JP2018190044A - Information processing apparatus and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To automatically add a new feature item based on a combination of a plurality of related items included in past data to a feature used when predicting a prediction subject value using machine learning.SOLUTION: A feature generation unit 26 generates a group of features of a group of past data based on the group of past data and feature definition information 22. A prediction unit 28 constructs a prediction model based on a feature of a group of learning data which is a part of the group of past data and a sale prediction value, and predicts a sale prediction value with respect to each of validation data based on the prediction model and a group of validation data which is a part of the group of past data. An error calculation unit 39 calculates an error between the sales prediction value and a sales actual value with respect to each of the validation data. A feature combination error calculation unit 34 calculates an average error of the plurality of validation data corresponding to the feature combination for each of feature combinations defined by the feature combination definition unit 32. A new feature item addition unit 36 generates a new feature item based on the specific feature combination including a maximum average error and adds it to the feature definition information 22.SELECTED DRAWING: Figure 1

Description

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

  Conventionally, a prediction target value is predicted using machine learning. As an example of a prediction process using machine learning, a learning process (so-called a so-called “target variable”) and based on past data that is past actual values for a plurality of related items (explanatory variables) related to the target item A process for predicting a prediction target value based on the prediction model and a value of a related item related to the prediction target value is known by performing supervised learning).

  The past data may have a large number of related items, and some of them may include items that do not significantly affect the value of the target item. Therefore, in order to improve the prediction accuracy of the prediction model, a prediction model may be constructed using values for items (feature items) selected from a plurality of related items included in past data. Note that the information including the plurality of feature items and values for them is called a feature.

  Conventionally, techniques for automatically selecting feature items from a plurality of related items included in past data have been proposed. 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. Generating a feature function for constructing a prediction model is described. Patent Document 2 also describes a process of extracting a combination of data items with the highest prediction accuracy from a plurality of data items (related items).

Japanese Patent Laying-Open No. 2005-092681 Japanese Patent Laying-Open 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 included in past data. Here, by using new feature items defined based on a combination of a plurality of related items, a more accurate prediction model can be constructed. For example, when the sales forecast of a certain store is performed and the related items included in the past data include “month (month)” and “week (week of the month)”, “12 By using a new feature item “whether or not it is the fourth week of the month”, a more accurate prediction model may be constructed. Of course, it is important that new feature items are generated based on appropriate combinations that improve the prediction accuracy of the prediction model, not based on irrelevant combinations of multiple related items included in past data. 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 included in past data to a feature used when predicting a prediction target value 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 performance 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 feature definition information in which the feature item group that is a candidate for the item to be 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 to be verified A prediction unit for prediction, an error calculation unit for calculating an error between each prediction value predicted by the prediction unit and an actual value for the target item for each of the verification data, and a feature item included in the feature definition information A plurality of feature items selected from the group and a plurality of feature combinations each including a value for the plurality of feature items, a plurality of previous items related to the plurality of verification data corresponding to the feature combination A feature combination error calculation unit for calculating a representative value of error, and a new feature defined by a specific feature combination identified based on the representative value of the error corresponding to each feature combination from the plurality of feature combinations A new feature item addition unit for adding an item to the feature definition information.

  The invention according to claim 2 is characterized in that the new feature item addition unit specifies the new feature item, and then an error relating to the verification data group calculated based on the provisional feature definition information in which the new feature item is provisionally added. When the average value of the new feature item is smaller than the average value of 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 added to the feature definition information. The information processing apparatus according to claim 1, wherein the information processing apparatus is added to the information processing apparatus.

  In the invention according to claim 3, the plurality of feature items included in the feature item group have a hierarchical relationship, and the feature combination error calculation unit includes 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 with another feature item belonging to a layer adjacent to a hierarchy to which the item belongs.

  The invention according to claim 4 is configured such that the past data group includes a plurality of past data arranged in time series, and the feature combination includes a value related to attention past data for the feature item selected from the feature item group, and The information processing apparatus according to claim 1, further comprising: a value related to past data other than the past past data for the feature item selected from the feature item group.

  In the invention according to claim 5, the predicting unit calculates, for each feature item included in the feature item group, a degree of contribution indicating a magnitude of influence of a value for each feature item on the value of the target item. The information combination 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 when defining the feature combination.

