partial prediction target corresponding to another learning model into a group. , Partial sum model generation unit 84 (for example, partial sum model generation unit 16) that generates a learning model with the sum of the partial prediction targets included in the group as the objective variable, and the predicted value of the group calculated using the learning model. It is provided with a prediction target calculation unit 85 (for example, a prediction target calculation unit 18) that calculates the value of the prediction target from the above.

With such a configuration, even if each of the plurality of prediction targets has different properties from each other, the desired prediction target is represented by the sum of the plurality of prediction targets. It is possible to realize a grouping of prediction targets that improves the prediction accuracy of.

Further, the information analyzer 80 includes an evaluation unit (for example, an evaluation unit 20) that evaluates a set of groups based on the difference between the calculated value of the prediction target and the actual value of the prediction target, and the already generated group. A control unit (for example, a control unit 24) that controls the set generation unit 83 so as to generate a set of groups having a configuration different from that of the set may be provided. Then, the set generation unit 83 may generate a set of groups according to the control of the control unit, and the evaluation unit may evaluate the set of each generated group.

Further, the control unit may control the estimation error calculation unit 82 so that the estimation error for the learning model generated by the partial sum model generation unit 84 is calculated. At this time, the estimation error calculation unit 82 calculates the estimation error when the value of the partial prediction target is estimated using the learning model generated by the partial sum model generation unit 84 along the time axis. Further, when controlling the set generation unit 83, the control unit may input the estimation error of the learning model generated by the partial sum model generation unit 84 into the set generation unit 83. At this time, the set generation unit 83 generates a set of groups by inputting the estimation error of the learning model generated by the partial sum model generation unit 84. With such a configuration, it becomes possible to gradually reduce the set of groups.

Further, the information analyzer 80 may include an output unit (for example, an output unit 22) that outputs a set of groups according to the evaluation of the evaluation unit. At this time, the evaluation unit uses an evaluation function (for example, the evaluation function W shown in the above equation 1) designed so that the smaller the difference between the calculated value of the prediction target and the actual value of the prediction target, the better the evaluation. It may be used to evaluate a set of groups and the output unit may output a set of groups with a better evaluation.

Further, in the evaluation unit, the smaller the difference between the calculated value of the prediction target and the actual value of the prediction target, the better the evaluation, and the smaller the number of groups constituting the group set, the better the evaluation. An evaluation function designed as described above (for example, the evaluation function W shown in the above equation 2) may be used to evaluate a set of groups.

In addition, the evaluation unit is included in each learning model so that the smaller the difference between the calculated value of the prediction target and the actual value of the prediction target, the better the evaluation, and each learning model corresponding to each group. A set of groups may be evaluated using an evaluation function (for example, the evaluation function W shown in Equation 3 above) designed so that the smaller the total number of explanatory variables having a non-zero coefficient, the better the evaluation.

FIG. 14 is a block diagram showing an outline of the information analysis system according to the present invention. In the information analysis system 90 according to the present invention, when the prediction target (for example, the prediction target Y) is represented by the sum of _{a plurality of partial prediction targets (for example, partial prediction targets y 1} , y ₂ , ..., Y _n). , The unit model generation unit 91 (for example, the unit model generation unit 12) that generates a learning model for each partial prediction target, and the estimation error when the value of the partial prediction target is estimated using the training model along the time axis. With the estimation error calculation unit 92 (for example, estimation error calculation unit 13) to be calculated and the estimation error as input, the similarity between the estimation error for one learning model and the estimation error for another learning model on the time axis Based on this, by grouping the partial prediction target corresponding to one learning model and the partial prediction target corresponding to another learning model into a group, the composition of the group set is different from that of the first group set and the second group set. For each of the set generation unit 93 (for example, the set generation unit 14a) that generates a set of groups, and the set of the first group and the set of the second group, the partial prediction targets included in the group. A learning model for each of the partial sum model generation unit 94 (for example, the partial sum model generation unit 16a) that generates a learning model with the sum as the objective variable, and the set of the first group and the set of the second group. The prediction target calculation unit 95 (for example, the prediction target calculation unit 18a) that calculates the value of the prediction target from the prediction value of the group calculated using the above, and the set of the first group and the set of the second group, respectively. On the other hand, an evaluation unit 96 (for example, an evaluation unit 20a) that evaluates a set of groups based on the difference between the value of the prediction target and the actual value of the prediction target is provided.

