JP6972641B2 - Information processing equipment and information processing programs - Google Patents

Description

The present invention relates to an information processing apparatus and an information processing program.

Conventionally, prediction target values have been predicted using machine learning. As an example of prediction processing using machine learning, learning processing (so-called) based on past actual values for a target item (objective variable) and a plurality of related items (explanatory variables) related to the target item. Supervised learning) is performed to construct a prediction model, and a process of predicting a prediction target value based on the prediction model and the values of related items related to the prediction target value is known.

Historical data may have a large number of related items, some of which may not significantly affect the value of the target item. Therefore, in order to improve the prediction accuracy of the prediction model, the prediction model may be constructed using the values for the items (feature items) selected from the plurality of related items of the past data. Information consisting of a plurality of feature items and values for them is called a feature.

Conventionally, a technique has been proposed in which a feature item is automatically selected from a plurality of related items possessed by past data. For example, in Patent Document 1, in order to construct a more effective prediction model, an appropriate attribute (feature item) is extracted from a plurality of attributes (related items) based on the weight of the support vector machine. , It is described to generate a feature function for building a predictive model. Further, Patent Document 2 also describes a process of extracting a combination of data items having the highest prediction accuracy from a plurality of data items (related items).

Japanese Unexamined Patent Publication No. 2005-09261 Japanese Unexamined Patent Publication No. 2015-072644

As described above, a more accurate prediction model can be constructed by selecting an appropriate feature item from a plurality of related items of the past data. Here, a more accurate prediction model can be constructed by using a novel feature item defined based on a combination of a plurality of related items. For example, when forecasting the sales of a certain store and having "month (month)" and "week (week of the month)" as related items included in the past data, "12". It may be possible to build a more accurate prediction model by using the new element item "whether or not it is the 4th week of the month". Of course, it is important that the new feature items are generated not based on the unreasonable combination of multiple related items included in the past data, but on the appropriate combination that improves the prediction accuracy of the prediction model. Become.

An object of the present invention is to automatically add a new feature item based on a combination of a plurality of related items possessed by past data to a feature used when predicting a prediction target value by using machine learning.

The invention according to claim 1 is a learning data group that is a part of a past data group including past actual values for a target item to be predicted and a related item group related to the target item, and the related item group. Based on the identity definition information in which the identity item group that is a candidate for the item used for prediction is defined, each predicted value for the target item of each verification data that is a part of the past data group and is the target of verification is set. For each of the prediction unit to be predicted, the error calculation unit for calculating the error between the predicted value predicted by the prediction unit and the actual value for the target item for each of the verification data, and the identity item included in the identity definition information. For each of the plurality of elemental combinations selected from the group and the plurality of elemental combinations consisting of the values for the plurality of elemental items, the elemental combination for calculating the representative value of the plurality of errors related to the plurality of verification data corresponding to the elemental combination. From the error calculation unit and the plurality of element combinations, a new element item defined by the specific element combination specified based on the representative value of the error corresponding to each element combination is generated, and the element definition information is generated. It is an information processing apparatus characterized by being provided with a new element item addition unit to be added to.

In the invention according to claim 2, the new feature item addition unit identifies the new feature item, and then the error regarding the verification data group calculated based on the pseudo-feature definition information to which the new feature item is provisionally added. When the average value of is smaller than the average value of the errors related to the verification data group calculated based on the feature definition information before the new feature item is provisionally added, the new feature item is referred to as the feature definition information. The information processing apparatus according to claim 1, wherein the information processing device is added to the above.

In the invention according to claim 3, a plurality of feature items included in the feature item group have a hierarchical relationship, and the feature combination error calculation unit has one feature item and the one feature in the hierarchical relationship. The information processing apparatus according to claim 1, wherein the feature combination is defined by combining the other feature item belonging to the layer adjacent to the layer to which the item belongs.

In the invention according to claim 4, the past data group is composed of a plurality of past data arranged in a time series, and the feature combination includes a value related to the past data of interest for a feature item selected from the feature item group. The information processing apparatus according to claim 1, wherein the information processing apparatus includes a value related to past data other than the attention past data for the feature item selected from the feature item group.

In the invention according to claim 5, the prediction unit calculates the degree of contribution indicating the magnitude of the influence of the value on each feature item on the value of the target item for each feature item included in the feature item group. The information processing apparatus according to claim 1, wherein the feature combination error calculation unit does not select a feature item whose contribution is equal to or less than a threshold value in defining the feature combination.

In the invention according to claim 6, the new feature item addition unit has the specific feature based on the representative value of the error corresponding to each feature combination and the contribution of a plurality of feature items included in each feature combination. The information processing apparatus according to claim 5, wherein the combination is specified.

The invention according to claim 7 comprises a learning data group which is a part of a past data group including past actual values for a target item to be predicted and a related item group related to the target item, and the above-mentioned invention. Each of the target items of each verification data that is a part of the past data group and is the target of verification based on the predisposition definition information in which the predisposition item group that is a candidate for the item used for prediction is defined among the related item group. Included in the prediction unit that predicts the predicted value, the error calculation unit that calculates the error between the predicted value predicted by the prediction unit and the actual value for the target item for each of the verification data, and the identity definition information. For each of the plurality of elemental combinations selected from the elemental item groups selected from the elemental item groups and the plurality of elemental combinations consisting of the values for the plurality of elemental items, the representative values of the plurality of errors related to the plurality of verification data corresponding to the elemental combination are calculated. A new element item defined by a specific element combination, which is specified based on the representative value of the error corresponding to each element combination, is generated from the element combination error calculation unit and the plurality of element combinations. It is an information processing program characterized by functioning as a new element item addition part to be added to the element definition information.