  The invention according to claim 6 is characterized in that the new feature item adding unit is configured to determine the specific feature based on a representative value of the error corresponding to each feature combination and contributions of a plurality of feature items included in each feature combination. The information processing apparatus according to claim 5, wherein a combination is specified.

  The invention according to claim 8 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; Based on the feature definition information in which the feature item group that is a candidate for the item to be used for prediction among the related item group is defined, Included in the feature definition information, a prediction unit that predicts a prediction value, 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, for each of the verification data A plurality of verification data corresponding to the feature combination for each of a plurality of feature items selected from a plurality of feature items and a plurality of feature combinations consisting of values for the plurality of feature items. A feature combination error calculating unit for calculating a representative value of a plurality of the errors related to each other, and a specific feature combination specified from the plurality of feature combinations based on the representative value of the error corresponding to each feature combination An information processing program that functions as a new feature item adding unit that adds a new feature item to the feature definition information.

  According to the invention which concerns on Claim 1 or 7, the new feature item based on the combination of the some related item which past data has is automatically added to the feature used when performing prediction of a prediction target value using machine learning. be able to.

  According to the second aspect of the present invention, it is possible to add the new feature item to the feature after confirming that the prediction accuracy is improved when the feature including the new feature item is used.

  According to the invention which concerns on Claim 3 or 5, compared with the case where the combination between all the feature items is defined, the number of defined feature combinations can be reduced.

  According to the invention which concerns on Claim 4, compared with the case where the value regarding past data other than attention past data is not included in a feature combination, the new feature item which can improve a prediction precision more can be added.

  According to the invention which concerns on Claim 6, compared with the case where a contribution is not considered, the new feature item which can improve a prediction precision more can be added.

1 is a schematic configuration diagram of an information processing apparatus according to an embodiment. It is a figure which shows the example of the content of past data DB. It is a figure which shows the example of the content of initial feature definition information. It is a figure which shows the example of the content of a feature. It is a figure which shows the example of the track record value and prediction value with respect to each verification data. It is a figure which shows the example of a feature combination. It is a figure which shows the example of the content of update feature definition information. It is a flowchart which shows the flow of a process 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 repetition process of a new feature item addition process. It is a figure which shows the example of the hierarchical relationship of each feature item. It is a figure which shows the example of the contribution with respect to each feature item.

  Hereinafter, embodiments of the present invention will be described.

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

  The information processing apparatus 10 is an apparatus that performs processing for predicting the future based on past performance data. In the present specification, the process of predicting sales on the prediction target date of a store (prediction target store) where the information processing apparatus 10 is present will be described as an example. Absent.

  The storage unit 12 includes, 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 apparatus 10. Alternatively, the storage unit 12 stores various control data or various processing data. 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.

  The past data DB 20 stores past performance values for target items that are prediction targets and past performance values for related item groups related to the target items. In the present embodiment, the target item is the sales of the prediction target store, and the past data DB 20 stores the past sales record of the prediction target store. In this embodiment, the related item group is various items related to sales (details will be described later), and the past data DB 20 stores each actual value for the related item group. 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 sales value of the prediction target store, and the related items related to the sales are year, month, day, weather, maximum temperature, minimum Stores actual values for temperature and other factors. Of course, the related items may have various items regardless of the influence on sales. For example, it may have various items such as day of the week, holiday or weekday, humidity, wind speed, store front traffic, average stock price, exchange rate and the like.

  In the present embodiment, one record in the past data DB 20 is data in which the actual value of sales on a certain day is associated with the actual value of that day for the related item group. In this specification, one record in the past data DB 20 is described as “past data”. That is, the past data group is stored in the past data DB 20 by sequentially accumulating past data. In the present embodiment, it is assumed that past data for three years from 2014 to 2016 related to the prediction target store is accumulated in the past data DB 20.

  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 details will be described later, the information processing apparatus 10 predicts the predicted sales value of the prediction target store on the prediction target date 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 prediction model for predicting the sales value from the value of each related item is constructed. Based on the model and the value for each related item on the prediction target date, the predicted sales value is predicted.

  The feature definition information 22 is information that defines a feature item group including a plurality of feature items that are candidates for items used to construct a prediction model. The feature item is an item based on a related item group included in past data. As described above, the related item group included in the past data may include items that are highly related to sales that are the target items and items that are lowly related. In general, when a prediction model is constructed in consideration of values of related items that are less relevant to the target item, the prediction accuracy of the prediction model is not very good. Conversely, if a prediction model is constructed based on appropriate values of 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, it is possible to construct a prediction model using an appropriate item from the related item group, thereby improving the prediction accuracy of the prediction model. Can do. 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 thus the prediction accuracy of the prediction processing 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, information whose new feature item has not been added by the information processing device 10 is “initial feature definition information 22 a”, and new feature item addition processing is performed by the processing of the information processing device 10. The latter is described separately from the “update feature definition information 22b”.

