JP7627906B2 - Address information processing device, address information processing method, and program - Google Patents

Description

The present invention relates to an address information processing device for processing address information.

Conventionally, there has been a technique for measuring the semantic similarity of texts including address information, which uses a method for calculating a similarity distance between texts based on the semantic similarity between words (Non-Patent Document 1).

Furthermore, there is a conventional technique for calculating the semantic distance between items in collaborative filtering based on k-nearest neighbors (Non-Patent Document 2).

Furthermore, a widely known technique has been to obtain a feature vector from text data including address information by one-hot encoding, and use the feature vector to perform machine learning learning processing or prediction processing.

In addition, there is a technology called federated learning, in which a server that acquires information from a mobile terminal (such as a smartphone or television) does not perform machine learning on the server, but instead performs machine learning on the mobile terminal and shares only the model parameters with the server (Non-Patent Document 3, Non-Patent Document 4).

Courtney Corley and Rada Mihalcea. Measuring the semantic similarity of texts. In Proceedings of the ACL Workshop on Empirical Modeling of Semantic Equivalence and Entailment, EMSEE ’05, p. 13?18, USA, 2005. Association for Computational Linguistics. Salmo M.S. J´unior and Marcelo G. Manzato. Collaborative filtering based on semantic distance among items. In Proceedings of the 21st Brazilian Symposium on Multimedia and the Web, WebMedia ’15, p. 53?56, New York, NY, USA, 2015. Association for Computing Machinery. Internet homepage: "Joint development of AI while maintaining confidentiality" (URL: https://chizaizukan.com/property/property169.html) Internet homepage "Federated Learning: Collaborative Machine Learning without Centralized Training Data" (URL: https://developers-jp.googleblog.com/2017/05/federated-learning-collaborative.html)

However, in the conventional technology, the processing cost is high when processing information including address information.

The address information processing device of the first invention is an address information processing device comprising: a target information acquisition unit that acquires target information having address information; a regional code acquisition unit that acquires a regional code corresponding to at least a part of the address information from a regional code storage unit in which one or more pieces of correspondence information indicating the correspondence between the regional code and at least a part of the address information are stored; a replacement information acquisition unit that acquires replacement information which is a vector including the regional code acquired by the regional code acquisition unit as an element; a replacement information processing unit that performs predetermined processing on the replacement information to acquire the processing result; and a processing result output unit that outputs the processing result.

With this configuration, the processing cost can be reduced when processing information including address information.

In addition, in the address information processing device of the second invention, compared to the first invention, the address information may include house number information or floor information, and the replacement information acquisition unit acquires a numerical value corresponding to house number information or floor information contained in a character string that does not correspond to an area code in the address information, and acquires replacement information which is a vector that also includes numerical values as elements.

With this configuration, the processing cost can be reduced when processing information including address information.

Furthermore, the address information processing device of the third invention is an address information processing device in which, compared to the first or second invention, the target information acquisition unit acquires two target information each including different address information, the regional code acquisition unit acquires a regional code corresponding to each of the two target information from the regional code storage unit, the replacement information acquisition unit acquires two replacement information which are two vectors including each regional code acquired by the regional code acquisition unit as an element, and the replacement information processing unit calculates the distance between the two vectors and acquires a processing result which is the distance.

With this configuration, it is possible to reduce the processing cost in the process of calculating the distance to information including address information.

In addition, the address information processing device of the fourth invention is an address information processing device in which, compared to the first or second invention, the target information includes other information other than address information, the target information acquisition unit acquires two pieces of target information each including different address information, the regional code acquisition unit acquires a regional code corresponding to each of the two target information from the regional code storage unit, the replacement information acquisition unit sets, for each of the two or more pieces of target information, a regional code corresponding to at least a part of the address information held by the target information as one or more elements of a vector and acquires one or more elements of the vector from the other information, and acquires replacement information which is a vector having one or more elements which are regional codes and one or more elements acquired from the other information, and the replacement information processing unit calculates the distance between the two vectors and acquires a processing result which is the distance.

With this configuration, it is possible to reduce the processing cost in the process of calculating the distance to information including address information and other information.

In addition, the address information processing device of the fifth invention is an address information processing device in which, compared to the first or second invention, the target information includes other information other than address information, the target information acquisition unit acquires two or more pieces of target information, the replacement information acquisition unit, for each of the two or more pieces of target information, sets a regional code corresponding to at least a part of the address information held by the target information as one or more elements of a vector and acquires one or more elements of the vector from the other information, and acquires replacement information which is a vector having one or more elements which are regional codes and one or more elements acquired from the other information, the replacement information processing unit performs a machine learning learning process on two or more pieces of teacher data which are the two or more pieces of replacement information acquired by the replacement information acquisition unit, acquires a processing result which is a learning model, and the processing result output unit accumulates the learning model.

With this configuration, processing costs can be reduced in machine learning learning processing for information including address information.

In addition, the address information processing device of the sixth invention is an address information processing device that includes a target information acquisition unit that acquires target information having address information, a regional code acquisition unit that acquires a regional code corresponding to at least a part of the address information from a regional code storage unit in which one or more correspondence information indicating the correspondence between the regional code and at least a part of the address information is stored, a replacement information acquisition unit that acquires replacement information that is a vector including the regional code acquired by the regional code acquisition unit as an element, a replacement information processing unit that performs machine learning prediction processing using the learning model and replacement information accumulated by the address information processing device of the fifth invention and acquires processing results, and a processing result output unit that outputs the processing results.

With this configuration, processing costs can be reduced in machine learning prediction processing for information including address information.

In addition, the address information processing device of the seventh invention is an address information processing device according to any one of the first to sixth inventions, wherein the regional code acquisition unit includes a regional code acquisition means for acquiring a regional code corresponding to at least a part of the address information held by the target information from a regional code storage unit, and a conversion code acquisition means for acquiring a conversion code paired with the regional code from a conversion code storage unit in which one or more pieces of code correspondence information managing the correspondence between the regional code and the conversion code are stored when the regional code acquired by the regional code acquisition means satisfies a conversion condition, and the replacement information acquisition unit acquires replacement information which is a vector including the conversion code acquired by the conversion code acquisition means as an element.

With this configuration, it is possible to improve the accuracy of processing information including address information.

According to the address information processing device of the present invention, it is possible to reduce the processing cost in processing information including address information.

Block diagram of address information processing device A in embodiment 1. A flowchart for explaining an example of the operation of the address information processing device A A flowchart illustrating an example of the replacement information acquisition process. A diagram showing the postal code management table The conversion code management table FIG. 2 is a diagram for explaining a processing example of the address information processing device A; FIG. 2 is a diagram for explaining a processing example of the address information processing device A; Figure showing data for US cities FIG. 1 shows the results of experiment 1. FIG. 2 shows the results of experiment 2. FIG. 3 shows the results of experiment 3. A diagram showing the restaurant management table A table showing the user's visited restaurants Overview of the computer system Block diagram of the computer system

Hereinafter, an embodiment of an address information processing device and the like will be described with reference to the drawings. Note that components with the same reference numerals in the embodiments perform similar operations, so repeated description may be omitted.

(Embodiment 1)
In this embodiment, we will describe an address information processing device that obtains a regional code from accepted address information, performs predetermined processing using replacement information, which is a vector whose elements are regional codes corresponding to at least a portion of the address information, and outputs the processing results.

In addition, in this embodiment, we will explain an address information processing device that obtains a regional code and a house number, etc. from address information, performs predetermined processing using replacement information, which is a vector whose elements are the regional code and a vector whose elements are the house number, etc., and outputs the processing results.

The above-mentioned predetermined process is, for example, a process of calculating a distance between two pieces of replacement information. The distance is, for example, a physical distance or a semantic distance, which will be described later. The predetermined process is, for example, a learning process or a prediction process of machine learning.

Further, in this embodiment, an address information processing device that performs correction processing for a unique area code such as an isolated area will be described.

1 is a block diagram of an address information processing device A according to the present embodiment. The address information processing device A includes a storage unit 1, a reception unit 2, a processing unit 3, and an output unit 4.

