Information Processing Systems

Abstract

We provide an information processing system that eliminates information loss due to discrete representation, enables pre-quantification of dynamic outcomes, and enhances intervention effectiveness. The information processing device 1000 of the information processing system 100 includes a processing device 900 for processing data, an input / output device 910, and a recording medium 920 that is a non-transitory computer-readable medium, and the processing device processes a persona snap ψ, which is a time slice of a space-time persona field ψ(x,t). t The system is equipped with a persona field storage unit 103 including a persona snap generation unit that executes processing including generating (x), and the spatiotemporal persona field ψ(x,t) is a representation of the psychological state of the subject as a continuous vector field with state coordinates x and time t as parameters, and is described using partial differential equations.

Description

[Technical Field]

[0001] The present invention relates to an information processing technique for human behavior analysis and intervention optimization, which is executed on an information processing system comprising a processor, a memory, and an input / output device. [Background technology]

[0002] In recent years, in fields such as direct marketing (DM), medicine, education, and finance, efforts to analyze a variety of observational data, such as target individuals' behavioral logs, vital signs, and social media posts, and to implement individually optimized intervention measures have rapidly expanded. Representative examples include methods that combine purchase history and browsing history to deliver retargeting advertisements, and behavior change apps that analyze step counts and sleep data to send notifications encouraging action.

[0003] In connection with such a method, attempts have been made to utilize artificial intelligence technology in generating personas. For example, Patent Document 1 discloses an information processing device including: a classification unit that classifies consumer consumption behavior data indicating the consumer's consumption behavior history and attributes into multiple segments; a prompt generation unit that generates, for at least some of the multiple segments, a prompt for generating a persona representing the segment based on the consumption behavior data belonging to the segment and consumer characteristics identified from the consumption behavior data; and a persona generation unit that generates a persona corresponding to the segment by inputting the prompt into a generation model, wherein the prompt generation unit extracts, from the multiple segments, segments whose level of focus as a target identified for each of the multiple segments satisfies a predetermined condition, and generates the prompt for each extracted segment.

[0004] As a technology for clustering and displaying customer personas on a GIS, Patent Document 2 discloses a regional characteristics analysis system for analyzing the regional characteristics of a specific area, which comprises: a behavior log memory unit that stores behavior log data of visitors to the specific area; a statistical information memory unit that stores statistical information of a base area that forms the basis for visitors to the specific area; and an analysis server that identifies visitors who have visited the specific area from the behavior log memory unit, acquires statistical information for the base area of ​​the identified visitor by referring to the statistical information memory unit, calculates the usage rate of the acquired statistical information according to the number of visitors, calculates virtual statistical data for the specific area according to the usage rate, and makes it available for analyzing the regional characteristics in the specific area.

[0005] Furthermore, uplift estimation is being performed using machine learning techniques to predict the impact of interventions such as advertisements and notifications on subjects. As one such technique, Patent Literature 3 discloses a system including at least a computer-readable memory that stores executable instructions and one or more processors programmed with the executable instructions, wherein the processor acquires first feature data regarding interactions between a first candidate set and content related to the item category after a communication about the item category is sent to the first candidate set, acquires second feature data regarding interactions between a second candidate set and content related to the same item category after the first communication is sent to the first candidate set and to which the first communication has not been sent, generates a treatment model using the first feature data, generates a control model using the second feature data, generates an uplift model using the treatment model and the control model, and generates an uplift model associated with the feature set based on the predicted effect. a treatment model that outputs a first probability that a candidate associated with a given feature set will interact with content related to the item category after receiving the communication related to the item category; the treatment model uses a weighting factor to correct for the effect on a subset of candidates that are over-represented in the first candidate set compared to a second candidate set; the control model outputs a second probability that the candidate will interact with content related to the item category if the candidate does not receive the communication; and the uplift model calculates a difference between the first and second probabilities output for the candidate associated with the feature set, the difference representing a predicted effect of the communication related to the item category on the candidate.

[0006] Furthermore, as a technology for optimizing advertisements using user influence scores, Patent Document 4 discloses a method for modeling causal relationships between users, the method including the following steps executed by a computing device: receiving communication between users as text data, the text data including sentiments expressed by one or more users regarding a certain subject in the communication; analyzing the text data to identify sentiments regarding the subject, the analysis including calculating a weighted average of one or more sentiment scores associated with sentiments toward the subject included in the communication; generating input data based on the weighted average of the one or more sentiment scores associated with the sentiments; receiving the input data as representing communication between users of social media; and simultaneously calculating one or more influence variables based on the input data. The method includes: determining causal relationships between users based on modeling, wherein the simultaneous modeling incorporates random fluctuations associated with the one or more influence variables; determining one or more influence variables from the one or more influence variables that influence the causal relationships between users, wherein the one or more influence variables include one or more endogenous variables and one or more exogenous variables, and wherein the exogenous variables moderate or mediate the influence of the endogenous variables on the causal relationships; generating a causal relationship model based on the influence variables and the causal relationships between the users; and controlling instances of content based on the causal relationship model.

[0007] As a learning device for determining the logical cause of an event based on the causal salience concept, Patent Document 5 discloses a method for causal learning, comprising: observing one or more events using a device, wherein the events are defined as occurring at specific relative times; selecting a subset of the events based on one or more criteria; and determining the logical cause of at least one of the events based on the selected subset.

[0008] Non-Patent Document 1 states that unstructured data such as text, images, and audio are not fully incorporated into numerical models, and that approximately 80% of such data held by companies remains unused. [Prior art documents] [Patent documents]

[0009] [Patent Document 1] Patent No. 7631605 [Patent Document 2] Patent No. 6943436 [Patent Document 3] U.S. Patent Publication No. 10699203 [Patent Document 4] U.S. Patent Publication No. 10,949,753 [Patent Document 5] Japanese Patent Application Laid-Open No. 2016-539407 [Non-patent literature]

[0010] [Non-Patent Document 1] IDC, "Extracting insights from complex, unstructured bigdata," IBM Think Blog, 2020-11-19. (https: / / www.ibm.com / think / insights / managing-unstructured-data) Summary of the Invention [Problem to be solved by the invention]

[0011] The technologies disclosed in Patent Documents 2 to 4 treat the customer state as a static model when predicting the effect of intervention, which is defined as sales activities for customers, including advertising and notifications, and therefore are unable to properly grasp the customer state at the timing of intervention. Therefore, it has been essentially difficult to improve prediction of the intervention effect.

[0012] The method disclosed in Patent Document 5 basically relies on correlation analysis, and there is a risk of overestimating or underestimating the intervention effect. Without pseudo-intervention analysis, statistical reversal phenomena such as Simpson's paradox become apparent, making it impossible to accurately determine whether sending coupons contributes to increased sales.