  FIG. 3 shows an example of the content 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 week, holiday or weekday as feature items. Each of these feature items is an item selected from the related item group included in the past data group. In the initial feature definition information 22a, possible values for each feature item are defined. As for the above-described feature items, since there are not so many values that can be taken in the past data, values that can be taken in the past data are defined as values that the feature item can take as it is.

  Further, the initial feature definition information 22a 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 rains. In the historical data group, the actual values for the maximum temperature and the minimum temperature can be continuous (that is, various) values, and the actual values for the weather can be various values such as clear, rainy, cloudy, fine and rainy. obtain. When building a prediction model, it is not appropriate to directly use such raw data that can take various values. That is, while the processing amount 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 indicating whether the maximum temperature is 15 degrees or more is defined based on a related item called the maximum temperature that the past data has, and the value that the feature item can take is 1 ( Two values of 15 degrees or more) and 0 (less than 15 degrees) are defined. The maximum temperature threshold (15 degrees in this example) may be determined by the user. As described above, regarding related items that can take various values, the number of values that each feature item can take can be reduced by appropriately devising the feature item. Thereby, the process of construction of a prediction model is simplified.

  Note that a feature item ID for uniquely identifying a feature item is attached to each feature item defined in the feature definition information 22.

  The control unit 14 includes, 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 It also functions as the feature item adding unit 36. When the control unit 14 performs these functions, a new feature item that can improve the prediction accuracy of the prediction model is added to the feature definition information 22. 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 learning data group is used to construct a prediction model, and includes a plurality of past data that is 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 is a part of the past data group. As described above, in the present embodiment, since the past data DB 20 stores past data groups for three years from 2014 to 2016, a plurality of past data for two years of 2014 and 2015 is stored. Is a learning data group, and a plurality of past data for 2016 is a verification data group. Of course, the classification method of the learning data group and the verification data group is not limited to this.

  The feature generation unit 26 generates a feature composed of a plurality of feature items and 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. Thereby, a feature group is generated. Note that the feature items included in the feature need not be all of the feature items included in the feature definition information 22, and may be partial 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 “20140401”. Based on each value for the related item included in the past data, a value for each feature item is determined. In the feature I1, a value “2014” is determined for the feature item “year”, and values for other feature items are similarly determined thereafter. The value “3” is determined for the feature item “day of the week” because it has the value “Tuesday” for the related item “day of the week” of the past data. Since the value for the related item “maximum temperature” in the past data is “17.4 degrees”, “1” indicating 15 degrees or more is included in the value for the feature item “whether the maximum temperature is 15 degrees or more”. It has been decided.

  In this manner, the feature generation unit 26 generates features for all past data groups including the learning data group and the verification data group. Each feature is assigned a past data ID so as to indicate a relationship between the feature and past data. The generated feature is passed to the prediction unit 28 and the feature combination error calculation unit 34.

  The prediction unit 28 first predicts 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 construct the prediction model. For example, an ensemble learning method for constructing a prediction model by combining a plurality of weak classifiers generated based on each feature and actual value, and the next weak classifier is referred to with reference to the learning result of one weak classifier The gradient boosting method of constructing a prediction model in consideration of the gradient of the loss function that defines the error between the predicted value and the actual value can be used. Alternatively, based on learning data sampled from the learning data group, by randomly selecting feature items used for identification (classification) in a non-terminal node, a plurality of decision trees with low correlation are created, and the plurality of decision trees It is possible to use a random forest method for constructing a prediction model using

  Next, based on the constructed prediction model and each feature generated by the feature generation unit 26 based on the verification data group, the prediction unit 28 predicts the target value of each verification data (sales prediction in this embodiment). Value). Here, each verification data already has a sales record value, but since the prediction unit 28 adds a new feature item to the feature definition information 22, the prediction unit 28 uses each of the known verifications. Predict the sales forecast value of the data.

