JP7139932B2 - Demand forecasting method, demand forecasting program and demand forecasting device - Google Patents

Description

The present invention relates to a demand forecasting method, a demand forecasting program, and a demand forecasting device.

The general method for predicting future demand for products is to forecast the future based on trends in past sales results. Sales cannot be predicted because it does not exist. For this reason, sales of new products are predicted by searching for similar products that have already been released in the past and using past sales data of the similar products.

For example, before or at the beginning of the launch of a new product, the prediction executor searches for multiple similar products and specifies the weight of each similar product, and uses that weight to calculate the weighted addition of past sales of similar products. , techniques for demand forecasting are known. Also known is a technology that extracts similar products that have been released in the past using social media data and product attribute data that include statements about products, and predicts the sales of new products using the results of these similar products. there is

JP 2008-186413 A JP-A-2000-3388 JP 2013-182415 A

However, with the above technology, it is difficult to improve the accuracy of demand prediction for new products that have not yet been put on the market.

For example, when a forecast executor specifies similar products and assigns weights, subjective and individual factors are strong, and it is possible to quantitatively determine what kind of content of similar products will affect demand forecasts for new products. Therefore, the accuracy of the demand forecast for new products is not necessarily high. In addition, weighted addition of past sales of similar products and social media data cannot accurately predict demand for new products that cannot be expressed by simple combinations of past similar products.

An object of one aspect is to provide a demand forecasting method, a demand forecasting program, and a demand forecasting apparatus capable of improving the accuracy of demand forecasting for new products.

In the first proposal, the demand forecasting method is such that a computer extracts information from each document describing the attributes of existing products that have started to be sold or new products that have not yet started to be sold, based on preset conditions. Execute processing for extracting feature words that indicate product attributes. In the demand forecasting method, a computer generates clustering information indicating a combination of degrees of having characteristic words for each product from the frequency of appearance of characteristic words included in each of the documents. In the demand forecasting method, a computer sets the generated clustering information as an explanatory variable, and uses learning data in which the sales performance of the existing product is set as an objective variable to learn a forecasting model that forecasts the demand for the new product. Execute the process.

According to one embodiment, it is possible to improve the accuracy of demand prediction for new products.

FIG. 1 is a diagram illustrating a demand forecasting device according to a first embodiment; FIG. FIG. 2 is a functional block diagram of a functional configuration of the demand prediction device according to the first embodiment; FIG. 3 is a diagram illustrating an example of a proposal stored in a proposal DB. FIG. 4 is a diagram showing an example of information stored in a sales information DB. FIG. 5 is a diagram showing an example of information stored in a monthly sales information DB. FIG. 6 is a diagram for explaining a learning phase according to the first embodiment; FIG. 7 is a diagram illustrating an application phase according to the first embodiment; FIG. 8 is a flowchart showing the flow of processing. FIG. 9 is a diagram for explaining a demand prediction device according to a second embodiment; FIG. 10 is a diagram for explaining the learning phase according to the second embodiment; FIG. 11 is a diagram illustrating an application phase according to the second embodiment; FIG. 12 is a diagram illustrating an example of smoothing. FIG. 13 is a diagram illustrating another example of smoothing. FIG. 14 is a diagram for explaining the smoothing result. FIG. 15 is a diagram explaining the effect. FIG. 16 is a diagram illustrating a comparative example of effects. FIG. 17 is a diagram illustrating a hardware configuration example.

Embodiments of the demand forecasting method, the demand forecasting program, and the demand forecasting apparatus disclosed in the present application will be described below in detail with reference to the drawings. In addition, this invention is not limited by this Example. Moreover, each embodiment can be appropriately combined within a range without contradiction.

[Description of demand forecasting device]
FIG. 1 is a diagram illustrating a demand prediction device 10 according to the first embodiment. A demand forecasting device 10 shown in FIG. 1 is an example of a computer device that executes demand forecasting for new products before they go on sale. This demand forecasting device 10 learns a forecast model in the learning phase, and executes demand forecasting using the learned forecast model in the application phase.

As shown in FIG. 1, in the learning phase, the demand forecasting device 10 uses text information such as proposals for existing products that have already been released and proposals for new products that are created during research and development before launch or during planning. , to extract words (keywords) that represent the content. Then, the demand prediction device 10 performs clustering using the extracted keywords to generate clusters. After that, the demand prediction device 10 learns a prediction model that performs demand prediction using learning data in which the cluster result of the existing product is an explanatory variable and the sales information of the existing product is an objective variable.

In the application phase after the completion of learning, the demand forecasting device 10 inputs the new product cluster result generated in the learning phase to the learned prediction model. Then, the demand prediction device 10 acquires the output result of the prediction model as a demand prediction.

In this way, the demand forecasting device 10 performs clustering by inputting text information associated with a product, obtains a cluster having semantic unity of a keyword group, and inputs this cluster result into a forecasting model as an explanatory variable. This makes it possible to quantitatively calculate the degree of impact of each cluster on sales. Therefore, the demand prediction device 10 can improve the accuracy of demand prediction for new products.

[Function configuration]
FIG. 2 is a functional block diagram showing the functional configuration of the demand prediction device 10 according to the first embodiment. As shown in FIG. 2 , the demand prediction device 10 has a communication section 11 , a storage section 12 and a control section 30 .

The communication unit 11 is a processing unit that controls communication between other devices, such as a communication interface. For example, the communication unit 11 receives various processing start instructions and various data from the administrator, and transmits learning results, prediction results, and the like to the administrator terminal.