Even with such a configuration, even if each of the plurality of prediction targets has different properties from each other, the desired prediction target is expressed by the sum of the plurality of prediction targets. It is possible to realize grouping of prediction targets so as to improve the prediction accuracy of the targets.

Some or all of the above embodiments may also be described, but not limited to:

(Appendix 1) When the prediction target is represented by the sum of a plurality of partial prediction targets, a unit model generation unit that generates a learning model for each partial prediction target and the partial prediction target using the learning model. The time axis of the estimation error calculation unit that calculates the estimation error when the value is estimated along the time axis, and the estimation error of one learning model and the estimation error of the other learning model by inputting the estimation error. Based on the similarity in the above, the set generation unit that generates a set of groups by grouping the partial prediction target corresponding to the one learning model and the partial prediction target corresponding to the other learning model into a group, and the above-mentioned A prediction target that calculates the value of the prediction target from the partial sum model generation unit that generates a learning model with the sum of the partial prediction targets included in the group as the objective variable and the prediction value of the group calculated using the learning model. An information analyzer characterized by having a calculation unit.

(Appendix 2) An evaluation unit that evaluates a set of groups based on the difference between the calculated value of the prediction target and the actual value of the prediction target, and a set of groups having a structure different from the already generated set of groups are generated. A control unit that controls the set generation unit is provided, the set generation unit generates a set of groups according to the control of the control unit, and the evaluation unit evaluates the set of each generated group. The information analyzer according to Appendix 1.

(Appendix 3) The control unit controls the estimation error calculation unit so that the estimation error for the learning model generated by the partial sum model generation unit is calculated, and the estimation error calculation unit is generated by the partial sum model generation unit. The estimation error when the value of the partial prediction target is estimated using the training model is calculated along the time axis, and the control unit controls the set generation unit, and the control unit generates the learning model generated by the partial sum model generation unit. The information analyzer according to Appendix 2, wherein the estimation error of the above is input to the set generation unit, and the set generation unit generates a set of groups by inputting the estimation error of the learning model generated by the partial sum model generation unit.

(Appendix 4) An output unit that outputs a set of groups according to the evaluation of the evaluation unit is provided, and the evaluation unit has a better evaluation as the difference between the calculated value of the prediction target and the actual value of the prediction target is smaller. The information analyzer according to Appendix 2 or Appendix 3, which evaluates a set of groups using the evaluation function designed in (1) and outputs a set of groups having a better evaluation.

(Appendix 5) In the evaluation section, the smaller the difference between the calculated value of the prediction target and the actual value of the prediction target, the better the evaluation, and the smaller the number of groups constituting the group set, the better the evaluation. The information analyzer according to Appendix 4, which evaluates a set of groups using an evaluation function designed to be.

(Appendix 6) The evaluation unit evaluates each learning model corresponding to each group so that the smaller the difference between the calculated value of the prediction target and the actual value of the prediction target, the better the evaluation. The information analyzer according to Appendix 4, which evaluates a set of groups using an evaluation function designed so that the smaller the total number of explanatory variables having a non-zero coefficient contained in is, the better the evaluation is.

(Appendix 7) The estimation error calculation unit standardizes the estimation error of each partial prediction target, and the set generation unit generates a set of groups by inputting the standardized estimation error. The information analyzer described in one.

(Appendix 8) When the prediction target is represented by the sum of a plurality of partial prediction targets, a unit model generation unit that generates a learning model for each partial prediction target and the partial prediction target using the learning model. The time axis of the estimation error calculation unit that calculates the estimation error when the value is estimated along the time axis, and the estimation error of one learning model and the estimation error of the other learning model by inputting the estimation error. Based on the similarity in the above, by grouping the partial prediction target corresponding to the one learning model and the partial prediction target corresponding to the other learning model into a group, the composition of the set of the groups is different from each other. Partial prediction included in the group for each of the set generator that generates the set of the first group and the set of the second group, and the set of the first group and the set of the second group, respectively. Calculated using the learning model for each of the partial sum model generation unit that generates a learning model with the sum of the objects as the objective variable, the set of the first group, and the set of the second group. For each of the prediction target calculation unit that calculates the value of the prediction target from the prediction value of the group, the set of the first group and the set of the second group, the value of the prediction target and the value of the prediction target An information analysis system including an evaluation unit that evaluates a set of the groups based on a difference from an actual value.