The storage unit 1 includes a region code storage unit 11 and a conversion code storage unit 12. The processing unit 3 includes a target information acquisition unit 31, a region code acquisition unit 32, a replacement information acquisition unit 33, and a replacement information processing unit 34. The region code acquisition unit 32 includes a region code acquisition means 321 and a conversion code acquisition means 322. The output unit 4 includes a processing result output unit 41.

Various types of information are stored in the storage unit 1. The various types of information include, for example, correspondence information (to be described later) and conversion code information (to be described later).

The area code storage unit 11 stores one or more pieces of correspondence information. The correspondence information is information indicating the correspondence between an area code and at least a part of address information. The correspondence information is, for example, information of a pair of an area code and at least a part of address information. The correspondence information is, for example, information of a link between an area code and at least a part of address information. The correspondence information may be information that allows an area code to be obtained from at least a part of address information.

An area code is information that specifies a spatial area. An area code corresponds to a part or all of address information. An area code is usually determined such that the closer an area specified by an area code is physically to another area specified by another area code, the smaller the difference between the area code and the other area code. Usually, the closer the physical distance between two areas is, the smaller the difference between the area codes of the two areas is determined.

However, there may be exceptions to the area code. For example, an area code corresponding to an exclave may be an exception. For example, there may be a large difference between the area code of the exclave and the area code of an area adjacent to the exclave, and the difference between the area code of the exclave and the area code of an area that is physically far from the exclave is small compared to the adjacent area.

The area code is, for example, a postal code. The postal code that is the area code may be a part of the postal code or the whole postal code. In other words, the area code may be, for example, only "603" out of the postal code "603-8555", or the whole code "603-8555". The area code is, for example, a part of a telephone number (for example, the area code).

The area code may be unique information that takes into consideration the location of the corresponding area. For example, the area code may be "Hokkaido...01", "Aomori...02", "Okinawa...47", etc.

Address information is information indicating an address. It is preferable that the address information includes all of the complete address information, but it may also be partial address information. In other words, it is preferable that the address information includes all information, such as "1-10 Nakadoji Myobu-cho, Shimogyo-ku, Kyoto City, Kyoto Prefecture," but it may also be partial address information, such as information only about the "prefecture" and "city/town/village," information up to the "prefecture," "city/town/village," and "block," or information only about the "city/town/village" and "ward." It is preferable that the address information includes a "banchi" (number). It is preferable that the address information includes "floor information" and "room number." It may also include a "street number," "street name," and the like. It may also include a "building name," and the like.

The conversion code storage unit 12 stores one or more pieces of code correspondence information. The code correspondence information is information that manages the correspondence between area codes and conversion codes. The code correspondence information is, for example, information on pairs of area codes and conversion codes. The code correspondence information may be information on links between area codes and conversion codes. In other words, the code correspondence information may be information that allows a conversion code to be obtained from an area code. A conversion code is a code that the address information processing device A uses when obtaining replacement information for a specific area code.

Furthermore, it is preferable that the area code storage unit 11 and the conversion code storage unit 12 are provided in the address information processing device A, but they may be present in an external device (not shown).

The reception unit 2 receives various instructions and information. The various instructions and information are, for example, target information and a start instruction. Target information is information to be processed. Target information has address information. Target information may include other information. Other information is information other than address information. The content and data structure of the other information are not important. A start instruction is an instruction to start operation of the address information processing device A.

Here, reception is a concept that includes reception of information input from input devices such as a keyboard, mouse, or touch panel, reception of information transmitted via a wired or wireless communication line, and reception of information read from recording media such as an optical disk, magnetic disk, or semiconductor memory.

The means for inputting various instructions and information may be any means, such as a touch panel, a keyboard, a mouse, or a menu screen.

The processing unit 3 performs various types of processing. The various types of processing are, for example, processing performed by a target information acquisition unit 31, a regional code acquisition unit 32, a replacement information acquisition unit 33, and a replacement information processing unit .

The target information acquisition unit 31 acquires one or more pieces of target information. The target information acquisition unit 31 acquires, for example, one or more pieces of target information accepted by the acceptance unit 2. The target information acquisition unit 31 acquires, for example, one or more pieces of target information from the storage unit 1. The target information acquisition unit 31 receives, for example, one or more pieces of target information from an external device (not shown). The method by which the target information acquisition unit 31 acquires the target information and the acquisition route of the target information are not important.

The target information acquisition unit 31 acquires, for example, two pieces of target information each including different address information.

The target information acquisition unit 31 acquires target information that is a set of two or more explanatory variables including, for example, address information. Note that the explanatory variables are information used in the learning process or prediction process of machine learning.

The regional code acquisition unit 32 acquires, from the regional code storage unit 11 , a regional code that corresponds to at least a part of the address information contained in the target information acquired by the target information acquisition unit 31 .

The area code acquisition unit 32 acquires, for example, from the area code storage unit 11, an area code included in corresponding information having information matching at least a part of the address information of the target information.

The area code acquisition unit 321 acquires, from the area code storage unit 11, an area code that corresponds to at least a part of the address information contained in the target information.

The area code acquisition unit 321 acquires, for example, an area code corresponding to each of two or more pieces of target information from an area code storage unit.

If the area code acquired by the area code acquisition means 321 satisfies the conversion condition, the conversion code acquisition means 322 acquires, from the conversion code storage unit 12, a conversion code that pairs with the area code acquired by the area code acquisition means 321.

The conversion condition is, for example, that the area code acquired by the area code acquisition means 321 exists in the conversion code storage unit 12. The conversion condition is, for example, that the area code acquired by the area code acquisition means 321 includes a predetermined numeric value or a predetermined character string.

The replacement information acquisition unit 33 acquires replacement information that is a vector including the area code acquired by the area code acquisition unit 32 as an element. A vector including an area code as an element may be a vector having an area code as one element, or may be a vector including two or more elements acquired from an area code. For example, if the area code is a postal code "603-8555", the vector may be a vector having [6038555] as one element, or may be a vector having two elements [603, 8555].

The replacement information acquisition unit 33 acquires, for example, a numerical value corresponding to the street address information or floor information contained in a character string that does not correspond to an area code in the address information, and acquires replacement information which is a vector including the numerical value as an element.

The replacement information acquisition unit 33 acquires replacement information, which is a vector having an element of a vector acquired from a postal code "600-8216" corresponding to "Higashi Shiokojicho, Shimogyo Ward, Kyoto City, Kyoto Prefecture" of the address information "721-1 Higashi Shiokojicho, Shimogyo Ward, Kyoto City, Kyoto Prefecture", an element of a vector acquired from the character string "Higashi Shiokojicho, Shimogyo Ward" by One-Hot Encoding, and an element of a vector acquired from the street address information "721-1". Such replacement information is, for example, a vector having [600, 8216][0, 0, 1, ..., 0, 0][721, 1]. Here, it is assumed that [0, 0, 1, ..., 0, 0] is a set of elements of a vector acquired from the character string "Higashi Shiokojicho, Shimogyo Ward, Kyoto City, Kyoto Prefecture" by One-Hot Encoding. Note that One-Hot Encoding is a well-known technique, and therefore a description thereof will be omitted. Note that such a method is the method of Proc1 described later.

For example, if the address information of the target information is "721-1 Higashi Shiokojicho, Shimogyo Ward, Kyoto City, Kyoto Prefecture," the replacement information acquisition unit 33 acquires replacement information having, for example, a postal code "600-8216" and address information "721-1" corresponding to "Higashi Shiokojicho, Shimogyo Ward, Kyoto City, Kyoto Prefecture." Note that it is preferable that the replacement information here is a vector (600, 8216, 721, 1). Note that this method is the method of Proc2 described later.

The replacement information acquisition unit 33 acquires, for example, two pieces of replacement information, which are two vectors including the area codes acquired by the area code acquisition unit 32 as elements.

For example, for each of two or more pieces of target information, the replacement information acquisition unit 33 sets a region code corresponding to at least a part of the address information of the target information as one or more elements of a vector, acquires one or more elements of the vector from the other information, and acquires replacement information that is a vector having one or more elements that are region codes and one or more elements acquired from the other information. Note that the process of acquiring one or more elements of the vector from the other information may be, for example, one-hot encoding, but is not essential. The process of acquiring one or more elements of the vector from the other information may be, for example, a process of acquiring a value corresponding to the other information from a database not shown.

