JP7742779B2 - Computer system and predictive model learning method - Google Patents

Description

The present invention relates to a system and method for predicting demand for items.

Item demand forecasting technology using predictive models is used to determine warehouse labor, inventory, and replenishment plans. The features handled by the predictive model are determined when the model is constructed.

As the items handled in a warehouse or the shipping patterns of the entire warehouse change over time, features may no longer contribute to predictions. Furthermore, when applying an existing prediction model to a new warehouse or new items, features must be adjusted to suit the target.

Adjusting the feature values requires a lot of effort. To address this issue, the technology described in Patent Document 1 is known.

Patent Document 1 states that "the demand forecasting method for predicting future demand includes the steps of: calculating an error between a predicted value of past demand, calculated by inputting actual measured values of explanatory variables into a forecasting model constructed based on actual measured values of demand and explanatory variables, which are external factors that affect increases or decreases in demand, and the actual measured values of past demand (S1); determining whether the calculated error is an abnormal value (S2); if it is determined that the error is an abnormal value, acquiring new explanatory variables (S3); updating the forecasting model based on the acquired new explanatory variables (S4); and newly predicting future demand using the updated forecasting model (S5)."

Japanese Patent Application Laid-Open No. 2017-016632

The technology described in Patent Document 1 acquires features input by the user or features from other existing models. User input has the problem of high cost. When acquiring features from other models, the features may not necessarily contribute to the prediction of the target.

The present invention provides technology related to a method for adjusting the features of a prediction model to improve prediction accuracy.

A representative example of the invention disclosed in this application is as follows. a first analysis period within the first period and a second analysis period within the second period, the first analysis period being a period in which there is a large error between the predicted shipment volume and the actual shipment volume, based on the demand forecast for the second item and the actual demand for the second item; a first analysis period within the first period and a second analysis period within the second period being a period in which there is a large error between the predicted shipment volume and the actual shipment volume , the first analysis period being a period in which there is a large error between the predicted shipment volume and the actual shipment volume, the second ...

The present invention makes it possible to generate models that include features that contribute to improving prediction accuracy. Issues, configurations, and advantages other than those described above will become clearer in the following examples.

FIG. 1 illustrates an example of a configuration of a demand forecasting system according to a first embodiment. FIG. 2 is a diagram illustrating details of a functional unit of a computer according to a first embodiment. FIG. 2 is a diagram illustrating an example of a data structure of shipping history management information according to the first embodiment. FIG. 10 is a diagram illustrating an example of a data structure of analysis period information according to the first embodiment. FIG. 4 is a diagram illustrating an example of a data structure of feature amount information according to the first embodiment. 10 is a flowchart illustrating an outline of a process executed by a feature amount adjustment unit according to the first embodiment. 10 is a flowchart illustrating an example of a change pattern generation process executed by a feature amount adjustment unit according to the first embodiment. 10 is a flowchart illustrating an example of a change pattern generation process executed by a feature amount adjustment unit according to the first embodiment. 10 is a flowchart illustrating an example of a verification process executed by a feature amount adjustment unit according to the first embodiment. FIG. 4 is a diagram illustrating an example of a screen displayed by a display unit according to the first embodiment. FIG. 4 is a diagram illustrating an example of a screen displayed by a display unit according to the first embodiment.

Below, examples of the present invention are described using the drawings. However, the present invention should not be construed as being limited to the description of the examples shown below. Those skilled in the art will readily understand that the specific configuration can be modified without departing from the spirit or intent of the present invention. In the configuration of the invention described below, identical or similar components or functions are designated by the same reference numerals, and redundant explanations will be omitted. Designations such as "first," "second," and "third" in this specification are used to identify components and do not necessarily limit the number or order. The position, size, shape, and range of each component shown in the drawings may not represent the actual position, size, shape, and range, etc., in order to facilitate understanding of the invention. Therefore, the present invention is not limited to the position, size, shape, and range, etc., disclosed in the drawings, etc.

Figure 1 is a diagram showing an example of the configuration of a demand forecasting system according to the first embodiment.

The demand forecasting system is a computer system consisting of at least one computer 100. The computer 100 has a CPU 110, an IO interface 111, a network interface 112, a main memory device 113, and a secondary memory device 114. Each piece of hardware is connected to each other via a bus.