[0013] An object of the present invention is to provide an information processing system that can improve the effectiveness of intervention. [Means for solving the problem]

[0014] In order to solve the above problems, the inventors have reviewed conventional techniques for enhancing intervention effectiveness and found that the following five limitations are bottlenecks in practical application.

[0015] i) Quantization error and discontinuity due to discrete categorization: Dividing a subject's state into coarse bins such as "low, medium, high" results in quantization error, which discards small continuous changes, and also creates artificial discontinuities at the category boundaries. As a result, it is impossible to capture the rate of change or gradient information, making it difficult to accurately predict state transitions (e.g., it is impossible to detect the signs of a sudden increase in interest).

[0016] ii) Insufficient quantification of qualitative data: Unstructured data such as text, images, and audio are not adequately incorporated into numerical models, and as described in Non-Patent Document 1, it has been reported that approximately 80% of such data held by companies is unused. As a result, sentiment scores from product reviews and the like cannot be used as explanatory variables for sales forecasts, and the accuracy of customer understanding reaches a plateau.

[0017] iii) Inadequate causal inference: Conventional methods relying on correlation analysis run the risk of overestimating or underestimating the effect of an intervention. Without pseudo-intervention analysis, statistical reversal phenomena such as Simpson's paradox become apparent, making it impossible to accurately determine whether sending coupons contributes to increased sales.

[0018] iv) Static intervention timing: When intervention times are fixed in advance on a calendar, it is difficult to implement real-time measures that match the peaks and valleys of the target person's state (e.g., sending emails once a week does not allow you to reach them at their emotional peak).

[0019] v) Lack of integration between modules: Data collection, analysis, and implementation are separated into separate systems (silos), making it impossible to operate a closed loop in real time. For example, because the CRM and ad distribution systems are separate, customer responses cannot be immediately fed back to measures, increasing maintenance and operational costs.

[0020] These limitations are due to the fact that human behavior is essentially a high-dimensional dynamics that is continuous in time and space and that combines qualitative and quantitative elements. Coarse Discrete Clustering Static rule base Overemphasis on correlation analysis This is due to the reliance on methods such as: (a) Missing the optimal timing of intervention (b) Imprecision in effect estimates (c) Increased system operation costs This results in insufficient improvement in outcomes such as customer experience, medical effectiveness, and learning outcomes.

[0021] As a result of intensive research to solve these problems, the inventors came up with the "space-time persona field" theory, which views the subject's state as a continuous vector field. By using this "space-time persona field" theory as its core, it is possible to provide an integrated intervention design technology centered on a persona field dynamic causal model that realizes quantitative conversion of qualitative data, causal inference, critical region detection, and module integration on the same platform.

[0022] In this specification, the "space-time persona field ψ(x, t)" refers to a vector field that maps the high-dimensional internal state of a subject (e.g., a client considering intervention) - psychological indicators, behavioral patterns, life events, environmental factors, etc. - onto a continuous domain of space and time.

[0023] where the input parameters x=(x1,x2,...,x d ) is a d-dimensional continuous space Ω(⊂R d ) constitutes a state coordinate (also referred to as the "evaluation axis" in this specification). Each component x i It expresses multiple evaluation criteria that quantitatively indicate the subject, such as purchasing willingness, disposable time, health consciousness, and sleep quality, as continuous values, enabling high-resolution understanding of state transitions while avoiding the quantization errors that occur in conventional discrete clustering.

[0024] Furthermore, the "persona field dynamic causal model" referred to in this specification is a dynamic causal inference model that integrates intervention causal inference based on a reaction-diffusion type partial differential equation (PDE) that governs the time evolution of the spatiotemporal persona field ψ(x,t) and Pearl's do-calculus into a single framework, and is also referred to as the "PF-DCM (Persona-Field Dynamic Causal Model)." PF-DCM estimates the intervention effect τ(x,t) on each evaluation axis online based on the continuously updated spatiotemporal persona field ψ(x,t), thereby optimizing the content and timing of measures.

[0025] The solution to the problem provided by this disclosure is outlined below. (1) High-resolution modeling (eliminating information loss due to discrete representation) Conventional discrete clustering methods reduce the subject's state to coarse divisions, resulting in the loss of subtle changes and intermediate transitions. As described above, by adopting a spatiotemporal persona field, which is a continuous vector field, the subject's state can be expressed as coordinates in "interest space" or "state space" (hereafter abbreviated as "spatial coordinates"). This allows the subject's state to be captured with high resolution, and from this perspective, the spatiotemporal persona field ψ(x,t) can be considered a digital twin of the subject's psychological state. This improves the accuracy of intervention design (measure content, delivery timing, etc.).

[0026] (2) Real-time prediction of causal effects (pre-quantification of dynamic outcomes) PF-DCM applies Pearl's do-calculus to online estimate the difference in outcomes between the presence and absence of a measure. This allows for a scientifically based evaluation of expected effects before implementing an intervention, eliminating the need to rely on empirical rules when selecting measures.

[0027] (3) Determining optimal timing (maximizing intervention effects based on critical phenomena) Applying critical phenomenon theory in statistical physics, we detect "critical points" where the spatiotemporal persona field undergoes abrupt state transitions. By intervening near critical points, we can maximize return on investment (ROI) while preventing excessive intervention.

[0028] (4) Utilizing qualitative data (integration of multimodal information) Unstructured data such as text, images, and audio are quantitatively transformed using large-scale language models (LLMs) and integrated into a spatiotemporal persona field, dramatically increasing the rate at which information is used and improving the accuracy of personalization.

[0029] (5) Dynamic adaptation (autonomous response to environmental changes) Specialized AI multi-agents work together to autonomously discover new data sources and trends and continuously update the PF-DCM, enabling intervention designs based on the latest situation at all times, enabling stable proposals for measures that enhance intervention effectiveness, and preventing obsolescence over the long term.

[0030] An example of a specific aspect of the present disclosure is as follows. [1] An information processing system comprising a processing device for data processing, an input / output device, and a recording medium which is a non-transitory computer-readable medium, wherein the processing device processes a persona snap ψ which is a time cross section of a space-time persona field ψ(x,t). t The information processing system is provided with a persona snap generation unit that executes processing including generating a spatiotemporal persona field ψ(x,t), in which the psychological state of a subject is expressed as a continuous vector field with state coordinates x and time t as parameters, and is described using partial differential equations.

[0031] [2] The information processing system described in [1] above, wherein the partial differential equation used to describe the space-time persona field ψ(x,t) is a reaction-diffusion partial differential equation.

[0032] [3] Persona Snap ψ t (x) is an information processing system described in [1] above, in which information on the time cross section of the space-time persona field ψ(x,t) is discretized into each of a plurality of cells formed by dividing the time cross section into a grid with a resolution Δx, and the discrete data includes data indicating the spatial differential operation result of the time cross section corresponding to the cell to which the discrete data belongs.