The replacement information acquisition unit 33 acquires replacement information that is, for example, a vector including, as an element, the conversion code acquired by the conversion code acquisition means 322. In other words, the replacement information acquisition unit 33 acquires replacement information in which, for example, at least a part of the address information in the target information is replaced with the conversion code acquired by the conversion code acquisition means 322.

The replacement information processing unit 34 performs a predetermined process on one or more pieces of replacement information acquired by the replacement information acquisition unit 33, and acquires a processing result.

The replacement information processing unit 34, for example, calculates the distance between the two vectors acquired by the replacement information acquisition unit 33, and acquires the processing result which is the distance. Note that the two vectors here are two pieces of replacement information.

The replacement information processing unit 34 performs a machine learning learning process on two or more pieces of teacher data, which are two or more pieces of replacement information acquired by the replacement information acquisition unit 33, and acquires a processing result, which is a learning model.

The replacement information processing unit 34 performs machine learning prediction processing using, for example, the learning model and replacement information accumulated by the address information processing device A, and obtains the processing result.

Specific examples of the predetermined processing will be described in detail below. The predetermined processing is not limited to a specific type, but as described above, the predetermined processing is, for example, a processing for calculating the distance between replacement information that is two vectors, a machine learning learning processing, and a machine learning prediction processing. (1) When the replacement information processing unit 34 calculates the distance

The replacement information processing unit 34 calculates the distance between the replacement information, which is two vectors, and obtains the processing result, which is the distance. The distance between the two vectors is, for example, the Euclidean distance, but may also be a distance based on cosine similarity, Mahalanobis distance, etc. The technology for calculating the distance between two vectors is a known technology, so a detailed description will be omitted.

Furthermore, when two pieces of replacement information are vectors consisting only of elements obtained from address information, the distance between the two pieces of replacement information is referred to as a physical distance. On the other hand, when two pieces of replacement information are vectors having elements obtained from address information and elements obtained from other information, the distance between the two pieces of replacement information is referred to as a semantic distance. Note that the semantic distance may also be referred to as a semantic similarity distance. When two pieces of replacement information also have other information, the distance between the two pieces of replacement information includes not only the physical distance between the regions but also the distance caused by the other information, and is therefore referred to as a semantic distance.
(2) When the replacement information processing unit 34 performs machine learning learning processing

The substitution information processing unit 34 performs a machine learning learning process on two or more pieces of teacher data, which are two or more vectors acquired by the substitution information acquisition unit 33, and acquires a processing result that is a learning model. Note that the teacher data is substitution information. Note that the learning model may also be called a learner, a classifier, or the like.

In addition, any machine learning algorithm can be used. For example, deep learning, decision tree, random forest, SVM, SVR, etc. can be used as an algorithm for performing the machine learning learning process, but the type of machine learning is not limited.

The substitution information processing unit 34 obtains a learning model by, for example, providing two or more pieces of substitution information to a module that performs a machine learning learning process and executing the module.

The replacement information processing unit 34, for example, provides replacement information that is one or more positive examples and replacement information that is one or more negative examples to a module that performs machine learning learning processing, and acquires a learning model by executing the module.

The learning processing modules are, for example, various functions in the TensorFlow library, fastText, tinySVM, and the random forest module of the R language.
(3) When the replacement information processing unit 34 performs machine learning prediction processing

The replacement information processing unit 34 performs machine learning prediction processing and acquires a processing result by using the learning model acquired by the learning processing in (2) and the vector acquired by the replacement information acquisition unit 33. Note that the vector acquired by the replacement information acquisition unit 33 is a vector that includes, as an element, the area code acquired by the area code acquisition unit 32 from the address information.

As with the learning process, any machine learning algorithm may be used to perform the prediction process. As with the learning process, any machine learning algorithm may be used to perform the prediction process, such as deep learning, decision tree, random forest, SVM, SVR, etc., but any type of machine learning may be used. Examples of such modules include various functions in the TensorFlow library, fastText, tinySVM, and the random forest module of the R language.

The output unit 4 outputs various information. Here, the various information is, for example, a processing result. The processing result is, for example, the distance, the learning model, and the prediction result in the prediction process.

Here, output is a concept that includes display on a display, projection using a projector, printing on a printer, sound output, transmission to an external device, storage on a recording medium, and handing over the processing results to other processing devices or other programs, etc.

The processing result output unit 41 outputs the processing result. For example, the processing result output unit 41 accumulates the learning model in the storage unit 1 or another device not shown. For example, the processing result output unit 41 displays a distance. For example, the processing result output unit 41 displays a prediction result.

The storage unit 1, the area code storage unit 11, and the conversion code storage unit 12 are preferably non-volatile recording media, but may also be realized as volatile recording media.

There is no restriction on the process by which information is stored in the storage unit 1, etc. For example, information may be stored in the storage unit 1, etc. via a recording medium, information transmitted via a communication line, etc. may be stored in the storage unit 1, etc., or information inputted via an input device may be stored in the storage unit 1, etc.

The reception unit 2 can be realized, for example, by a device driver for an input means such as a touch panel or a keyboard, or by control software for a menu screen.

The processing unit 3, the target information acquisition unit 31, the area code acquisition unit 32, the replacement information acquisition unit 33, the replacement information processing unit 34, the area code acquisition means 321, and the conversion code acquisition means 322 can usually be realized by a processor, a memory, etc. The processing procedure of the processing unit 3, etc. is usually realized by software, and the software is recorded in a recording medium such as a ROM. However, it may also be realized by hardware (dedicated circuit). The processor is, for example, a CPU, an MPU, a GPU, etc., and the type is not important.

The output unit 4 and the processing result output unit 41 may be considered to include or not include an output device such as a display, a speaker, etc. The output unit 4 etc. may be realized by driver software for an output device, or a combination of driver software for an output device and an output device, etc.

Next, an example of the operation of the address information processing device A will be described with reference to the flowchart of FIG.

(Step S201) The reception unit 2 judges whether or not a start instruction has been received. If a start instruction has been received, the process proceeds to step S202, and if a start instruction has not been received, the process returns to step S201.

(Step S202) The target information acquisition unit 31 acquires one or more pieces of target information. Note that the target information may be included in the start instruction, may be present in the storage unit 1, or may be received from an external device (not shown), etc.

(Step S203) The area code obtaining unit 32 assigns 1 to a counter i.

(Step S204) The area code acquisition unit 32 judges whether or not the i-th target information exists among the target information acquired in step S202. If the i-th target information exists, the process proceeds to step S205, and if the i-th target information does not exist, the process proceeds to step S209.

(Step S205) The area code obtaining unit 32 obtains address information from the i-th target information.

(Step S206) The area code acquiring unit 32 acquires, from the area code storage unit 11, an area code paired with all or part of the address information acquired in step S205.

(Step S207) The replacement information acquisition unit 33 acquires replacement information using the area code acquired in step S206, and temporarily stores the information in a buffer (not shown). An example of the replacement information acquisition process will be described with reference to the flowchart of FIG.

(Step S208) The area code obtaining unit 32 increments the counter i by 1. The process returns to step S204.

(Step S209) The replacement information processing unit 34 performs a predetermined process on the one or more pieces of replacement information temporarily stored in step S207, and obtains a process result.

(Step S210) The processing result output unit 41 outputs the processing result acquired in step S209. Then, the process returns to step S201.

In the flowchart of FIG. 2, the specified processing in step S209 is, for example, the distance calculation processing, the machine learning learning processing, and the machine learning prediction processing described above, but other processing may also be used as long as it is processing using one or more pieces of replacement information.

2, after step S206, the conversion code acquisition means 322 judges whether the area code acquired in step S206 satisfies the conversion condition, and if the conversion condition is satisfied, the conversion code acquisition means 322 preferably acquires a conversion code corresponding to the area code from the conversion code storage unit 12. In this case, it is preferable that the replacement information acquisition unit 33 acquires replacement information using the acquired conversion code instead of the area code.

Furthermore, in the flowchart of FIG. 2, the process ends when the power is turned off or an interrupt for ending the process occurs.

Next, an example of the replacement information acquisition process in step S207 will be described with reference to the flowchart in FIG.

(Step S301) The replacement information acquisition unit 33 judges whether or not one or more pieces of extra-code information are present in the i-th address information acquired in step S205. If one or more pieces of extra-code information are present, the process proceeds to step S302. If one or more pieces of extra-code information are not present, the process proceeds to step S307.