(Appendix 9) An output unit that outputs a set of groups according to the evaluation of the evaluation unit is provided, and the evaluation unit has a better evaluation as the difference between the calculated value of the prediction target and the actual value of the prediction target is smaller. The information analysis system according to Appendix 8, which evaluates a set of groups using the evaluation function designed in 1 and outputs a set of groups having a better evaluation.

(Appendix 10) When the prediction target is represented by the sum of a plurality of partial prediction targets, a learning model is generated for each of the partial prediction targets, and the value of the partial prediction target is estimated using the learning model. Is calculated along the time axis, and the estimation error is used as an input, and the estimation error for one learning model and the estimation error for another learning model are based on the similarity on the time axis. A set of groups is generated by grouping the partial prediction target corresponding to the learning model of the above and the partial prediction target corresponding to the other learning model into a group, and the sum of the partial prediction targets included in the group is used as the objective variable. An information analysis method characterized in that a learning model is generated and the value of the prediction target is calculated from the prediction value of a group calculated using the learning model.

(Appendix 11) One or more sets of groups having a configuration different from the already generated set of groups are generated, and the set of each group is created based on the difference between the calculated value of the prediction target and the actual value of the prediction target. The information analysis method according to Appendix 10 to be evaluated.

(Appendix 12) When the prediction target is represented by the sum of a plurality of partial prediction targets, a learning model is generated for each of the partial prediction targets, and the value of the partial prediction target is estimated using the learning model. Is calculated along the time axis, and the estimation error is used as an input, and the estimation error for one learning model and the estimation error for another learning model are based on the similarity on the time axis. By grouping the partial prediction target corresponding to the training model of the above and the partial prediction target corresponding to the other learning model into a group, the composition of the set of the groups is different from that of the first group and the second group. A set is generated, and a learning model in which the sum of the partial prediction targets included in the group is the objective variable is generated for each of the set of the first group and the set of the second group. For each of the set of the first group and the set of the second group, the value of the prediction target is calculated from the predicted value of the group calculated by using the learning model, and the value of the prediction target is calculated for each of the first group. An information analysis method for evaluating each of the set and the set of the second group based on the difference between the value of the prediction target and the actual value of the prediction target.

(Appendix 13) A group with a better evaluation is evaluated by using an evaluation function designed so that the smaller the difference between the calculated value of the prediction target and the actual value of the prediction target, the better the evaluation. The information analysis method according to Appendix 12, which outputs a set of.

(Appendix 14) When the prediction target is represented by the sum of a plurality of partial prediction targets on a computer, a unit model generation process that generates a learning model for each partial prediction target, and the partial prediction using the learning model. Estimated error calculation processing that calculates the estimated error when estimating the target value along the time axis, the time between the estimated error for one learning model and the estimated error for the other learning model using the estimated error as an input A set generation process for generating a set of groups by grouping a partial prediction target corresponding to the one learning model and a partial prediction target corresponding to the other learning model based on the similarity on the axis. Partial sum model generation processing that generates a learning model with the sum of the partial prediction targets included in the group as the objective variable, and prediction that calculates the value of the prediction target from the prediction value of the group calculated using the learning model. An information analysis program for executing the target calculation process.

(Appendix 15) An evaluation process in which a computer evaluates a set of groups based on the difference between the calculated value of the prediction target and the actual value of the prediction target, and a group having a configuration different from the already generated set of groups. A control process for controlling the set generation process is executed so as to generate a set of, a set of groups is generated according to the control by the control process in the set generation process, and each group generated in the evaluation process is generated. The information analysis program according to Appendix 14 for evaluating a set.

