JP2021012649A - Information processing device, information processing method, and program - Google Patents

Abstract

To allow for efficiently extracting feature quantities.SOLUTION: An information processing device is provided, comprising a memory and a processor. The processor is configured to perform steps of: searching for multiple feature quantities that can be extracted using information contained in multiple tables; granting metadata for the multiple feature quantities; and determining at least one feature quantity from among the multiple feature quantities based on the metadata.SELECTED DRAWING: Figure 2

Description

Embodiments of the present invention relate to information processing devices, information processing methods and programs.

Extracting and acquiring features from a plurality of related tables is performed, for example, in the field of machine learning. This is executed by associating data or the like based on an ID (identifier: Identifier) common to a plurality of tables and calculating a feature amount from the associated data or the like. However, a table generally used for such statistical processing, machine learning, etc. has a huge amount of data. For this reason, extracting features in all tables due to the number of operations used for synthesis or the amount of data itself leads to a shortage of computer resources or an increase in cost, and features that were originally desired to be acquired. The amount may not be available.

J. M. Kanter, et.al., "Deep feature synthesis: Towards automating data science endeavors," 2015 IEEE International Conference on Data Science and Advanced Analytics (DSAA), 19-21 Oct. 2015 H. T. Lam, et.al., "One button machine for automating feature engineering in relational databases," arXiv.org, arXiv: 1706.0327, https://arxiv.org.abs/1706.00327v1, 1 Jun., 2017

One embodiment provides an information processing device, an information processing method, and a program for efficiently extracting feature quantities.

According to one embodiment, the information processing device comprises a memory and a processor. The processor searches for a plurality of feature quantities that can be extracted using information contained in a plurality of tables, adds metadata to the plurality of feature quantities, and the plurality of features based on the metadata. It is configured to determine at least one feature from the features of and to perform.

An example of multiple tables. The block diagram which shows the function of the feature amount extraction apparatus which concerns on one Embodiment. A flowchart showing a feature amount extraction process according to an embodiment. The figure which shows the implementation example of the feature amount extraction apparatus which concerns on one Embodiment.

Hereinafter, embodiments will be described with reference to the drawings. The table used in the description is shown as an example, and does not limit the present embodiment.

For example, it may be a simpler table than the table in the drawing, or it may be a table having more complicated mutual relationships. Further, the number of rows and the number of columns are not limited to this, and the number of rows and columns may be further increased. Further, examples for purchasing a product are given, but the present invention is not limited to this, and for example, there may be a plurality of tables that can be learned by machine learning. Further, although it is divided into a plurality of tables, the present invention is not limited to this, and a table may be a table in which a plurality of tables are combined in one table.

FIG. 1 is a diagram showing an example of a plurality of tables (table groups) to be extracted. In the present embodiment, effective features are extracted from such a plurality of tables.

The customer table is a table that associates customer IDs with customer attributes. A customer ID is assigned to each customer, for example, and the customer ID is associated with gender and age data.

The order table is a table that associates the order ID with the customer ID. This order ID is, for example, an ID given for each order when the customer places one order. The order is, for example, an order for the purchase of a product and may include the purchase of one or more products per order.

The processing table is a table that associates the processing ID with the order ID and the product ID. This processing ID is an ID given to each ordered product, and one product ID is associated with the ordered ID. When one order includes purchase orders for a plurality of products, a processing ID is assigned to each product ID. For example, in the figure, product IDs 3 and 1 are associated with processing IDs 1 and 2, and the same order ID 1 is associated with these two processes. This means that the order for order ID1 has placed a purchase order for the products with product IDs 1 and 3.

The product table is a table that associates the product ID with the price. For each product ID, the price of the product indicated by the product ID is stored.

Features are extracted from such a plurality of tables. The feature quantity is, for example, statistical processing from the information in these tables, such as mean, mode, minimum value, variance, mode, median, covariance, and other statistical processing, or differentiation, integration, and FFT (Fast Fourier Transform). ), Etc., or a combination thereof.

FIG. 2 is a block diagram showing the functions of the feature amount extraction device according to the present embodiment. The feature amount extraction device 1 includes an input unit 10, a storage unit 12, a feature amount search unit 14, a metadata addition unit 16, a feature amount determination unit 18, a feature amount extraction unit 20, and an output unit 22. To be equipped.