The non-code information is information in the address information, and does not correspond to the regional code acquired by the regional code acquisition unit 32. For example, when the address information is "721-1 Higashi Shiokojicho, Shimogyo Ward, Kyoto City, Kyoto Prefecture" and the regional code (here, the postal code) acquired by the regional code acquisition unit 32 is "600-8216", the non-code information is "721-1" obtained by removing "Higashi Shiokojicho, Shimogyo Ward, Kyoto City, Kyoto Prefecture" which corresponds to the postal code "600-8216" from the address information.

(Step S302) The replacement information acquiring unit 33 assigns 1 to a counter i.

(Step S303) The replacement information acquisition unit 33 judges whether or not the i-th extra-code information is present among the extra-code information acquired in step S301. If the i-th extra-code information is present, the process proceeds to step S304, and if the i-th extra-code information is not present, the process proceeds to step S307.

(Step S304) The replacement information acquisition unit 33 judges whether or not to acquire an element constituting a vector from the i-th code extra information. If an element is to be acquired, the process proceeds to step S305, and if an element is not to be acquired, the process proceeds to step S306. It is assumed that whether or not the code extra information becomes source information constituting an element of a vector is determined by the code extra information. For example, it is determined that the "banchi" and "banchi" of the address information become elements of a vector, while the "building name (e.g., apartment name)" and the "room number" do not become elements of a vector. Also, for example, it is determined that the numeric values "721" and "1" of the code extra information "721-1" become elements of a vector, while the character "-" does not become an element of a vector. However, it is preferable that which information of the address information becomes an element of a vector differs depending on the process performed by the replacement information processing unit 34.

(Step S305) The replacement information acquisition unit 33 acquires elements constituting a vector from the i-th extra-code information. Note that the replacement information acquisition unit 33 acquires, for example, "721" or "1" from "721-1" including two extra-code information.

(Step S306) The replacement information acquisition unit 33 increments the counter i by 1. The process returns to step S303. Note that, by the loop from step S303 to step S306, for example, the elements "721" and "1" of the vector are acquired.

(Step S307) The replacement information acquisition unit 33 judges whether or not one or more pieces of other information are present in the i-th target information in step S204. If one or more pieces of other information are present, the process proceeds to step S308, and if one or more pieces of other information are not present, the process proceeds to step S313.

(Step S308) The replacement information acquiring unit 33 assigns 1 to a counter j.

(Step S309) The replacement information acquisition unit 33 judges whether or not the jth other information exists among the one or more other information in the i-th target information. If the jth other information exists, the process proceeds to step S310, and if the jth other information does not exist, the process proceeds to step S313.

(Step S310) The replacement information acquisition unit 33 uses the j-th other information to determine whether or not to acquire an element of the vector. If an element of the vector is to be acquired, the process proceeds to step S311. If an element of the vector is not to be acquired, the process proceeds to step S312. It is assumed that it is predetermined whether or not the other information is to become the original information of the element of the vector.

(Step S311) The replacement information acquisition unit 33 acquires an element of a vector using the j-th other information. The replacement information acquisition unit 33 may directly use the j-th other information as an element of a vector, or may acquire a value paired with the j-th other information from a table (not shown). There is no restriction on the method of acquiring an element of a vector using the j-th other information.

(Step S312) The replacement information obtaining unit 33 increments the counter j by 1. The process returns to step S309.

(Step S313) The replacement information acquisition unit 33 constructs a vector having the area code acquired in step S206 as one or more elements. This vector may have an element acquired in step S305. This vector may also have an element acquired in step S311. If a conversion code has been acquired, the replacement information acquisition unit 33 constructs a vector having the conversion code as one or more elements instead of the area code acquired in step S206.

A specific example of the operation of the address information processing device A in this embodiment will be described below. Four specific examples will be described below. In these specific examples, it is assumed that the area code is a postal code.

(Specific Example 1)
Assume now that the postal code management table shown in Fig. 4 is stored in the area code storage unit 11. The postal code management table here is a table that manages the postal codes for each area in Japan. The postal code management table manages a large number of records, each of which has an "ID,""addressinformation," and a "postal code." The "ID" is information that identifies the record.

It is also assumed that the conversion code storage unit 12 stores a conversion code management table shown in Fig. 5. The conversion code management table here manages a large number of records each having an "ID", "zip code", and "conversion code". The "ID" is information that identifies the record. It is also assumed that the "zip code" managed here is the zip code of an exclave.

In the specific example 1, a processing example of the address information processing device A will be described with reference to Fig. 6. In the specific example 1, the address information processing device A performs a process of calculating the distance between two points from two pieces of address information. The distance here is calculated using only the address information, and therefore can be said to be the above-mentioned physical distance.

In the first specific example, three methods will be described: a conventional method using only one-hot encoding, method 1 (Proc1) of the address information processing device A, and method 2 (Proc2).

And the two address information given are Address 1 "721-1 Higashishiokojicho, Shimogyo Ward, Kyoto City, Kyoto Prefecture" and Address 2 "1-10 Nakadoji Myobucho, Shimogyo Ward, Kyoto City, Kyoto Prefecture".
[Conventional method]

The conventional method is a method of vectorizing all information constituting address information by one-hot encoding. That is, in the conventional method, it is assumed that a vector [1, 1, 0, ..., 0] is obtained by one-hot encoding from the character string "Kyoto City, Kyoto Prefecture" in address 1. It is also assumed that a vector [0, 0, 1, ..., 1, ..., 0, 0] is obtained by one-hot encoding from the character string "Higashi Shiokojicho, Shimogyo Ward" in address 1. It is also assumed that a vector [0, 1, ..., 1, ..., 1] is obtained from "721-1" in address 1. It is then assumed that the obtained vectors are combined to obtain replacement information 1 [1, 1, 0, ..., 0, 0, 0, 1, ..., 1, ..., 0, 0, 1, ..., 1, ...] (see 601). Note that, here,
Although 610 is shown as two vectors, a single combined vector is used when calculating the distance, which will be described later.

In addition, in the conventional method, it is assumed that a vector [1, 1, 0, ..., 0] is obtained by One-Hot Encoding from the character string "Kyoto City, Kyoto Prefecture" in Address 2. Also, a vector [0, 0, 1, ..., 1, ..., 1, 0] is obtained by One-Hot Encoding from the character string "Nakadoji Myobu-cho, Shimogyo Ward" in Address 2. Furthermore, it is assumed that a vector [0, 1, ....1, ....1, ....0] is obtained from "1-10" in Address 2. Then, it is assumed that the obtained vectors are combined to obtain replacement information 2 [1, 1, 0, ..., 0, 0, 0, 1, ..., 1, ..., 0, 1, ....1, ....1 ....0].

Next, in the conventional method, the Euclidean distance (X ₁ ) between replacement information 1 and replacement information 2, which are vectors, is calculated, and the distance (X ₁ ) is output.
[Proc 1]

Proc1 is a method for constructing a vector consisting of an element obtained from a postal code corresponding to address information, a large number of elements obtained by one-hot encoding from the character string of the ward, city, town, or village included in the address information, and elements obtained from the house number information (lot number) and residential address indication included in the address information.

Assume that the reception unit 2 of the address information processing device A, which performs processing using Proc1, receives two pieces of address information: Address 1 "721-1 Higashishiokojicho, Shimogyo-ku, Kyoto City, Kyoto Prefecture" and Address 2 "1-10 Nakadoji Myobu-cho, Shimogyo-ku, Kyoto City, Kyoto Prefecture."

Next, the target information acquisition unit 31 acquires the address 1 and the address 2 accepted by the acceptance unit 2 .

Next, the area code obtaining unit 32 obtains the postal code "600-8216" that pairs with "Higashi Shiokojicho, Shimogyo Ward, Kyoto City, Kyoto Prefecture" of Address 1 from the postal code management table of FIG.

Next, the replacement information acquisition unit 33 acquires a vector [600, 8216] from the postal code "600-8216". It is also assumed that the replacement information acquisition unit 33 acquires a vector [0, 0, 1, . . ., 1, . . ., 0, 0] for "Higashi Shiokojicho, Shimogyo Ward" of address 1 by one-hot encoding. The replacement information acquisition unit 33 also acquires a vector [0, 1, . . ., 1, . . .] from "721-1" of address 1.