The storage unit 12 is an example of a storage device that stores various data, programs executed by the control unit 30, and the like, and is, for example, a memory or a hard disk. The storage unit 12 has a proposal DB 13, a sales information DB 14, a monthly sales information DB 15, a text information DB 16, a weight information DB 17, a cluster DB 18, a learning data DB 19, a learning result DB 20, and a prediction result DB 21.

The proposal DB 13 is a database that stores proposal data for existing products that have already been released and proposal data for new products that have not yet been released and are generated at the stage of research and development or planning. Specifically, the proposal DB 13 stores data of each proposal containing keywords representing information about products such as material names, target years, product details, and the like.

FIG. 3 is a diagram showing an example of a proposal stored in the proposal DB 13. As shown in FIG. As shown in FIG. 3, the proposal includes item a and item b representing the features of the product, the description of the product, and the like. Item a includes a document 1a that specifically describes information about item a, and item b includes a document 1b that specifically describes information about item b. The documents 1a and 1b also contain keywords. For example, item a is an item describing product features, and item b is an item describing the target of the product.

The sales information DB 14 is a database that stores sales information of existing products. Specifically, the sales information DB 14 stores the sales of existing products on the release date. FIG. 4 is a diagram showing an example of information stored in the sales information DB 14. As shown in FIG. As shown in FIG. 4, the sales information DB 14 stores "product, release date, sales" in association with each other.

The "product" stored here is the name of an existing product that has been released, the "release date" is the date on which the sale was started, and the "sales" is the number of sales. The example in FIG. 4 indicates that product 1 was launched on June 10, 2018 and sold 100 units that day. It should be noted that the sales information DB 14 can store not only the sales on the sales start date but also the sales on a specific sales date. In this embodiment, the product 1, product 2, and product 3 are described as existing products.

The monthly sales information DB 15 is a database that stores monthly sales information of existing products. FIG. 5 is a diagram showing an example of information stored in the monthly sales information DB 15. As shown in FIG. As shown in FIG. 5, the sales information DB 14 associates and stores "merchandise, 1st month, 2nd month, 3rd month".

The "product" stored here is the product name of an existing product that has been released, and the "1st month" and the like are the number of sales for each month, two months, and three months from the start of sales. be. In the example of FIG. 5, product 1 sold 1,000 units in the first month after its release, 300 units in the first and second months, and 100 units in the second and third months. indicates that It should be noted that it is also possible to use information on a daily basis, a yearly basis, and the like, instead of the monthly basis.

The text information DB 16 is a database that stores data on proposals for existing products and new products. Specifically, the text information DB 16 stores what kind of document is included in each item a and item b for each product.

For example, the text information DB 16 stores "product, item a, item b" in association with each other. The "product" stored here is the name of the product, and the "item a" and "item b" are information indicating the location of the document from which the keyword used for cluster classification is extracted. . For example, if the document 1a included in the item a of the product 1 and the document 1b included in the item b are the extraction sources, the text information DB 16 stores "product 1, item b" as "product, item a, item b". Document 1a, Document 1b" are stored.

The weight information DB 17 is a database that stores information on the weight of keywords included in text information. Specifically, the weight information DB 17 stores the weight of each keyword extracted from a proposal or the like.

The cluster DB 18 is a database that stores information on clusters into which products including existing products and new products are classified. Specifically, the cluster DB 18 stores the result of clustering each product with a keyword. That is, the cluster DB 18 stores a cluster ID assigned to each cluster for each product.

The learning data DB 19 is a database that stores learning data used for learning a prediction model for each month. Specifically, the learning data DB 19 stores a plurality of pieces of learning data in which the cluster ID of each product is set as an explanatory variable, and the monthly sales of each product after its release are set as objective variables. For example, the learning data DB 19 includes learning data for learning a prediction model for sales prediction for the first month, learning data for learning a prediction model for sales prediction for the second month, Stores learning data for learning each prediction model for sales prediction.

The learning result DB 20 is a database that stores the learning results of each prediction model for each month. For example, the learning result DB 20 stores the determination result (classification result) of the learning data by the control unit 30, various parameters learned by multiple regression analysis, machine learning, and the like. For example, the learning result DB 20 contains various parameters for configuring a forecast model for sales forecast for the first month, a forecast model for sales forecast for the second month, and a forecast model for sales forecast for the third month. etc. to remember.

The prediction result DB 21 is a database that stores sales prediction results of new products predicted using a learned prediction model. Specifically, the forecast result DB 21 stores the sales forecast result for the first month, the sales forecast result for the second month, and the sales forecast result for the third month for each new product.

The control unit 30 is a processing unit that controls the entire demand prediction device 10, such as a processor. This control unit 30 has a learning processing unit 40 and a prediction processing unit 50 . Note that the learning processing unit 40 and the prediction processing unit 50 are an example of an electronic circuit possessed by a processor and an example of a process executed by the processor.

The learning processing unit 40 has a word extraction unit 41, a weight calculation unit 42, a selection unit 43, a clustering unit 44, a learning data generation unit 45, and a learning unit 46, and is a processing unit that learns a prediction model for each month.

The word extraction unit 41 is a processing unit that extracts keywords appearing in each item included in proposals for existing products and new products. For example, the word extraction unit 41 performs morphological analysis on the document 1a described in the item a of the proposal for the product 1, and extracts K1a, K2a, K3a, K4a, etc. as keywords. The word extraction unit 41 also performs morphological analysis and the like on the document 1b described in the item b of the proposal for the product 1, and extracts keywords K1b, K2b, K3b, K4b, K5b, and the like.