[0033] [4] The processing device includes a qualitative data conversion unit that performs processing including converting unstructured data including information about the subject input via the input / output device into an initial persona vector φ(x) including a numerical vector indicating the persona of the subject at a predetermined time, and a persona field update unit that performs processing including updating the spatiotemporal persona field ψ(x,t) using the initial persona vector φ(x), and the persona snap generation unit generates the persona snap ψ of the spatiotemporal persona field ψ(x,t) updated by the persona field update unit. t The information processing system according to any one of [1] to [3] above, which generates (x).

[0034] [5] The processing device generates data including a covariate Z, an intervention variable A, and an outcome variable Y based on information acquired via the input / output device, and the persona snap ψ generated by the persona snap generation unit. t (x) and (x) are input, and the information processing system described in any one of [1] to [3] further comprises a causal inference unit that performs processing including calculating the intervention effect τ defined by the following formula. τ=E1-E0 Here, E1 is the expected value when intervention is performed, and E0 is the expected value when no intervention is performed.

[0035] [6] The processing device is t The information processing system described in [5] above, further comprising a critical region detection unit that executes processing including detecting a region in a critical state as a critical region in (x), and further comprising an intervention candidate determination unit that executes processing including determining the critical region in which the intervention effect τ is positive as an intervention candidate.

[0036] [7] The critical region detection unit detects the person snap ψ t The information processing system according to [6] above, wherein the minimum eigenvalue is calculated from the Hessian matrix of (x), and the region where the minimum eigenvalue is within a predetermined numerical range is determined to be the critical region.

[0037] [8] The information processing system described in [7] above, wherein the processing device further includes a policy generation unit that executes processing including determining at least one of the intervention content, intervention medium, and intervention timing for the critical area determined to be an intervention candidate.

[0038] [9] The information processing system described in [4] above, wherein the process of updating the spatiotemporal persona field ψ(x,t) includes at least one of a weighted kernel integration and an extended Kalman filter assimilation process using the initial persona vector φ(x).

[0039]

[10] The information processing system described in [4] above, wherein the qualitative data conversion unit comprises a preprocessor, an LLM encoder, and a reduction block, wherein the preprocessor performs processing including converting the unstructured data, including data other than text, into a multimodal token sequence as input, the LLM encoder performs processing including encoding the multimodal token sequence as input to a large-scale language model into a high-dimensional embedding vector, and the reduction block performs processing including compressing the high-dimensional embedding vector using a projection matrix to obtain the initial persona vector φ(x).

[0040]

[11] The information processing system described in

[10] above, wherein the qualitative data conversion unit includes a quality evaluation unit that estimates the reliability of the high-dimensional embedding vector and masks components that do not satisfy a predetermined lower limit condition to generate a masked embedding vector, and the reduction block compresses the masked embedding vector instead of the high-dimensional embedding vector.

[0041]

[12] The information processing system described in [5] above, further comprising a triplet event creation unit that executes processing including evaluating the information acquired via the input / output device from the perspective of an explanatory variable describing the state before intervention to set the covariate Z, evaluating it from the perspective of whether or not intervention is performed to set the intervention variable A, and evaluating it from the perspective of whether or not an outcome is performed to set the outcome variable Y.

[0042]

[13] The information processing system described in [5] above, wherein the causal graph referred to by the causal inference unit when calculating the intervention effect τ has a causal path from the covariate Z to the intervention variable A (Z → A), a causal path from the covariate Z to the outcome variable Y (Z → Y), and a causal path from the intervention variable A to the outcome variable Y (A → Y).

[0043]

[14] The information processing system described in

[13] above, wherein the causal graph further has a causal path (ψ→A) from the space-time persona field ψ(x,t) to the intervening variable A.

[0044]

[15] The information processing system described in [5] above, wherein the causal inference unit estimates the intervention effect τ using a double correction estimation formula that combines a propensity score model and an outcome model.

[0045]

[16] The information processing system described in [8] above, wherein the processing device further includes an update learning unit that performs processing including updating parameters of the partial differential equation describing the space-time persona field ψ(x,t) using information including the response of the intervention measure implemented based on the decision of the measure generation unit as input.

[0046]

[17] The information processing system described in

[16] above, wherein the update learning unit executes processing including updating parameters used in the causal inference unit.

[0047]

[18] An information processing system described in any one of [1] to [3] above, wherein the processing device further includes a persona map generation unit that generates map data for displaying on the input / output device a persona map in which the norm ∥ψ(x,t)∥ or the gradient norm ∥∇ψ(x,t)∥ for at least a portion of the space-time persona field ψ(x,t) is plotted in a two- or more-dimensional coordinate space including the state coordinate x as a coordinate axis. [Effects of the Invention]

[0048] According to the present invention, an information processing system is provided that can improve the effectiveness of intervention by eliminating information loss due to discrete representation and realizing prior quantification of dynamic outcomes. [Brief explanation of the drawings]