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

  The error calculation unit 30 includes the actual value of the target item included in the verification data (sales actual value in the present embodiment) and the predicted value of the target item predicted by the prediction unit 28 for the verification data (the predicted sales value in the present embodiment). And the error is calculated. The error calculation unit 30 calculates an error for each verification data. Thereby, a plurality of errors corresponding to each verification data are calculated. For example, in the example of FIG. 5, the error calculation unit calculates the error “2” for the verification data indicated by the past data ID “20160401” and the error “4” for the verification data indicated by the past data ID “20160402”. In the same manner, an error for each verification data is calculated.

  When calculating a plurality of errors for each verification data, the error calculation unit 30 calculates an average value of 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 becomes a judgment material when finishing the process of adding a new feature item.

  A plurality of errors for each piece of 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 includes 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 combinations shown in FIG. 6 are 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 combinations shown in FIG. 6 are an example. In the present embodiment, the feature combination definition unit 32 defines all possible combinations between the feature items included in the feature definition information 22 and the possible values. To do. However, the combination of the same feature items shall not be defined. For example, a combination of a value “1” for the feature item “month” and a value “2” for the same feature item “month” is not defined. As shown in FIG. 6, each feature combination is assigned 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 a feature combination error calculation unit 34.

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

  The verification data corresponding to the feature combination is verification data having values for the feature items included in the feature combination. For example, when the feature combination is a combination of a value “12” for the feature item “Month” and a value “3” for the feature item “Week”, the verification data corresponding to the feature combination is related item “Month” "Is" December ", and the value for the related item" week "is" 3 ". That is, in this embodiment, 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 is 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 errors related to the plurality of corresponding verification data for each feature combination. In this embodiment, an average value is calculated as a representative value, but the representative value may be a median value, for example. 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. Is identified. Next, a plurality of errors associated with the specified plurality of past data IDs are extracted from the plurality of errors passed from the error calculation unit 30. Then, an average value of the plurality of extracted errors is calculated. The average value is the average error of the feature combinations (hereinafter referred to as “average error”). In this manner, 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 adding unit 36 includes a plurality of features defined by the feature combination defining unit 32 based on the representative value of errors of each feature combination (average error in this embodiment) calculated by the feature combination error calculating unit 34. A specific feature combination is identified from the combinations. Although a specific feature combination specifying method will be described later, a feature combination having a large average error is basically a specific feature combination.

  In the present embodiment, the new feature item adding 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 set as a specific feature combination. . That is, one feature combination having the largest average error among a plurality of feature combinations is set as a specific feature combination. The new feature item adding unit 36 may specify the specific feature combination by a specifying 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, identify a plurality of feature combinations that are higher in the ranking (for example, 1st to 3rd) It may be a feature combination. Alternatively, an error threshold may be provided in advance, and all feature combinations for which an average error equal to or greater than the error threshold is calculated may be specified as the specific feature combination. When such a specifying method is adopted, a plurality of feature combinations can be specified as specific feature combinations.

  Next, the new feature item adding unit 36 generates a new feature item based on the specific feature combination. In the present embodiment, a new feature item is generated indicating whether or not a specific feature combination is applicable. 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 adding unit 36 reads “December A new feature item “is the fourth week?” Is generated. Then, “1 (Yes)” and “0 (No)” are generated as possible values for the new feature item.

  Then, the new feature item adding unit 36 adds the generated new feature item to the feature definition information 22. Thereby, the update feature definition information 22b is generated. In addition, when a plurality of specific feature combinations are specified, the new feature item adding unit 36 generates a plurality of new feature items respectively corresponding to the plurality of specific feature combinations and a plurality of possible values for them, and the feature definition It adds to the information 22.

  FIG. 7 shows an example of the 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.

  As described above, the new feature item adding unit 36 adds a new feature item defined by the feature combination that is a combination of a plurality of feature items defined in the initial feature definition information 22 a and values thereof to the feature definition information 22. Is done. When a new prediction model is constructed by the prediction unit 28 based on the updated feature definition information 22b, the new prediction model is improved in prediction accuracy over the old prediction model constructed based on the initial feature definition information 22a. Expected.

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

  Also, according to the feature combination specifying method in the present embodiment, one specific feature combination is specified in one new feature item adding process, and one new feature item is added accordingly. Thereby, the change of the feature definition information 22 in one new feature item addition process is minimized, that is, the amount of fluctuation of the prediction model in one new feature item addition process is minimized. This is based on the assumption that the prediction accuracy of the prediction model may not be improved by adding a new feature item once, assuming that the prediction model is repeatedly changed by repeatedly performing the new feature item addition processing. And gradually changing the prediction model while minimizing the amount of variation leads to improving the prediction accuracy of the prediction model earlier.