According to the invention of claim 1 or 7, a new feature item based on a combination of a plurality of related items possessed by past data is automatically added to the features used when predicting a prediction target value by using machine learning. be able to.

According to the invention of claim 2, the new feature item can be added to the feature after confirming that the prediction accuracy when the feature including the new feature item is used is improved.

According to the invention of claim 3 or 5, the number of defined feature combinations can be reduced as compared with the case of defining combinations among all feature items.

According to the invention of claim 4, it is possible to add a new feature item that can further improve the prediction accuracy as compared with the case where the feature combination does not include the value related to the past data other than the past data of interest.

According to the invention of claim 6, it is possible to add a new feature item that can further improve the prediction accuracy as compared with the case where the contribution degree is not taken into consideration.

It is a block diagram of the information processing apparatus which concerns on this embodiment. It is a figure which shows the content example of the past data DB. It is a figure which shows the content example of the initial feature definition information. It is a figure which shows the content example of a feature. It is a figure which shows the example of the actual value and the predicted value for each verification data. It is a figure which shows the example of the feature combination. It is a figure which shows the content example of the update feature definition information. It is a flowchart which shows the process flow of the information processing apparatus which concerns on 1st Embodiment. It is a graph which shows the average value of the error with respect to the iterative processing of a new feature item addition processing. It is a figure which shows the example of the hierarchical relation of each feature item. It is a figure which shows the example of the degree of contribution to each feature item.

Hereinafter, embodiments of the present invention will be described.

<First Embodiment>
FIG. 1 shows a schematic configuration diagram of the information processing apparatus 10 according to the present embodiment. The information processing device 10 may be a general computer, for example, a server or a personal computer. As shown in FIG. 1, the information processing apparatus 10 includes a storage unit 12 and a control unit 14. Further, although not shown in FIG. 1, the information processing apparatus 10 is composed of, for example, a network adapter or the like, and is a communication unit for communicating with another apparatus via a communication line such as the Internet, for example, a liquid crystal panel or the like. It is composed of a display unit for displaying the processing content of the information processing apparatus 10 (for example, a predicted value described later), for example, a mouse, a keyboard, a touch panel, or the like, and inputs an instruction from a user. It may include an input unit for the purpose.

The information processing device 10 is a device that performs processing for predicting the future based on past actual data. In this specification, the process of predicting the sales on the prediction target date of the store (prediction target store) in which the information processing device 10 is located will be described as an example, but the information processing device 10 predicts only this. No.

The storage unit 12 is composed of, for example, a hard disk, a ROM (Read Only Memory), a RAM (Random Access Memory), or the like. The storage unit 12 stores an information processing program for operating each unit of the information processing device 10. Alternatively, various control data, various processing data, and the like are stored in the storage unit 12. Further, as shown in FIG. 1, the past data DB 20 is defined in the storage unit 12, and the feature definition information 22 is stored.

In the past data DB 20, past actual values for the target item to be predicted and past actual values for the related item group related to the target item are accumulated. In the present embodiment, the target item is the sales of the forecast target store, and the past data DB 20 stores the actual value of the past sales of the forecast target store. Further, in the present embodiment, the related item group is various items related to sales (details will be described later), and each actual value for the related item group is stored in the past data DB 20. These data may be stored by the user in the past data DB 20, or may be automatically collected and stored.

FIG. 2 shows an example of the contents of the past data DB 20. In FIG. 2, the past data DB 20 is shown in a table format, but the data format of the past data DB 20 is not limited to this. In the present embodiment, data (records) are accumulated in the past data DB 20 every day. As shown in FIG. 2, the past data DB 20 stores the actual value of the sales of the forecast target store, and the related items related to the sales include the year, month, day, weather, maximum temperature, and minimum. Actual values for temperature etc. are stored. Of course, as related items, various items may be included regardless of the magnitude of the influence on sales. For example, it may have various items such as day of the week, holiday or weekday, humidity, wind speed, traffic volume in front of the store, average stock price, and exchange rate.

In the present embodiment, one record in the past data DB 20 is data in which the actual value of sales for a certain day and the actual value of the related item group for that day are associated with each other. In this specification, one record in the past data DB 20 is referred to as "past data". That is, the past data group is stored in the past data DB 20 by sequentially accumulating the past data. In the present embodiment, it is assumed that the past data DB 20 stores the past data for three years from 2014 to 2016 regarding the store to be predicted.

In the past data DB 20 in the present embodiment, a past data ID that can uniquely identify the past data is attached to each past data.

Although the details will be described later, in the information processing apparatus 10, the forecasted sales value on the forecast target date of the forecast target store is predicted based on the past data group stored in the past data DB 20. Specifically, by learning the relationship between the actual value of each related item in each past data and the actual value of sales, a forecast model for predicting the sales value from the value of each related item is constructed, and the forecast is made. Forecast sales values are predicted based on the model and the values for each related item on the forecasted date.

The feature definition information 22 is information that defines a feature item group composed of a plurality of feature items that are candidates for items used for constructing a prediction model. The feature item is an item based on the related item group included in the past data. As described above, the related item group included in the past data may include items that are highly related to the sales that are the target items and items that are not related to the sales. In general, when a prediction model is constructed in consideration of the values of related items that are not related to the target item, the prediction accuracy of the prediction model does not improve so much. On the contrary, if the prediction model is constructed based on the values of appropriate related items, the prediction accuracy of the prediction model can be improved. That is, by defining an appropriate feature item group in the feature definition information 22, a prediction model can be constructed using an appropriate item from the related item group, so that the prediction accuracy of the prediction model can be improved. Can be done. As described above, the feature item group defined in the feature definition information 22 is a factor greatly related to the prediction accuracy of the prediction model and, by extension, the prediction accuracy of the prediction process in the information processing apparatus 10. In some cases, all the related items included in the past data may be defined as feature items.