(Appendix 16) When the prediction target is represented by the sum of a plurality of partial prediction targets on a computer, a unit model generation process that generates a learning model for each partial prediction target, and the partial prediction using the learning model. Estimated error calculation processing that calculates the estimated error when estimating the target value along the time axis, the time between the estimated error for one learning model and the estimated error for the other learning model using the estimated error as an input By grouping the partial prediction target corresponding to the one learning model and the partial prediction target corresponding to the other learning model into groups based on the similarity on the axis, the composition of the set of the groups is different from each other. A set generation process for generating a set of one group and a set of a second group, respectively, and the partial prediction included in the group for each of the set of the first group and the set of the second group. A partial sum model generation process that generates a learning model with the sum of the objects as the objective variable, a group calculated using the learning model for each of the first group set and the second group set. For each of the prediction target calculation process for calculating the value of the prediction target from the prediction value of the first group and the set of the first group and the set of the second group, the value of the prediction target and the value of the prediction target An information analysis program for executing an evaluation process that evaluates a set of the groups based on the difference from the actual value.

(Appendix 17) The computer is made to execute an output process that outputs a set of groups according to the evaluation in the evaluation process, and the smaller the difference between the calculated value of the prediction target and the actual value of the prediction target in the evaluation process, the smaller the difference. The information analysis program according to Appendix 16, wherein a set of groups is evaluated by using an evaluation function designed to give a good evaluation, and a set of groups having a better evaluation is output by the output process.

10 Reception unit 12 Unit model generation unit 13 Estimation error calculation unit 14, 14a Set generation unit 16, 16a Partial sum model generation unit 18, 18a Prediction target calculation unit 20, 20a Evaluation unit 22 Output unit 24 Control unit 30 Storage unit 100 Information Analyzer 200 Information analysis system

Claims (10)

