Information processing device, information processing method, and program

Abstract

To provide an information processing device, information processing method, and program that appropriately estimate causal relationships. [Solution] In an information processing system in which an information processing device, a terminal device, and an AI processing device are connected to each other via a network so as to be able to communicate with each other, the information processing device 10 includes: an acquisition unit that acquires at least two pieces of event information relating to an event; a provisional estimation unit that provisionally estimates the causal relationship between each event based on at least two pieces of event information; a first AI processing unit that inputs instruction information including instructions for identifying a third factor between each event based on the causal relationship to a first AI model and acquires the third factor identified by the first AI model; a second AI processing unit that inputs instruction information including instructions for estimating the causal relationship between each event based on the causal relationship and the third factor to a second AI model and acquires the causal relationship estimated by the second AI model; and an output unit that outputs information including the estimated causal relationship between each event.

Description

Technical Field

[0001] The disclosed technology relates to an information processing apparatus, an information processing method, and a program.

Background Art

[0002] Conventionally, a causal search program that causes a computer to execute the following processing has been disclosed: by a plurality of processes, based on first sample data including a numerical sequence for each of a plurality of variables to be calculated, calculation of a first statistical information amount between each of the plurality of variables and another variable is executed in parallel processing; the first statistical information amount for each of the plurality of variables is shared among the plurality of processes; by each of the plurality of processes, based on the first statistical information amount for each of the plurality of variables, a first variable that is the first in a causal order, in which the plurality of variables are arranged such that variables that can be a cause in a causal relationship between the plurality of variables come first, is determined; by each of the plurality of processes, based on the first statistical information amount for each of the plurality of variables, a second variable that is next to the first variable in the causal order among the plurality of variables is estimated; and by each of the plurality of processes, it is determined to execute: calculation of a second statistical information amount between each of the remaining variables obtained by excluding the first variable from the plurality of variables and another remaining variable, based on second sample data obtained by removing the contribution of the first variable from the first sample data, and calculation of a third statistical information amount between each of the remaining variables obtained by excluding the first variable and the second variable from the plurality of variables and another remaining variable, based on third sample data obtained by removing the contributions of the first variable and the second variable from the first sample data.

Prior Art Literature

Patent Literature

[0003]

Patent Literature 1

Summary of Invention

Problem to be Solved by Invention

[0004] However, the causal search program described in Patent Document 1 sometimes failed to properly estimate causal relationships due to noise and errors included in the sample data, as well as the influence of events for which data could not be measured as part of the sample data.

[0005] Therefore, one of the objectives of disclosure technology is to appropriately estimate causal relationships. [Means for solving the problem]

[0006] An information processing device according to one aspect of the present invention comprises: an acquisition unit that acquires at least two pieces of event information relating to an event; a provisional estimation unit that provisionally estimates the causal relationship between each event based on at least two pieces of event information; a first AI processing unit that inputs instruction information including instructions for identifying a third factor between each event based on the causal relationship to a first AI model and acquires the third factor identified by the first AI model; a second AI processing unit that inputs instruction information including instructions for estimating the causal relationship between each event based on the causal relationship and the third factor to a second AI model and acquires the causal relationship estimated by the second AI model; and an output unit that outputs information including the estimated causal relationship between each event. [Effects of the Invention]

[0007] According to the disclosure technology, causal relationships can be appropriately estimated. [Brief explanation of the drawing]

[0008] [Figure 1] This figure shows an example of the configuration of the information processing system according to this embodiment. [Figure 2] This figure shows an example of the hardware configuration of the information processing device according to this embodiment. [Figure 3] This figure shows an example of the functional block configuration of the information processing device according to this embodiment. [Figure 4] This figure shows an example of an event information database according to this embodiment. [Figure 5] This figure shows an example of a third-factor information database according to this embodiment. [Figure 6] This figure shows an example of an estimated result information database according to this embodiment. [Figure 7] This is an example of a screen displayed on the terminal device according to this embodiment. [Figure 8] This figure shows an example of the functional block configuration of the terminal device in this embodiment. [Figure 9] This flowchart shows an example of the processing procedure of the information processing device according to this embodiment. [Figure 10] This flowchart shows another example of the processing procedure of the information processing device according to this embodiment. [Modes for carrying out the invention]

[0009] A preferred embodiment of the disclosed technology will be described with reference to the attached drawings. In each drawing, components denoted by the same reference numerals have the same or similar configuration.

[0010] In this embodiment, "causal relationship" includes a relationship where the effect changes when the cause changes. The direction of the arrow indicates the direction of causality, and a causal relationship in which event A (cause) influences event B (effect) is shown in the form A→B. The cause may include explanatory variables, and the effect may include dependent variables.

[0011] <Information Processing System 1> The following describes the information processing system 1 in the disclosed technology. Figure 1 is a diagram showing an example of the configuration of the information processing system 1 according to this embodiment. The information processing system 1 shown in Figure 1 includes an information processing device 10, one or more terminal devices 20, and an AI processing device 30.

[0012] The information processing device 10, the terminal device 20, and the AI processing device 30 may be communicably connected to each other via a network N. The network N may be a network for communication between the information processing device 10, the terminal device 20, and the AI processing device 30. For example, the network N may be any one of the Internet, an intranet, a LAN, a mobile communication network, a leased line, a packet communication network, a telephone line, a corporate network, other communication lines, combinations thereof, and the like. Further, the network N may be either wired or wireless.

[0013] <Information processing apparatus 10> The information processing device 10 is configured by a server, a personal computer, or the like. The information processing device 10 is a device that assumes part of the functions of information processing provided by the information processing system 1, for example, acquiring event information related to an event. The information processing device 10 may be configured using a virtual server, a cloud server, or the like. The information processing device 10 may also be referred to as a computer.

[0014] <Terminal device 20> The terminal device 20 is, for example, a device used by a user, and is a mobile phone terminal (including a smartphone), a tablet, or a personal computer. For example, the user may output event information to the information processing device 10 by operating the terminal device 20. Further, the user may display information output by the information processing device 10 via the terminal device 20.