The CPU 110 executes programs stored in the main memory device 113. By executing processing in accordance with the programs, the CPU operates as a functional unit (module) that realizes a specific function. In the following explanation, when processing is described using a functional unit as the subject, this indicates that the CPU is executing a program that realizes that functional unit.

The IO interface 111 is an interface for connecting to external devices. The computer 100 connects to the input device 101 and output device 102 via the IO interface 111. The input device 101 is a keyboard, mouse, touch panel, etc. The output device 102 is a display, printer, etc.

The network interface 112 is an interface for connecting to external devices via the network 103. The network 103 may be a WAN (Wide Area Network) or a LAN (Local Area Network), etc. The connection type of the network 103 may be either wired or wireless.

The main memory device 113 stores the programs executed by the CPU 110 and the data used by the programs. The main memory device 113 also includes a work area temporarily used by the programs. The main memory device 113 is, for example, a dynamic random access memory (DRAM).

The main memory device 113 stores programs that implement the input unit 120, prediction unit 121, feature adjustment unit 122, and display unit 123. The main memory device 113 also stores programs (not shown) such as an OS (Operating System), but these are omitted in this embodiment.

The input unit 120 accepts various inputs from the user via the input device 101 or the network 103. The prediction unit 121 predicts the demand for an item over a prediction period and outputs the prediction results. The prediction period is a time range for predicting the demand for an item, such as a year, month, or week. Note that the prediction period is not limited to these. It is assumed that the prediction period is set in advance. The feature adjustment unit 122 adjusts the features of an existing prediction model when the prediction accuracy of the existing prediction model deteriorates, or when an existing prediction model is used to generate a prediction model to be applied to a new prediction target. The display unit 123 presents the prediction results and information related to the features to the user via the output device 102. Note that the computer 100 may accept input from the user by receiving data from a user terminal communicatively connected via the network 103, and output data to the user terminal by transmitting the output to be output by the output device 102 to the user terminal.

The secondary storage device 114 permanently stores data. Examples of the secondary storage device 114 include a hard disk drive (HDD) and a solid state drive (SSD).

The secondary storage device 114 stores shipping history management information 130, input data management information 131, and prediction model management information 132. Note that the secondary storage device 114 also stores information not shown, such as information for managing item characteristics, but this information is omitted in this embodiment.

Shipping history management information 130 stores data (shipping history) related to past shipments of items. The shipping history includes the amount of items shipped during a measurement period. Measurement periods can be, for example, days, weeks, or months. However, the measurement period is not limited to these. The data structure of shipping history management information 130 will be explained using Figure 2.

Input data management information 131 stores the input data to be input into the prediction model. The input data includes values for items such as item type, item arrival quantity, item shipment quantity, and whether or not there is a sale. Input data may be input directly to the prediction model, or input data that has undergone preprocessing may be input. The shipping history itself may also be treated as input data. In this case, input data management information 131 can be considered the same as shipping history management information 130.

Prediction model management information 132 stores definition information for mathematical models (prediction models) that predict item demand. Prediction model management information 132 stores information related to feature quantities, item characteristics such as item type and category, and model definition information according to warehouse characteristics. Prediction models are, for example, decision trees and neural networks. Note that the present invention is not limited to the type and structure of the prediction model.

The demand forecasting system may be composed of multiple computers 100. In this case, functional units are distributed across the multiple computers 100. For example, the input unit 120 is located on one computer 100, the prediction unit 121 is located on another computer 100, the feature adjustment unit 122 is located on another computer 100, and the display unit 123 is located on another computer 100.

The programs and data stored in the main memory device 113 may also be stored in the secondary memory device 114. In this case, the CPU 110 reads the programs and data from the secondary memory device 114 and stores them in the main memory device 113.

Figure 2 is a diagram illustrating the details of the functional units of the computer 100 of Example 1.

The input unit 120 inputs input data information 211, which includes input data for the second item during the first period, to the prediction unit 121; inputs shipping history information 212, which includes the shipping history for the second item during the second period, to the feature adjustment unit 122; and inputs a prediction model 213 for the first item to the prediction unit 121 and the feature adjustment unit 122. Note that the measurement period of the input data information 211 is the same as the measurement period of the shipping history information 212. The time span of the first period is the same as the time span of the second period. The time span of the first period is, for example, one year.