As will be described in detail later, in the present embodiment, a new feature item (new feature item) is automatically added to the feature definition information 22 by the processing of the information processing apparatus 10. In the present specification, among the feature definition information 22, the one to which the new feature item is not added by the information processing device 10 is referred to as "initial feature definition information 22a", and the new feature item is added by the processing of the information processing device 10. The latter one is described separately from the "update feature definition information 22b".

FIG. 3 shows an example of the contents of the initial feature definition information 22a in the present embodiment. The initial feature definition information 22a is generated by the user. As shown in FIG. 3, the initial feature definition information 22a includes year, month, day, week, day of the week, holiday or weekday as feature items. Each of these feature items is an item selected from the related item group possessed by the past data group. Further, in the initial feature definition information 22a, possible values are defined for each feature item. Since the above-mentioned feature items are originally not so many values that can be taken in the past data, the values that can be taken in the past data are defined as the values that the feature items can take as they are.

Further, the initial feature definition information 22a also includes feature items such as whether the maximum temperature is 15 degrees or higher, the minimum temperature is less than 10 degrees, or whether it has rained. In the past data group, the actual values for the maximum and minimum temperatures can take continuous (that is, various) values, and the actual values for the weather also take various values such as fine, rain, cloudy, and then rain. obtain. It is not appropriate to directly use raw data that can take such various values when constructing a predictive model. That is, while the amount of processing can be enormous, the effect of improving the prediction accuracy cannot be expected so much. Therefore, in the initial feature definition information 22a, for example, a feature item such as whether the maximum temperature is 15 degrees or higher is defined based on the related item of the maximum temperature possessed by the past data, and the value that the feature item can take is 1 ( Two values are defined: 15 degrees or more) and 0 (less than 15 degrees). The maximum temperature threshold (15 degrees in this example) may be determined by the user. As described above, for related items that can take various values, the number of values that can be taken by each feature item can be reduced by appropriately devising the feature items. This simplifies the process of building a predictive model.

Each feature item defined in the feature definition information 22 is assigned a feature item ID for uniquely identifying the feature item.

The control unit 14 is composed of, for example, a CPU (Central Processing Unit) or a microcontroller. The control unit 14 controls each unit of the information processing apparatus 10 based on the information processing program stored in the storage unit 12. Further, as shown in FIG. 1, the control unit 14 includes a past data classification unit 24, a feature generation unit 26, a prediction unit 28, an error calculation unit 30, a feature combination definition unit 32, a feature combination error calculation unit 34, and a new unit. It also functions as a feature item addition unit 36. When the control unit 14 exerts these functions, a new feature item that can improve the prediction accuracy of the prediction model is added to the feature definition information 22. The details of each of these functions will be described below.

The past data classification unit 24 classifies the past data group stored in the past data DB 20 into a learning data group and a verification data group. The training data group is used for constructing a prediction model, and consists of a plurality of past data that are a part of the past data group. On the other hand, the verification data group is used to verify the prediction accuracy of the constructed prediction model, and is composed of a plurality of past data that are a part of the past data group. As described above, in the present embodiment, since the past data DB 20 stores the past data group for three years from 2014 to 2016, a plurality of past data for two years 2014 and 2015 are stored. Is used as a training data group, and a plurality of past data for 2016 is used as a verification data group. Of course, the method of classifying the training data group and the verification data group is not limited to this.

The feature generation unit 26 generates a plurality of feature items and a feature consisting of a plurality of values for each feature item based on the past data stored in the past data DB 20 and the feature definition information 22. The feature generation unit 26 generates a plurality of features corresponding to each past data included in the past data group. This creates a feature group. It should be noted that the feature items included in the features do not have to be all of the feature items included in the feature definition information 22, and may be some feature items of a plurality of feature items defined in the feature definition information 22. ..

FIG. 4 shows an example of a feature group generated by the feature generation unit 26. FIG. 4 shows an example of a feature group generated based on each past data shown in FIG. 2 and the feature definition information 22 shown in FIG.

For example, the feature I1 shown in FIG. 4 corresponds to the past data (see FIG. 2) indicated by the past data ID “201440401”. The value for each feature item is determined based on each value for the related item in the past data. In the feature I1, the value "2014" is determined for the feature item "year", and thereafter, the value for each of the other feature items is determined in the same manner. The value "3" is determined for the feature item "day of the week" because it has the value of "Tuesday" for the related item "day of the week" of the past data. In addition, since the value for the related item "maximum temperature" of the past data is "17.4 degrees", the value for the feature item "whether the maximum temperature is 15 degrees or more" is "1" indicating 15 degrees or more. It has been decided.

In this way, the feature generation unit 26 generates features for all past data groups including the training data group and the verification data group. A past data ID is attached to each feature so as to indicate the relationship between the feature and the past data. The generated features are passed to the prediction unit 28 and the feature combination error calculation unit 34.

First, the prediction unit 28 makes a prediction based on each feature generated by the feature generation unit 26 based on the learning data group and the actual value (sales actual value in this embodiment) of the target item included in each learning data. Build a model.

Various methods can be used to build the predictive model. For example, it is an ensemble learning method that constructs a prediction model by combining a plurality of weak classifiers generated based on each element and actual value, and the next weak classifier is referred to by referring to the learning result of one weak classifier. While learning, the gradient boosting method can be used to build a prediction model considering the slope of the loss function that defines the error between the predicted value and the actual value. Alternatively, based on the training data sampled from the training data group, by randomly selecting the feature items used for identification (classification) in the non-terminal node, a plurality of decision trees with low correlation are created, and the plurality of decision trees are created. A random forest method can be used to build a predictive model using.