[0015] Further, an application program (app) for using various functions provided by the information processing device 10 may be installed in the terminal device 20. The application may be web browsing software. The application may cause the terminal device 20 to execute at least part of the processing disclosed in the embodiments described below among the various functions provided by the information processing device 10. By executing the application, the terminal device 20 may, for example, access the information processing device 10 and transmit and receive information used for executing the application.

[0016] <AI processing device 30> The AI processing device 30 is, for example, a computer that operates an AI (Artificial Intelligence) model for executing processing on an input character string including natural language. The AI processing device 30 stores a large amount of data used in the AI model, and executes processing associated with the AI model using the data stored in the database.

[0017] The AI model is a language model specialized for natural language processing, and may be a type of so-called Generative Artificial Intelligence (Generative AI) that generates and outputs sentences based on instruction information (a prompt) when the instruction information is input. The AI model may be, for example, a general-purpose language model (e.g., a large language model) adaptable to natural language processing such as information extraction, text summarization, text generation, and question answering.

[0018] Furthermore, the instruction information may be an input sentence for the AI model. The instruction information input to the AI model includes, for example, instruction information containing an instruction related to specifying a third factor between each event based on a tentatively estimated causal relationship.

[0019] The AI processing device 30 may configure generative AI using an AI model. Existing generative AI can be used as the generative AI. For example, the AI processing device 30 can transmit information to and receive information from the information processing device 10 via an API of the generative AI.

[0020] The AI processing device 30 includes, for example, services that can be provided on the cloud. The AI processing device 30 may be, for example, a generative AI server that provides a cloud-based service using an AI model. Note that the generative AI server is not limited to this, and may be any generative AI server that executes natural language processing using an AI model that provides the same functions.

[0021] In this embodiment, the information processing device 10 is described as a device equipped with functions for acquiring event information, and the AI ​​processing device 30 is described as a device equipped with generation AI functions. However, the information processing device 10 and the AI ​​processing device 30 may be configured as the same information processing device, or they may be configured to be provided as a single service.

[0022] <Hardware Configuration> Figure 2 shows an example of the hardware configuration of the information processing device 10 according to this embodiment. The information processing device 10 includes a processor 11 such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit), a storage device 12 such as memory (for example, RAM (Random Access Memory) or ROM (Read Only Memory)), an HDD (Hard Disk Drive) and / or an SSD (Solid State Drive), a communication interface 13 for wired or wireless communication, an input device 14 for receiving input operations, and an output device 15 for outputting information.

[0023] The input device 14 is, for example, a keyboard, touch panel, mouse, camera, and / or microphone. The output device 15 is, for example, a display, touch panel, and / or speaker. The input device 14 receives various types of information from the user. The output device 15 may also include a display unit that displays various types of information to the user.

[0024] The hardware configuration described above is merely an example. The information processing device 10 within the information processing system 1 may omit some of the hardware shown in Figure 2, or it may include hardware not shown in Figure 2. Furthermore, the hardware shown in Figure 2 may be composed of one or more devices. Also, if the information processing device 10 is composed of multiple devices, each device may include at least some of this hardware. In addition, the same hardware configuration as described above may apply to at least one of the terminal device 20 and the AI ​​processing device 30. Furthermore, at least one of the terminal device 20 and the AI ​​processing device 30 may omit some of the hardware shown in Figure 2, or it may include hardware not shown in Figure 2.

[0025] <Functional Block Configuration> (Information processing device 10) Figure 3 shows an example of the functional block configuration of the information processing device 10 according to this embodiment. The information processing device 10 includes a storage unit 100 and a control unit 110. The control unit 110 includes an acquisition unit 111, a provisional estimation unit 112, a first AI processing unit 113, a second AI processing unit 114, and an output unit 115. The control unit 110 may further include a determination unit 116 and a correction unit 117.

[0026] The memory unit 100 can be implemented using the storage device 12 provided by the information processing device 10. The control unit 110 can be implemented by the processor 11 of the information processing device 10 executing a program stored in the storage device 12.

[0027] Furthermore, the program may be stored in a storage medium. The storage medium on which the program is stored may be a non-transitory computer-readable medium. The non-transitory storage medium is not particularly limited, but may be, for example, a USB (Universal Serial Bus) memory or a CD-ROM (Compact Disc-Read Only Memory).

[0028] The storage unit 100 stores the data necessary for the information processing device 10 to perform information processing. The storage unit 100 includes, for example, an event information DB 100a and another event information DB 100a. It is possible to add data items to each DB as needed.

[0029] Figure 4 shows an example of the event information DB 100a according to this embodiment. The event information DB 100a manages event information. The event information DB 100a may store event information IDs associated with each event information (e.g., diet, exercise habits, income, and health).

[0030] The event information ID includes information that identifies the row of the event information. Each event information includes information corresponding to each event. In the event information DB 100a, for example, quantified rank information may be stored by defining a predetermined rank (e.g., three ranks: rank 1 (low), rank 2 (medium), and rank 3 (high)) for each event (diet, exercise habits, income, and health), and one data set may be formed by the information associated with the event information ID. The example shown in the event information DB 100a is an example in which the acquisition unit 111 acquires at least four event information, and the event information DB 100a may store the event information as a single dataset.

[0031] Figure 5 shows an example of the third-factor information DB100b according to this embodiment. The third-factor information DB100b manages information on third factors. The third-factor information DB100b may store, for example, a third-factor ID, the name of the third factor (B), the name of the structure, and the structure information in association with each other.

[0032] Furthermore, the third factor may include auxiliary factors that influence the causal structure, or other factors that influence the cause-and-effect relationship. The third factor may be an auxiliary factor in the causal structure, or it may be a variable that influences the causal relationship.

[0033] The third factor ID contains information that identifies each third factor. The name of the third factor (B) contains the name information of each third factor. In this embodiment, the name of the third factor includes confounding factors, mediating factors, and modulating factors.