Here, the combination of item and period is one of the following:
(Pattern 1) The first item and the second item are the same, but the first period and the second period are different.
(Pattern 2) The first period and the second period are the same, but the first item and the second item are different.
(Pattern 3) The first item and the second item are different, and the first period and the second period are different.

Pattern 1 is intended for situations where features need to be adjusted due to a decline in the predictive accuracy of an existing model. Patterns 2 and 3 are intended for situations where a predictive model for a new item is to be generated.

The prediction unit 121 obtains the prediction model 213 for the first item, inputs the input data information 211 into the prediction model 213, generates predicted shipping quantity information 214, and outputs it to the feature adjustment unit 122.

The analysis period identification unit 201 of the feature adjustment unit 122 extracts an analysis period based on the shipping history information 212 and the forecast shipping volume information 214, and generates analysis period information 215. The analysis period is a period shorter than the measurement period, and is the period during which analysis related to features is performed.

The change pattern generation unit 202 generates change patterns for features from the existing model based on the differences between the two shipping patterns during the analysis period, and outputs them as a change pattern list 216.

The verification unit 203 changes the features of the prediction model 213 based on the prediction model 213 and the change pattern list 216, and performs learning of the prediction model 213. The verification unit 203 also causes the prediction unit 121 to perform predictions using the updated prediction model 213. Based on the shipment volume prediction results, the verification unit 203 identifies features that contribute to improving prediction accuracy and selects features to be used in a new prediction model. The verification unit 203 outputs the processing results to the display unit 123 as feature information 217.

The display unit 123 presents information about the feature amount based on the feature amount information 217 output from the feature amount adjustment unit 122.

Regarding each functional unit of the computer 100, multiple functional units may be combined into one functional unit, or one functional unit may be divided into multiple functional units for each function. For example, the change pattern generation unit 202 may include an analysis period identification unit 201.

Instead of the input data information 211, the input unit 120 may input shipping history information of the second item, which has a small error with respect to the prediction using the prediction model 213, to the feature adjustment unit 122. In this case, the analysis period identification unit 201 extracts the analysis period based on the two shipping histories.

Figure 3 is a diagram showing an example of the data structure of the shipping history management information 130 in Example 1.

Shipping history management information 130 stores the shipping history of multiple items. Specifically, shipping history management information 130 stores entries including item ID 301, shipping date 302, and shipping quantity 303. One entry corresponds to one shipping history. The shipping history management information 130 shown in Figure 3 stores shipping history for a measurement period of one day.

Item ID 301 is a field that stores the item's identification information. Shipping Date 302 is a field that stores the item's shipping date. Shipping Amount 303 is a field that stores the item's shipping amount per day.

Figure 4 shows an example of the data structure of the analysis period information 215 in Example 1.

The analysis period information 215 stores entries including a period ID 401, an analysis period 402, and a difference pattern 403. One entry corresponds to one analysis period.

Period ID 401 is a field that stores identification information for the analysis period. Analysis period 402 is a field that stores the analysis period. Difference pattern 403 is a field that stores a value that indicates the pattern of differences between the time series of two shipments. Difference pattern 403 stores either "first" or "second." "First" indicates that there is a characteristic peak in the time series of the shipment forecast for the second item. "Second" indicates that there is a characteristic peak in the time series of the shipping history for the second item.

In the following description, an analysis period in which difference pattern 403 is "first" will be referred to as a first type analysis period, and an analysis period in which difference pattern 403 is "second" will be referred to as a second type analysis period.

Figure 5 shows an example of the data structure of feature information 217 in Example 1.

Feature information 217 stores entries including a feature name 501, a feature description 502, a corresponding feature name 503, and a new flag 504. One entry corresponds to one feature.

Feature name 501 is a field that stores the name of the feature. Feature description 502 is a field that stores an explanation of the feature, such as the meaning of the feature and the values that the feature can take. Corresponding feature name 503 is a field that stores the name of the feature used when generating the feature. New flag 504 is a field that stores a flag indicating whether the feature is newly generated. New flag 504 stores either "1" indicating a newly generated feature or "0" indicating an existing feature.