[0049] [Figure 1] FIG. 1 is a diagram illustrating a specific example of a configuration of an information processing system according to an embodiment of the present disclosure. [Figure 2] FIG. 1 is a functional block diagram of an information processing system according to an embodiment of the present disclosure. [Figure 3] FIG. 10 is a functional block diagram illustrating an example of a specific operation of a qualitative data conversion unit. [Figure 4] FIG. 10 is a functional block diagram illustrating an example of a specific operation of a persona field storage unit. [Figure 5] 10 is a graph showing a specific example of the time progression of a one-dimensional space-time persona field ψ(x, t). [Figure 6] FIG. 1 is a contour map showing a specific example of the time progression of a two-dimensional space-time persona field ψ(x, y, t). [Figure 7] FIG. 7 is a diagram showing the gradient vector ∇ψ for the time progression of the two-dimensional space-time persona field ψ(x, y, t) shown in FIG. 6. [Figure 8] This is a map in which the gradient intensity is shown in gray scale on the diagram showing the gradient vector ∇ψ of FIG. 7 . [Figure 9]FIG. 7 is a wireframe representation of the time progression of the two-dimensional space-time persona field ψ(x, y, t) shown in FIG. 6. [Figure 10] 1 is a structural equation graph used in the information processing system according to the present embodiment. [Figure 11] FIG. 10 is a functional block diagram illustrating an example of a specific operation of a causal inference unit. [Figure 12A] 1 is a scatter plot showing the relationship between the latent expression φ1 and the true propensity score e. [Figure 12B] FIG. 12B is a diagram plotting the difference (residual) between the plotted values ​​of the propensity score e shown in FIG. 12A and the fitting curve. [Figure 13A] 1 is a graph showing the results of estimating the intervention effect using the online double correction estimation formula. [Figure 13B] 13B is a graph plotting the absolute value of the difference between the cumulative average estimated intervention effect τhatcum(t) and the true value of the intervention effect τ in FIG. 13A. [Figure 14] 1 is a histogram of DR estimates. [Figure 15] These are simulation results showing the relationship between the causal effect inference method for calculating the intervention effect τ and the estimation error (root mean square error, RMSE). [Figure 16] 10 is a flowchart illustrating processing executed in an intervention candidate detection unit. [Figure 17] FIG. 2 is a 2-column by 2-row block diagram showing the processing flow performed by the intervention candidate detection unit. [Figure 18] 18 is a diagram showing the output results of each functional block shown in FIG. 17. FIG. [Figure 19] 19 is a graph showing the distribution of each block shown in FIG. 18 in three dimensions. [Figure 20] 10 is a graph showing a simulation result of the dependency of the calculation time and the critical cell number on the resolution Δx. [Figure 21] FIG. 10 is a functional block diagram illustrating an example of a specific operation of a policy generation unit. [Figure 22]10 is a flowchart showing an example of processing using an AI agent executed in the information processing system according to the present embodiment. [Figure 23] 10 is a graph showing the lunar age dependency of the intensity of the space-time persona field ψ for each item according to the embodiment. [Figure 24] FIG. 10 is a diagram showing a persona map in which the intensity of the persona snap ψt(x) according to the embodiment is plotted on two-dimensional state coordinates. [Figure 25] FIG. 10 is a diagram showing a persona map in which the gradient norm ∥∇ψt(x)∥ of the persona snap ψt(x) according to the embodiment is plotted on two-dimensional state coordinates. DETAILED DESCRIPTION OF THE INVENTION

[0050] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. 1. Overall structure 1 is a diagram showing a specific example of the configuration of an information processing system according to an embodiment of the present disclosure. As shown in Fig. 1, the information processing system 100 according to an embodiment of the present disclosure includes a processing device 900 that performs data processing, an input / output device 910 that inputs data related to data to be processed by the processing device 900 and displays data including the processing results of the processing device 900, and a non-transitory computer-readable medium (also referred to as a "recording medium" in this specification) 920 that stores instructions to be executed by the processing device 900 and saves data to be processed by the processing device 900, data in the middle of processing, and / or data resulting from the processing.

[0051] Processing unit 900 may include one or more single-chip or multi-chip microprocessors designed and / or manufactured by Intel Corporation, Advanced Micro Devices, Inc. (AMD), Arm Holdings (Arm), Apple Computer, Inc., etc. Examples of microprocessors include Intel Corporation's CELERON®, PENTIUM®, COREi3, COREi5, COREi7 ("INTELCORE" is a registered trademark of Intel Corporation), AMD's AMDOPTERON®, PHENOM®, ATHLON®, TURION®, RYZEN®, and Arm's CORTEX-A, CORTEX-R, CORTEX-M ("CORTEX" is a registered trademark of Arm).

[0052] The processing unit 900 may include a processor (GPU, APU) suitable for machine learning. Examples of GPUs and APUs may include one or more single-chip or multi-chip graphics processing units (GPUs) designed and / or manufactured by NVIDIA Corporation, Advanced Micro Devices, Inc. (AMD), Intel Corporation, Apple Inc., etc. Examples of GPUs include NVIDIA Corporation's GEFORCE®, QUADRO®, TESLA®, and RTX series (e.g., GEFORCE® 3080, RTX 4090, etc.), AMD's RADEON®, RADEON PRO®, FIRE PRO®, and Instinct series (e.g., MI250X, etc.), Intel Corporation's ARC series (e.g., ARCA770, etc.), and Apple Inc.'s Apple GPUs (e.g., M1 GPU, M2 GPU, etc.).

[0053] 1, the processing device 900 has a configuration in which a processor is housed in a housing. In the information processing system 100, data communication is possible at least between the processing device 900 and an input / output device 910, and between the processing device 900 and a recording medium 920.

[0054] 1, information processing system 100 is configured by, but is not limited to, a plurality of devices that communicate with each other via a network NW such as the Internet. For example, at least two of processing device 900, input / output device 910, and recording medium 920 may be incorporated into a single device. Alternatively, at least one of processing device 900, input / output device 910, and recording medium 920 may be provided in a plurality of devices.

[0055] Fig. 2 is a functional block diagram of an information processing system according to an embodiment of the present disclosure. In the example shown in Fig. 2, a processing device 900, an input / output device 910, and a recording medium 920 are provided in a single information processing device 1000. Therefore, the information processing system 100 according to this embodiment is implemented in the information processing device 1000. In Fig. 2, the information processing device 1000, which is a specific example of a personal computer, is shown as a rectangle drawn with two dotted lines, and its components include the processing device 900 and recording medium 920 incorporated into a housing (shown as a rectangle drawn with solid lines) of the personal computer (information processing device 1000), and the input / output device 910 provided separately from the housing.

[0056] Each functional block will be explained below. The data collection unit 101 is used to acquire input data such as behavior logs, vital signs, and SNS posts of customers and other people who are candidates for intervention (hereinafter abbreviated as "subjects") and to perform preprocessing. Preprocessing includes masking personal identification information and creating a triplet stream for causal inference, which will be described later. <Z t ,A t ,Y t > This includes generating

[0057] The data that has undergone preprocessing in the data collection unit 101 may be output to a subsequent functional block (e.g., the qualitative data conversion unit 102), or may be temporarily stored in a recording medium 920 so that the subsequent functional block can read the data. In one specific example, the information processing system 100 according to this embodiment uses a messaging system for asynchronously transmitting event data to send and receive data between functional blocks. This messaging system may be, for example, a distributed stream processing platform (e.g., Apache Kafka). The source of input data for the data collection unit 101 is not limited. In FIG. 2, input data can be obtained from an external database 930 or a customer's information terminal 940 via a network NW.

[0058] The qualitative data conversion unit 102 uses a trained generative model such as a large-scale language model (hereinafter referred to as "LLM") to vectorize the unstructured data as persona elements, and generates an initial persona vector φ(x) that is used as the initial value of the spatiotemporal persona field ψ(x,t) or as base data for updating the parameters of the spatiotemporal persona field ψ(x,t). Here, the state coordinate (evaluation axis) x, specifically categories and context tags, becomes the spatial coordinate axis of the state space when the spatiotemporal persona field ψ(x,t) is displayed as a persona map.