Next, the replacement information acquisition unit 33 combines the acquired vectors to acquire replacement information 1 [600, 8216, 0, 0, 1, . . ., 0, 0, 0, 1, . . ., 1, . . . 1] (6
See 02).

Similarly, the area code acquisition unit 32 acquires the postal code "600-8533" paired with "Nakadoji Myobu-cho, Shimogyo-ku, Kyoto City, Kyoto Prefecture" of Address 2 from the postal code management table of FIG.

Next, the replacement information acquisition unit 33 acquires a vector [600, 8533] from the postal code "600-8533". It is also assumed that the replacement information acquisition unit 33 acquires a vector [0, 0, 1, . . ., 1, . . ., 1, 0] for "Nakadoji Myobu-cho, Shimogyo-ku" of address 2 by one-hot encoding. The replacement information acquisition unit 33 also acquires a vector [0, 1, . . ., 1, . . ., 0] from "1-10" of address 1.

Next, the replacement information acquisition unit 33 combines the acquired vectors to acquire replacement information 2 [600, 8533, 0, 0, 1, ..., 1, 0, 1, ... 1, ... 1, ... 0].

Next, the replacement information processing unit 34 calculates the Euclidean distance (X ₂ ) between the replacement information 1, which is a vector, and the replacement information 2. Next, the processing result output unit 41 outputs the distance (X ₂ ).
[Proc 2]

Proc2 is a method for constructing a vector that includes elements obtained from a postal code corresponding to address information and elements obtained from address information (information including house number information) excluding the character string (prefecture, ward, city, town, or village) corresponding to the postal code.

Assume that the reception unit 2 of the address information processing device A, which performs processing using Proc2, receives two pieces of address information: Address 1 "721-1 Higashishiokojicho, Shimogyo-ku, Kyoto City, Kyoto Prefecture" and Address 2 "1-10 Nakadoji Myobu-cho, Shimogyo-ku, Kyoto City, Kyoto Prefecture."

Next, the target information acquisition unit 31 acquires the address 1 and the address 2 accepted by the acceptance unit 2 .

Next, the area code obtaining unit 32 obtains the postal code "600-8216" that pairs with "Higashi Shiokojicho, Shimogyo Ward, Kyoto City, Kyoto Prefecture" of Address 1 from the postal code management table of FIG.

Next, the replacement information acquisition unit 33 acquires a vector [600, 8216] from the postal code "600-8216". The replacement information acquisition unit 33 also acquires a vector [721, 1] from "721-1" of the address 1.

Next, the replacement information acquisition unit 33 combines the acquired vectors to acquire replacement information 1 [600, 8216, 721, 1] (see 603).

Similarly, the area code acquisition unit 32 acquires the postal code "600-8533" paired with "Nakadoji Myobu-cho, Shimogyo-ku, Kyoto City, Kyoto Prefecture" of Address 2 from the postal code management table of FIG.

Next, the replacement information acquisition unit 33 acquires a vector [600, 8533] from the postal code "600-8533". The replacement information acquisition unit 33 also acquires a vector [1, 10] from "1-10" of the address 1.

Next, the replacement information acquisition unit 33 combines the acquired vectors to acquire replacement information 2 [600, 8533, 1, 10].

Next, the replacement information processing unit 34 calculates the Euclidean distance (X _{3 ) between the replacement information 1 [600, 8216, 721, 1], which are vectors, and the replacement information 2 [600, 8533, 1, 10]. Next, the processing result output unit 41 outputs the distance (X 3 )} .

Next, the processing result output unit 41 outputs the evaluation value acquired by the replacement information processing unit 34 .

By the above processing, in Proc1 and Proc2, part of the address information is replaced with an area code, and in particular, in Proc2, a vector with a very small number of dimensions is constructed and the vector is used to calculate the distance between two points, enabling high-speed distance calculation processing.

(Specific Example 2)
In specific example 2, a process for calculating the distance between two pieces of address information in the United States will be described with reference to Fig. 7. In specific example 2, similar to specific example 1, three methods will be described: a conventional method using only One-Hot Encoding, method 1 (Proc1) of the address information processing device A, and method 2 (Proc2).
[Conventional method]

If Address 1 is given as “29 E 2nd Ave, New York” (701 in FIG. 7 ), when it is vectorized by One-Hot Encoding in the same manner as the processing of the conventional method in concrete example 1, the result is [0,...,1,...,1,1,1,0,0,...,0,1,1,0,...,0] (702).

Also, suppose that when the given address 2 “205 E, Houston, St, New York, NY” (703) is vectorized by One-Hot Encoding, it becomes [0,...,1,...,1,0,0,1,1,...,0,1,1,0,...,0] (704).

Furthermore, suppose that when the given address 3 “182 2nd Ave, New York, NY” (705) is vectorized by One-Hot Encoding, it becomes [0,...,1,...,0,1,1,0,0,...,0,1,1,0,...,0] (706).

Then, it is assumed that the distance (Y ₁ ) between the point of address 1 and the point of address 2 is calculated as the Euclidean distance between vector 702 and vector 704 using the conventional method, and the distance (Y ₁ ) is output.

Also, it is assumed that the distance (Y ₂ ) between the point of Address 1 and the point of Address 3 is calculated as the Euclidean distance between the vector 702 and the vector 706 using the conventional method, and the distance (Y ₂ ) is output.
[Proc 1]

It is assumed that the reception unit 2 of the address information processing device A, which performs processing by Proc1, receives the above address 1, address 2, and address 3.

Next, the target information acquisition unit 31 acquires Address 1, Address 2, and Address 3 accepted by the acceptance unit 2.

Next, the area code acquisition unit 32 acquires "E 2nd
The zip code "10,0003" that is paired with "1000 1st Avenue, New York" is obtained from the zip code management table. The zip code management table in the second specific example has the same structure as that in FIG.

Next, the replacement information acquisition unit 33 acquires a vector [0,...,1,...,0] from "29" in "#address" of address 1. Furthermore, the replacement information acquisition unit 33 acquires a vector [1,1,1,0,0,...,1] from "E 2nd Ave, New York" in address 1 by One-Hot Encoding.
, 0] is acquired. The replacement information acquisition unit 33 also acquires a vector [10, 003] from the zip code “10, 003” acquired by the area code acquisition unit 32.

Next, the replacement information acquisition unit 33 combines the acquired vectors to acquire replacement information 1 [0, . . . , 1, . . . , 0, 1, 1, 1, 0, 0, . . . , 0, 10, 003] (see 707).

Similarly, the area code acquisition unit 32 acquires the zip code "10-002" that pairs with "E, Houston, St, New York, NY" of Address 2 from the postal code management table.

Next, the replacement information acquisition unit 33 acquires a vector [0,...,1,...,0] from "205" of "#address" of address 2. Also, it is assumed that the replacement information acquisition unit 33 acquires a vector [1,0,0,1,1,...,0] by One-Hot Encoding for "E, Houston, St, New York, NY" of address 2. Also, the replacement information acquisition unit 33 acquires a vector [10,002] from zip code "10,002".

Next, the replacement information acquisition unit 33 combines the acquired vectors to acquire replacement information 2 [0,...,1,...,0,1,0,0,1,1,...,0,10,002] (see 708).

Similarly, the area code acquisition unit 32 acquires the zip code "10-003" that pairs with "2nd, Ave, New York, NY" of Address 3 from the postal code management table.

Next, the replacement information acquisition unit 33 acquires a vector [0,...,1,...,0] from "182" of "#address" of address 3. Also, it is assumed that the replacement information acquisition unit 33 acquires a vector [0,1,1,0,0,...,0] by One-Hot Encoding for "2nd, Ave, New York, NY" of address 3. Also, the replacement information acquisition unit 33 acquires a vector [10,003] from zip code "10,003".

Next, the replacement information acquisition unit 33 combines the acquired vectors to acquire replacement information 3 [0, . . . , 1, . . . , 0, 1, 1, 0, 0, . . . , 0, 10, 003] (see 709).

Next, the replacement information processing unit 34 calculates the Euclidean distance (Y ₃ ) between the replacement information 1, which is a vector, and the replacement information 2. Next, the processing result output unit 41 outputs this distance (Y ₃ ).