Figure 6 is a flowchart outlining the processing performed by the feature amount adjustment unit 122 in Example 1.

The feature adjustment unit 122 acquires the predicted shipping quantity information 214 for the second item and the shipping history information 212 for the second item (step S101).

Next, the analysis period identification unit 201 of the feature amount adjustment unit 122 identifies an analysis period based on the predicted shipping volume information 214 and the shipping history information 212, and generates analysis period information 215 (step S102). Specifically, the following processing is performed:

(S102-1) The analysis period determination unit 201 sets the variable t1 to the start date and time of the first period in the forecast shipping volume information 214, and sets the variable t2 to the start date and time of the second period in the shipping history information 212.

(S102-2) The analysis period determination unit 201 acquires a shipping forecast time series for the processing time range from variable t1 from the forecast shipping quantity information 214, and acquires a shipping history time series for the processing time range from variable t2 from the shipping history information 212. Here, the processing time range is a period shorter than the first period and can be set arbitrarily. The processing time range is, for example, one month.

(S102-3) The analysis period identification unit 201 calculates the absolute value of the error in the shipment volume for each measurement period within the processing time range, and also calculates the sum of the absolute values of the errors.

(S102-4) The analysis period identification unit 201 determines whether the sum of the absolute values of the errors is greater than the first threshold and whether the maximum absolute value of the errors in the shipment volume during the measurement period is greater than the second threshold. The first threshold and the second threshold can be set arbitrarily. For example, the first threshold can be set to twice the average error.

(S102-5) If the above conditions are met, the analysis period identification unit 201 registers the time range as the analysis period in the analysis period information 215. The value of the difference pattern 403 is set based on the error. That is, if there is a peak in the shipping history time series, the difference pattern 403 is "second," and if there is a peak in the shipping forecast time series, the difference pattern 403 is "first."

(S102-6) The analysis period identification unit 201 adds the processing time range to variables t1 and t2.

(S102-7) The analysis period identification unit 201 determines whether the variable t1 has not passed the end time of the first period and whether the variable t2 has not passed the end time of the second period. If the above conditions are not met, the analysis period identification unit 201 returns to S102-2 and executes the same processing. If the above conditions are not met, the analysis period identification unit 201 ends the processing of step S102.

As described above, in this embodiment, a period in which there is a large error and a characteristic peak exists in either the shipping history time series or the shipping forecast time series for the second item is extracted as the analysis period.

Next, the change pattern generation unit 202 of the feature adjustment unit 122 executes a change pattern generation process (step S103). Details of the change pattern generation process will be explained using Figures 7A and 7B.

Next, the verification unit 203 of the feature amount adjustment unit 122 executes a verification process (step S104) and outputs the feature amount information 217 to the display unit 123 (step S105). Details of the verification process will be explained using Figure 8.

Figures 7A and 7B are flowcharts illustrating an example of the change pattern generation process executed by the feature amount adjustment unit 122 in Example 1.

The change pattern generation unit 202 references the analysis period information 215 and determines whether only the first type analysis period exists (step S201).

If only a first type analysis period exists, the change pattern generation unit 202 starts loop processing of the first type analysis period (step S202). Specifically, the change pattern generation unit 202 selects one first type analysis period.

The change pattern generation unit 202 generates additional candidate features based on the time series of shipment forecasts for the selected first type analysis period in the forecast shipment quantity information 214 (step S203).

Specifically, the change pattern generation unit 202 generates a feature related to the date and time when the shipping volume is at its highest (peak date and time) as an additional candidate feature. This feature is, for example, a flag indicating a sale day, and is set to "1" if the current date and time is a sale day, and "0" if the current date and time is not a sale day.

The change pattern generation unit 202 determines whether processing has been completed for all first type analysis periods (step S204).

If processing has not been completed for all first-type analysis periods, the change pattern generation unit 202 returns to step S202 and performs the same processing.

When processing has been completed for all type-1 analysis periods, the change pattern generation unit 202 generates a change pattern (step S205). The change pattern is composed of identification information and identification information for one or more feature quantities. The change pattern generation unit 202 registers the generated change pattern in the change pattern list 216.