[0059] The persona field storage unit 103 is a module that stores the spatiotemporal persona field ψ(x, t) that represents the state of the subject, in a form in which its time evolution is defined by a reaction-diffusion partial differential equation (hereinafter referred to as "reaction-diffusion PDE"). The persona field storage unit 103 stores the data generated by the qualitative data conversion unit 102. Initial persona vector φ(x) The data including the above is used as input, and the spatiotemporal persona field ψ(x, t) is successively updated by combining the reaction-diffusion PDE and the extended Kalman filter (EKF) assimilation process, and data indicating the latest spatiotemporal persona field ψ(x, t) is retained. Such data may be stored on the recording medium 920.

[0060] The persona field storage unit 103 also stores a persona snapshot ψ(x, t) that is generated by processing the information contained in the latest spatiotemporal persona field ψ(x, t) so that subsequent functional blocks (causal inference unit 104, intervention candidate detection unit 105) can use the information. t This data may be output from the persona field storage unit 103 to the subsequent function blocks (causal inference unit 104, intervention candidate detection unit 105), or may be temporarily stored in the recording medium 920 so that the subsequent function blocks can read it.

[0061] The persona field storage unit 103 updates and stores the spatiotemporal persona field ψ(x, t) in real time, and the persona snap ψ t The timing of generating the data indicating (x) is arbitrary. The process may be started when the qualitative data conversion unit 102 generates the initial persona vector φ(x), or it may be an online process in which the process is repeated at a separately determined cycle.

[0062] The causal inference unit 104 calculates the persona snapshot ψ generated by the persona field storage unit 103. t Data indicating (x) and the latest observation data via the data collection unit 101 (specifically, triplet stream<Zt,At,Yt> The data includes the following example.) is used as input, and the intervention effect τ(x,t) is estimated using PF-DCM. The estimated result is<userID,τ,timestamp> The data is generated as a keyed stream in the form of a key and transmitted to the intervention candidate detection unit 105. A recording medium 920 may be used for this transmission. The processing of the causal inference unit 104 is performed by using the persona snapshot ψ in the persona field storage unit 103. s This may be performed in synchronization with the generation of (x, t), or may be online processing in which the processing is repeated at a separately determined cycle.

[0063] The intervention candidate detection unit 105 extracts critical cells by free energy analysis, and performs a logical AND with cells where the intervention effect τ(x, t) is positive to obtain intervention candidate data.<userID,λmin,τ,t> is generated and registered in the intervention queue control unit 106.

[0064] The intervention queue control unit 106 receives the intervention candidate data generated by the intervention candidate detection unit 105 as input, manages the intervention candidate data using three functions: buffering, rate control, and duplication prevention, and pops the intervention candidate data to the policy generation unit 107 according to the transmission capacity. In other words, the intervention queue control unit 106 rectifies the implementation triggers of the subscription policies to ensure the processing stability of the subsequent functional blocks.

[0065] The measure generation unit 107 receives as input data including candidate intervention data and intervention timing, and determines an intervention measure including at least one of the intervention content, intervention medium, and intervention implementation schedule using at least one AI agent, generates intervention measure data including information about the intervention measure, and outputs the generated intervention measure data to the distribution and response measurement unit 108.

[0066] The distribution and response measurement unit 108 outputs instruction data including intervention measure data so that the intervention measure determined by the measure generation unit 107 can be distributed through at least one channel (specific examples of channels include app notification, email, API integration, and visits by sales representatives). In Fig. 2, an external information processing device 950 including an email transmission server and an information terminal 960 of a sales representative who implements the intervention measure are exemplified as targets that receive the instruction data from the distribution and response measurement unit 108.

[0067] Furthermore, the delivery and response measurement unit 108 collects real-time response data (opening, clicks, purchases, etc.) obtained by implementing the intervention measures, and outputs the data to the update learning unit 109 as a response log. A recording medium 920 may be involved in this process. In FIG. 2, this flow is indicated by a dashed line. In FIG. 2, this flow is indicated by a dashed arrow. The data collection unit 101 may be involved in the collection of response data. In other words, the data collection unit 101 and the delivery and response measurement unit 108 may be functionally integrated to form a data input / output unit.

[0068] The update learning unit 109 obtains embedded statistical data from the qualitative data conversion unit 102 and obtains a reaction log from the distribution and reaction measurement unit 108, and re-learns (by reinforcement learning) the parameters included in the reaction-diffusion PDE used when updating the spatiotemporal persona field ψ(x, t) in the persona field storage unit 103 and the parameters used in the causal inference unit 104, and updates these parameters. Data including the updated parameters is output from the update learning unit 109 to the persona field storage unit 103 and the causal inference unit 104. In Figure 2, this data flow is indicated by dotted arrows.

[0069] Specific examples of the response log obtained from the distribution and response measurement unit 108 include ROI (Return on Investment), click-through rate, and verification results of the intervention effect τ (for example, the average value of the relative error between the predicted value and the actual value).The update learning unit 109 uses the obtained embedded statistical data and response log to quantitatively evaluate the drift of the embedded statistical data and response log using, for example, the KS test or the population stability index (PSI).In addition, it optimizes the parameters used in the persona field memory unit 103 and the causal inference unit 104 using Bayesian optimization and stochastic gradient descent (SGD).

[0070] In this way, the update learning unit 109 constantly monitors the data, model, and behavior (intervention effect) by obtaining information on embedded statistics and the results of policy implementation, and continuously makes corrections to prevent degradation of system performance.

[0071] The logical blocks in Figure 2 can be implemented regardless of whether they are hardware or software. This information processing system 100 can be adapted to scalable architectures such as server-client, peer-to-peer, and cloud-native distributed architectures. Streams between blocks are connected asynchronously via message queues or event buses, and the processing unit 900, input / output device 910, and recording medium 920 can be distributed to any node (including the physical layout example in Figure 2).

[0072] Fig. 3 is a functional block diagram illustrating an example of a specific operation of the qualitative data conversion unit 102. As shown in Fig. 3, the qualitative data conversion unit 102 includes, as functional blocks, a preprocessor 111, an LLM encoder 112, a quality evaluation unit 113, a reduction block 114, and an initial persona generation unit 115.

[0073] In this example, since the input includes data other than text but the output is text, an LLM is used as the generated learning model. Specific examples of qualitative data (unstructured data) that can be input include text, images, audio, and application operation logs. The qualitative data conversion unit 102 uses the LLM to convert the qualitative data into a numerical vector r∈Rm in real time and incorporates it into the spatiotemporal persona field ψ(x,t), making it possible to utilize unstructured information that could not previously be used as an explanatory variable.

[0074] The preprocessor 111 included in the qualitative data conversion unit 102 normalizes the input data by, for example, BPE (BytePairEncoding) for text, by, for example, CLIP patching for images, and by, for example, converting audio into MEL tokens, to create a multimodal token sequence T. That is, in step S111 executed in the preprocessor 111, a normalization process is performed to convert unstructured data into the token sequence T.