In this manner, the word extraction unit 41 extracts keywords from the existing product proposals and the new product proposals, stores the extraction results in the text information DB 16 , and outputs them to the weight calculation unit 42 .

The weight calculator 42 is a processor that calculates the weight of each keyword. Specifically, the weight calculator 42 calculates the TFIDF (Term Frequency Inverse Document Frequency) of each keyword extracted by the word extractor 41 . In the above example, the weight calculator 42 calculates the TGIDF of the keyword "K1a" in the document 1a of the item a for the product 1 as the weight of K1a.

In this way, the weight calculator 42 calculates the weight of each keyword extracted from the existing product proposal and each keyword extracted from the new product proposal, and stores the calculation results in the weight information DB 17. , to the selection unit 43 .

The selection unit 43 is a processing unit that executes keyword selection. Specifically, the selection unit 43 excludes, from among the keywords extracted by the word extraction unit 41, the keywords whose weight is less than a predetermined value, the keywords corresponding to the stop word list, and the keywords corresponding to the part of speech to be excluded. .

In this way, the selection unit 43 executes selection from each keyword of the existing product and each keyword of the new product, and outputs the selection result to the clustering unit 44 . The stop keyword is a list of words that are excluded as keywords, and the parts of speech to be excluded are particles, etc. These are set in advance by the administrator, etc., or a general dictionary is used. be able to.

The clustering unit 44 is a processing unit that clusters products using the keywords selected by the selection unit 43 . That is, the clustering unit 44 generates clustering information indicating a combination of degrees of characteristic words for each product.

For example, the clustering unit 44 clusters each product and assigns a cluster ID to each cluster. For example, the clustering unit 44 sets the ID "C1a" to the cluster having K1a and K3a among the keywords belonging to the item a, and sets the ID "C2a" to the cluster having K2a and K3a. Similarly, the clustering unit 44 sets the ID "C1b" to the cluster having K1b and K2b among the keywords belonging to the item b, and sets the ID "C2b" to the cluster having K2b and K3b.

Then, the clustering unit 44 associates the product with the cluster ID and generates "item a (C1a, C2a), item b (C1b, C2b)" for each product. For example, when the keyword in item a of product 1 is "K2a, K3a" and the keyword in item b is "K1b, K2b", the clustering unit 44 determines "item a (C1a, C2a), item b ( C1b, C2b)” to generate “0, 1, 1, 0”.

In this manner, the clustering unit 44 clusters products including existing products and new products, stores the clustering results in the cluster DB 18 , and outputs them to the learning data generation unit 45 . Various general methods can be used as the clustering method.

The learning data generation unit 45 is a processing unit that generates learning data using the clustering result. Specifically, the learning data generation unit 45 generates learning data using the clustering results of the existing products among the clustering results as an explanatory variable and the monthly sales information as an objective variable, and stores the learning data in the learning data DB 19 .

In the above example, the learning data generation unit 45 uses “0, 1, 1, 0”, which are “item a (C1a, C2a), item b (C1b, C2b)” of product 1, as explanatory variables, Learning data is generated with the first month sales "1000", the second month sales "300", and the third month sales "100" stored in the information DB 15 as objective variables.

That is, the learning data generation unit 45 generates "(0, 1, 1, 0), 1000" as learning data for the prediction model that performs sales prediction for the first month, and generates sales prediction for the second month. Generate "(0, 1, 1, 0), 300" as learning data for the prediction model that performs sales forecast for the first month, and generate "(0, 1, 1, 0 ), producing 100”. In this way, the learning data generation unit 45 generates each learning data for the prediction model for each month using the cluster ID to which the existing product belongs.

The learning unit 46 is a processing unit that executes learning of the prediction model for each month. Specifically, the learning unit 46 uses the monthly learning data stored in the learning data DB 19 to perform supervised learning to generate monthly prediction models. For example, the learning unit 46 performs learning processing using multiple regression analysis.

Then, the learning unit 46 stores the learning result in the learning result DB 20 . The timing to end the learning process is arbitrary, such as when learning using a predetermined number or more of learning data is completed, or when the error between the target variable (label) and the output result of the prediction model is less than the threshold. can be set.

The prediction processing unit 50 is a processing unit that executes sales prediction for new products using a learned prediction model. For example, the prediction processing unit 50 reads various parameters from the learning result DB 20 and constructs a prediction model for the first month, a prediction model for the second month, and a prediction model for the third month. Then, the prediction processing unit 50 inputs the cluster ID "1, 0, 1, 0" associated with the new product to each prediction model as a feature amount, acquires each output result, and stores it in the prediction result DB 21. . In this manner, the prediction processing unit 50 performs sales prediction for the first month, second month, and third month for the new product.

[Concrete example]
Next, specific examples of the learning phase and the application phase will be described with reference to FIGS. 6 and 7. FIG. FIG. 6 is a diagram for explaining a learning phase according to the first embodiment; FIG. 7 is a diagram illustrating an application phase according to the first embodiment;

(learning phase)
As shown in FIG. 6, the text information DB 16 stores documents belonging to item a and documents belonging to item b for each product. Specifically, the text information DB 16 stores "product 1, document 1a, document 1b", "product 2, document 2a, document 2b", "product 3, document 3a, document 3b", and "new product, document a, document b".