Specifically, the change pattern generation unit 202 generates a change pattern that includes an existing feature and one or more candidate features to add. If there are multiple candidate features to add, two or more change patterns are generated. Note that an upper limit may be set on the number of change patterns to be generated. Note that the change pattern includes the existing feature as is.

If it is determined in step S201 that an analysis period other than a first type analysis period is included, the change pattern generation unit 202 determines whether only a second type analysis period exists (step S206).

If only second type analysis periods exist, the change pattern generation unit 202 starts loop processing of the second type analysis periods (step S207). Specifically, the change pattern generation unit 202 selects one second type analysis period.

The change pattern generation unit 202 identifies candidate features for deletion based on the time series of shipping history for the selected second type analysis period in the shipping history information 212 (step S208).

Specifically, the change pattern generation unit 202 identifies feature quantities related to the date and time when the shipping volume is at its highest (peak date and time) as feature quantities to be deleted.

The change pattern generation unit 202 determines whether processing has been completed for all second type analysis periods (step S209).

If processing has not been completed for all second type analysis periods, the change pattern generation unit 202 returns to step S207 and performs the same processing.

When processing has been completed for all second-type analysis periods, the change pattern generation unit 202 generates a change pattern (step S210). The change pattern generation unit 202 registers the generated change pattern in the change pattern list 216.

Specifically, the change pattern generation unit 202 generates a change pattern by deleting one or more deletion candidate features from existing features. If there are multiple deletion candidate features, two or more change patterns are generated. Note that an upper limit may be set on the number of change patterns to be generated. Note that the change pattern includes existing features other than the deletion candidate features as they are.

If it is determined in step S206 that the first type analysis period and the second type analysis period are included, the change pattern generation unit 202 generates a period pair by combining the first type analysis period and the second type analysis period (step S211).

The change pattern generation unit 202 starts loop processing of period pairs (step S212). Specifically, the change pattern generation unit 202 selects one period pair.

The change pattern generation unit 202 generates candidate features to add and identifies candidate features to delete (step S213). The method for generating candidate features to add is the same as that in step S203, and the method for identifying candidate features to delete is the same as that in step S208. In generating candidate features to add, the change pattern generation unit 202 may generate candidate features to add based on candidate features to delete. For example, if the feature related to the first day of the second type analysis period is a candidate feature to delete, the change pattern generation unit 202 generates the feature related to the first day of the first type analysis period as a candidate feature to add.

The change pattern generation unit 202 determines whether processing has been completed for all period pairs (step S214).

If processing has not been completed for all period pairs, the change pattern generation unit 202 returns to step S212 and performs the same processing.

When processing has been completed for all period pairs, the change pattern generation unit 202 generates a change pattern (step S215). The change pattern generation unit 202 registers the generated change pattern in the change pattern list 216.

Specifically, the change pattern generation unit 202 generates a change pattern by adding one or more feature quantities candidate for addition and deleting one or more feature quantities candidate for deletion from existing feature quantities. Note that an upper limit may be set on the number of change patterns to be generated. Note that the change pattern includes existing feature quantities other than feature quantities candidate for deletion as they are.

Figure 8 is a flowchart illustrating an example of the verification process performed by the feature amount adjustment unit 122 in Example 1.

The verification unit 203 acquires the prediction model 213 and the change pattern list 216 (step S301), and also acquires the input data information 211 and shipping history information for the second item (step S302). The shipping history information for the second item is treated as the correct answer data for the prediction.

The verification unit 203 starts loop processing of the change patterns (step S303). Specifically, the verification unit 203 selects one change pattern.

The verification unit 203 sets features based on the change pattern and executes a learning process for the prediction model 213 (step S304). This generates a new prediction model from the existing prediction model 213. Note that a well-known learning method can be used, so a detailed explanation will be omitted.

The verification unit 203 inputs the new prediction model and input data information 211 to the prediction unit 121 and instructs it to perform a prediction (step S305). The verification unit 203 obtains predicted shipping volume information 214 from the prediction unit 121.

The verification unit 203 calculates an evaluation index for the new prediction model based on the shipping history information and forecast shipping quantity information 214 for the second item (step S306). For example, the verification unit 203 calculates the mean square error between the shipping record and the shipping forecast.

The verification unit 203 determines whether processing has been completed for all change patterns (step S307).

If processing has not been completed for all change patterns, the verification unit 203 returns to step S303 and performs the same processing.