The input unit 10 receives data input and a request from the user. The input unit 10 connects the feature amount extraction device 1 and the outside via, for example, an interface connected to a network or the like, or a user interface that accepts input from a user.

The storage unit 12 stores data, programs, and the like necessary for extracting the feature amount in the feature amount extraction device 1. Further, the storage unit 12 may include a memory for temporarily storing a composite value of various numerical values in the feature amount extraction process.

The feature amount search unit 14 searches for a feature amount that can be acquired from the data of a plurality of tables. The search for features may be performed by a general method. For example, in the data of a plurality of tables, a combination that can be executed for the data among the prepared statistical processing, numerical processing, and the like is searched, and the feature amount that can be extracted is searched.

The metadata addition unit 16 adds metadata to the feature amount searched by the feature amount search unit 14. The metadata given is, for example, the operation used to synthesize the features, the number of tables to be followed for synthesis (the number of paths to the table), the data of the source table, for example, the uniqueness of the data. It is information about the type of data such as distribution, continuous value, and categorical value. A plurality of these data may be given in combination.

Hereinafter, the feature amount is described as "synthesis", which means that the feature amount is synthesized from a plurality of variables, and is synonymous with the "extraction" of the hidden feature amount. That is, the apparatus and method according to the present embodiment search for a plurality of feature quantities that can be extracted from the data of the table group, narrow down the extracted plurality of feature quantities based on the priority, and obtain a desired feature quantity. In other words, it is extracted further.

The feature amount determination unit 18 selects and determines one or a plurality of feature amounts to be extracted from the feature amounts searched by the feature amount search unit 14 based on the metadata given by the metadata addition unit 16. The priority calculated based on the metadata may be used for determining (selecting) the feature amount. For example, a predetermined rule, function, or the like may be applied to the calculation or the like in the extraction process to calculate the priority. The feature amount determination unit 18 calculates the priority of the feature amount searched under the preset conditions, and determines the feature amount to be extracted based on the priority. Further, the feature amount determination unit 18 may extract the feature amount in the order of the calculated priority.

The priority may be preset by the user. For example, in the case of statistical processing, the priority is set as average> variance> maximum = minimum> median. The priority may be set for another amount in the same manner. For example, the shorter the path length, the higher the priority. The storage unit 12 may receive and store these priority assignment settings from the user via the input unit 10. The feature amount determination unit 18 calculates the priority for each searched feature amount based on the preset processing and the priority for each amount.

As for the above priority, the user may determine a priority suitable for a desired task such as demand forecasting. Further, as described below, the feature amount determining unit 18 may select and determine the extraction target based on the calculated importance.

The feature amount determination unit 18 may acquire a function for calculating the priority of the searched feature amount by learning. Specifically, the priority is acquired based on the importance of the searched features.

For example, based on the concept of permutation importance, when a task using a plurality of features for a certain table group, for example, statistical processing or processing by a neural network exists (is set), it is used in the processing. Calculate the importance of the feature amount. For this importance, for example, a certain feature amount is focused on, and the content of the feature amount is shuffled for processing. Those whose processing results do not change much before and after shuffling are less important, and those whose processing results change significantly before and after shuffling are more important. Features may be extracted with this importance as a priority.

Further, the processing is not limited to shuffle, and the processing may be performed by fixing the value to an inappropriate value of the feature amount based on the average value, the dispersion value, or the like of the feature amount. In this case as well, the importance may be higher as the result changes before and after the value is fixed.

In the above, it may be determined that the result does not change much / greatly changes, for example, based on statistics such as the average value and the variance value of the processing results before shuffling. For example, it may depend on whether or not the value of the processing after shuffling is within ± 1σ from the average value of the processing before shuffling. 1σ is not given as an example and is not limited. The present invention is not limited to this, as long as it is appropriately determined under predetermined conditions, and as another example, it may be based on statistics such as the average value and variance of the processing after shuffling. Further, it may be based on a method equivalent to the commonly used method of permutation importance.

The above example does not necessarily have to be independent. That is, a feature amount to be an extraction candidate is selected based on the priority calculated by a predetermined rule, a function, or the like, and the value is shuffled or the like to obtain the importance of the selected feature amount. The priority may be updated based on the importance. The priority is not limited to this, and the priority may be calculated by taking into consideration both the value calculated based on the predetermined rule and the above-mentioned importance as long as the process can appropriately calculate the priority. ..