Then, the word extraction unit 41 extracts keywords from the document of each product (S1). For example, the word extraction unit 41 extracts keywords "K1a, K2a, K3a" from the document of item a, and extracts keywords "K1b, K2b, K3b" from the document of item b.

Subsequently, the weight calculator 42 calculates the TFIDF of each keyword, and the selector 43 executes keyword selection (S2). For example, for product 1, K1a is not applicable, K2a is weight (0.7), K3a is weight (0.1), K1b is weight (0.8), K2b is weight (0.6), K3b is Selected as not applicable. For product 2, K1a is weighted (0.8), K2a is not applicable, K3a is weighted (0.7), K1b is not applicable, K2b is weighted (0.1), K3b is weighted (0.8) is selected.

For product 3, K1a is not applicable, K2a is weight (0.8), K3a is weight (0.3), K1b is not applicable, K2b is weight (0.3), K3b is weight (0.5) is selected. For new products, K1a has a weight of 0.7, K2a does not apply, K3a has a weight of 0.9, K1b has a weight of 0.7, K2b has a weight of 0.6, and K3b is not applicable is selected.

Then, the clustering unit 44 clusters existing products and new products using the results of weight calculation and keyword selection (S3). For example, for item a, the clustering unit 44 classifies products containing K1a and K3a into cluster C1a, and classifies products containing K2a and K3a into cluster C2a. For item b, the clustering unit 44 also classifies products containing K1b and K2b into cluster C1b, and classifies products containing K2b and K3b into cluster C2b.

As a result, product 1 belongs to C2a with respect to item a and belongs to C1b with respect to item b, so "0, 1, 1, 0" is generated as the clustering result "C1a, C2a, C1b, C2b". Product 2 belongs to C1a with respect to item a and belongs to C2b with respect to item b, so "1, 0, 0, 1" is generated as the clustering result "C1a, C2a, C1b, C2b".

Since product 3 belongs to C2a with respect to item a and C2b with respect to item b, "0, 1, 0, 1" is generated as the clustering result "C1a, C2a, C1b, C2b". Since the new product belongs to C1a with respect to item a and C1b with respect to item b, "1, 0, 1, 0" is generated as the clustering result "C1a, C2a, C1b, C2b".

After that, the learning data generation unit 45 generates learning data using the clustering results of the existing products among the clustering results including the existing products and the new products and the sales results of the existing products (S4 and S5).

For example, the learning data generator 45 generates three pieces of learning data for product 1 . That is, the learning data generation unit 45 sets the clustering result "0, 1, 1, 0" of the product 1 as explanatory variables, and generates learning data in which the first month's sales "1000" is set as the objective variable. . The learning data generation unit 45 sets the clustering result "0, 1, 1, 0" of the product 1 as explanatory variables, and generates learning data in which the second month's sales "300" is set as the objective variable. The learning data generating unit 45 sets the clustering result "0, 1, 1, 0" of the product 1 as explanatory variables, and generates learning data in which the third month's sales "100" is set as the objective variable.

Similarly, the learning data generation unit 45 sets the clustering result "1, 0, 0, 1" of the product 2 as explanatory variables, and generates learning data in which the first month's sales "1500" is set as the objective variable. do. The learning data generation unit 45 sets the clustering result "1, 0, 0, 1" of the product 2 as explanatory variables, and generates learning data in which the second month's sales "1000" is set as the objective variable. The learning data generation unit 45 sets the clustering result "1, 0, 0, 1" of the product 2 as explanatory variables, and generates learning data in which the third month's sales "700" is set as the objective variable.

Similarly, the learning data generation unit 45 sets the clustering result "0, 1, 0, 1" of the product 3 as the explanatory variable, and generates the learning data in which the first month's sales "5000" is set as the objective variable. do. The learning data generation unit 45 generates learning data by setting the clustering result "0, 1, 0, 1" of the product 3 as an explanatory variable and setting the second month's sales "2000" as an objective variable. The learning data generating unit 45 sets the clustering result "0, 1, 0, 1" of the product 3 as an explanatory variable, and generates learning data in which the third month's sales "500" is set as an objective variable.

Then, the learning unit 46 learns each prediction model using each generated learning data (S6). That is, the learning unit 46 learns a prediction model (learning model) using the clustering result as vector data representing product features and monthly sales information as correct information.

For example, the learning unit 46 learns the learning data "(0, 1, 1, 0), 1000" for the product 1, the learning data "(1, 0, 0, 1), 1500" for the product 2, and the A prediction model for predicting sales for the first month is learned by multiple regression analysis using learning data “(0, 1, 0, 1), 5000”.

Similarly, the learning unit 46 obtains learning data “(0, 1, 1, 0), 300” for product 1, learning data “(1, 0, 0, 1), 1000” for product 2, product 3 learning data "(0, 1, 0, 1), 2000" are used to learn a prediction model that predicts sales for the second month by multiple regression analysis.

Similarly, the learning unit 46 obtains learning data “(0, 1, 1, 0), 100” for product 1, learning data “(1, 0, 0, 1), 700” for product 2, product 3 learning data "(0, 1, 0, 1), 500" are used to learn a prediction model for predicting sales for the third month by multiple regression analysis.

(application phase)
In the application phase, as shown in FIG. 7, the prediction processing unit 50 selects "1, 0, 1,0” is extracted.