[0034] The structure name includes structure name information including the third factor. In this embodiment, the name of the third factor includes chain structure, branch structure, and collision structure. The structure information includes structure information including the third factor. The structure information may also include an image diagram of the structure information including the third factor. B in the structure information indicates the third factor.

[0035] As shown in the third-factor information DB100b, intermediate factors include factors in the cause-to-effect pathway and are associated with a chain structure (A→B→C). Intermediate factors may be mediating factors in the cause-to-effect pathway, or they may be modifying factors that change the strength of the cause-and-effect relationship.

[0036] Confounding factors include elements that influence both cause and effect, and are associated with a branching structure (A←B→C). Furthermore, confounding factors include elements that distort true causal relationships.

[0037] Collision factors include elements that are influenced by both cause and effect, and correspond to a collision structure (A←B→C). Furthermore, collision factors include elements that influence a common outcome.

[0038] Figure 6 shows an example of the estimation result information DB100c according to this embodiment. The estimation result information DB100c manages information on the third factor. The estimation result information DB100c may store, for example, the estimation result ID, the first event name, the second event name, and the third factor ID in association with each other.

[0039] The estimated result ID includes information that identifies the result estimated by the information processing device 10. The first event name includes the event name of the causal relationship event estimated as the cause. The second event name includes the event name of the causal relationship event estimated as the result. The first event name and the second event name may be stored as a pair in the estimated result information DB 100c. The third factor ID includes information that identifies the third factor in the relationship between the identified first event name and the second event name. If multiple third factors are identified, multiple third factor IDs are stored; if no third factors are identified, no third factor ID is stored.

[0040] Returning to Figure 3, let's continue the explanation. The acquisition unit 111 acquires at least two pieces of event information related to an event. The user may, for example, operate the terminal device 20 to output a dataset (rows: data for each event information, columns: each event information name) to the information processing device 10, with four event names (e.g., advertisement (number of flyers distributed), season, discount rate, sales) associated as event information. Alternatively, the acquisition unit 111 may acquire the event information output from the terminal device 20. Note that the event information may be in the form of a dataset as described above, or it may consist only of the event names for each event.

[0041] The provisional estimation unit 112 estimates the causal relationships between events based on at least two or more pieces of event information. The provisional estimation unit 112 may input instruction information, including instructions for provisionally estimating the causal relationships between events based on at least two or more pieces of event information, into a predetermined AI model, and obtain the causal relationships between events provisionally estimated by the predetermined AI model. Note that provisional estimation includes making assumptions and provisionally determining the causal relationships between event information.

[0042] The provisional estimation unit 112 may input instruction information generated by incorporating acquired event information into instruction information prepared in advance by the operator of the information processing device 10 to the AI ​​processing device 30 having a predetermined AI model. The AI ​​processing device 30 may have the predetermined AI model provisionally estimate the causal relationships between each event based on the instruction information input by the provisional estimation unit 112. The AI ​​processing device 30 may then output the causal relationships between each event provisionally estimated by the predetermined AI model to the information processing device 10, thereby allowing the provisional estimation unit 112 to obtain the provisionally estimated causal relationships between each event.

[0043] The provisional estimation unit 112 may input instruction information, which includes instructions for provisionally estimating causal relationships between events based on event information and causal inference theory, into a predetermined AI model, and obtain the causal relationships between events provisionally estimated by the predetermined AI model. Causal inference theory includes theories that aim to statistically clarify the cause-and-effect relationship in which one event causes another event, based on data corresponding to the events. Furthermore, causal inference theory includes theoretical knowledge of causal inference, which includes knowledge for determining whether or not a causal relationship exists between events.

[0044] The provisional estimation unit 112 can make provisional estimations, including determining whether the relationship between events is a causal relationship, by making provisional estimations of causal relationships between each event (e.g., investment in business A → sales of business A) based on event information (e.g., investment in business A → sales of business A) and causal inference theory.

[0045] Furthermore, the provisional estimation unit 112 may input instruction information to the AI ​​model, which includes instructions for provisionally estimating the causal relationships between each event in a way that avoids a cyclical structure, based on the temporal order of at least two pieces of event information, and obtain the causal relationships between each event that the AI ​​model has provisionally estimated. This makes it possible to provisionally estimate causal relationships that suppress the occurrence of cyclical structures.

[0046] Furthermore, the provisional estimation unit 112 is not limited to provisionally estimating causal relationships between events using an AI model. The provisional estimation unit 112 may also provisionally estimate causal relationships between events using a learning model that has learned the relationships between previously estimated causal relationships, the causal events corresponding to those causal relationships, and the resulting events. Alternatively, the provisional estimation unit 112 may, for example, input each event into a learning model and obtain the provisional estimation results output by the learning model. Alternatively, the provisional estimation unit 112 may provisionally estimate causal relationships between events using a definition table (rule-based) that defines the relationships between previously estimated causal relationships, the causal events corresponding to those causal relationships, and the resulting events.

[0047] The first AI processing unit 113 inputs instruction information to the first AI model, which includes instructions for identifying third factors between events based on the causal relationships between each event that have been tentatively estimated, and obtains the third factors identified by the first AI model. The first AI processing unit 113 may, for example, input instruction information to the first AI model, which includes instructions for identifying multiple third factors between each event based on the causal relationships between each event that have been tentatively estimated.

[0048] Furthermore, the first AI processing unit 113 may input instruction information to the first AI model, which includes instructions for identifying a third factor between each event, based on the causal relationship between each event that has been tentatively estimated and domain information related to the event information, and obtain the third factor identified by the first AI model.

[0049] Furthermore, the first AI processing unit 113 may input instruction information to the first AI model, which includes instructions to supplement third factors not included in the event information based on domain knowledge regarding each event information, and obtain the third factors supplemented by the first AI model. This makes it possible to obtain third factors that cannot be obtained by statistical causal exploration alone.

[0050] Furthermore, domain knowledge includes knowledge for estimating relationships between events, including third factors, from empirical rules and prior research; knowledge for explaining how multiple third factors influence events; and knowledge for understanding the mechanisms of influence between events, including related third factors.