The feature amount extraction unit 20 synthesizes and extracts the feature amount determined by the feature amount determination unit 18 from a plurality of tables. When the feature amount determining unit 18 determines a plurality of feature amounts as extraction targets, the feature amount extracting unit 20 may synthesize and extract the feature amounts that can be synthesized in a predetermined time set in advance, for example. Then, a number of features up to a predetermined maximum value set in advance may be combined and extracted.

If there is a predetermined time set in advance, the feature amount extraction unit 20 may measure the time in itself or in another unit, and the feature amount that can be acquired when the predetermined time has elapsed may be extracted. Alternatively, up to the feature amount being acquired after a predetermined time has elapsed may be extracted.

When there is a predetermined maximum value set in advance, the feature amount determination unit 18 may determine the feature amount equal to or less than the maximum value as an extraction target, and the feature amount extraction unit 20 may extract these feature amounts. As another example, the feature amount extraction unit 20 may extract the feature amounts up to a predetermined maximum value in descending order of priority from the feature amounts determined by the feature amount determination unit 18. In this case, the feature amount determination unit 18 may output the priority to the feature amount extraction unit 20, or the feature amount determination unit 18 rearranges the feature amounts to be extracted in order of priority, and the feature amount extraction unit 20 may perform the feature amount extraction unit 20. The features may be extracted in the sorted order.

The output unit 22 outputs the feature amount extracted by the feature amount extraction unit 20. Here, in the present embodiment, the output is a concept including not only the output to the outside of the feature amount extraction device 1 but also the storage in the storage unit 12.

FIG. 3 is a flowchart showing a processing flow of the feature amount extraction device 1 according to the present embodiment. Hereinafter, the feature amount extraction process according to the present embodiment will be described with reference to FIG.

First, the feature amount extraction device 1 receives input of table information and the like via the input unit 10 (S100). The information to be accepted is not only the information in the table, but also the type of metadata to be given, the function for converting the metadata to the priority, and the required information among the above-mentioned predetermined time and predetermined maximum value. The above numerical values and the like may be set in advance in the feature amount extraction device 1 depending on the situation.

Next, the feature amount search unit 14 searches for a feature amount that can be synthesized from a plurality of table information (S102). The feature amount search unit 14 searches for a feature amount that can be synthesized from a plurality of tables based on, for example, preset statistical processing, numerical processing, and the like.

For example, when the four tables as shown in FIG. 1 are the tables for which the feature amount is to be extracted, the feature amount such as the average price for each order and the average price of the purchased product in one order for the customer, For the product, the gender, age, information about the product ordered together with one order, etc. of the customer who purchased the product can be synthesized as a feature quantity. The above is shown as an example, and there are many other quantities that can be synthesized as feature quantities. It is also possible to acquire information such as the number of passes, the type of operation to be used, and the number of each feature at the same timing.

Returning to FIG. 3, the metadata addition unit 16 then adds metadata (S104). The metadata addition unit 16 adds metadata to the feature amount based on the information about the feature amount searched by the feature amount search unit 14. As described above, the metadata is given to each of the searched features based on the information of the operator, the number of passes, and the data of the table.

Next, the feature amount determination unit 18 calculates the priority to be extracted for the feature amount searched by the feature amount search unit 14 based on the added metadata (S106). As described above, as an example, the priority is preset by the user for each element. The priority for each feature is calculated by summing the priorities for each element set by the user. For example, the priority of the calculation used for synthesizing the features is set as average = 5, variance = 4, maximum = minimum = 3, ..., And the priority is max (6) with respect to the length of the path. -It is assumed that the path length is set to 0). In this case, the length of the path is 1, and the priority of the feature amount synthesized by the average may be calculated as 5 + 5 = 10. Instead of the sum, it may be calculated as the product (25), or it may be calculated as the maximum value (5) for each element. Furthermore, when using the maximum and the average, the method of calculating the priority, such as taking the sum of these and taking the product with other elements, is trained to be optimal by the user or by machine learning. You may decide freely by the method. For example, when the combination of certain operations is important, the priority of the feature quantity including the combination of the operations may be particularly high.