Then, the prediction processing unit 50 assigns the clustering result "1, 0, 1, 0" of the new product to the prediction model for one month, the prediction model for two months, and the prediction model for three months. input. After that, the prediction processing unit 50 outputs the output value “2500” of the prediction model for one month, the output value “20” of the prediction model for two months, and the output value “300” of the prediction model for three months. get.

As a result, the prediction processing unit 50 predicts the sales for the first month after the new product goes on sale as "2500", the sales for the first and second months as "20", and the sales for the second month. The sales forecast from the first month to the third month is predicted to be "300".

[Process flow]
FIG. 8 is a flowchart showing the flow of processing. As shown in FIG. 8, when the learning processing unit 40 is instructed to start processing (S101: Yes), it reads the proposal data stored in the proposal DB 13 (S102).

Subsequently, the learning processing unit 40 extracts keywords from the data of each proposal (S103), and calculates the weight of the keywords (S104). Then, the learning processing unit 40 selects keywords using the stop word list and weights (S105).

After that, the learning processing unit 40 performs clustering using the selected keywords (S106), and generates a clustering result in which products and cluster IDs are associated with each other (S107).

Then, the learning processing unit 40 extracts the clustering results of the existing products from the clustering including the new products as explanatory variables (S108), and generates learning data using the monthly sales information of each product (S109). Subsequently, the learning processing unit 40 executes learning of the prediction model using the learning data (S110).

After that, when the learning is completed (S111: Yes), the prediction processing unit 50 inputs the clustering result of the new product among the clustering results in the learning phase to the learned prediction model (S112) to acquire the prediction result. (S113).

In the first embodiment, an example in which clustering is performed including new products has been described, but this method is an effective method when a plan for new products is created at the time of learning. However, it is conceivable that there is no proposal for a new product at the time of learning. Therefore, in the second embodiment, an example will be described in which a prediction model is learned using proposals for existing products at the time of learning, and prediction is performed using proposals for new products at the time of prediction.

[Description of the demand forecasting device according to the second embodiment]
FIG. 9 is a diagram for explaining the demand prediction device 10 according to the second embodiment. As shown in FIG. 9, in the learning phase, the demand prediction device 10 extracts keywords representing content from proposals of existing products that have already been put on the market. Then, the demand prediction device 10 clusters existing products using the extracted words to generate clusters. After that, the demand prediction device 10 sets the cluster result of the existing product as an explanatory variable, and uses learning data in which the sales information of the existing product is set as an objective variable to learn a prediction model that performs demand prediction.

In the application phase after the completion of learning, the demand prediction device 10 extracts keywords from the proposals of the new products, and calculates the similarity between the texts of each product and the new product using the number of matching keywords. Then, the demand prediction device 10 calculates the weighted addition for each cluster ID of the existing product using the similarity between the texts of each product and the new product. That is, the demand prediction device 10 calculates the degree of dependence of the new product on the cluster to which each product belongs. After that, the demand forecasting device 10 inputs the results of the weighted addition to each forecasting model and acquires the output result of the forecasting model as a demand forecast.

[Concrete example]
Next, specific examples of the learning phase and the application phase will be described with reference to FIGS. 10 and 11. FIG. FIG. 10 is a diagram for explaining the learning phase according to the second embodiment; FIG. 11 is a diagram illustrating an application phase according to the second embodiment;

(learning phase)
A different point from Example 1 is that only existing products are clustered using only existing product information without using new product information. Specifically, as shown in FIG. 10, the text information DB 16 stores "product 1, document 1a, document 1b", ""Product 2, Document 2a, Document 2b" and "Product 3, Document 3a, Document 3b" are stored.

Then, the word extraction unit 41 extracts a keyword from the document of each product (S10). For example, the word extraction unit 41 extracts keywords "K1a, K2a, K3a" from the document of item a, and extracts keywords "K1b, K2b, K3b" from the document of item b.

Subsequently, the weight calculator 42 calculates the TFIDF of each keyword, and the selector 43 executes keyword selection (S11). For example, for product 1, K1a is not applicable, K2a is weight (0.7), K3a is weight (0.1), K1b is weight (0.8), K2b is weight (0.6), K3b is Selected as not applicable. For product 2, K1a is weighted (0.8), K2a is not applicable, K3a is weighted (0.7), K1b is not applicable, K2b is weighted (0.1), K3b is weighted (0.8) is selected. For product 3, K1a is not applicable, K2a is weight (0.8), K3a is weight (0.3), K1b is not applicable, K2b is weight (0.3), K3b is weight (0.5) is selected.

Then, the clustering unit 44 performs clustering of existing products using the results of weight calculation and keyword selection (S12). For example, as in the first embodiment, the clustering unit 44 classifies item a into clusters C1a and C2a, and classifies item b into clusters C1b and C2b.

As a result, for product 1, "0, 1, 1, 0" is generated as the clustering result "C1a, C2a, C1b, C2b". For product 2, "1, 0, 0, 1" is generated as the clustering result "C1a, C2a, C1b, C2b". For product 3, "0, 1, 0, 1" is generated as the clustering result "C1a, C2a, C1b, C2b".

After that, the learning data generation unit 45 generates learning data using the clustering result of the existing products and the actual sales of the existing products (S13). For example, the learning data generation unit 45 sets the clustering result "0, 1, 1, 0" of the product 1 as explanatory variables for the product 1, the first month sales "1000", the second month sales Three sets of learning data are generated with each of "300" and third-month sales "100" set as objective variables.