When the prediction target is represented by the sum of multiple partial prediction targets,
A unit model generation unit that generates a learning model for each partial prediction target,
An estimation error calculation unit that calculates the estimation error when the value of the partial prediction target is estimated using the learning model along the time axis.
Using the estimation error as an input, the partial prediction target corresponding to the one learning model and the other are based on the similarity on the time axis between the estimation error for one learning model and the estimation error for the other learning model. A set generator that generates a set of groups by grouping the partial prediction targets corresponding to the learning model of
A partial sum model generation unit that generates a learning model with the sum of the partial prediction targets included in the group as the objective variable, and
An information analysis device including a prediction target calculation unit that calculates a prediction target value from a group prediction value calculated using the learning model. An evaluation unit that evaluates a set of groups based on the difference between the calculated value of the prediction target and the actual value of the prediction target,
It is provided with a control unit that controls a set generation unit so as to generate a set of groups having a configuration different from that of a group that has already been generated.
The set generation unit generates a set of groups according to the control of the control unit.
The information analyzer according to claim 1, wherein the evaluation unit evaluates a set of each generated group. The control unit controls the estimation error calculation unit so that the estimation error for the learning model generated by the partial sum model generation unit is calculated.
The estimation error calculation unit calculates the estimation error when the value of the partial prediction target is estimated using the learning model generated by the partial sum model generation unit along the time axis.
When controlling the set generation unit, the control unit inputs the estimation error for the learning model generated by the partial sum model generation unit into the set generation unit.
The information analyzer according to claim 2, wherein the set generation unit generates a set of groups by inputting an estimation error of the learning model generated by the partial sum model generation unit. Equipped with an output unit that outputs a set of groups according to the evaluation of the evaluation unit
The evaluation unit evaluates a set of groups using an evaluation function designed so that the smaller the difference between the calculated value of the prediction target and the actual value of the prediction target, the better the evaluation.
The information analyzer according to claim 2 or 3, wherein the output unit outputs a set of groups having a better evaluation. In the evaluation section, the smaller the difference between the calculated value of the prediction target and the actual value of the prediction target, the better the evaluation, and the smaller the number of groups that make up the set of groups, the better the evaluation. The information analyzer according to claim 4, which evaluates a set of groups using a designed evaluation function. When the prediction target is represented by the sum of multiple partial prediction targets,
A unit model generation unit that generates a learning model for each partial prediction target,
An estimation error calculation unit that calculates the estimation error when the value of the partial prediction target is estimated using the learning model along the time axis.
Using the estimation error as an input, the partial prediction target corresponding to the one learning model and the other are based on the similarity on the time axis between the estimation error for one learning model and the estimation error for the other learning model. By grouping the partial prediction targets corresponding to the learning model of, the set generator that generates the set of the first group and the set of the second group, which have different configurations of the set of the groups, and a set generation unit.
For each of the set of the first group and the set of the second group, a partial sum model generation unit that generates a learning model with the sum of the partial prediction targets included in the group as an objective variable, and
For each of the set of the first group and the set of the second group, a prediction target calculation unit that calculates the value of the prediction target from the prediction value of the group calculated by using the learning model, and a prediction target calculation unit.
For each of the set of the first group and the set of the second group, an evaluation unit that evaluates the set of the group based on the difference between the value of the prediction target and the actual value of the prediction target. An information analysis system characterized by being equipped. When the prediction target is represented by the sum of multiple partial prediction targets,
A learning model is generated for each of the partial prediction targets,
The estimation error when the value of the partial prediction target is estimated using the learning model is calculated along the time axis.
Using the estimation error as an input, the partial prediction target corresponding to the one learning model and the other are based on the similarity on the time axis between the estimation error for one learning model and the estimation error for the other learning model. A set of groups is generated by grouping the partial prediction targets corresponding to the learning model of
A learning model with the sum of the partial prediction targets included in the group as the objective variable is generated.
An information analysis method characterized in that the value of the prediction target is calculated from the prediction value of the group calculated using the learning model. When the prediction target is represented by the sum of multiple partial prediction targets,
A learning model is generated for each of the partial prediction targets,
The estimation error when the value of the partial prediction target is estimated using the learning model is calculated along the time axis.
Using the estimation error as an input, the partial prediction target corresponding to the one learning model and the other are based on the similarity on the time axis between the estimation error for one learning model and the estimation error for the other learning model. By grouping the partial prediction targets corresponding to the learning model of, a first group set and a second group set having different configurations of the group sets are generated.
For each of the set of the first group and the set of the second group, a learning model in which the sum of the partial prediction targets included in the group is the objective variable is generated.
For each of the set of the first group and the set of the second group, the value of the prediction target is calculated from the prediction value of the group calculated by using the learning model.
For each of the set of the first group and the set of the second group, the set of the group is evaluated based on the difference between the value of the prediction target and the actual value of the prediction target. Information analysis method to do. On the computer
When the prediction target is represented by the sum of multiple partial prediction targets,
Unit model generation process that generates a learning model for each partial prediction target,
Estimation error calculation processing that calculates the estimation error when the value of the partial prediction target is estimated using the learning model along the time axis.
Using the estimation error as an input, the partial prediction target corresponding to the one learning model and the other are based on the similarity on the time axis between the estimation error for one learning model and the estimation error for the other learning model. A set generation process that generates a set of groups by grouping the partial prediction targets corresponding to the learning model of
A partial sum model generation process that generates a learning model with the sum of the partial prediction targets included in the group as the objective variable, and
An information analysis program for executing a prediction target calculation process that calculates the value of the prediction target from the prediction value of the group calculated using the learning model. On the computer
When the prediction target is represented by the sum of multiple partial prediction targets,
Unit model generation process that generates a learning model for each partial prediction target,
Estimation error calculation processing that calculates the estimation error when the value of the partial prediction target is estimated using the learning model along the time axis.
Using the estimation error as an input, the partial prediction target corresponding to the one learning model and the other are based on the similarity on the time axis between the estimation error for one learning model and the estimation error for the other learning model. A set generation process that generates a set of a first group and a set of a second group, which have different configurations of the sets of the groups, by grouping the partial prediction targets corresponding to the learning model of.
A partial sum model generation process for generating a learning model with the sum of the partial prediction targets included in the group as an objective variable for each of the set of the first group and the set of the second group.
For each of the set of the first group and the set of the second group, a prediction target calculation process for calculating the value of the prediction target from the prediction value of the group calculated using the learning model, and a prediction target calculation process.
For each of the first group set and the second group set, an evaluation process is executed to evaluate the set of the group based on the difference between the value of the prediction target and the actual value of the prediction target. Information analysis program to let you.

Images