In this way, any operation can be used as long as it can be calculated so that the priority that the user wants to increase is higher. Of course, the priority may be set for the table data information in the same manner, and the priority may be calculated based on the composition method, the number of passes, and the table data information, or from at least one of these three. The priority may be calculated.

It should be noted that although the extraction is performed when the priority is high, the design may be such that the extraction is performed with priority when the priority is low. In this case, the user sets the priority so that the process to be prioritized becomes smaller, and based on this setting, the calculation method is determined so that the priority of the feature amount to be extracted with priority becomes smaller.

Next, the feature amount determination unit 18 determines the feature amount to be extracted (S108). For example, the feature quantities to be extracted are determined in the order of priority calculated in S106. As described above, the feature amount determination unit 18 calculates the priority, transmits the priority information to the feature amount extraction unit 20, and sets the extraction target based on the priority received by the feature amount extraction unit 20. The feature amount may be determined.

Next, the feature amount extraction unit 20 synthesizes and extracts the priorities in the determined order (S110). For example, in the case of the plurality of tables shown in FIG. 1, the average price for one order for each customer is calculated. For example, the case of calculating the customer with the customer ID 2 will be described.

The feature amount extraction unit 20 acquires data necessary for obtaining an average value from each table. First, the order ID ordered by the customer is extracted from the order table. Order IDs 1 and 4 are eligible. Next, the products purchased under order IDs 1 and 4 are extracted from the processing table. For order ID 1, product 3 and product 1 are purchased, and for order ID 4, product 3 is purchased.

Based on this result and the price in the product table, the average value for each purchased order of the customer ID 2 which is the feature amount to be extracted is calculated. In this case, it can be calculated as follows with reference to each table.

In this way, the feature amount extraction unit 20 synthesizes the feature amount to be extracted by directly referring to the information in other tables and calculating the feature amount to be extracted. That is, the feature amount extraction unit 20 calculates another feature amount (in the above example, a feature amount such as the sum of each order ID) in each table, and the feature amount to be extracted without holding it as a table. Is synthesized. Similarly, for other customer IDs, the average purchase price per order is calculated and extracted as a feature amount.

Returning to FIG. 3, the feature amount extraction unit 20 then outputs the feature amount extracted via the output unit 22 (S112). The extracted feature amount may be stored in a storage or the like inside or outside the feature amount extraction device 1.

The feature amount extraction device 1 determines whether or not the feature amount extraction is completed (S114), and if the feature amount extraction is completed (S114: YES), ends the process. When the extraction of the feature amount is not completed (S114: NO), the next feature amount is extracted based on the priority.

As described above, according to the present embodiment, when data is selected over a plurality of tables and feature quantities are synthesized and extracted from the selected data, all the searched feature quantities are calculated and temporarily stored. It is possible to extract important features without holding them in the region. In this way, by synthesizing and extracting the features with high priority without preserving the features during the synthesis, it is possible to significantly reduce the memory used during the feature extraction. .. In addition, since it is possible to select and synthesize high-priority feature quantities without synthesizing other searched feature quantities, it is possible to reduce the processing time.

As described above, by using the feature amount extraction device 1 according to the present embodiment, the important feature amount extraction can be executed on a flexible schedule, and the problem that the important feature amount cannot be extracted without being able to overcome the memory can be solved. It becomes possible to deal with it. In addition, since it is possible to narrow down the extraction to important features, it is possible to avoid the extraction of unnecessary features that may cause noise in machine learning.

In the feature amount extraction device 1 in the above-described embodiment and modification, each function may be a circuit composed of an analog circuit, a digital circuit, or an analog / digital mixed circuit. Further, a control circuit for controlling each function may be provided. The mounting of each circuit may be by ASIC (Application Specific Integrated Circuit), FPGA (Field Programmable Gate Array) or the like.

In all the above descriptions, at least a part of the feature amount extraction device 1 may be composed of hardware or software, and may be executed by a CPU or the like by information processing of the software. When it is composed of software, the feature extraction device 1 and a program that realizes at least a part of the functions are stored in a storage medium such as a flexible disk or a CD-ROM, read by a computer, and executed. You may. The storage medium is not limited to a removable one such as a magnetic disk or an optical disk, and may be a fixed storage medium such as a hard disk device or a memory. That is, information processing by software may be concretely implemented using hardware resources. Further, the processing by software may be implemented in a circuit such as FPGA and executed by hardware. The generation of the model and the processing after inputting to the model may be performed by using an accelerator such as a GPU, for example.