[0051] Furthermore, if the first AI processing unit 113 obtains multiple third factors, it may input instruction information to the first AI model, including instructions for organizing each third factor based on the multiple third factors (e.g., deleting duplicate third factors), and obtain the third factors organized by the first AI model.

[0052] Furthermore, the first AI processing unit 113 may input instruction information to the first AI model that includes instructions for identifying multiple third factors between each event based on causal relationships. If the first AI processing unit 113 obtains multiple third factors, it may input instruction information to the first AI model that includes instructions for calculating the similarity of the multiple third factors and integrating third factors having a similarity of a predetermined value or higher. The first AI processing unit 113 may also arbitrarily employ known methods such as machine learning models to calculate the similarity. For example, the first AI processing unit 113 may convert each third factor into a vector representation using a dense vector model such as BERT (Bidirectional Encoder Representations from Transformers) or Word2Vec and calculate the cosine similarity. The first AI processing unit 113 may then integrate the third factors based on the calculated similarity.

[0053] The second AI processing unit 114 inputs instruction information to the second AI model, which includes instructions for estimating the causal relationships between each event based on causal relationships and third factors, and obtains the causal relationships estimated by the second AI model. The second AI processing unit 114 may also have the second AI processing unit 114 analyze the influence of third factors and estimate causal relationships based on the results of that analysis.

[0054] Furthermore, the second AI processing unit 114 may input instruction information to the second AI model, including instructions for estimating an influence mechanism including a third factor based on the causal relationship and the third factor, and obtain the influence mechanism estimated by the second AI model. Subsequently, the second AI processing unit 114 may input instruction information to the second AI model, including instructions for estimating the causal relationship between each event based on the causal relationship, the third factor, and the influence mechanism, and obtain the causal relationship estimated by the second AI model.

[0055] Furthermore, the second AI processing unit 114 may input instruction information to the second AI model, which includes instructions for estimating causal relationships between events based on causal relationships, third factors, and domain information related to event information, and obtain the causal relationships estimated by the second AI model.

[0056] Hereinafter, unless otherwise specified, the first AI processing unit 113 and the second AI processing unit 114 will simply be referred to as the AI ​​processing unit. Similarly, unless otherwise specified, the first AI model and the second AI model will simply be referred to as the AI ​​model, and unless otherwise specified, the first AI model group and the second AI model group will simply be referred to as the AI ​​model group.

[0057] The output unit 115 outputs information including the causal relationships between each estimated event. The output unit 115 may also output to the terminal device 20, in a predetermined format (e.g., list format), the third factor associated with each event (e.g., (event A, event B), (event B, event C)) and the estimated result information of the event (e.g., event A (cause), event B (effect))).

[0058] Through the above processing, the information processing device 10 can appropriately estimate causal relationships regardless of the acquired event information (e.g., data quality). Furthermore, by confirming the results of the estimated causal relationships, the user can understand the causal relationships between events, including third factors not included in the event information. In addition, the information processing device 10 can appropriately extract domain knowledge from the first AI model and the second AI model by having them perform causal inference processing in stages. Furthermore, the information processing device 10 can appropriately execute causal inference processing by concentrating the first AI model and the second AI model into a single process at each stage. Moreover, even when there is a high possibility of noise or confounding in the acquired event information, the information processing device 10 can estimate causal relationships that are close to the ideal causal structure and are highly reliable.

[0059] Furthermore, the second AI processing unit 114 may input instruction information to the second AI model, which includes instructions for estimating the causal relationship between each event information and the third factor by selecting at least one of the following structures: a chain structure, a branching structure, and a collision structure, based on the causal relationship and the third factor, and obtain the causal relationship estimated by the second AI model.

[0060] The second AI processing unit 114 may input instruction information to the second AI model, which includes instructions for estimating the causal relationship between each event information and the third factor by selecting at least one of the following structures: a chain structure, a branching structure, and a collision structure, based on the causal relationship (e.g., event A → event B) and the third factor (e.g., factor C), and obtain the causal relationship estimated by the second AI model.

[0061] The second AI processing unit 114 may obtain the selected structure (e.g., chain structure) and the causal relationship (e.g., event A → factor C → event B) as causal relationships estimated by the second AI model, by associating them. The second AI processing unit 114 is not limited to selecting at least one of the chain structure, branch structure, and collision structure, but may also select from a predetermined causal inference theory structure pattern.

[0062] Furthermore, the second AI processing unit 114 may obtain the reason for selecting the causal relationship estimated by the second AI model (e.g., for event A → event B, factor C exists as a collision structure. The reason is xxx).

[0063] Furthermore, the output unit 115 may output the selected structure (e.g., collision structure) and the causal relationship (e.g., event A → factor C ← event B) in association.

[0064] Through the above processing, the information processing device 10 can obtain causal relationships estimated based on at least one of the following structures: a chain structure, a branching structure, and a collision structure. Furthermore, by having an AI model estimate the causal relationships, the information processing device 10 can reduce the amount of time required for manual verification by experts and efficiently output causal relationships.

[0065] Furthermore, if multiple third factors are obtained, the second AI processing unit 114 may input instruction information to the second AI model, including instructions for selecting the third factor that has the most influence on the causal relationship based on the causal relationship and each third factor, and obtain the third factor selected by the second AI model. Then, the second AI processing unit 114 may input instruction information to the second AI model, including instructions for estimating the causal relationship between each event based on the causal relationship and only the third factor selected by the second AI model, and obtain the causal relationship estimated by the second AI model.

[0066] If multiple third factors (e.g., third factor A, third factor B, third factor C, third factor D) are obtained for a single causal relationship (e.g., event A → event B), the second AI processing unit 114 may input instruction information to the second AI model that includes instructions to take each third factor one by one based on the causal relationship and each third factor, analyze how it affects the causal relationship between events, and determine whether or not a causal relationship exists. The second AI processing unit 114 may then obtain the third factors that the second AI model has determined to have a causal relationship.