Next, the prediction unit 28 predicts the target item of each verification data based on the constructed prediction model and each feature generated by the feature generation unit 26 based on the verification data group (sales forecast in this embodiment). Value) is predicted. Here, each verification data already has the actual sales value, but the forecasting unit 28 adds a new feature item to the feature definition information 22, so that the forecasting unit 28 has each known verification. Forecast the sales forecast value of the data.

When the forecasting process is performed by the forecasting unit 28, a plurality of sales forecast values corresponding to each verification data are obtained. FIG. 5 shows an example of a sales forecast value corresponding to the actual sales value of each verification data. The plurality of sales forecast values predicted by the forecast unit 28 are associated with a past ID indicating the corresponding past data and the actual sales value of the past data, and are passed to the error calculation unit 30.

The error calculation unit 30 has an actual value of the target item possessed by the verification data (actual sales value in the present embodiment) and a predicted value of the target item predicted by the forecasting unit 28 for the verification data (sales forecast value in the present embodiment). Calculate the error with. The error calculation unit 30 calculates an error for each verification data. As a result, a plurality of errors corresponding to each verification data are calculated. For example, in the example of FIG. 5, the error calculation unit calculates an error "2" with respect to the verification data indicated by the past data ID "20160401", and an error "4" with respect to the verification data indicated by the past data ID "20160402". ], And the error for each verification data is calculated in the same manner below.

When the error calculation unit 30 calculates a plurality of errors for each verification data, the error calculation unit 30 calculates the average value of the errors for all the verification data (hereinafter referred to as “total average error”). Needless to say, the total average error is calculated by dividing the total value of each error by the number of verification data. As will be described later, the total average error can be used as a judgment material when the process of adding a new feature item is completed.

A plurality of errors for each verification data calculated by the error calculation unit 30 are associated with the past data ID of each verification data and passed to the feature combination error calculation unit 34.

The feature combination definition unit 32 defines a plurality of feature combinations based on the feature definition information 22. The feature combination is composed of a plurality of feature items defined in the feature definition information 22 and a plurality of values for the plurality of feature items. FIG. 6 shows a specific example of the feature combination. The feature combination shown in FIG. 6 is defined based on the feature definition information 22 shown in FIG.

For example, the feature combination P1 shown in FIG. 6 is a combination of a value "12" for the feature item "month" and a value "3" for the feature item "week". The feature combination shown in FIG. 6 is an example, and in the present embodiment, the feature combination definition unit 32 defines all the possible combinations between the feature item and the possible value of the feature definition information 22. do. However, the combination of the same feature items shall not be defined. For example, the combination of the value "1" for the feature item "month" and the value "2" for the same feature item "month" is not defined. As shown in FIG. 6, each feature combination is given a combination ID that uniquely identifies the feature combination. A plurality of feature combinations defined by the feature combination definition unit 32 are passed to the feature combination error calculation unit 34.

The feature combination error calculation unit 34 calculates representative values of a plurality of errors related to a plurality of verification data corresponding to the feature combination in the verification data group for each of the plurality of feature combinations defined by the feature combination definition unit 32.

The verification data corresponding to the feature combination is the verification data having a value for the feature item included in the feature combination. For example, when the feature combination is a combination of the value "12" for the feature item "month" and the value "3" for the feature item "week", the verification data corresponding to the feature combination is the related item "month". The value for "December" is "December", and the value for the related item "Week" is "3", which is the verification data. That is, in the present embodiment, the seven verification data corresponding to the third week of December 2016 are a plurality of verification data corresponding to the feature combination. In this way, a plurality of corresponding verification data are specified for each feature combination. The verification data corresponding to the feature combination is specified based on each feature generated by the feature generation unit 26 based on the verification data group.

When the feature combination error calculation unit 34 specifies a plurality of verification data corresponding to each feature combination, the feature combination error calculation unit 34 calculates a representative value of the error related to the plurality of corresponding verification data for each feature combination. In the present embodiment, the average value is calculated as the representative value, but the representative value may be, for example, the median value. Specifically, the feature combination error calculation unit 34 selects a certain feature combination, and based on each feature passed from the feature generation unit 26, a plurality of past data IDs of a plurality of verification data corresponding to the feature combination. To identify. Next, among the plurality of errors passed from the error calculation unit 30, a plurality of errors associated with the specified plurality of past data IDs are extracted. Then, the average value of the extracted plurality of errors is calculated. The average value of the pieces is the average value of the error of the feature combination (hereinafter referred to as "average error"). In this way, the feature combination error calculation unit 34 calculates the average error of all the feature combinations defined by the feature combination definition unit 32.

The new feature item addition unit 36 has a plurality of features defined by the feature combination definition unit 32 based on the representative value of the error of each feature combination (average error in this embodiment) calculated by the feature combination error calculation unit 34. Specify a specific feature combination from the combinations. The method for specifying the specific feature combination will be described later, but basically, the feature combination having a large average error is regarded as the specific feature combination.

In the present embodiment, the new feature item addition unit 36 ranks a plurality of feature combinations so that the average error of each feature combination is in descending order, and the first feature combination is the specific feature combination. .. That is, one feature combination having the largest average error among a plurality of feature combinations is defined as a specific feature combination. The new feature item addition unit 36 may specify the specific feature combination by a specific method other than the above. For example, after ranking a plurality of feature combinations so that the average error of each feature combination is in descending order, a plurality of feature combinations (for example, 1st to 3rd place) that are higher in the ranking are specified. It may be a combination of features. Alternatively, an error threshold value may be set in advance, and all feature combinations for which an average error equal to or higher than the error threshold value has been calculated may be specified as a specific feature combination. When such a specific method is adopted, a plurality of feature combinations can be specified as specific feature combinations.