For example, the computer can be made into the device of the above-described embodiment by reading the dedicated software stored in the storage medium readable by the computer. The type of storage medium is not particularly limited. Further, by installing the dedicated software downloaded via the communication network on the computer, the computer can be used as the device of the above embodiment. In this way, information processing by software is concretely implemented using hardware resources.

FIG. 4 is a block diagram showing an example of a hardware configuration according to an embodiment of the present invention. The feature amount extraction device 1 includes a processor 71, a main storage device 72, an auxiliary storage device 73, a network interface 74, and a device interface 75, and these are connected as a computer device 7 via a bus 76. realizable.

Although the computer device 7 of FIG. 4 includes one component, the computer device 7 may include a plurality of the same components. Further, although one computer device 7 is shown in FIG. 4, software may be installed on a plurality of computer devices, and each of the plurality of computer devices may execute a part of processing different from the software. ..

The processor 71 is an electronic circuit (processing circuit, processing circuitry) including a control device and an arithmetic unit of a computer. The processor 71 performs arithmetic processing based on data and programs input from each apparatus of the internal configuration of the computer apparatus 7, and outputs the arithmetic result and the control signal to each apparatus and the like. Specifically, the processor 71 controls each component constituting the computer device 7 by executing an OS (operating system) of the computer device 7, an application, or the like. The processor 71 is not particularly limited as long as it can perform the above processing. Each component of the feature amount extraction device 1 is realized by the processor 71.

The main storage device 72 is a storage device that stores instructions executed by the processor 71, various data, and the like, and the information stored in the main storage device 72 is directly read by the processor 71. The auxiliary storage device 73 is a storage device other than the main storage device 72. Note that these storage devices mean arbitrary electronic components capable of storing electronic information, and may be memory or storage. Further, the memory includes a volatile memory and a non-volatile memory, but any of them may be used. The memory for storing various data in the feature amount extraction device 1 may be realized by the main storage device 72 or the auxiliary storage device 73. For example, the storage unit 12 may be mounted on the main storage device 72 or the auxiliary storage device 73. As another example, when an accelerator is provided, the storage unit 12 may be mounted in the memory provided in the accelerator. Further, a plurality of processors may be connected to one memory, or a single processor may be connected.

The network interface 74 is an interface for connecting to the communication network 8 wirelessly or by wire. As the network interface 74, one conforming to the existing communication standard may be used. Information may be exchanged by the network interface 74 with the external device 9A which is communicated and connected via the communication network 8.

The external device 9A includes, for example, a camera, motion capture, an output destination device, an external sensor, an input source device, and the like. Further, the external device 9A may be a device having some functions of the components of the feature amount extraction device 1. Then, the computer device 7 may receive a part of the processing result of the feature amount extraction device 1 via the communication network 8 like a cloud service.

The device interface 75 is an interface such as a USB (Universal Serial Bus) that directly connects to the external device 9B. The external device 9B may be an external storage medium or a storage device. The storage unit 12 may be realized by the external device 9B.

The external device 9B may be an output device. The output device may be, for example, a display device for displaying an image, a device for outputting audio, or the like. For example, there are LCD (Liquid Crystal Display), CRT (Cathode Ray Tube), PDP (Plasma Display Panel), speaker and the like, but the present invention is not limited to these.

The external device 9B may be an input device. The input device includes devices such as a keyboard, a mouse, and a touch panel, and gives the information input by these devices to the computer device 7. The signal from the input device is output to the processor 71. In addition, some or all of the processes, means, etc. of the present disclosure may be executed or functioned based on at least one of a control circuit and a storage circuit provided on the cloud via a network.

As used herein, the expressions "at least one of a, b and c (one)" or "at least one of a, b or c (one)" are a, b, c, ab, ac, bc, Includes any combination of abc. It also covers combinations with multiple instances of any element such as a-a, a-b-b, a-a-b-b-c-c. In addition, it covers the addition of other elements other than a, b and / or c, such as having a-b-c-d.