[0075] The LLM encoder 112 included in the qualitative data conversion unit 102 encodes (maps) the token sequence T generated by the preprocessor 111 into a d-dimensional embedding vector E corresponding to the space of interest (step S112). The number of dimensions d is, for example, 4096. This converts discrete input data such as language, images, and audio into a continuous numeric vector space. That is, in step S112 executed by the LLM encoder 112, an embedding process is performed to convert discrete data into a continuous numeric vector.

[0076] Note that the embedding process may be performed using an external WebAPI in step S112 executed by the LLM encoder 112. However, when using an external WebAPI, it is preferable to take appropriate measures to protect privacy and comply with laws and regulations related to personal information protection, such as the General Data Protection Regulation (GDPR) and the California Consumer Privacy Act (CCPA). Specific examples of such measures to protect privacy are shown below.

[0077] Before sending the raw token sequence T to the external LLM, the preprocessor 111 hashes the personal identification attributes and adds differential privacy noise (ε dp ≒2). Communication with external WebAPIs is performed using encrypted signals, for example using TLS 1.3, and signals from WebAPIs are immediately stored in an on-premise environment rather than in a cloud environment. Only metadata will be stored long-term in the external WebAPI; the content sent to the external WebAPI will not be retained in the external WebAPI.

[0078] The quality evaluation unit 113 in the qualitative data conversion unit 102 receives the generation probability p1 output by the LLM encoder 112 and the consistency probability p2 output by the external fact-checking model for the high-dimensional embedding vector E generated by the LLM encoder 112, and calculates the geometric mean p conf Calculate =√(p1·p2). conf is low, in other words, a component that does not satisfy a predetermined lower limit condition (for example, when the threshold value is 0.8 and it is less than 0.8) is determined to have low reliability, and masking is performed to obtain a masked embedding vector E'. That is, step S113 executed by the quality evaluation unit 113 is not a process of estimating the embedding vector E itself, but rather evaluates the quality of the generated vector and removes unnecessary information.

[0079] The reduction block 114 included in the qualitative data conversion unit 102 reduces the dimension of the masked embedding vector E'. For example, a numerical vector r is obtained using a projection matrix W of m×d (m is sufficiently smaller than d (m<<d)). The projection matrix W is learned using principal component analysis, an autoencoder, or the like, and is set as a matrix with a spectral norm of 1 (∥W∥2=1), which is used as a normalized linear discriminator. That is, in step S114 executed in the reduction block 114, a compression process is performed on the masked embedding vector E'.

[0080] The sensitivity coefficient k in this compression process may be automatically adjusted in proportion to the Fisher information of the intervention effect τ, which is the causal estimation result obtained by the causal inference unit 104. Specifically, if the intervention effect τ is easily affected by the parameters, the sensitivity coefficient k is increased, and conversely, if it is not easily affected by the parameters, the sensitivity coefficient k is decreased. Note that, when the intervention effect τ is used as input in step S114, the latest intervention effect τ can be obtained from the causal inference unit 104 or the recording medium 920.

[0081] The initial persona generation unit 115 included in the qualitative data conversion unit 102 generates an initial persona vector φ(x) including the numerical vector r obtained by compressing the masked embedding vector E' in the reduction block 114. The initial persona vector φ(x) is a static vector indicating the subject's persona state at the generation time t0, and since the formula itself does not include the time parameter t, it has no time dependency and is used as a boundary value when updating the space-time vector field ψ(x,t). Furthermore, when a new space-time vector field ψ(x,t) is generated, it functions as its initial condition. That is, in step S115 executed in the initial persona generation unit 115, a process is performed in which the numerical vector r is input to generate the initial persona vector φ(x).

[0082] The initial persona vector φ(x) may include additional information other than the numerical vector r. One example of such additional information is tag data attached to the numerical vector r. The tag data includes the state coordinates (evaluation axes) x = (x1, x2, , x) of the spatiotemporal persona field ψ(x, t), such as category, emotion, and purchase intention. d ) may include a label (hereinafter referred to as an "evaluation axis tag") corresponding to the initial persona vector φ(x). When the initial persona vector φ(x) has an evaluation axis tag, using the tag facilitates clustering and searching of the initial persona vector φ(x), and may also improve the efficiency of processing that uses the initial persona vector φ(x) as input (e.g., time calculation processing such as differential calculation in the causal inference unit 104 and the intervention candidate detection unit 105).

[0083] This point will be explained in detail. In the information processing system 100 according to this embodiment, a spatiotemporal continuous spatiotemporal persona field ψ(x,t) is generated based on discrete information from a subject, and this is used to evaluate the intervention effect τ and the like. However, when performing processing to obtain information from such a continuum, handling the calculation processing as it is in the continuum is not easy with current calculation capabilities, as it would result in excessively long processing times. Therefore, as one of the countermeasures, the information processing system 100 according to this embodiment extracts information discretely from the continuum, and executes calculation processing using the obtained data group. Specifically, data that is a time cross-section of the spatiotemporal persona field ψ(x,t), which is a continuum (persona snapshot ψ, which will be described later) is used. t (x)) is prepared and used as the calculation target. In addition, although the time slice is a spatial continuum, in this disclosure, as an example, t (x) is treated as a collection of multiple cells discretized with a given spatial resolution Δx. t By including the coordinate system (division axis) information when treating (x) as a collection of multiple cells in the evaluation axis tag of the initial persona vector φ(x), the persona snap ψ t (x) can be treated as a group of data in which feature data (persona strength, gradient information, etc.) is set for each cell, making subsequent processing easier.

[0084] The initial persona generation unit 115 may output data including the generated initial persona vector φ(x) directly to the persona field memory unit 103, or may output it to the recording medium 920 so that other functional blocks such as the persona field memory unit 103 can read the initial persona vector φ(x) from the recording medium 920.

[0085] The initial persona generation unit 115 generates embedded statistical data, such as the time required for statistical calculations and the mean and variance of the embedding vector in relation to the embedding process performed in generating the initial persona vector φ(x), and outputs this as information for updating the parameters to the update learning unit 109. A recording medium 920 may be included in this data flow.

[0086] By carrying out such processing, the following effects are expected. i) Improved utilization of qualitative information: Unstructured data (text, images, audio, operation logs, etc.) can be incorporated into the spatiotemporal persona field ψ(x, t) as a vector r, making it possible to utilize a variety of information sources that could not previously be used as explanatory variables.

[0087] ii) Real-time reflection: A series of processes from tokenization (step S111), encoding (step S112), compression (step S114), and φ generation (step S115) can be completed in a short time. This allows subsequent measures (causal estimation, criticality judgment, and policy generation) to be operated on the latest spatiotemporal persona field ψ(x, t).