  When the new feature item addition process is repeatedly performed, in the second and subsequent processes, it is possible to define a feature combination including the new feature item added in the first process. 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 determines, for example, whether the feature combination is “fourth week in December?” It is possible to define a feature combination of a value “1” with respect to and a value “1” with respect to 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 necessarily improved from the prediction accuracy of the prediction model constructed based on the initial feature definition information 22a. Accordingly, the new feature item adding unit 36 confirms that the prediction accuracy of the prediction model has 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 adding unit 36 identifies a specific feature combination having the maximum average error from among a plurality of feature combinations defined by the feature combination defining unit 32 by the same processing as described above, and sets the specific feature combination to the specific feature combination. A new feature item is generated based on this. Then, the new feature item addition unit 36 provisionally adds the generated new feature item to the feature definition information 22. Thereby, provisional feature definition information to which a new feature item is provisionally added is generated.

  Then, each part below the feature generation part 26 performs the process similar to the above-mentioned using the said temporary feature definition information. Specifically, the feature generation unit 26 generates a provisional feature based on the provisional feature definition information and the past data group, and the prediction unit 28 uses the provisional prediction model based on the provisional feature and the sales performance value of the learning data group. And a predicted value for each verification data is calculated based on the temporary features of the temporary prediction model and 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 predicted sales value predicted using the provisional prediction model. Then, the error calculation unit 30 calculates an average value of a plurality of errors corresponding to all the verification data (hereinafter referred to as “temporary total average error”).

  Here, the new feature item adding unit 36 is predicted using the sales model value of each verification data calculated in the process of generating the new feature item and the prediction model constructed based on the initial feature definition information 22a. In addition, an average value of a plurality of errors with each sales forecast value (hereinafter referred to as “total average error before processing”) is compared with a provisional total average error. If the provisional total average error is smaller than the pre-processing total average error, it means that the prediction accuracy of the prediction model has been improved by adding a new feature item, so the provisionally added new feature item is formally added to the feature definition information 22. To do. On the other hand, if the provisional total average error is equal to or greater than the pre-processing total average error, it means that the prediction accuracy of the prediction model has decreased due to the addition of the new feature item, so the new feature item adding unit 36 has provisionally added the new feature item. Is deleted. 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 creates a feature including 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 S <b> 16, the prediction unit 28 constructs a prediction model based on each feature generated based on the learning data group in step S <b> 14 and the sales record value included in each learning data.

  In step S18, the prediction unit 28 predicts a sales predicted 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 sales record value and the sales prediction value calculated in step S18 for each verification data.

  In step S22, the error calculation unit 30 calculates an average value of a plurality of errors for all the verification data calculated in step S20 this time (hereinafter referred to as “current total average error”) 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, and is smaller than the average value of the plurality of errors for all the verification data (hereinafter referred to as “previous total average error”). Judging. If the condition is satisfied, the process is terminated; otherwise, the process proceeds to step S24. This time, since there is no previous average error value because of the initial processing, the process proceeds to step S24. Details of step S22 will be described later.

  In step S24, the feature combination definition unit 32 defines a plurality of feature combinations including a plurality of feature items and 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, for the feature combination defined in step S24, among a plurality of errors corresponding to each verification data calculated in step S20, a plurality of verification data corresponding to the feature combination. Are extracted, and an average value (average error) of the extracted errors is calculated. This is performed for each feature combination defined in step S24, and an average error corresponding to each feature combination is calculated.

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

  In step S <b> 30, the new feature item adding unit 36 adds the new feature item generated in step S <b> 28 and possible values for the new feature item to the feature definition information 22. Thereby, the update feature definition information 22b is generated.

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

  As described above, in the present embodiment, the processing from step S <b> 14 to step S <b> 30 is repeated, and new feature items are sequentially added to the feature definition information 22. Thereby, it is expected that the prediction accuracy of the prediction model constructed by the prediction unit 28 is sequentially improved.

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

  This is based on the fact that the relationship between the number of repetitions of new feature item addition processing 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. When the number of items becomes relatively large, the total average error does not change much by one new feature item addition process. Therefore, when the difference between the current total average error and the previous total average error is less than or equal to the threshold value, it can be determined that the total average error will not decrease drastically any more, so the new feature item addition process is repeated. .