Based on all the above descriptions, additions, effects or various modifications of the embodiments may be conceived, but are not limited to the individual embodiments described above. Various additions, changes and partial deletions can be made without departing from the conceptual idea and purpose of the present invention derived from the contents defined in the claims and their equivalents. For example, in all the above-described embodiments, the mathematical formulas and numerical values used in the explanation are shown as examples, and are not limited thereto.

1: Feature extraction device,
10: Input section,
12: Memory,
14: Feature search unit,
16: Metadata assignment department,
18: Feature determination unit,
20: Feature extraction unit,
22: Output section

Claims (21)

With one or more memories
It comprises one or more processors connected to the one or more memories.
The one or more processors
Searching for multiple features that can be extracted using information contained in multiple tables,
Adding metadata for the plurality of features and
Determining at least one feature from the plurality of features based on the metadata
Is configured to run,
Information processing device. The one or more processors
To calculate the priority for each of the plurality of features based on the metadata,
Determining at least one feature from the plurality of features based on the calculated priority.
Is configured to run further,
The information processing device according to claim 1. The one or more processors
To calculate the priority for the plurality of features based on a predetermined rule,
Is configured to run further,
The information processing device according to claim 2. Extracting the determined feature amount from the information contained in the plurality of tables, and
Is configured to run further,
The information processing device according to claim 2 or 3. Performing machine learning using the extracted features,
Is configured to run further,
The information processing device according to claim 4. The metadata includes at least the processing used for extracting the feature amount, the number of paths to the table provided with the data necessary for extracting the feature amount, and information about the data in the table from which the feature amount is extracted. Including one,
The information processing device according to claim 4 or 5. The process used for extracting the feature amount is a statistical process.
The information processing device according to any one of claims 4 to 6. The one or more processors
To determine the order of extracting the feature quantities in descending order of the calculated priority.
Is configured to run further,
The information processing device according to any one of claims 4 to 7. The one or more processors
To calculate the priority based on a predetermined value for the extraction process of the plurality of feature quantities.
Is configured to run further,
The information processing device according to claim 8. The one or more processors
When a process using an extractable feature amount is preset, the priority is calculated based on the result of executing the process by changing the feature amount.
Is configured to run further,
The information processing device according to claim 8 or 9. The one or more processors
Acquiring information across the plurality of tables based on the identifier given to the information,
Extracting the feature amount based on the acquired information,
Is configured to run further,
The information processing device according to any one of claims 4 to 10. The one or more processors
Extracting the feature amount that can be extracted in a predetermined time,
Is configured to run,
The information processing device according to any one of claims 4 to 11. The one or more processors
Extracting the number of features up to a predetermined maximum value,
The information processing apparatus according to any one of claims 4 to 12, which is configured to execute the above. The one or more processors
Extracting the feature amount without calculating the feature amount other than the feature amount to be extracted and storing it as a table in the memory.
Is configured to run,
The information processing device according to any one of claims 4 to 13. One or more processors
Search for multiple features that can be extracted using the information contained in multiple tables,
Metadata is added to the plurality of features.
Based on the metadata, at least one feature quantity is determined from the plurality of feature quantities.
Information processing method that prepares for that. The one or more processors
Based on the metadata, the priority is calculated for each of the plurality of features.
The at least one feature amount is determined from the plurality of feature amounts based on the calculated priority.
The information processing method according to claim 15, further comprising the above. The one or more processors
The determined feature amount is extracted from the information contained in the plurality of tables.
The information processing method according to claim 15 or 16, further comprising the above. The metadata includes at least the processing used for extracting the feature amount, the number of paths to the table provided with the data necessary for extracting the feature amount, and information about the data in the table from which the feature amount is extracted. Including one,
The information processing method according to any one of claims 15 to 17. The one or more processors
The feature amount to be extracted is extracted without calculating the feature amount other than the feature amount and storing it as a table in the memory.
The information processing method according to any one of claims 15 to 18. On the computer
A means for searching for multiple features that can be extracted using information contained in multiple tables,
A means for imparting metadata for the plurality of features,
A means for determining at least one feature quantity from the plurality of feature quantities based on the metadata.
A program that functions as. When executed by a computer, the computer
Steps to search for multiple features that can be extracted using information contained in multiple tables,
A step of adding metadata for the plurality of features,
A step of determining at least one feature from the plurality of features based on the metadata.
A non-transitory computer-readable medium that stores a program that runs a program.

Images