Similarly, the learning data generation unit 45 sets the clustering result “1, 0, 0, 1” of the product 2 as explanatory variables, the first month sales “1500”, the second month sales “1000” , and third month sales "700" are set as objective variables to generate three sets of learning data. Similarly, the learning data generation unit 45 sets the clustering result “0, 1, 0, 1” of product 3 as explanatory variables, the first month sales “5000” and the second month sales “2000”. , and third month sales "500" are set as objective variables to generate three learning data.

Then, the learning unit 46 learns each prediction model using each generated learning data (S14). For example, the learning unit 46 learns the learning data "(0, 1, 1, 0), 1000" for the product 1, the learning data "(1, 0, 0, 1), 1500" for the product 2, and the A prediction model for predicting sales for the first month is learned by multiple regression analysis using learning data “(0, 1, 0, 1), 5000”.

Similarly, the learning unit 46 obtains learning data “(0, 1, 1, 0), 300” for product 1, learning data “(1, 0, 0, 1), 1000” for product 2, product 3 learning data "(0, 1, 0, 1), 2000" are used to learn a prediction model that predicts sales for the second month by multiple regression analysis.

Similarly, the learning unit 46 obtains learning data “(0, 1, 1, 0), 100” for product 1, learning data “(1, 0, 0, 1), 700” for product 2, product 3 learning data "(0, 1, 0, 1), 500" are used to learn a prediction model for predicting sales for the third month by multiple regression analysis.

(application phase)
In the application phase, unlike the first embodiment, the information of the new product is used to calculate the similarity between the texts of the new product and the existing product, and the feature quantity that indicates how much the new product is related to the classified clusters. Calculate Then, prediction is executed with the feature amount of the new product as an input.

As shown in FIG. 11, the text information DB 16 stores "new product, document a, document b" as the new product proposal data "product, item a, item b". In this state, the prediction processing unit 50 extracts keywords from each document of the proposal data of the new product, calculates the weight of each keyword using TGIDF, and selects the keywords (S20). For example, the prediction processing unit 50 extracts the keywords "K1a, K3a" from the document a of the new product item a, and extracts the keywords "K1b, K2b" from the document b of the item b. Then, the prediction processing unit 50 calculates the weight of K1a "0.7", the weight of K3a "0.9", the weight of K1b "0.7", and the weight of K2b "0.6". Assume that K2a and K3b are not extracted here.

Subsequently, the prediction processing unit 50 uses a technique such as cosine similarity to calculate the inter-text similarity between the weight information of the new product and the weight information of each existing product (S21). For example, the prediction processing unit 50 calculates the inter-text similarity “0.11” for the item a and the inter-text similarity “1.00” for the item b for the existing product 1 . Similarly, for the existing product 2, the prediction processing unit 50 calculates the inter-text similarity of item a of “0.98” and the inter-text similarity of item b of “0.08”. In addition, the prediction processing unit 50 calculates the inter-text similarity of item a “0.28” and the inter-text similarity of item b “0.33” for the existing product 3 . It should be noted that not only text-to-text similarity using cosine similarity but also a general similarity calculation method such as matching rate among all keywords can be used.

After that, the prediction processing unit 50 weights and adds the dummy variable of the cluster ID of the existing product with the similarity between the texts. (S22). In the above example, the prediction processing unit 50 refers to the clustering result generated in S12 of FIG. ”, and the inter-text similarity “1.00” is set for the cluster C1b of the item b to which the product 1 belongs.

Similarly, the prediction processing unit 50 refers to the clustering result generated in S12 of FIG. A similarity between texts of “0.08” is set for the cluster C2b of item b to which 2 belongs. Also, the prediction processing unit 50 refers to the clustering result generated in S12 of FIG. A similarity between texts of "0.33" is set for the cluster C2b of the item b to which .

After that, the prediction processing unit 50 adds the inter-text similarity set for the cluster ID of each existing product to generate a feature amount (feature vector) of the new product. In the above example, the prediction processing unit 50 sets the similarity between texts of product 2 to C1a of item a to be “0.98”, and the similarity between texts of product 1 to C2a of item a to be “0.98”. "0.39" is set by adding the degree "0.11" and the text-to-text similarity "0.28" of product 3. Similarly, the prediction processing unit 50 sets the inter-text similarity of product 1 to “1.00” for C1b of item b, and sets the inter-text similarity of product 2 to “0.00” for C2b of item b. 0.08” and the text-to-text similarity of product 3 “0.33” are added to set “0.42”.

Then, the prediction processing unit 50 uses the addition result “0.98, 0.39, 1.00, 0.42” as an explanatory variable, and uses the learned prediction model for one month and the prediction model for two months. Input to the model and prediction model for three months, and obtain the prediction result (S23). In the above example, the prediction processing unit 50 sets the sales forecast for the first month after the new product is released to "2500 units" and the sales forecast for the first month to the second month as "20 units". , the sales forecast from the second month to the third month is predicted to be "300".

By the way, in Embodiment 1 and Embodiment 2, the case of performing monthly demand forecasting using the monthly forecasting model has been described, but the present invention is not limited to this. For example, by performing smoothing using the monthly demand forecast results, it is possible to estimate the forecast results for periods when no forecast is made.

For example, when the time granularity (for example, four weeks) of the learning data is larger than the desired prediction time granularity (for example, weekly), the prediction processing unit 50 of the demand prediction device 10 obtains the prediction result, then Divide equally according to the particle size. For example, the prediction processing unit 50 divides the prediction values for four weeks into four and converts them into stepwise weekly prediction values.