Next, the new feature item addition unit 36 generates a new feature item based on the specific feature combination. In the present embodiment, a new feature item of whether or not it corresponds to a specific feature combination is generated. For example, when a specific feature combination consisting of a combination of a value "12" for the feature item "month" and a value "4" for the feature item "week" is specified, the new feature item addition unit 36 is set to "December. Generate a new feature item "Is it the 4th week?" Then, "1 (yes)" and "0 (no)" are generated as possible values for the new feature item.

Then, the new feature item addition unit 36 adds the generated new feature item to the feature definition information 22. As a result, the update feature definition information 22b is generated. When a plurality of specific feature combinations are specified, the new feature item addition unit 36 generates a plurality of new feature items corresponding to each of the plurality of specific feature combinations and a plurality of possible values for them, and defines the features. Add to information 22.

FIG. 7 shows an example of updated feature definition information 22b to which a new feature item is added. The updated feature definition information 22b shown in FIG. 7 is obtained by adding one new feature item in the above example to the initial feature definition information 22a shown in FIG.

In this way, the new feature item addition unit 36 adds a new feature item defined from a plurality of feature items defined in the initial feature definition information 22a and a feature combination that is a combination of values for them to the feature definition information 22. Will be done. When a new prediction model is constructed by the prediction unit 28 based on the updated feature definition information 22b, the prediction accuracy of the new prediction model is improved as compared with the old prediction model constructed based on the initial feature definition information 22a. It will be expected.

In particular, since new feature items based on specific feature combinations with a large average error are added from among a large number of feature combinations that can be defined, it can be expected that the prediction accuracy of the new prediction model will exceed the prediction accuracy of the old prediction model. .. For example, when the initial feature definition information 22a includes the feature items "month" and "week", the forecast model based on the initial feature definition information 22a has a monthly fluctuation in sales and a weekly sales fluctuation. It is built with separate consideration and not with the combination of a particular month and a particular week. In this case, for example, sudden sales fluctuations in a specific week of a specific month due to an annual specific event will not be suitably reflected in the forecast model. According to the present embodiment, for example, a new feature item based on a feature combination of a feature item "12" for a feature item "month" and a value "4" for a feature item "week" is added, so that the prediction unit 28 can perform the prediction unit 28. It is possible to build a forecast model that takes into account sudden sales fluctuations that occur only in the fourth week of December.

Further, according to the method for specifying the feature combination in the present embodiment, one specific feature combination is specified in one new feature item addition process, whereby one new feature item is added. As a result, the change of the feature definition information 22 in one feature item addition process is minimized, that is, the fluctuation amount of the prediction model in one feature feature item addition process is minimized. This is because, on the premise that the prediction model is repeatedly changed by repeatedly performing the new element item addition process, the prediction accuracy of the prediction model may not be improved by one new element item addition process. And, by minimizing the fluctuation amount of the prediction model and gradually changing it, it leads to improving the prediction accuracy of the prediction model at an earlier stage.

When the new feature item addition process is repeated, it is possible to define a feature combination including the new feature item added in the first process in the second and subsequent processes. For example, when the updated feature definition information 22b as shown in FIG. 7 is generated in the first process, in the second process, the feature combination definition unit 32 is, for example, the feature combination “4th week of December?”. It is possible to define a feature combination of a value "1" for the feature item "is the maximum temperature 15 degrees or higher?" And a value "1" for the feature item "is the maximum temperature 15 degrees or higher?".

As described above, the prediction accuracy of the prediction model constructed based on the updated feature definition information 22b is not always higher than the prediction accuracy of the prediction model constructed based on the initial feature definition information 22a. Therefore, the new feature item addition unit 36 confirms that the prediction accuracy of the prediction model is improved by adding the new feature item defined based on the specific feature combination to the initial feature definition information 22a. The new feature item may be added to the initial feature definition information 22a. The specific processing flow is as follows.

First, the new feature item addition unit 36 identifies a specific feature combination having the maximum average error from a plurality of feature combinations defined by the feature combination definition unit 32 by the same processing as described above, and uses the specific feature combination as the specific feature combination. Generate a new feature item based on. Then, the new feature item addition unit 36 temporarily adds the generated new feature item to the feature definition information 22. As a result, the pseudo-feature definition information to which the new feature item is provisionally added is generated.

After that, each part below the feature generation unit 26 performs the same processing as described above using the pseudo-feature definition information. Specifically, the element generation unit 26 generates a pseudo element based on the pseudo element definition information and the past data group, and the prediction unit 28 generates a tentative prediction model based on the pseudo element and the sales actual value of the training data group. Is constructed, and the predicted value for each verification data is calculated based on the tentative prediction model and the tentative nature of the verification data group. The error calculation unit 30 calculates a plurality of errors between the actual sales value of each verification data and each sales forecast value predicted using the provisional prediction model. Then, the error calculation unit 30 calculates the average value of a plurality of errors corresponding to all the verification data (hereinafter referred to as "provisional total average error").

Here, the new element item addition unit 36 is predicted using the actual sales value of each verification data calculated in the process of generating the new element item and the prediction model constructed based on the initial element definition information 22a. The average value of a plurality of errors with each sales forecast value (hereinafter referred to as "pre-processed total average error") is compared with the provisional total average error. If the tentative total average error is smaller than the unprocessed total average error, it means that the prediction accuracy of the prediction model has improved due to the addition of the new feature item, so the tentatively added new feature item is officially added to the feature definition information 22. do. On the other hand, if the provisional total average error is greater than or equal to the pre-processing total average error, it means that the prediction accuracy of the prediction model has deteriorated due to the addition of new feature items. To delete. That is, the new feature item is not added to the feature definition information 22.