When processing has been completed for all change patterns, the verification unit 203 selects a feature based on the evaluation index (step S308). For example, the verification unit 203 may select a feature based on the change pattern that most improves prediction accuracy, or the verification unit 203 may select a feature by combining multiple change patterns that improve prediction accuracy.

Figures 9A and 9B are diagrams showing examples of screens displayed by the display unit 123 in Example 1.

The display unit 123 generates display information for displaying the screen 900 based on the feature information 217.

Screen 900 displays feature information 217. When the user double-clicks a candidate feature for addition, the display unit 123 displays screen 910 as shown in Figure 9B.

Screen 910 displays a graph showing the shipping forecast for the item. Symbols indicating the dates and times corresponding to the candidate features to be added and symbols indicating the dates and times corresponding to the existing features used to generate the candidate features to be added are superimposed on the graph.

In addition, the change patterns and evaluation indices may be displayed on screen 900, allowing the user to select features.

According to the present invention, the feature adjustment unit 122 can select features that contribute to improving the prediction accuracy of item shipment volume over a given period, and generate a prediction model with high prediction accuracy.

The present invention is not limited to the above-described embodiments, but includes various modifications. Furthermore, for example, the above-described embodiments provide detailed descriptions of the configurations in order to clearly explain the present invention, and the present invention is not necessarily limited to those that include all of the described configurations. Furthermore, it is possible to add, delete, or replace part of the configuration of each embodiment with another configuration.

Furthermore, the above-described configurations, functions, processing units, processing means, etc. may be implemented in part or in whole in hardware, for example, by designing them as integrated circuits. The present invention can also be realized by software program code that implements the functions of the embodiments. In this case, a storage medium on which the program code is recorded is provided to a computer, and a processor in the computer reads the program code stored on the storage medium. In this case, the program code read from the storage medium itself implements the functions of the above-described embodiments, and the program code itself and the storage medium on which it is stored constitute the present invention. Examples of storage media for providing such program code include flexible disks, CD-ROMs, DVD-ROMs, hard disks, SSDs (Solid State Drives), optical disks, magneto-optical disks, CD-Rs, magnetic tape, non-volatile memory cards, and ROMs.

Furthermore, the program code that realizes the functions described in this embodiment can be implemented in a wide range of programming or scripting languages, such as assembler, C/C++, Perl, Shell, PHP, Python, and Java (registered trademark).

Furthermore, the software program code that realizes the functions of the embodiments may be distributed via a network and stored on a storage means such as a computer's hard disk or memory, or on a storage medium such as a CD-RW or CD-R, and the processor of the computer may read and execute the program code stored on the storage means or storage medium.

In the above examples, the control lines and information lines shown are those deemed necessary for explanation, and not all control lines or information lines in the product are necessarily shown. All components may be interconnected.

100 Computer 101 Input device 102 Output device 103 Network 110 CPU
111 IO interface 112 Network interface 113 Main memory device 114 Sub-memory device 120 Input unit 121 Prediction unit 122 Feature amount adjustment unit 123 Display unit 130 Shipping history management information 131 Input data management information 132 Prediction model management information 201 Analysis period identification unit 202 Change pattern generation unit 203 Verification unit 211 Input data information 212 Shipping history information 213 Prediction model 214 Predicted shipping amount information 215 Analysis period information 216 Change pattern list 217 Feature amount information 900, 910 Screen

Claims (10)