[0088] iii) Hallucination Suppression: In this embodiment, hallucinations in the LLM output are suppressed by a two-stage confidence evaluation and an error filter. First step (S113): The geometric mean p is calculated from the LLM probability p1 and the external fact model probability p2. conf Calculate =√(p1·p2) and p conf The vector components with low σ are masked to obtain the masked embedding vector E' (error filter). Second step (S114): When generating the initial persona vector φ(x) from the masked embedding vector E', the sensitivity coefficient k is adjusted in proportion to the Fisher information of the intervention effect τ. This automatically reduces the weighting of information with low reliability or small effect, minimizing the propagation of misinformation into the spatiotemporal persona field.

[0089] 2. Time-Space Persona Field Here, the spatiotemporal persona field ψ(x,t) will be explained in detail. (1) Concept definition In this specification, as described above, the "space-time persona field ψ(x, t)" is a vector field that maps the high-dimensional internal state of a target person such as a customer onto a continuous space-time domain, and is a transformation function that receives a d-dimensional real-valued vector and a non-negative real value as input, and outputs a k-dimensional real-valued vector that indicates probability, intensity, etc., as shown in the following equation.

number

[0090] In this way, by embedding the discrete cluster number (real-valued vector) in a continuous space, partial differential operations can be applied to the resulting spatiotemporal persona field ψ(x,t). As a result, it is possible to define the energy functional and perform its convergence analysis. It also becomes possible to directly handle multidimensional gradients and local peaks mathematically. Furthermore, it is possible to retain O(nlog(N / k)) bits of information that are lost as quantization error in conventional k-class classification, making it possible to capture state gradients and local precursors with high resolution.

[0091] The spatiotemporal persona field ψ(x,t) centrally manages the psychological state, behavioral tendencies, life events, environmental factors, etc. of each individual subject (customer, patient, learner, etc.) as a continuous vector field, thereby constructing a "state map" (digital twin) that changes smoothly over time. Therefore, for example, subjects' states, which were previously categorized into boxes such as "30s, female, hobby nails," can be replaced with a topographical map showing their internal state using the spatiotemporal persona field ψ(x,t), visualizing the subject's peaks (high interest), valleys (low interest), and flow lines (trend changes).

[0092] (2) Time evolution model The spatiotemporal persona field ψ(x,t) treats the subject's high-dimensional state as a "continuous field like a terrain." The mathematical rules that describe its evolution over time are partial differential equations (PDEs). By defining PDEs, it becomes possible to analyze, predict, and control "where, to what extent, and what external stimuli will cause the state to change."

[0093] The PDE of the space-time persona field ψ(x,t) is expressed as the following equation as an extension of the reaction-diffusion equation.

number

[0094] Here, the diffusion term D∇ 2 ψ represents the spatial spread of interests, knowledge, and behavioral patterns. Specifically, since the diffusion coefficient D is non-negative, it represents the natural diffusion of interest and motivation, such as the "bleeding" of states between similar users (interests average out through word of mouth). The larger the diffusion coefficient D, the greater the degree of diffusion, and it can represent, for example, a state in which interest is quickly lost.

[0095] The flow term -v ∇ψ expresses the situation where the state is "carried" in one direction like a tide (shifting as an overall trend due to seasonal or event effects). The flow velocity vector v can indicate seasonality and trends.

[0096] The response term f(ψ,u(t),x,t) expresses the response to external input u(t) such as advertising, notification, or treatment, and the stronger the stimulus from the external input u(t), the steeper the slope. Interventions such as advertising presentations and external measures are also a type of external input u(t), and the response to the intervention is also expressed by the response term.

[0097] The noise term η(x,t) is due to individual differences and measurement noise, and typically follows an Ornstein-Uhlenbeck process with a white Gaussian process as input.

[0098] The parameters included in these terms (such as the diffusion coefficient D and the flow velocity vector v) are parameters that can be updated in the update learning unit 109.

[0099] (3) Boundary / initial conditions The method for setting the boundary condition of the space-time persona field ψ(x,t) is not limited, and a Neumann condition that the rate of change in the normal direction of the space-time persona field ψ(x,t) is zero on the boundary surface may be imposed to realize reflective behavior with no inflow or outflow from the boundary. Alternatively, a Dirichlet condition that fixes the reference value ψref of the space-time persona field ψ(x,t) on the boundary surface may be imposed to set the variable field within the system to match the boundary control value.

[0100] The initial condition of the spatiotemporal persona field ψ(x,t), i.e., ψ(x,0), can be calculated using past data statistics, or hidden knowledge and processes related to the parameters of the spatiotemporal persona field ψ(x,t) can be input as a prior distribution.

[0101] (4) Details of the persona field memory unit 103 The persona field storage unit 103 is a module that stores the spatiotemporal persona field ψ(x, t), which represents the subject's latent state, in a form defined by the reaction-diffusion PDE described above. The persona field storage unit 103 stores the subject's latent state as a continuous vector field and provides a common platform that can be shared by all downstream processes, such as causal inference (causal inference unit 104), critical cell detection (intervention candidate detection unit 105), and visualization (persona map generation unit 110). In other words, the persona field storage unit 103 is a functional block that realizes a continuous field digital twin of the user's state.

[0102] Specifically, because it is a continuous vector field, it is possible to construct a demand map that maintains gradient and curvature. Furthermore, because the causal inference (causal inference unit 104) and critical cell detection (intervention candidate detection unit 105) refer to a common spatiotemporal persona field ψ(x,t), system consistency can be ensured. Furthermore, by immediately reflecting parameter update information for the reaction-diffusion PDE generated by online re-learning (update learning unit 109), it is possible to suppress drift.

[0103] Fig. 4 is a functional block diagram illustrating an example of a specific operation of the persona field storage unit 103. As shown in Fig. 4, the persona field storage unit 103 includes, as functional blocks, an initial persona acquisition unit 131, an update parameter acquisition unit 132, a persona field update unit 133, and a persona snap generation unit 134.

[0104] The initial persona acquisition unit 131 included in the persona field storage unit 103 acquires the initial persona vector φ(x) generated by the initial persona generation unit 115. The initial persona vector φ(x) may be acquired directly from the initial persona generation unit 115, or the latest initial persona vector φ(x) stored by the initial persona generation unit 115 on the recording medium 920 may be acquired. Therefore, the processing of step S131 performed by the initial persona acquisition unit 131 is to acquire the initial persona vector φ(x).

[0105] The update parameter acquisition unit 132 included in the persona field storage unit 103 acquires update data for the parameters generated by the update learning unit 119. It also acquires the latest values ​​of the parameters used when updating the spatiotemporal persona field ψ(x,t), which are stored in a recording medium 920 inside or outside the persona field storage unit 103. The update data generated by the update learning unit 119 may be acquired directly from the update learning unit 119, or the latest update data stored in the recording medium 920 by the update learning unit 119 may be acquired. Therefore, the process of step S132 performed by the update parameter acquisition unit 132 is to acquire the update parameters.