Furthermore, the replacement information processing unit 34 calculates the Euclidean distance (Y ₄ ) between the replacement information 1, which is a vector, and the replacement information 3. Next, the processing result output unit 41 outputs this distance (Y ₄ ).
[Proc 2]

It is assumed that the reception unit 2 of the address information processing device A, which performs processing by Proc2, receives the above address 1, address 2, and address 3.

Next, the target information acquisition unit 31 acquires Address 1, Address 2, and Address 3 accepted by the acceptance unit 2.

Next, the area code acquisition unit 32 acquires "E 2nd
The zip code "10,0003" that corresponds to "1000 1 Street Ave, New York" is obtained from the postal code management table.

Next, the replacement information acquisition unit 33 acquires a vector [029] from the "29" in "#address" of address 1. The replacement information acquisition unit 33 also acquires a vector [2] from the street number "2nd". The replacement information acquisition unit 33 also acquires a vector [10,003] from the zip code "10,003".

Next, the replacement information acquisition unit 33 combines the acquired vectors to acquire replacement information 1 [029, 2, 10, 003] (see 710).

Similarly, the area code acquisition unit 32 acquires the zip code "10-002" that pairs with "E, Houston, St, New York, NY" of Address 2 from the postal code management table.

Next, the replacement information acquisition unit 33 acquires a vector [205] from "205" in "#address" of address 2. Furthermore, since there is no street number for address 2, the replacement information acquisition unit 33 acquires a vector [0] corresponding to non-existence. Furthermore, the replacement information acquisition unit 33 acquires a vector [10,002] from the zip code "10,002".

Next, the replacement information acquisition unit 33 combines the acquired vectors to acquire replacement information 2 [205, 0, 10, 002] (see 711).

Similarly, the area code acquisition unit 32 acquires the zip code "10-003" that pairs with "2nd, Ave, New York, NY" of Address 3 from the postal code management table.

Next, the replacement information acquisition unit 33 acquires a vector [182] from "182" in "#address" of address 3. The replacement information acquisition unit 33 also acquires a vector [2] from the street number "2nd" of address 3. The replacement information acquisition unit 33 also acquires a vector [10,003] from the zip code "10,003".

Next, the replacement information acquisition unit 33 combines the acquired vectors to acquire replacement information 3 [182, 2, 10, 003] (see 712).

Next, the replacement information processing unit 34 calculates the Euclidean distance (Y ₅ ) between the replacement information 1, which is a vector, and the replacement information 2. Next, the processing result output unit 41 outputs this distance (Y ₅ ).

Furthermore, the replacement information processing unit 34 calculates the Euclidean distance (Y ₆ ) between the replacement information 1, which is a vector, and the replacement information 3. Next, the processing result output unit 41 outputs this distance (Y ₆ ).

By the above processing, in Proc1 and Proc2, part of the address information is replaced with an area code, and in particular, in Proc2, a vector with a very small number of dimensions is constructed and the vector is used to calculate the distance between two points, enabling high-speed distance calculation processing.
[Experimental Results]

The experimental results of the second specific example are described below. In this experiment, 6,720 pieces of POI data from six cities in the United States (New York, Washington, Boston, San Francisco, Los Angeles, and Seattle) were used. The data for each city in the United States is shown in FIG. 8. In FIG. 8, the state code, city code range, and
, and the number of POI data (#POIs).

Under these conditions, the following three experiments will be described.
[Experiment 1]

In experiment 1, the accuracy of distance calculation in each of the conventional method, the method of Proc1, and the method of Proc2 described above was measured.

To evaluate the accuracy of the distance, first, the actual distance between each POI was calculated using the latitude and longitude information of each POI. Then, the error between the actual distance and the conventional method (Euclidean distance), the error between the actual distance and the Proc1 method (Euclidean distance), and the error between the actual distance and Proc2 (Euclidean distance) were calculated, and each error was subjected to Min-Max normalization and compared using the MSE value.

In addition, to evaluate the accuracy of the distance, secondly, we calculated the distance between each POI using the cosine similarity of the conventional method and the proposed method, and further performed Min-Max normalization on each, and compared the MSE values.

The results of Experiment 1 are shown in Fig. 9. (a) in Fig. 9 is the conventional method, (b) is Proc1, and (c) is Proc2. The values for each city and each method in Fig. 9 are the average values of the Euclidean distance from the actual distance and the distance based on the cosine similarity for each method.

As shown in FIG. 9, Proc2(c), which vectorizes address information without using one-hot encoding and has the smallest number of dimensions, had an average error of 0.1131 for the distances between all cities, showing the highest accuracy.
[Experiment 2]

In experiment 2, the processing speed of distance calculation for each of the conventional method, the Proc1 method, and the Proc2 method was measured.

Specifically, for each of the six U.S. cities mentioned above, we measured the processing speed of each method when calculating Euclidean distance while increasing the number of data (number of POIs) by 100. In this experiment, we measured the time required to calculate the distance by Min-Max normalization using the python time library.

The results of such measurements are shown in Figure 10. Graph 6 ((a) to (f)) in Figure 10 is a graph showing the processing time for each number of POIs in each city, with the horizontal axis representing the number of POIs and the vertical axis representing the processing time.

10, there is no significant difference between the graphs of the results of the conventional method (1001, 1004, 1007, 1010, 1013, 1016) and the graphs of the results of Proc1 (1002, 1005, 1008, 1011, 1014, 1017), and it is understood that the processing speeds are roughly the same. On the other hand, in FIG. 10, the graphs of the results of Proc2 (1003, 1006, 1009, 1012, 1015, 1018) show that the method of Proc2 can perform processing at a much higher speed than the other two methods.

This shows that the load of vectorizing text data such as addresses and street names using One-Hot Encoding, constructing vectors with a large number of dimensions, and processing them is heavy.
[Experiment 3]

In experiment 3, we measured the memory usage efficiency of the conventional method, the Proc1 method, and the Proc2 method. In this experiment, we used a python module called memory profiler to analyze the memory consumption of the program line by line, and measured the memory usage efficiency of the distance calculation of each method.

The results of such measurements are shown in Fig. 11. The vertical axis of the graph in Fig. 11 is memory usage (MiB), and the horizontal axis is time (seconds). In Fig. 11, 1101 is a graph of the conventional method, 1102 is a graph of Proc1, and 1103 is a graph of Proc2.

In FIG. 11, it can be seen that in the conventional method, Proc1 continued to use about 23,000 MiB of memory for about 230 seconds (see 1101 and 1102 in FIG. 11), whereas Proc2 completed processing in about 25 seconds, which is about 1/10 of the time, with the same memory usage.

From the above results, it was confirmed that Proc2 showed the best results, with memory usage efficiency about 90% higher than the conventional method and Proc1.

The above three experiments demonstrated the merits of the Proc2 technique in particular.

(Specific Example 3)
Specific example 3 is a case where the replacement information processing unit 34 performs a machine learning learning process. In specific example 3, a process of forming a vector using address information by the method of Proc2 that does not use One-Hot Encoding is adopted.

Currently, the restaurant management table of FIG. 12 is stored in storage unit 1. The restaurant management table is a table that manages values corresponding to the type of restaurant. The restaurant management table manages two or more records each having an "ID," "type," and "value." Note that "type" is the type of restaurant, and "value" is a numerical value that can be replaced with the type of restaurant.

It is also assumed that the area code storage unit 11 stores a postal code management table shown in Fig. 4. It is also assumed that the conversion code storage unit 12 stores a conversion code management table shown in Fig. 5.

Furthermore, it is assumed that the storage unit 1 stores a user-visited restaurant management table shown in FIG. 13. The user-visited restaurant management table is a table that manages information on restaurants that the user has visited so far. The user may be a specific user or may be multiple users. The user-visited restaurant management table has many records having "ID", "address information", "store attribute", and "rating value". The "store attribute" here has "type", "average unit price", "number of menu items", and "size". "Type" is the type of restaurant. "Average unit price" is the average unit price per customer. "Number of menu items" is the number of dishes (including courses) that can be served at the restaurant. It is assumed that the records in the user-visited restaurant management table are target information that is the target of machine learning. In addition, the circumstances and method of storing each record in the user-visited restaurant management table are not important. Each record in the user-visited restaurant management table is, for example, information that a user makes a reservation, actually visits, and inputs a rating value after visiting.