  It should be noted that conditions other than those described above can be adopted as the termination condition for the repetition process of the new feature item addition process. For example, a condition may be that the number of feature items included in the feature definition information 22 has reached a predetermined number, or a condition that the number of repetitions 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 possible combinations between the feature items included in the feature definition information 22 and possible values. According to this, all feature combinations can be used as candidates for specific feature combinations without omission, while the number of defined feature combinations may be enormous, and a feature combination defining unit 32, a feature combination error calculating unit 34, or the amount of processing of the new feature item adding unit 36 may increase.

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

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

  FIG. 10 shows an example of the hierarchical relationship in the present embodiment. In the example of FIG. 10, the feature item “year” is defined as the first layer, the feature item “month” as the second layer, the feature item “week” as the third layer, and the feature item (day of the week) as the fourth layer. . As described above, the hierarchical relationship between a plurality of feature items is a layer that conforms to 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 modes.

  The feature combination definition unit 32 defines a feature combination based on the hierarchical relationship defined in the feature definition information 22. Specifically, first, the feature combination defining unit 32 selects one feature item included in the feature combination. Next, in the hierarchical relationship, a feature item belonging to a layer adjacent to the layer to which the one feature item belongs is selected as the other feature item. Then, a feature combination is defined by combining the one feature item (and a value for the feature item) and the other feature item (and a value for the feature item). In other words, in the second embodiment, the feature combination defining 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. For example, the feature combination defining unit 32 defines a feature combination of a feature item “year” belonging to the first layer and a feature item “month” belonging to the second layer, but a feature item “year” belonging to the first layer. And the feature combination of 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 the past data, the combination of related items “year” and “week”, which are items separated from each other in the hierarchy, particularly affects sales (for example, sales fluctuate suddenly only in the fourth week of each month in 2015) It can be said that there is relatively little. Therefore, it can be said that there is a low possibility that the average error of the feature combinations including feature items whose hierarchies are separated from each other will increase. Accordingly, the feature combination definition unit 32 does not define feature combinations that include feature items that are separated from each other in the hierarchy, and thus feature combinations that are defined while leaving a feature combination that is likely to have a large average error. Can be reduced.

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

  Thereby, the feature combination of the feature item “March 2015” in the second layer and the feature item “week” in the third layer becomes possible by the addition processing of the new feature item. In this manner, feature combinations including three or more feature items can be defined 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 a feature combination are defined as values related to the same past data (attention past data). For example, the feature combination P1 illustrated in FIG. 6 has a value of “12” for the feature item “month” of a certain target past data, and a value of “3” for the feature item “week”. It means that.

  When the past data group is composed of a plurality of past data arranged in 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 every day as in the present embodiment, for example, whether the day before the prediction target day is a holiday or a weekday, or whether the day after the prediction target day is a holiday or a weekday Etc. may affect the sales forecast on the forecast date. In such a case, for example, the prediction accuracy of the prediction model can be further improved by constructing the prediction model in consideration of whether 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 uses a value related to the past featured data for the feature item selected from the feature item group defined in the feature definition information 22, and also from the feature item group. A feature combination including a value related to past data different from the past past data of interest for the selected feature item is defined.

  For example, in the third embodiment, the feature combination defining unit 32 sets the value “0 (weekday)” of the attention past data (today) for the feature item “holiday or weekday” and the previous day for the feature item “whether it is holiday or weekday”. A feature combination of value “1 (weekday)” is defined. Further, the two feature items included in the feature combination need not be the same. For example, the value “0 (weekday)” for the feature item “holiday or weekday”, the previous day value “ A feature combination of “1 (Sunday)” may be defined. Further, the plurality of values included in the feature combination need not relate to values of past data that are continuous in time series, and may be values of the current day and values of, for example, two days ago and one week ago.