Hereinafter, the processing flow of the information processing apparatus 10 according to the present embodiment will be described with reference to the flowchart shown in FIG.

In step S10, the user stores the past data group in the past data DB 20 and stores the initial feature definition information 22a in the storage unit 12.

In step S12, the past data classification unit 24 classifies the past data group stored in the past data DB 20 into a learning data group and a verification data group.

In step S14, the feature generation unit 26 obtains a feature consisting of a plurality of feature items and a plurality of values for each feature item for each past data stored in the past data DB 20 in step S10 based on the initial feature definition information 22a. Generate.

In step S16, the prediction unit 28 builds a prediction model based on each feature generated based on the learning data group in step S14 and the actual sales value included in each learning data.

In step S18, the prediction unit 28 predicts the sales forecast value for each verification data based on the prediction model constructed in step S16 and each feature generated based on the verification data group in step S14.

In step S20, the error calculation unit 30 calculates an error between the actual sales value and the predicted sales value calculated in step S18 for each verification data.

In step S22, the error calculation unit 30 calculates the average value of a plurality of errors (hereinafter referred to as “this time total average error”) for all the verification data calculated in this step S20 in the previous step S20. Whether or not the difference between the current total average error and the previous total average error is less than the threshold value, which is smaller than the average value of multiple errors for all the verification data (hereinafter referred to as "previous total average error"). To judge. If the condition is satisfied, the process is terminated, and if the condition is not satisfied, the process proceeds to step S24. Since this time is the initial processing and the value of the previous average error does not exist, the process proceeds to step S24. The details of step S22 will be described later.

In step S24, the feature combination definition unit 32 defines a plurality of feature items and a plurality of feature combinations consisting of a plurality of values for the plurality of feature items based on the initial feature definition information 22a.

In step S26, the feature combination error calculation unit 34 has a plurality of verification data corresponding to the feature combination among the plurality of errors corresponding to each verification data calculated in step S20 for the feature combination defined in step S24. Multiple errors are extracted, and the average value (average error) of the extracted multiple errors is calculated. This is performed for each feature combination defined in step S24, and the average error corresponding to each feature combination is calculated.

In step S28, the new feature item addition unit 36 identifies the feature combination having the largest average error calculated in step S26 among the plurality of feature combinations defined in step S24. Then, based on the specified feature combination, a new feature item and a possible value for the new feature item are generated.

In step S30, the new feature item addition unit 36 adds the new feature item generated in step S28 and the possible values for the new feature item to the feature definition information 22. As a result, the update feature definition information 22b is generated.

After the processing of step S30, the process returns to step S14 again, and in step S14 again, the feature generation unit 26 generates new features for each past data based on the updated feature definition information 22b generated in step S30. After that, the same processing is performed again from step S16 to step S22 based on the new features.

As described above, in the present embodiment, the processes of steps S14 to S30 are repeated, and new feature items are sequentially added to the feature definition information 22. As a result, it is expected that the prediction accuracy of the prediction model constructed by the prediction unit 28 will be gradually improved.

The condition shown in step S22 is the end condition of the iterative process. As described above, the condition of step S22 is whether or not the current total average error is smaller than the previous total average error and the difference between the current total average error and the previous total average error is equal to or less than the threshold value. In other words, the total average error this time is smaller than the previous total average error, but when the total average error does not change much from the previous time, the iterative process ends.

This is based on the fact that the relationship between the number of repetitions of the new feature item addition process and the total average error generally has the relationship shown in FIG. As shown in FIG. 9, when the number of repetitions of the new feature item addition process is relatively small, the total average error is relatively greatly reduced by one new feature item addition process, but the number of repetitions of the new feature item addition process is relatively small. When the number of new features becomes relatively large, the total average error does not change much by one new feature item addition process. Therefore, when the difference between the total average error this time and the total average error last time is less than the threshold value, it can be judged that the total average error does not decrease dramatically, and the iterative processing of the new feature item addition process is terminated. ..

Conditions other than the above can be adopted as the end condition of the iterative process of the new feature item addition process. For example, it may be a condition that the number of feature items included in the feature definition information 22 has reached a predetermined number, or it may be a condition that the number of times of repeated processing of the new feature item addition process has reached a predetermined number.

<Second Embodiment>
Although the second embodiment is based on the first embodiment, the processing content of the feature combination definition unit 32 is different from that of the first embodiment. In the first embodiment, the feature combination definition unit 32 defines all the possible combinations between the feature item and the possible value of the feature definition information 22. According to this, while all feature combinations can be candidates for specific feature combinations without omission, the number of defined feature combinations may become enormous, and the feature combination definition unit 32 and feature combination error calculation unit The processing amount of 34 or the new feature item addition unit 36 may increase.

In view of this, in the second embodiment, the feature combination definition unit 32 does not define the feature combinations that are unlikely to have a large average error, so that the number of defined feature combinations is reduced. As a result, the amount of processing in the new feature item addition processing is reduced. Hereinafter, the processing of the feature combination definition unit 32 in the second embodiment will be described.

First, in the second embodiment, the feature definition information 22 defines a hierarchical relationship among a plurality of feature items. The definition of the hierarchical relationship may be performed by a user or the like in advance.

FIG. 10 shows an example of the hierarchical relationship in this embodiment. In the example of FIG. 10, the feature item "year" is defined as the first layer, the feature item "month" is defined as the second layer, the feature item "week" is defined as the third layer, and the feature item (day of the week) is defined as the fourth layer. .. In this way, the hierarchical relationship between the plurality of feature items is a hierarchy that matches the actual concept of each feature item. Of course, the hierarchical relationship shown in FIG. 10 is an example, and the hierarchical relationship between a plurality of feature items can be defined in various ways.