Next, the prediction processing unit 50 assumes that the sales after the release change over time according to a specific probability distribution. Since sales generally decline, probability distributions such as Weibull distribution, logarithmic normal distribution, and logarithmic logistic distribution are used.

Then, the prediction processing unit 50 calculates the parameters of the probability distribution using a general method so as to fit the entire plurality of prediction values, and uses the obtained parameters to make a new prediction at an arbitrary time. Calculate the value (interpolated/corrected predicted value). That is, substituting the time yields a new predicted value.

A specific example will now be described with reference to FIGS. 12 to 14. FIG. FIG. 12 is a diagram for explaining an example of smoothing, and FIG. 13 is a diagram for explaining another example of smoothing. FIG. 14 is a diagram for explaining the smoothing result.

FIG. 12 illustrates the case where smoothing is used to reduce random fluctuations. For example, since weekly sales are forecasted independently, forecast results may vary if irregular changes (forecast errors) are included in the forecast results. Therefore, irregular changes can be suppressed by interpolating or correcting predicted values by comprehensively judging prediction results for a plurality of weeks.

In the diagram shown in FIG. 12, a weekly forecast model is learned using the weekly sales performance as the objective variable, and the weekly (each week) sales forecast is calculated from the release date using the learned weekly forecast model. , is the result of smoothing the weekly predicted values. By doing so, the predicted value can be corrected, so the accuracy of the demand prediction can be improved. Note that the unit of the input predicted value is not limited to weekly. It is also possible to perform the same processing on the monthly predicted values and the predicted values for four weeks.

FIG. 13 illustrates a case of adjusting the gap between the time granularity of learning data and the time granularity of prediction. For example, there may be a gap between the unit of data used for learning the prediction model (for example, 4 weeks) and the unit for which prediction is required (for example, weekly).

For example, there is no weekly data because data is acquired and managed only for four weeks. In addition, although there is weekly data, the amount of data is insufficient due to the small amount of data, and it is not possible to create a prediction model as it is. For example, when the need arises. In these cases, general methods can only obtain prediction results in the same unit as the training data.

Therefore, by interpolating and correcting the prediction result, the prediction result is calculated with an arbitrary time granularity different from that of the learning data. FIG. 13 shows the result of smoothing the prediction values for four weeks obtained by learning with the data for four weeks, and obtaining new prediction values for each week. In other words, the four weeks' worth of predicted values are divided into four and once converted into stepwise weekly predicted values.

In this way, a prediction value that cannot be obtained from the prediction result can be estimated, so demand prediction can be performed without depending on learning data. Note that the calculation can be performed in arbitrary units such as daily, not limited to weekly. Also, the unit of input predicted values is not limited to four weeks, and the same processing can be performed on monthly predicted values and weekly predicted values.

By applying such a smoothing method to Example 1 and Example 2, the results shown in FIG. 14 can be obtained. Specifically, the forecast value for each week can be estimated using the forecast result for each month obtained from the forecast model for each month.

Although the embodiments of the present invention have been described so far, the present invention may be implemented in various different forms other than the embodiments described above.

[Effect of Example]

The effect of the above embodiment will be described with reference to FIGS. 15 and 16. FIG. FIG. 15 is a diagram explaining the effect. FIG. 16 is a diagram illustrating a comparative example of effects. As shown in FIG. 15, the method according to the above embodiment sets the results of executing text mining and clustering using the aims, target demographics, characteristics, campaign information, etc. described in the proposal as explanatory variables, Execute machine learning using learning data with attributes such as past sales performance set as objective variables.

Through such learning, the method according to the above embodiment can learn not only the initial demand and the transition of demand for a new product, but also the cumulative increase and the trend (pattern) of demand. That is, the method according to the above embodiment can learn by decomposing the demand pattern of similar products into a combination of elements.

As a result, as shown in FIG. 16, the error rate can be reduced compared to the general method A that performs constant prediction and the general method B that performs machine learning using only standard information and numerical information. According to simulations, the median error rate can be reduced to 19.3% by using the method according to the embodiment.

[Data, figures, etc.]
The numerical values, data examples, number of data, label setting contents, etc. used in the above embodiments are merely examples, and can be arbitrarily changed. A keyword is an example of a feature word. Also, the existing product is a product in the past, and may be a product whose sales have ended at the present stage, or may be a product that is currently being sold. In addition to sales, the number of contracts for mobile phones and the like can also be used as the objective variable. Also, in the above embodiment, an example of generating a monthly forecast model using monthly sales information has been described, but the present invention is not limited to this, and daily, weekly, and yearly sales information may be used. can generate a variety of predictive models.

Also, an example using items a and b of the proposal data has been described, but the present invention is not limited to this, and one or more items can be used, and the entire proposal can be used as one item. can. In addition to the proposal, a product manual or the like can also be used.

[system]
Information including processing procedures, control procedures, specific names, and various data and parameters shown in the above documents and drawings can be arbitrarily changed unless otherwise specified.

Also, each component of each device illustrated is functionally conceptual, and does not necessarily need to be physically configured as illustrated. That is, the specific forms of distribution and integration of each device are not limited to those shown in the drawings. That is, all or part of them can be functionally or physically distributed and integrated in arbitrary units according to various loads and usage conditions. For example, the learning processing unit 40 and the prediction processing unit 50 can be realized by separate devices.

Further, each processing function performed by each device may be implemented in whole or in part by a CPU and a program analyzed and executed by the CPU, or implemented as hardware based on wired logic.