[0067] Subsequently, the second AI processing unit 114 may input instruction information to the second AI model, which includes instructions for estimating the causal relationships between each event based on the causal relationships and only the third factors that the second AI model has determined to have a causal relationship, and obtain the causal relationships estimated by the second AI model.

[0068] Through the above processing, the information processing device 10 can subdivide the complex processing task of estimating causal relationships and have the second AI model analyze multiple third factors in a stepwise manner, thereby estimating causal relationships that take into account the influence of these third factors. Furthermore, because the information processing device 10 has multiple third factors estimated individually, it can estimate causal relationships more appropriately compared to when the second AI model processes multiple third factors at once.

[0069] Furthermore, the first AI model may be a group of first AI models that includes multiple first AI models. The first AI processing unit 113 may use a predetermined framework to determine priorities based on causal relationships and input instruction information to the group of first AI models, including instructions for identifying third factors between each event and the priority of those third factors, and obtain the third factors identified by the group of first AI models. The first AI processing unit 113 may also obtain evaluation results of the third factors identified by the group of first AI models (e.g., the results of evaluating each third factor on a scale of 10 for causal relationship A → B, and evaluation comments).

[0070] The predetermined framework for determining priorities may be a framework for tasks with no choices, and one example is NGT (nominal group technique). The information processing device 10 can have the first AI model group discuss using NGT, thereby causing the first AI model group to generate ideas divergently.

[0071] Furthermore, the specified framework may also be a set of rules that structure the overall flow of the discussion and determine the order in which the AI ​​model should make decisions.

[0072] Furthermore, if the information processing device 10 determines that there is no choice in the discussion, it may input instruction information to the first AI model group, including instructions to have the first AI model group engage in discussion using NGT as a predetermined framework. Alternatively, the first AI processing unit 113 may have the first AI models engage in divergent discussion and obtain the ideas generated independently by each first AI model.

[0073] Furthermore, the first AI processing unit 113 may input instruction information to the first AI model group, including instructions to discuss using NGT and independently propose candidate third factors, and obtain the candidate third factors proposed by the first AI model group. The output unit 115 may also output the obtained candidate third factors in list format. This allows for the appropriate estimation of causal relationships by analyzing each third factor individually according to priority when multiple third factors exist.

[0074] The second AI model may be a group of second AI models that includes multiple AI models. The second AI processing unit 114 may input instruction information to the group of second AI models, which includes instructions for discussing using a predetermined framework for reaching a consensus based on causal relationships and third factors, and for estimating causal relationships between each event, and obtain the causal relationships estimated by the group of second AI models.

[0075] The predetermined framework for reaching a consensus includes the Delphi method. The Delphi method includes a framework for anonymously collecting the opinions of multiple experts and reaching a consensus. The information processing device 10 may also use the Delphi method to have the AI ​​models engage in discussions in discussions with multiple choices.

[0076] Furthermore, the second AI processing unit 114 may input instruction information to the second AI model group, which includes instructions for estimating the causal relationship between each event information and the third factor by selecting at least one of the following structures—a chain structure, a branch structure, and a collision structure—using the Delphi method based on the causal relationship and the third factor, and obtain the causal relationship estimated by the second AI model group. This allows the second AI model group to conduct a convergent discussion to select the appropriate one from the options (chain structure, branch structure, and collision structure) based on causal inference theory, thereby realizing a discussion similar to that which experts use to reach a consensus.

[0077] Furthermore, the second AI processing unit 114 may input instruction information to the second AI model group, including instructions for determining the third factor using a predetermined framework for discussing and reaching a consensus based on causal relationships and third factors, and obtain the third factor determined by the second AI model group. The second AI processing unit 114 can achieve appropriate judgment based on causal inference theory by ultimately selecting the third factor by majority vote based on the Delphi method.

[0078] Furthermore, the information processing device 10 may use NGT to extract diverse ideas from the first AI model group for divergent tasks, and use the Delphi method to have the second AI model group reach a consensus for convergent tasks, thereby controlling the discussion according to the characteristics of each task. In addition, the information processing device 10 may adopt theoretical insights of causal inference as the judgment criteria for each discussion.

[0079] Specifically, the first AI processing unit 113 may acquire prioritized third-factor candidates, and the second AI processing unit 114 may determine whether or not to select each of the prioritized third-factor candidates and acquire the selected third-factor. Furthermore, the second AI processing unit 114 may estimate the most appropriate causal relationship based on the selected third-factor.

[0080] The information processing device 10 can acquire deep knowledge about each third factor from the AI ​​model group by having it discuss multiple third factors one by one in sequence. As a result, it can more accurately understand the influence mechanisms between events based on the acquired information. Furthermore, since it can comprehensively and systematically acquire third factors not included in event information (e.g., confounding factors that affect events), it can complement high-quality domain knowledge such as that provided by experts, and estimate causal relationships more appropriately.

[0081] Through the above processing, the information processing device 10 can appropriately estimate causal relationships by having the first AI model group engage in divergent discussions and then the second AI model group engage in convergent decision-making, thereby progressing the discussions step by step at each stage of processing and comprehensively and efficiently identifying a variety of third factors.

[0082] Furthermore, the information processing device 10 can control the first AI model group and the second AI model group to follow a structured argument framework, thereby enabling sophisticated causal analysis and domain knowledge supplementation similar to that performed by experts.

[0083] Furthermore, by clearly defining the tasks to be discussed at each stage, the information processing device 10 enables the AI ​​model to process intensively within the scope of the given tasks. As a result, the information processing device 10 can appropriately extract domain knowledge from the AI ​​model, even for information between events consisting only of numerical data, without depending on the information present in the context.

[0084] Furthermore, instead of having the AI ​​model group analyze multiple third factors at once, the information processing device 10 has the AI ​​model group discuss each third factor sequentially, thereby extracting knowledge about each third factor and separating the two tasks of "extraction of third factors" and "causal judgment considering each third factor." In addition, by discussing each third factor individually, information not included in the event information can be obtained from the AI ​​model group.