Furthermore, a scale value management table for managing values corresponding to each piece of "scale" information is stored in storage unit 1. "Scale" is the size of a restaurant, and in the scale value management table, the information "large" is associated with a value of "3," the information "medium" with a value of "2," and the information "small" with a value of "1."

In this situation, it is assumed that the user of the address information processing device A inputs, to the address information processing device A, an instruction to start the machine learning learning process.

Then, the reception unit 2 of the address information processing device A receives the start instruction.

Next, the target information acquisition unit 31 acquires the "address information", "store attributes", and "evaluation value" of all the records in Fig. 13. Note that information having the "address information", "store attributes", and "evaluation value" is corresponding information.

Next, the regional code acquisition unit 32 acquires address information from each piece of target information acquired by the target information acquisition unit 31. Then, the regional code acquisition unit 32 acquires a postal code paired with a part of each piece of acquired address information (here, a prefecture, ward, city, town, or village, which is a part of the address information corresponding to a postal code) from the postal code management table of FIG.

Next, the replacement information acquisition unit 33 acquires two elements of a vector (eg, [600, 8216]) from the acquired postal code (eg, 600-8216) for each piece of target information.

Next, the replacement information acquisition unit 33 judges whether or not one or more pieces of extra-code information (e.g., a street address, a building number) are present in the acquired address information for each piece of target information. If extra-code information is present (e.g., if the address is "10-1"), the replacement information acquisition unit 33 uses the extra-code information to acquire one or more elements of a vector (here, [10, 1]).

Next, the replacement information acquisition unit 33 obtains a "value" that is paired with the "type" information of each correspondence information, as follows:
The value is obtained from the restaurant management table in FIG. 12, and the "value" is set as one element of the vector.

Next, the replacement information acquiring unit 33 acquires the "average unit price" of each piece of correspondence information as one element of a vector corresponding to each piece of correspondence information.

Next, the replacement information acquiring unit 33 acquires the "number of menu items" included in each piece of correspondence information as one element of a vector corresponding to each piece of correspondence information.

Furthermore, the replacement information acquisition unit 33 acquires the "value" that pairs with the "scale" of each piece of correspondence information from the storage unit 1, and sets the "value" as one element of the vector.

Next, the replacement information acquisition unit 33 sets a value corresponding to the store attribute information of each piece of correspondence information as one element of a vector.

Furthermore, the replacement information acquisition unit 33 sets the "evaluation value" of each piece of correspondence information as one element of a vector.

As a result, the replacement information acquisition unit 33 constructs a vector having a structure of (two elements acquired from the postal code, one or more elements acquired from the extra-code information of the address information, a "value" paired with the "type" information, average unit price, number of menu items, a "value" paired with the "scale", ..., evaluation value) for each piece of target information. Each vector is training data with (two elements acquired from the postal code, a "value" paired with the "type" information, average unit price, number of menu items, a "value" paired with the "scale", ...) as explanatory variables and the "evaluation value" as the objective variable. Here, the training data is composed of the same number of records as in FIG. 13.

Next, the replacement information processing unit 34 performs a machine learning learning process on all the teacher data acquired by the replacement information acquisition unit 33 to acquire a learning model. Note that the learning model is an example of a processing result acquired by the replacement information processing unit 34.

Next, the processing result output unit 41 accumulates the acquired learning model in, for example, the storage unit 1 .

Through the above process, a learning model for estimating evaluation values using store address information and store attribute information as input was constructed. Furthermore, through the above process, a part of the address information is replaced with an area code to construct a vector with a very small number of dimensions, and this vector is used for machine learning learning processing, enabling high-speed learning processing.

(Specific Example 4)
Specific example 4 is a case where the replacement information processing unit 34 performs machine learning prediction processing. In specific example 4, a process of forming a vector using address information by the method of Proc2 that does not use One-Hot Encoding is adopted.

It is now assumed that the learning model constructed by the processing of the specific example 3 is stored in the storage unit 1 of the address information processing device A.

In such a situation, it is assumed that the user inputs the "address information," "type," "average unit price," "number of menu items," "size," and the like, of the store that he or she wishes to use.

The receiving unit 2 of the address information processing device A then receives the target information consisting of the input "address information," "type," "average unit price," "number of menu items," "size," and the like.

Next, the target information acquisition unit 31 acquires the target information.

Next, the area code acquisition unit 32 acquires the "address information" contained in the target information. Next, the area code acquisition unit 32 acquires a postal code paired with a part of the acquired address information from the postal code management table of FIG.

Next, the replacement information acquisition unit 33 acquires two elements of a vector (for example, [603, 8555]) from the postal code acquired by the area code acquisition unit 32 (for example, 603-8555).

Next, the replacement information acquisition unit 33 judges whether or not one or more pieces of extra-code information (e.g., a street address, a building number) are present in the acquired address information for each piece of target information. If extra-code information is present (e.g., if the address is "3, 8"), the replacement information acquisition unit 33 uses the extra-code information to acquire one or more elements of the vector (here, [3, 8]).

Next, the replacement information acquisition unit 33 acquires a "value" that pairs with the "type" information of the acquired correspondence information from the restaurant management table of FIG. 12, and sets the "value" as one element of a vector.

Next, the replacement information acquisition unit 33 acquires the "average unit price" of the corresponding information as one element of a vector corresponding to each corresponding information.

Next, the replacement information acquiring unit 33 acquires the "number of menu items" contained in the corresponding information as one element of a vector corresponding to each corresponding information.

Furthermore, the replacement information acquisition unit 33 acquires, from the storage unit 1, a "value" that pairs with the "scale" of the corresponding information, and sets the "value" as one element of the vector.

Furthermore, the replacement information acquisition unit 33 sets a value corresponding to the store attribute information of the one correspondence information as one element of the vector.

As a result, the replacement information acquisition unit 33 constructs a vector having a structure of (two elements acquired from the postal code, one or more elements acquired from the extra-code information of the address information, a "value" paired with the "type" information, average unit price, number of menu items, a "value" paired with the "scale", ...). Here, the replacement information acquisition unit 33 constructs, for example, a vector [603, 8555, 3, 8, 1, 5000, 10, 3, ...].

Next, the replacement information processing unit 34 acquires the learning model from the storage unit 1. Then, the replacement information processing unit 34 provides a vector having the structure acquired by the replacement information acquisition unit 33 (two elements acquired from the postal code, a "value" paired with the "type" information, an average unit price, the number of menu items, a "value" paired with the "scale", ...) and the learning model to a module that performs machine learning prediction processing, executes the module, and acquires a predicted evaluation value.

Next, the processing result output unit 41 outputs the evaluation value acquired by the replacement information processing unit 34 .

The above process allows us to estimate the evaluation value of a store using the store's address information and store attribute information. In addition, the above process replaces part of the address information with a regional code to create a vector with a very small number of dimensions, and uses this vector to perform machine learning prediction processing, enabling high-speed prediction processing.

As described above, according to this embodiment, the processing cost can be reduced in processing information including address information.

Furthermore, according to this embodiment, it is possible to reduce the processing cost in the process of calculating the distance between two points using information including address information.

Furthermore, according to this embodiment, processing costs can be reduced in machine learning learning processing for information including address information.

Furthermore, according to this embodiment, processing costs can be reduced in machine learning prediction processing for information including address information.

Furthermore, according to this embodiment, even if the area indicated by the address information is unique, such as an isolated area, it is possible to process address information with high accuracy.

The address information processing device A in this embodiment is suitable for real-time processing because it can process target information including address information at high speed. The address information processing device A is also suitable for machine learning processing of target information including address information, and is useful for federated learning (see Internet homepage URL: "https://developers-jp.googleblog.com/2017/05/federated-learning-collaborative.html") that enables smartphones to learn a shared prediction model in collaboration.