[hardware]
FIG. 17 is a diagram illustrating a hardware configuration example. As shown in FIG. 17, the demand prediction device 10 has a communication device 10a, a HDD (Hard Disk Drive) 10b, a memory 10c, and a processor 10d. 17 are interconnected by a bus or the like.

The communication device 10a is a network interface card or the like, and communicates with other servers. The HDD 10b stores programs and DBs for operating the functions shown in FIG.

The processor 10d reads from the HDD 10b or the like a program that executes the same processing as each processing unit shown in FIG. 2 and develops it in the memory 10c, thereby operating the process of executing each function described with reference to FIG. 2 and the like. That is, this process performs the same function as each processing unit of the demand forecasting device 10 . Specifically, the processor 10d reads a program having the same functions as those of the learning processing unit 40, the prediction processing unit 50, and the like from the HDD 10b and the like. Then, the processor 10d executes a process for executing processing similar to that of the learning processing unit 40, the prediction processing unit 50, and the like.

Thus, the demand forecasting device 10 operates as an information processing device that executes a demand forecasting method by reading and executing a program. Further, the demand prediction device 10 can read the program from the recording medium by the medium reading device, and execute the read program to realize the same function as the embodiment described above. Note that the programs referred to in other embodiments are not limited to being executed by the demand forecasting device 10 . For example, the present invention can be applied in the same way when another computer or server executes the program, or when they cooperate to execute the program.

10 demand prediction device 11 communication unit 12 storage unit 13 proposal DB
14 Sales information DB
15 Monthly sales information DB
16 Text information DB
17 Weight information DB
18 Cluster DB
19 Learning data DB
20 Learning result DB
21 Prediction result DB
30 control unit 40 learning processing unit 41 word extraction unit 42 weight calculation unit 43 selection unit 44 clustering unit 45 learning data generation unit 46 learning unit 50 prediction processing unit

Claims (10)

the computer
extracting feature words indicating the attributes of each product based on preset conditions from each document describing the attributes of existing products that have started selling or new products that have not started selling,
generating clustering information indicating a combination of the degree of having characteristic words for each product from the appearance frequency of the characteristic words contained in each of the documents;
setting the generated clustering information as an explanatory variable, and using learning data in which the sales performance of the existing product is set as an objective variable, and executing a process of learning a prediction model for predicting demand for the new product. demand forecasting method. The extracting process extracts the feature word from each document corresponding to each of a plurality of existing products and the document corresponding to the new product,
The generating process clusters the plurality of existing products and the new product using feature words extracted from each to generate the clustering information;
The learning process generates a plurality of learning data using clustering information corresponding to each of the plurality of existing products among the clustering information, and executes a process of learning the prediction model using the plurality of learning data. 2. The demand forecasting method according to claim 1, wherein: The processing of inputting the clustering information corresponding to the new product out of the clustering information to a trained prediction model and acquiring the output result of the learned prediction model as a demand prediction for the new product, on the computer. 3. The demand forecasting method according to claim 2, wherein The extracting process extracts the feature word from each document corresponding to each of a plurality of existing products,
The generating process clusters the plurality of existing products using characteristic words extracted from each to generate the clustering information;
3. The process of learning is characterized in that a process of generating a plurality of learning data using the clustering information of a plurality of existing products and learning the prediction model using the plurality of learning data. 1. The demand forecasting method according to 1. extracting the characteristic word from the document corresponding to the new product;
calculating the degree of similarity between the new product and each of the plurality of existing products using the characteristic word of the new product and the characteristic word of each of the plurality of existing products;
Using the similarity, associate the new product with each cluster in which the plurality of existing products are clustered,
Inputting the result of associating the new product with each cluster into a learned prediction model, and obtaining the output result of the learned prediction model as a demand prediction for the new product;
5. The demand forecasting method according to claim 4, wherein said computer executes the processing. In the learning process, the clustering information is set as an explanatory variable, and the sales performance of the existing product for each predetermined period is set as an objective variable. 2. The demand forecasting method according to claim 1, wherein each forecasting model for forecasting is learned. The processing of inputting the clustering information corresponding to the new product to each learned prediction model and acquiring the output result of each learned prediction model as the demand prediction for the new product for each predetermined period 7. The demand forecasting method according to claim 6, wherein the demand forecasting method is computer-implemented. 3. The computer executes a process of performing smoothing using the result of the demand forecast for the new product forecasted for each of the predetermined periods and interpolating the forecast result for each predetermined period. 8. The demand forecasting method according to 7. to the computer,
extracting feature words indicating the attributes of each product based on preset conditions from each document describing the attributes of existing products that have started selling or new products that have not started selling,
generating clustering information indicating a combination of the degree of having characteristic words for each product from the appearance frequency of the characteristic words contained in each of the documents;
setting the generated clustering information as an explanatory variable, and using learning data in which the sales performance of the existing product is set as an objective variable, to execute a process of learning a prediction model that predicts demand for the new product. demand forecasting program. an extraction unit for extracting characteristic words indicating the attributes of each product based on preset conditions from documents describing the attributes of existing products that have started to be sold or new products that have not started to be sold;
a generation unit that generates clustering information indicating a combination of degrees of having characteristic words for each product from the frequency of appearance of the characteristic words included in each of the documents;
a learning unit that learns a prediction model that predicts demand for the new product by using learning data in which the generated clustering information is set as an explanatory variable and the sales performance of the existing product is set as an objective variable. demand forecasting device.

Images