[0085] The determination unit 116 integrates the estimated causal relationships and determines whether the integrated causal relationship is a cyclic structure. The determination unit 116 may also determine whether the integrated causal relationship is a cyclic structure after integrating the individually estimated causal relationships into a consistent causal structure as a whole. Hereinafter, the integrated causal relationship will also be referred to as the overall causal relationship.

[0086] The determination unit 116 may represent each estimated causal relationship as a causal graph, such as a directed acyclic graph or a partial ancestral graph. In this embodiment, the estimated causal relationships are represented by a directed acyclic graph. A directed acyclic graph consists of nodes (including at least one of events and third factors) and edges (including arrows indicating causal relationships between nodes), and the direction of the edges represents which nodes influence other nodes. A characteristic of a directed acyclic graph is that it is directed but does not contain cycles, which clarifies the direction of causal relationships and makes it possible to visualize how intervention in one node affects other nodes.

[0087] Furthermore, a cyclic structure includes a state on a causal graph showing a causal relationship in which event A influences event B, event B influences event C, and event C then influences event A in a chain reaction, forming a closed loop (event A → event B → event C → event A). Similarly, a cyclic structure includes a structure on a causal graph in which event A influences event B, and event B then influences event A in a chain reaction, forming a closed loop.

[0088] The determination unit 116 may also input instruction information, including instructions on whether or not the integrated causal relationship has a cyclic structure, into a predetermined AI model, obtain the determination result made by the predetermined AI model, and determine whether or not it has a cyclic structure based on the determination result.

[0089] If the modification unit 117 determines that the integrated causal relationship is a cyclic structure, it will modify it to become a non-cyclic structure. If the modification unit 117 determines that the integrated causal relationship is a cyclic structure, it may modify the integrated causal relationship to become a non-cyclic structure using a predetermined algorithm.

[0090] The modification unit 117 may detect cycles using depth-first search and calculate all possible patterns by removing edges connecting nodes so that all cycles are eliminated. The modification unit 117 may then calculate the Bayes information criterion (BIC) for all patterns and select the structure with the smallest BIC to modify the integrated causal relationships into a non-cyclic structure.

[0091] Alternatively, the modification unit 117 may consider only the edges included in the cyclic structure as deletion candidates, calculate the BIC for each deletion pattern, select the deletion pattern with the lowest BIC, and delete the edges based on that deletion pattern, thereby modifying the integrated causal structure to become a non-cyclic structure. Note that by considering only the edges included in the cyclic structure as deletion candidates and not modifying other edges, the modification unit 117 can eliminate the cyclic structure included in the integrated causal relationship while ensuring the overall reliability of the causal structure.

[0092] Furthermore, the output unit may output the modified overall causal relationships in the form of a predetermined graph structure. For example, the output unit may input instruction information, including instructions for generating a directed acyclic graph based on the modified overall causal relationships, into a predetermined AI model, obtain the directed acyclic graph generated by the predetermined AI model, and output it to the terminal device 20.

[0093] Through the above processing, the information processing device 10 can eliminate cycles contained in the integrated causal structure and modify it into a causal structure that satisfies the application conditions of causal estimation methods such as directed acyclic graphs (DAGs), which are based on the premise of acyclicity. This suppresses the problem of indeterminate estimation due to the inability to define the causal order because of the presence of cycles, and enables the estimation of a more reliable causal structure.

[0094] Figure 7 shows an example of a screen displayed on the terminal device 20 according to this embodiment. In the example shown in Figure 7, the terminal device 20 displays graph A100, which shows causal relationships output from the information processing device 10. By checking graph A100 output by the information processing device 10, the user can visually grasp complex causal relationships. Furthermore, by checking graph A100, the user can clarify the relationships between events and use it as a basis for policy planning and evaluation in practical work.

[0095] Graph A100 represents the relationships between nodes using a directed acyclic graph. In the example of graph A100, events include seasonality, advertising, discounts, and sales, while third factors include temperature and inventory. The output unit may also output to the terminal device 20 in a way that allows identification of nodes related to events and nodes related to third factors.

[0096] The relationships between nodes are represented by arrows connecting the edges. The arrows are directional and indicate the influence between nodes. In the example of graph A100, the event (seasonality) influences the event (advertising) and the event (sales). Furthermore, the event (advertising) is influenced by the event (seasonality) and influences the event (sales). Also, the event (discount) influences the event (advertising) and the event (sales).

[0097] Furthermore, in the example of Graph A100, the third factor (temperature) is a confounding factor and has an influence on the event (seasonality) and the event (advertising). The third factor (inventory) is an intermediate factor and is influenced by the event (seasonality) and has an influence on the event (advertising). Note that Graph A100 is just one example, and other third factors may be included.

[0098] (Terminal device 20) Figure 8 shows an example of the functional block configuration of the terminal device 20 in this embodiment. The terminal device 20 includes a storage unit 200 and a control unit 210. The storage unit 200 can be implemented using a storage device 12 provided in the terminal device 20.

[0099] Furthermore, the control unit 210 can be realized by the processor 11 of the terminal device 20 executing a program stored in the storage device 12. This program can be stored in a storage medium. The storage medium containing the program may be a non-transitory computer-readable medium. The non-transitory storage medium is not particularly limited, but may be, for example, a USB memory stick or a CD-ROM.

[0100] The control unit 210, in cooperation with the information processing device 10, provides various functions necessary for, for example, receiving information output from the information processing device 10 and displaying it on the screen of the terminal device 20. For example, the control unit 210 provides functions such as acquiring various information (image data, text data, etc.) from the information processing device 10 for drawing on the screen of the terminal device 20. The control unit 210 includes, for example, a communication unit 211 and a UI (User Interface) unit 212.

[0101] The memory unit 200 stores various programs and data necessary for the control unit 210 to perform this information processing.

[0102] The communication unit 211 has the function of performing various types of communication with the information processing device 10 using the communication IF 13.

[0103] The UI unit 212 has, for example, a function to receive various inputs from the user and a function to display screens output by the information processing device 10 on the display.