The processing in this embodiment may be realized by software. This software may be distributed by software download or the like. This software may be recorded on a recording medium such as a CD-ROM and distributed. This also applies to other embodiments in this specification. The software for realizing the address information processing device A in this embodiment is a program as follows. That is, this program includes a computer that can access a region code storage unit in which one or more pieces of correspondence information indicating the correspondence between a region code and at least a part of address information are stored, a target information acquisition unit that acquires target information having address information, a region code acquisition unit that acquires a region code corresponding to at least a part of the address information from the region code storage unit, a replacement information acquisition unit that acquires replacement information that is a vector including the region code acquired by the region code acquisition unit as an element, a replacement information processing unit that performs a predetermined process on the replacement information and acquires a processing result,
The program is for causing the processing result output unit to function as a processing result output unit that outputs the processing result.

14 shows the appearance of a computer that executes the program described in this specification to realize the address information processing device A of the various embodiments described above. The above-mentioned embodiments can be realized by computer hardware and a computer program executed thereon. FIG. 14 is an overview of this computer system 300, and FIG. 15 is a block diagram of the computer system 300.

In FIG. 14, a computer system 300 includes a computer 301 including a CD-ROM drive, a keyboard 302 , a mouse 303 , and a monitor 304 .

15, in addition to a CD-ROM drive 3012, a computer 301 includes an MPU 3013, a bus 3014 connected to the CD-ROM drive 3012 etc., a ROM 3015 for storing programs such as a boot-up program, a RAM 3016 connected to the MPU 3013 for temporarily storing instructions of application programs and providing temporary storage space, and a hard disk 3017 for storing application programs, system programs, and data. Although not shown here, the computer 301 may further include a network card for providing connection to a LAN.

A program that causes computer system 300 to execute the functions of address information processing device A of the above-described embodiment may be stored in CD-ROM 3101, inserted into CD-ROM drive 3012, and further transferred to hard disk 3017. Alternatively, the program may be transmitted to computer 301 via a network (not shown) and stored on hard disk 3017. The program is loaded into RAM 3016 when executed. The program may be loaded directly from CD-ROM 3101 or the network.

The program does not necessarily include an operating system (OS) or a third party program that causes the computer 301 to execute the functions of the address information processing device A of the above-mentioned embodiment. The program only needs to include an instruction portion that calls appropriate functions (modules) in a controlled manner to obtain a desired result. How the computer system 300 operates is well known, and a detailed description will be omitted.

The program may be executed by a single computer or a plurality of computers. That is, the program may be executed by a centralized processing or a distributed processing.

Furthermore, in each of the above embodiments, it goes without saying that two or more communication means present in one device may be realized physically by one medium.

In each of the above embodiments, each process may be realized by centralized processing in a single device, or may be realized by distributed processing in a plurality of devices.

The present invention is not limited to the above-described embodiment, and various modifications are possible, and it goes without saying that these modifications are also included within the scope of the present invention.

INDUSTRIAL APPLICABILITY As described above, the address information processing device according to the present invention has an effect of reducing the processing cost in processing information including address information, and is useful as an address information processing device, etc.

A Address information processing device 1 Storage unit 2 Acceptance unit 3 Processing unit 4 Output unit 11 Area code storage unit 12 Conversion code storage unit 31 Target information acquisition unit 32 Area code acquisition unit 33 Replacement information acquisition unit 33 Replacement information acquisition unit 34 Replacement information processing unit 41 Processing result output unit 321 Area code acquisition means 322 Conversion code acquisition means

Claims (9)

a target information acquisition unit for acquiring target information having address information;
a region code acquisition unit that acquires a region code corresponding to at least a part of the address information from a region code storage unit that stores one or more pieces of correspondence information indicating a correspondence between a region code and at least a part of the address information;
a replacement information acquisition unit that acquires replacement information that is a vector including the area code acquired by the area code acquisition unit as an element;
a replacement information processing unit that performs a predetermined process on the replacement information and acquires a processing result;
and a processing result output unit for outputting the processing result. The address information may include street address information or floor information;
The replacement information acquisition unit
2. The address information processing device according to claim 1, further comprising: acquiring, from the address information, a numerical value corresponding to the street address information or floor information contained in a character string not corresponding to the area code; and acquiring replacement information which is a vector including the numerical value as an element. The target information acquisition unit,
Two pieces of target information each including different address information are obtained;
The area code acquisition unit
Obtaining a region code corresponding to each of the two pieces of target information from the region code storage unit;
The replacement information acquisition unit
Obtaining two pieces of replacement information, which are two vectors including the respective area codes acquired by the area code acquisition unit as elements;
The replacement information processing unit includes:
3. The address information processing apparatus according to claim 1, further comprising: a processor that calculates a distance between the two vectors and obtains a processing result that is the distance. The target information includes information other than address information,
The target information acquisition unit,
Two pieces of target information each including different address information are obtained;
The area code acquisition unit
Obtaining a region code corresponding to each of the two pieces of target information from the region code storage unit;
The replacement information acquisition unit
For each of the two or more pieces of target information, the area code corresponding to at least a part of the address information contained in the target information is set as one or more elements of a vector, and one or more elements of the vector are obtained from the other information, and replacement information is obtained as a vector having the one or more elements which are the area code and the one or more elements obtained from the other information;
The replacement information processing unit includes:
3. The address information processing apparatus according to claim 1, further comprising: a processor that calculates a distance between the two vectors and obtains a processing result that is the distance. The target information includes information other than address information,
The target information acquisition unit,
Acquire two or more pieces of target information;
The replacement information acquisition unit
For each of the two or more pieces of target information, the area code corresponding to at least a part of the address information contained in the target information is set as one or more elements of a vector, and one or more elements of the vector are obtained from the other information, and replacement information is obtained as a vector having the one or more elements which are the area code and the one or more elements obtained from the other information;
The replacement information processing unit includes:
A machine learning learning process is performed on two or more pieces of teacher data, which are the two or more pieces of replacement information acquired by the replacement information acquisition unit, to acquire a processing result, which is a learning model;
The processing result output unit
3. The address information processing apparatus according to claim 1, further comprising: storing said learning model. a target information acquisition unit for acquiring target information having address information;
a region code acquisition unit that acquires a region code corresponding to at least a part of the address information from a region code storage unit that stores one or more pieces of correspondence information indicating a correspondence between a region code and at least a part of the address information;
a replacement information acquisition unit that acquires replacement information that is a vector including the area code acquired by the area code acquisition unit as an element;
a replacement information processing unit that performs machine learning prediction processing using a learning model stored in the address information processing device according to claim 5 and the replacement information, and acquires a processing result;
and a processing result output unit for outputting the processing result. The area code acquisition unit
a region code acquisition unit that acquires a region code corresponding to at least a part of the address information included in the target information from the region code storage unit;
a conversion code acquisition means for acquiring a conversion code paired with the regional code acquired by the regional code acquisition means from a conversion code storage unit in which one or more pieces of code correspondence information for managing the correspondence between the regional code and the conversion code are stored, when the regional code acquired by the regional code acquisition means satisfies a conversion condition;
The replacement information acquisition unit
The address information processing device according to claim 1 , wherein the conversion code acquisition means acquires replacement information that is a vector including the acquired conversion code as an element. An address information processing method implemented by a regional code storage unit storing one or more pieces of correspondence information indicating a correspondence between a regional code and at least a part of address information, a target information acquisition unit, a regional code acquisition unit, a replacement information acquisition unit, a replacement information processing unit, and a processing result output unit,
a target information acquisition step in which the target information acquisition unit acquires target information having address information; and a regional code acquisition step in which the regional code acquisition unit acquires a regional code corresponding to at least a part of the address information from the regional code storage unit.
a replacement information acquisition step in which the replacement information acquisition unit acquires replacement information that is a vector including the area code acquired in the area code acquisition step as an element;
a replacement information processing step in which the replacement information processing unit performs a predetermined process on the replacement information and acquires a processing result;
The address information processing method further comprises a processing result output step in which the processing result output unit outputs the processing result. A computer that can access an area code storage unit that stores one or more pieces of correspondence information indicating a correspondence between an area code and at least a part of address information,
a target information acquisition unit for acquiring target information having address information;
a region code acquisition unit that acquires a region code corresponding to at least a part of the address information from the region code storage unit;
a replacement information acquisition unit that acquires replacement information that is a vector including the area code acquired by the area code acquisition unit as an element;
a replacement information processing unit that performs a predetermined process on the replacement information and acquires a processing result;
A program for causing the processing result output unit to function as a processing result output unit that outputs the processing result.

Images