The feature combination definition unit 32 defines the feature combination based on the hierarchical relationship defined in the feature definition information 22. Specifically, first, the feature combination definition unit 32 selects one feature item included in the feature combination. Next, in the hierarchical relationship, the feature item belonging to the layer adjacent to the layer to which the one feature item belongs is selected as the other feature item. Then, the feature combination is defined by combining the one feature item (and the value for the feature item) and the other feature item (and the value for the feature item). In other words, in the second embodiment, the feature combination definition unit 32 does not define a feature combination of one feature item and a feature item belonging to a layer not adjacent to the layer to which the one feature item belongs.

This will be described with reference to the example of FIG. The feature combination definition unit 32 defines, for example, the feature combination of the feature item "year" belonging to the first layer and the feature item "month" belonging to the second layer, but the feature item "year" belonging to the first layer. And the feature combination with the feature item "week" belonging to the third layer is not defined. Specifically, the feature combination of the value "2015" for the feature item "year" and the value "3" for the feature item "month" is defined, but the value "2015" for the feature item "year" and the feature item. The combination with the value "4" for "week" is not defined.

For example, in historical data, the combination of related items "year" and "week", which are items that are separated from each other, has a particular effect on sales (for example, sales fluctuate suddenly only in the 4th week of every month in 2015). It can be said that there are relatively few things. Therefore, it can be said that it is unlikely that the average error of the feature combinations including the feature items whose layers are separated from each other becomes large. Therefore, the feature combination definition unit 32 does not define the feature combinations including the feature items whose layers are separated from each other, so that the feature combinations defined while leaving the feature combinations with a high possibility that the average error becomes large. The number of can be reduced.

In the second embodiment, when a new feature item is added to the feature definition information 22, the layer of the added new feature item is the deeper layer of the two feature items included in the new feature combination. .. For example, a new feature item "March 2015" based on a combination of features "2015" for the feature item "year" in the first layer and "3" for the feature item "month" in the second layer. Is defined, the layer of the new feature item is the second layer, which is the deeper layer of the first layer and the second layer.

As a result, the feature item "March 2015?" In the second layer and the feature item "week" in the third layer can be combined by further additional processing of the feature item. In this way, it is also possible to define a feature combination including three or more feature items along the hierarchical relationship.

<Third Embodiment>
Although the third embodiment is based on the first embodiment, the processing content of the feature combination definition unit 32 is different from that of the first embodiment and the second embodiment. In the first embodiment and the second embodiment, a plurality of values included in the feature combination are defined as values relating to the same past data (attention past data). For example, in the feature combination P1 shown in FIG. 6, the value for the feature item "month" of a certain attention past data is "12", and the value for the feature item "week" is "3". It means that.

When the past data group is composed of a plurality of past data arranged in a time series, the prediction accuracy may be improved by predicting the predicted value in consideration of the values for the related items of the plurality of past data. When the past data group is composed of a plurality of past data accumulated daily as in the present embodiment, for example, whether the day before the forecast target date is a holiday or a weekday, or the day after the forecast target date is a holiday or a weekday. Etc. may affect the sales forecast on the forecast target date. In such a case, the prediction accuracy of the prediction model can be further improved by constructing the prediction model in consideration of whether or not the day before or after the prediction target day is a holiday.

In view of this, in the third embodiment, the feature combination definition unit 32 has the values related to the past data of interest for the feature items selected from the feature item group defined in the feature definition information 22, and also from the feature item group. Define a feature combination including a value related to past data different from the featured past data for the selected feature item.

For example, in the third embodiment, the feature combination definition unit 32 has a value "0 (weekday)" of attention past data (on the day) for the feature item "holiday or weekday", and the previous day for the feature item "holiday or weekday". A feature combination with a value of "1 (weekday)" is defined. In addition, the two feature items included in the feature combination do not have to be the same. For example, the value of the day "0 (weekday)" for the feature item "holiday or weekday" and the value of the previous day for the feature item "day of the week" " You may define the feature combination of "1 (Sunday)". Further, the plurality of values included in the feature combination do not have to be related to the values of consecutive past data in the time series, and may be the values of the current day and the values of, for example, two days ago and one week ago.

If such a feature combination is defined and the average error for the feature combination is maximized by the feature combination error calculation unit 34, the new feature item "whether the current day is a weekday and the previous day is a holiday" based on the feature combination is defined as a feature. Added to information 22. As a result, the prediction unit 28 can build a prediction model with improved prediction accuracy in consideration of values for related items in a plurality of past data such as "whether the current day is a weekday and the previous day is a holiday".

<Fourth Embodiment>
Although the fourth embodiment is based on the first embodiment, the magnitude of the influence on the predicted value of the target item (sales in this embodiment) for each feature item defined in the feature definition information 22. It is different from the first embodiment in that the processing based on the contribution degree is performed.

The degree of contribution to each feature item is calculated by the prediction unit 28. The basic method for calculating the contribution is as follows. First, as described above, the prediction unit 28 constructs a prediction model based on the element group related to the training data group generated by the element generation unit 26 and the actual sales value included in the training data group, and verifies the prediction model. The sales forecast value for the verification data is calculated based on the nature of the data. Next, in the features related to the verification data, the values for the feature items of interest are randomly changed to generate the changed features. Then, the forecasting unit 28 calculates the sales forecast value for the verification data based on the forecast model and the change feature. The larger the difference between the sales forecast value calculated here based on the change feature and the pre-calculated sales forecast value, the higher the contribution of the feature feature item. Therefore, the forecasting unit 28 calculates the contribution of the feature feature item as the difference between the sales forecast value based on the change feature and the sales forecast value calculated in advance is larger.