  When such a feature combination is defined and the average error for the feature combination is maximized by the feature combination error calculation unit 34, a new feature item “whether the current day is a weekday and the previous day is a holiday” or the like based on the feature combination is defined. It is added to the information 22. Accordingly, the prediction unit 28 can construct 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 influence of each feature item defined in the feature definition information 22 on the predicted value of the target item (sales in this embodiment) is large. This is different from the first embodiment in that the processing based on the contribution degree indicating 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 feature group related to the learning data group generated by the feature generation unit 26 and the sales result value included in the learning data group. Based on the data feature, a sales forecast value for the verification data is calculated. Next, in the feature related to the verification data, a change feature is generated by randomly changing the value for the feature item of interest. And the prediction part 28 calculates the sales prediction value with respect to the said verification data based on a prediction model and a change feature. The greater the difference between the sales forecast value calculated here based on the changed feature and the sales forecast value calculated in advance, the higher the contribution of the feature item of interest. Therefore, the prediction unit 28 calculates the contribution degree of the feature item of interest as the difference between the sales prediction value based on the changed feature and the sales prediction value calculated in advance increases.

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

  The feature combination definition unit 32 generates a feature combination based on the contribution degree of each feature item. Specifically, a contribution threshold is set in advance, and when defining a feature combination, feature items whose contribution is less than or equal to the contribution threshold are not selected. When the feature combination error calculation unit 34 calculates an average error for a feature combination including a feature item having a low contribution, it can be said that the average error is unlikely to increase. Therefore, by not defining feature combinations that include feature items that are equal to or less than the contribution threshold value, it is possible to reduce the number of feature combinations that are defined while leaving feature combinations that are likely to have a large average error.

  In the first embodiment, the new feature item adding unit 36 specifies a specific feature combination based on an average error calculated for each feature combination. In the fourth embodiment, each feature combination is further added. The specific feature combination is specified based on the contribution degree of the feature item included in the. For example, a product of an average error of a certain feature combination and an average value or a sum 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 feature combination method based on an index value, for example, a feature combination having a maximum index value among a plurality of feature combinations may be set as a specific feature combination, or a plurality of feature values may be arranged in descending order. After ranking the feature combinations, a plurality of feature combinations (for example, the first to third ranks) at the top of the ranking may be specified feature combinations, or an index threshold value may be provided in advance. All feature combinations for which index values equal to or greater than the index threshold are calculated may be specified feature combinations.

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

  As mentioned above, although embodiment which concerns on this invention was described, this invention is not limited to the said embodiment, A various change is possible unless it deviates from the meaning of this invention.

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

Claims (7)

A learning data group that is a part of a past data group including a past actual value for a target item to be predicted and a related item group related to the target item, and candidate items to be used for prediction among the related item group A prediction unit that predicts each prediction value for a target item of each verification data that is a part of the past data group and is a verification target, based on the feature definition information in which the feature 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;
For each of a plurality of feature items selected from a feature item group included in the feature definition information and a plurality of feature combinations composed of values for the plurality of feature items, a plurality of the verification data corresponding to the feature combination A feature combination error calculation unit for calculating a representative value of error;
A new feature item adding unit that adds, to the feature definition information, a new feature item defined by a specific feature combination that is specified based on a representative value of the error corresponding to each feature combination from the plurality of feature combinations When,
An information processing apparatus comprising: The new feature item adding unit, after specifying the new feature item, an average value of errors related to the verification data group calculated based on the provisional feature definition information to which the new feature item is provisionally added is the new feature item Adding the new feature item to the feature definition information when the error is smaller than the average value of 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. The 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 time series,
The feature combination includes a value relating to attention past data for the feature item selected from the feature item group and a value relating to past data other than the attention past data for the feature item selected from the feature item group.
The information processing apparatus according to claim 1. The prediction unit calculates, for each feature item included in the feature item group, a degree of contribution indicating a magnitude of an influence that a value for each feature item has on the value of the target item,
The feature combination error calculation unit does not select a feature item whose contribution is equal to or less than a threshold when defining the feature combination.
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 contributions of a plurality of feature items included in each feature combination.
The information processing apparatus according to claim 5. Computer
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 a verification data group that is a part of the past data group Each of the verification data included in the verification data group based on each value for the related item group and feature definition information in which the feature item group that is a candidate for the item used for prediction is defined. A prediction unit that predicts each prediction value for a target item of each verification data that is a part of the past data group and is a target of verification;
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;
For each of a plurality of feature items selected from a feature item group included in the feature definition information and a plurality of feature combinations composed of values for the plurality of feature items, a plurality of the verification data corresponding to the feature combination A feature combination error calculation unit for calculating a representative value of error;
A new feature item adding unit that adds, to the feature definition information, a new feature item defined by a specific feature combination that is specified based on a representative value of the error corresponding to each feature combination from the plurality of feature combinations When,
An information processing program that functions as a computer program.

Images