A computer system comprising at least one computer,
The system holds model information for managing a model for predicting the demand for an item and actual result information for managing the actual demand for the item,
the model information includes information about a first model that uses, as input, feature quantities calculated using input data including items related to shipment of a first item and predicts the demand for the first item;
The at least one computer
calculating a demand forecast for the second item in the first period by inputting into the first model a feature calculated using input data including items related to shipment of the second item in the first period;
From the performance information, a demand performance of the second item during a second period having the same time width as the first period is obtained;
identifying a first analysis period within the first period and a second analysis period within the second period in which an error between the predicted shipping amount and the actual shipping amount is large based on the demand forecast for the second item and the actual demand for the second item;
Analyzing a peak occurrence time of the shipping amount of the first item in the first analysis period and a peak occurrence time of the shipping amount of the second item in the second analysis period;
A computer system characterized in that model learning is performed using, as input, feature quantities that have been changed based on the results of the analysis. 2. The computer system of claim 1,
The at least one computer
Calculating an error between the demand forecast and the actual demand for the second item for each unit time interval along the time series;
identifying a partial period of the first period in which the error is greater than a threshold and a peak in the shipping volume of the second item exists as the first analysis period;
a partial period of the second period in which the error is greater than a threshold value and in which a peak in the shipment volume of the second item occurs is identified as the second analysis period. 3. The computer system according to claim 2,
The at least one computer
identifying, based on the first model, a feature associated with a time when a peak in the shipment volume of the first item occurs during the first analysis period as a deletion candidate feature;
generating, as an additional candidate feature, a feature associated with a time when a peak in the shipment volume of the second item occurs during the second analysis period based on the actual demand for the second item;
generating a feature change pattern based on the deletion candidate feature and the addition candidate feature;
A computer system that learns a model using, as an input, feature amounts changed based on the feature amount change pattern. 4. The computer system according to claim 3,
The at least one computer
generating a pair of the first analysis period and the second analysis period;
a computer system that generates the additional candidate feature based on a feature associated with a time when a peak in shipping volume of the first item occurs during the first analysis period; 4. The computer system according to claim 3,
The at least one computer
Evaluating the predictive accuracy of the model generated by the learning;
Selecting the feature quantity to be handled by the model based on the result of the evaluation;
A computer system that presents the selected feature quantities. A method for training a predictive model executed by a computer system having at least one computer, comprising:
the computer system holds model information for managing a model for predicting demand for an item, and performance information for managing actual demand performance for the item;
the model information includes information about a first model that uses, as input, feature quantities calculated using input data including items related to shipment of a first item and predicts the demand for the first item;
a first step in which the at least one computer calculates a demand forecast for the second item in the first time period by inputting, into the first model, feature quantities calculated using input data including items related to shipment of the second item in the first time period ;
a second step in which the at least one computer acquires, from the performance information, actual demand data for the second item during a second period having the same time duration as the first period;
a third step in which the at least one computer identifies a first analysis period within the first period and a second analysis period within the second period in which there is a large error between the shipping volume prediction result and the shipping volume actual result , based on the demand forecast for the second item and the demand actual result for the second item;
a fourth step in which the at least one computer analyzes a peak occurrence time of the shipping amount of the first item in the first analysis period and a peak occurrence time of the shipping amount of the second item in the second analysis period;
a fifth step in which the at least one computer executes model training using the feature values modified based on the results of the analysis as input . The prediction model learning method according to claim 6,
The third step includes:
a step of calculating an error between the demand forecast and actual demand for the second item for each unit time interval along a time series by the at least one computer;
the at least one computer identifying, as the first analysis period, a partial period of the first period in which the error is greater than a threshold and in which a peak in the shipment volume of the second item occurs;
and the at least one computer specifies as the second analysis period a partial period of the second period in which the error is greater than a threshold and in which a peak in the shipment volume of the second item occurs. The prediction model learning method according to claim 7,
The fourth step includes:
a sixth step in which the at least one computer identifies, based on the first model, a feature associated with a time when a peak in the shipment volume of the first item occurs in the first analysis period, as a deletion candidate feature;
a seventh step in which the at least one computer generates, as additional candidate features, feature quantities related to the occurrence time of a peak in the shipment volume of the second item in the second analysis period, based on the actual demand for the second item;
an eighth step of generating, by the at least one computer, a feature change pattern based on the deletion candidate feature and the addition candidate feature;
The fifth step is a method for training a prediction model, characterized in that it includes a step in which the at least one computer trains a model using, as input, features modified based on the feature modification pattern. The prediction model learning method according to claim 8,
The seventh step includes:
generating a pair of the first analysis period and the second analysis period by the at least one computer;
and generating the additional candidate features based on features related to the timing of a peak in the shipment volume of the first item during the first analysis period. The prediction model learning method according to claim 8,
a step of evaluating, by the at least one computer, the predictive accuracy of the model generated by the training;
selecting, by the at least one computer, the feature quantities to be handled by the model based on the results of the evaluation;
and a step in which the at least one computer presents the selected feature quantities.