By such processing, information indicating the contribution degree of each feature item is attached to the feature definition information 22, and the feature definition information 22-2 with the contribution degree is generated. FIG. 11 shows an example of feature definition information 22-2 with contribution.

The feature combination definition unit 32 generates feature combinations based on the contribution of each feature item. Specifically, the contribution threshold is set in advance, and when defining the feature combination, the feature item whose contribution is equal to or less than the contribution threshold is not selected. When the feature combination error calculation unit 34 calculates the average error for the feature combination including the feature item having a low contribution, it can be said that the possibility that the average error becomes large is low. Therefore, by not defining the feature combinations including the feature items equal to or less than the contribution threshold, the number of the defined feature combinations can be reduced while leaving the feature combinations with a high possibility that the average error becomes large.

Further, in the first embodiment, the new feature item addition unit 36 specified the specific feature combination based on the average error calculated for each feature combination, but in the fourth embodiment, each feature combination is further specified. A specific feature combination is specified based on the contribution of the feature items included in. For example, the product of the average error of a certain feature combination and the average value or the total value of a plurality of contributions for a plurality of feature items included in the feature combination is calculated as an index value of the feature combination. Then, the specific feature combination is specified based on the index value calculated for each feature combination. As a specific element combination method based on the index value, for example, among a plurality of element combinations, the element combination having the maximum index value may be used as the specific element combination, or a plurality of element combinations so that the index values of each element combination are in descending order. After ranking the element combinations, a plurality of element combinations (for example, 1st to 3rd place) higher in the ranking may be used as a specific element combination, or an index threshold value may be set in advance and the index threshold may be set. All the element combinations for which the index value equal to or higher than the index threshold is calculated may be used as the specific element combination.

By specifying the specific feature combination in consideration of the contribution degree, the new feature item generated based on the feature combination including the feature item having a high contribution degree can be easily added to the feature definition information 22. As a result, the prediction accuracy of the prediction model constructed by the prediction unit 28 can be further improved.

Although the embodiments according to the present invention have been described above, the present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the spirit of the present invention.

10 Information processing device, 12 Storage unit, 14 Control unit, 20 Past data DB, 22 Feature definition information, 24 Past data classification unit, 26 Feature generation section, 28 Prediction section, 30 Error calculation section, 32 Feature combination definition section, 34 Feature combination error calculation unit, 36 new feature item addition section.

Claims (7)

A training data group that is a part of a past data group including past actual values for a target item to be predicted and a related item group related to the target item, and a candidate for an item to be used for prediction among the related item groups. A prediction unit that predicts each predicted value for the target item of each verification data that is a part of the past data group and is the target of verification, based on the prescriptive definition information in which the predisposition item group is defined.
For each of the verification data, an error calculation unit that calculates an error between each prediction value predicted by the prediction unit and an actual value for the target item, and
For each of the plurality of feature items selected from the feature item group included in the feature definition information and the plurality of feature combinations consisting of the values for the plurality of feature items, the plurality of verification data corresponding to the feature combination are described above. The feature combination error calculation unit that calculates the representative value of the error, and the feature combination error calculation unit.
From the plurality of feature combinations, a new feature item defined by the specific feature combination specified based on the representative value of the error corresponding to each feature combination is generated and added to the feature definition information. Item addition part and
An information processing device characterized by being equipped with. In the new feature item addition unit, after the new feature item is specified, the average value of the errors related to the verification data group calculated based on the pseudo-feature definition information to which the new feature item is provisionally added is the new feature. The new feature item is added to the feature definition information when it is smaller than the average value of the errors related to the verification data group calculated based on the feature definition information before the item is provisionally added.
The information processing apparatus according to claim 1. A plurality of feature items included in the feature item group have a hierarchical relationship.
The feature combination error calculation unit defines the feature combination by combining one feature item and the other feature item belonging to a layer adjacent to the layer to which the one feature item belongs in the hierarchical relationship.
The information processing apparatus according to claim 1. The past data group is composed of a plurality of past data arranged in a time series.
The feature combination includes a value related to past data of interest for a feature item selected from the feature item group and a value related to past data other than the past data of interest for a feature item selected from the feature item group.
The information processing apparatus according to claim 1. The prediction unit calculates the degree of contribution indicating the magnitude of the influence of the value on each feature item on the value of the target item for each feature item included in the feature item group.
In defining the feature combination, the feature combination error calculation unit does not select a feature item whose contribution is equal to or less than the threshold value.
The information processing apparatus according to claim 1. The new feature item addition unit identifies the specific feature combination based on the representative value of the error corresponding to each feature combination and the contribution of a plurality of feature items included in each feature combination.
The information processing apparatus according to claim 5. Computer,
A training data group that is a part of a past data group including past actual values for a target item to be predicted and a related item group related to the target item, and a verification data group that is a part of the past data group. The verification data included in the verification data group is based on the respective values for the related item group and the identity definition information in which the identity item group that is a candidate for the item used for prediction in the related item group is defined. A prediction unit that predicts each predicted value for the target item of each verification data that is a part of the past data group and is the target of verification, and
For each of the verification data, an error calculation unit that calculates an error between each prediction value predicted by the prediction unit and an actual value for the target item, and
For each of the plurality of feature items selected from the feature item group included in the feature definition information and the plurality of feature combinations consisting of the values for the plurality of feature items, the plurality of verification data corresponding to the feature combination are described above. The feature combination error calculation unit that calculates the representative value of the error, and the feature combination error calculation unit.
From the plurality of feature combinations, a new feature item defined by the specific feature combination specified based on the representative value of the error corresponding to each feature combination is generated and added to the feature definition information. Item addition part and
An information processing program characterized by functioning as.

Images