[0104] Regarding the functional block configuration described above, it is also possible to configure the information processing device 10 to have all or part of the storage unit 100 and the control unit 110, all of which are included in the information processing device 10, in at least one of the terminal device 20 and the AI ​​processing device 30. In other words, the various processes according to this embodiment may be executed by the processor of the information processing device 10, by at least one of the processors of the terminal device 20 and the AI ​​processing device 30, or by the processors of the information processing device 10, the terminal device 20 and the AI ​​processing device 30 working together to execute them.

[0105] <Operation of the information processing device 10> Next, the operation of the information processing device 10 according to this embodiment will be described. Figure 9 is a flowchart showing an example of the processing procedure of the information processing device 10 according to this embodiment.

[0106] In step S101, the acquisition unit 111 acquires at least two pieces of event information related to the event.

[0107] In step S102, the provisional estimation unit 112 provisionally estimates the causal relationships between each event based on at least two or more pieces of event information. In addition, if the acquired event information (e.g., seasonality, economic conditions, advertising campaigns, sales) is such that the provisional estimation unit 112 provisionally estimates the overall causal structure based on the temporal order, such as seasonality → economic conditions → advertising campaigns → sales. This helps to suppress the cyclical structure.

[0108] In step S103, the first AI processing unit 113 inputs instruction information to the first AI model, which includes instructions for identifying third factors between each event based on causal relationships, and obtains the third factors identified by the first AI model.

[0109] In step S104, the second AI processing unit 114 inputs instruction information to the second AI model, which includes instructions for estimating the causal relationships between each event based on the causal relationships and third factors, and obtains the causal relationships estimated by the second AI model.

[0110] In step S105, the output unit 115 outputs information including the causal relationships between each of the estimated events.

[0111] Through the above processing, causal relationships can be appropriately estimated regardless of the quality of the acquired event information (e.g., number of data points, presence or absence of bias). Furthermore, the information processing device 10 can appropriately extract domain knowledge from the first AI model and the second AI model by having them perform causal inference processing step by step, thereby enabling them to appropriately perform causal inference processing.

[0112] Next, the operation of the information processing device 10 according to this embodiment will be described. Figure 10 is a flowchart showing another example of the processing procedure of the information processing device 10 according to this embodiment.

[0113] In step S201, the acquisition unit 111 acquires at least two pieces of event information related to the event.

[0114] In step S202, the control unit determines the causal order of the acquired event information. The control unit performs processing to represent the macro structure of the n event information in causal order. The causal order includes an order that abstracts causal relationships in one dimension. The control unit may determine the causal order of the acquired event information based on an existing method (e.g., CausalOrder).

[0115] Step S202 is executed when the number of event information items is n>=3 (e.g., diet, exercise habits, income, health (n=4)), and is skipped if n is not >=3.

[0116] In step S203, the provisional estimation unit 112 may take two of the acquired event information in a permutation (nP2) and provisionally estimate the causal relationship between each event. If the order of the acquired event information is eating and exercise habits, the provisional estimation unit 112 may provisionally estimate that eating is the cause and exercise habits are the effect. Furthermore, when the provisional estimation unit 112 takes two of the acquired n event information in a permutation (nP2), it may exclude combinations that contradict the causal order. That is, the event that precedes the causal order is e i , the subsequent event e j When this is the case, using the rank function π(e) i )<π(e j The pair (e i , e j We take ) as the object of causal estimation. This relationship can be expressed by the following equation (1).

number

[0117] In step S204, the first AI processing unit 113 inputs instruction information to the first AI model, which includes instructions for identifying third factors between each event based on the tentatively estimated causal relationships, and obtains the third factors identified by the first AI model. Specifically, the first AI processing unit 113 may input instruction information to the first AI model, which includes instructions for identifying 0 to n third factors between each event for each of the chain structure, branch structure, and collision structure, based on the tentatively estimated causal relationships, and obtains the third factors identified by the first AI model.

[0118] In step S205, the second AI processing unit 114 inputs instruction information to the second AI model, which includes instructions for estimating the causal relationship between each event based on the tentatively estimated causal relationship and the identified third factor, and obtains the causal relationship estimated by the second AI model. Specifically, if multiple third factors have been obtained, the second AI processing unit 114 may input instruction information to the second AI model, which includes instructions for selecting the third factor that has the most influence on the causal relationship based on the causal relationship and each third factor, and obtain the third factor selected by the second AI model. Then, the second AI processing unit 114 may input instruction information to the second AI model, which includes instructions for estimating the causal relationship between each event based on the causal relationship and only the third factor selected by the second AI model, and obtains the causal relationship estimated by the second AI model.

[0119] The second AI processing unit 114 may input instruction information to the second AI model, which includes instructions to select one third factor and analyze the relationship between that third factor and the tentatively estimated causal relationship. Subsequently, the second AI processing unit 114 may input instruction information to the second AI model, which includes instructions to verify the results of the relationship analysis. The second AI processing unit 114 may store the verification results verified by the second AI model in the storage unit 100. The second AI processing unit 114 may perform the relationship analysis process, the relationship analysis result verification process, and the verification result saving process for each identified third factor.

[0120] Alternatively, the second AI processing unit 114 may input instruction information to the second AI model, including instructions for selecting the verification result that best explains the tentatively estimated causal relationship, and then select the third factor that best influences the tentatively estimated causal relationship by obtaining the verification result selected by the second AI model.

[0121] In step S206, the determination unit 116 integrates the causal relationships between the estimated events and determines whether the integrated causal relationships have a cyclic structure. If the determination unit 116 determines that the overall causal relationship has a cyclic structure (S206: YES), the process proceeds to step S207. If the determination unit 116 determines that the overall causal relationship does not have a cyclic structure (S206: NO), the process proceeds to step S208.

[0122] In step S207, the modification unit 117 modifies the overall causal relationship to a non-cyclic structure. This ensures the consistency of the entire causal structure (e.g., A100 in Figure 7) while respecting the causal relationships between each estimated event, by modifying the cyclic structure to a more plausible structure.

[0123] In step S208, the output unit 115 may output information including the overall causal relationship. The output unit 115 may output a causal graph (e.g., A100 in Figure 7) to the terminal device 20 as information including the overall causal relationship.

[0124] Through the above processing, causal relationships can be estimated more appropriately, regardless of the quality of the acquired event information (e.g., number of data points, presence or absence of bias). By checking the information output by the information processing device 10, which includes the overall causal relationships, the user can easily grasp complex causal relationships.

[0125] Although embodiments of the present invention have been described above, these embodiments or examples are provided to facilitate understanding of the present invention and are not intended to limit it. The present invention can be modified or improved without departing from its spirit, and equivalents thereof are also included. Furthermore, the present invention can form various disclosures by appropriately combining the multiple components disclosed in the above embodiments or examples. For example, some components may be deleted from all the components shown in the embodiments. Moreover, components may be appropriately combined in different embodiments. [Explanation of Symbols]

[0126] 1... Information processing system, 10... Information processing device, 11... Processor, 12... Memory device, 13... Communication interface, 14... Input device, 15... Output device, 20... Terminal device, 30... AI processing device, 100... Memory unit, 110... Control unit, 111... Acquisition unit, 112... Provisional estimation unit, 113... First AI processing unit, 114... Second AI processing unit, 115... Output unit, 116... Judgment unit, 117... Correction unit, 200... Memory unit, 210... Control unit, 211... Communication unit, 212... UI unit

Claims

1. An acquisition unit that acquires at least two pieces of event information relating to an event, which include information identifying the event or information quantifying the event. A provisional estimation unit that provisionally determines the causal relationship between each event based on the information of at least two events, thereby provisionally estimating the causal relationship between each event, A first AI processing unit inputs instruction information to a first AI model, which includes instructions for identifying a third factor between each of the events that is a factor that affects the cause-and-effect relationship, based on the tentatively estimated causal relationship, and acquires the third factor identified by the first AI model. A second AI processing unit inputs instruction information to a second AI model, which includes instructions for estimating the causal relationships between each of the events based on the tentatively estimated causal relationships and the third factor, and acquires the causal relationships estimated by the second AI model. The system includes an output unit that outputs information including the causal relationships between the events estimated by the second AI model, Information processing device.

2. The second AI processing unit selects, based on the tentatively estimated causal relationship and the third factor, whether the third factor corresponds to a chain structure, a branching structure, or a collision structure, inputs instruction information to the second AI model, which includes instructions for estimating the causal relationship between each event and the third factor based on the selected structure, and obtains the causal relationship estimated by the second AI model. The information processing apparatus according to claim 1.

3. When multiple third factors are obtained, the second AI processing unit inputs instruction information to the second AI model, which includes instructions for selecting the third factor that has the most influence on the provisionally estimated causal relationship based on the provisionally estimated causal relationship and each third factor, and obtains the third factor selected by the second AI model. The information processing apparatus according to claim 1.

4. The first AI processing unit inputs instruction information, including instructions for identifying the third factors between each of the events, to the first AI model based on the tentatively estimated causal relationship and domain information relating to the event information, which indicates knowledge for estimating the relationship between events including the third factors, and obtains the third factors identified by the first AI model. The information processing apparatus according to claim 1.

5. The second AI processing unit inputs instruction information to the second AI model, which includes instructions for estimating the causal relationships between each of the events, based on the tentatively estimated causal relationships, the third factor, and domain information relating to the event information, which indicates knowledge for estimating the relationships between events including the third factor, and obtains the causal relationships estimated by the second AI model. The information processing apparatus according to claim 1.

6. The aforementioned first AI model is a group of first AI models, which includes multiple first AI models. The first AI processing unit discusses using a predetermined framework for determining priorities based on the tentatively estimated causal relationships, inputs instruction information including instructions for identifying third factors between each event and the priority of those third factors into the first AI model group, and obtains the third factors identified by the first AI model group. The information processing apparatus according to claim 1.

7. The aforementioned second AI model is a group of second AI models that includes multiple second AI models. The second AI processing unit discusses using a predetermined framework for reaching a consensus based on the tentatively estimated causal relationships and the third factors, inputs instruction information to the second AI model group, which includes instructions for estimating the causal relationships between each of the events, and obtains the causal relationships estimated by the second AI model group. The information processing apparatus according to claim 1.

8. A determination unit that integrates the causal relationships between the events estimated by the second AI model and determines whether the integrated causal relationships have a cyclical structure, The system further includes a modification unit that modifies the structure to a non-cyclic structure if it is determined to be a cyclic structure. The information processing apparatus according to claim 1.

9. Computers Obtain at least two pieces of event information relating to an event, which include information identifying the event or information quantifying the event. Based on the information of at least two events mentioned above, the causal relationships between each event are tentatively determined, thereby tentatively estimating the causal relationships between each event. Based on the provisionally estimated causal relationships, instruction information including instructions for identifying a third factor between each of the events, which is a factor that influences the cause-and-effect relationship, is input to the first AI model, and the third factor identified by the first AI model is obtained. Instruction information, including instructions for estimating the causal relationship between each of the events based on the tentatively estimated causal relationship and the third factor, is input to the second AI model, and the causal relationship estimated by the second AI model is obtained. The second AI model outputs information including the causal relationships between each of the events estimated by the second AI model, An information processing method that performs the following.

10. On the computer, Obtain at least two pieces of event information relating to an event, which include information identifying the event or information quantifying the event. Based on the information of at least two events mentioned above, the causal relationships between each event are tentatively determined, thereby tentatively estimating the causal relationships between each event. Based on the provisionally estimated causal relationships, instruction information including instructions for identifying a third factor between each of the events, which is a factor that influences the cause-and-effect relationship, is input to the first AI model, and the third factor identified by the first AI model is obtained. Instruction information, including instructions for estimating the causal relationship between each of the events based on the tentatively estimated causal relationship and the third factor, is input to the second AI model, and the causal relationship estimated by the second AI model is obtained. The second AI model outputs information including the causal relationships between each of the events estimated by the second AI model, A program that executes the command.

Images