[0106] At the start of operation of the information processing system 100 according to this embodiment, no data to be updated exists, but in this case, the update parameter acquisition unit 132 may acquire data to be used as initial values. The data to be used as initial values ​​may be provided by the update learning unit 119 or may be stored in the recording medium 920.

[0107] The persona field update unit 133 sequentially updates the spatiotemporal persona field ψ(x, t) using the reaction-diffusion PDE updated with the parameters received from the update parameter acquisition unit 132. As an update method, either or both of the following (a) weighted kernel integration and (b) EKF (Extended Kalman Filter) assimilation may be applied.

[0108] (a) Weighted kernel integral For the initial persona vector φ(x'), the kernel integral of the following equation using a Gaussian kernel K(x,x';σ) is performed to recalculate the continuous field ψ(x,t), updating it with high resolution while preserving the gradient and curvature. ψ(x,t)=∫ΩK(x,x';σ)φ(x')dx' K(x,x';σ)=exp[-∥x-x'∥ 2 / (2σ 2 )] where: ψ(x,t): Space-time persona field at time t φ(x'): initial persona vector K(x,x';σ): Gaussian kernel (bandwidth:σ) Ω: integral domain in state coordinate space is.

[0109] (a) EKF (Extended Kalman Filter) assimilation Based on the updated reaction-diffusion PDE, a predicted value of the initial persona vector φ(x') at time t is generated, and this is compared with the observed value obtained by the weighted kernel integral above (a) and an EKF assimilation process (extended Kalman filter assimilation process) is performed to update the spatiotemporal persona field ψ(x,t).

[0110] The updated spatiotemporal persona field ψ(x,t) is stored in the recording medium 920, replacing the previous spatiotemporal persona field ψ(x,t). In addition, the latest estimation error covariance matrix Σ(d×d) obtained in the assimilation process is carried over using the hot-swap method at the time of the next update. Therefore, step S133 performed by the persona field update unit 133 is a sequential update process of the spatiotemporal persona field ψ(x,t).

[0111] The persona snap generation unit 134 extracts a time cross section at a predetermined time t0 from the latest spatiotemporal persona field ψ(x, t) updated by the persona field update unit 133, and generates a persona snap ψ, which is an instantaneous vector. t0 (x). Here, the initial persona vector φ(x) is used as the initial condition for generating and updating the space-time persona field ψ(x,t), and the persona snap ψ t It is not directly involved in the extraction of (x) itself.

[0112] The space-time persona field ψ(x,t) is a continuous vector field defined by a reaction-diffusion PDE, so the time slice of the persona field ψ t (x) is also essentially a continuum. However, since numerical processing of the continuum as is would impose a large computational load, in this embodiment, discretization is performed using the following procedure.

[0113] Step 1: Spatial discretization State coordinates (evaluation axes) x = (x1, x2, , xd ) along with a given resolution Δx=(Δx1,Δx2,...,Δx d ) and Persona Snap ψ t (x) is gridded and multiple cells C k (where k = 1 … N). If the state coordinate (evaluation axis) x is d-dimensional, the grid width (resolution) is also the same d-dimensional vector Δx = (Δx1, Δx2, …, Δx d ) for each dimension.

[0114] Step 2: Adding cell attributes Each cell C k The cell set {C k} is the data structure to be calculated.

[0115] With the above, Persona Snap ψ t This allows processing within a realistic calculation time while maintaining the high-resolution characteristics of (x).The number of dimensions d of the evaluation axis x that defines the cell coordinate system can be any positive integer.

[0116] Therefore, in one example of this embodiment, persona snap ψ t Each cell C of (x) k (position coordinate x k ), the cell identification information (index k or position coordinate x k ), and the persona snap ψ as the persona strength within each cell. t The norm of (x) (e.g., the L2 norm) ∥ψ t (x k )∥ is a data structure that holds the cell resolution Δx=(Δx1,Δx2,...,Δx d ) can be set appropriately for each evaluation axis. In this way, the grid width (resolution Δx) of the spatiotemporal persona field ψ(x,t) can be set arbitrarily, allowing for flexible adjustment of the trade-off between calculation accuracy and processing time, which is one of the factors supporting the high practicality of this information processing system.

[0117] This resolution Δx corresponds to the resolution of the target user's status, and its size is set by considering the balance between the computation load and the critical cell capture rate. For example, in a small-scale campaign (target users≦10 4 ) when the resolution Δx is set to 0.1 to 0.2, large-scale distribution (target users ≧ 10 7 ), it may be preferable to set the resolution Δx relatively large, to around 0.4 to 0.5. In reality, it is desirable to automatically switch the resolution Δx while monitoring the inference latency (the delay time from input to output) and memory usage.

[0118] As an example of operation, calculations are first performed with a coarse resolution Δx^(coarse) to grasp the general distribution of the spatiotemporal persona field ψ(x,t). Next, fine resolution Δx^(fine) (<Δx^(coarse)) is applied only to areas where detailed analysis is deemed necessary, and the desired information, such as inference results, is obtained. This two-stage mesh strategy makes it possible to maintain a critical cell capture rate while reducing the computational load.

[0119] Furthermore, Persona Snap ψ t (x) includes information on the spatial differential operation result (for example, gradient (first-order differential), Laplacian (second-order differential), etc.) for each cell. min In some cases, it may be easier to calculate.

[0120] By setting and managing the spatiotemporal persona field ψ(x,t) as a continuous vector field in this way, it is possible to represent the minute gradients and spatial repercussions of interests, motivations, and health risks without loss, and provide highly accurate and continuous input (persona snap ψ) for the subsequent causal effect estimation means and intervention timing decision means. t (x)). In addition, since the space-time persona field ψ(x,t), which is a continuum, is not limited by the grid width Δx or sampling interval Δt of the initial persona vector φ(x,t), which is the input at the time of setting, it is possible to analyze the space-time persona field ψ(x,t) at any resolution after the fact.

[0121] (5) Numerical implementation example The nonlinear reaction-diffusion partial differential equation mentioned above is integrated using the ADI (Alternating Direction Implicit) method or the Crank-Nicolson method, with a spatial mesh (grid width Δx) of 0.05 and a time step Δt of 1.0 s. The numerical stability conditions are as follows:

number

[0122] Figure 5 is a graph showing a specific example of the time progression of a one-dimensional spatiotemporal persona field ψ(x,t). In Figure 5, the horizontal axis represents the spatial coordinate system, i.e., the axis of the state coordinate x, and the vertical axis represents the strength (norm ∥ψ(x,t)∥) of the spatiotemporal persona field ψ(x,t). Note that in Figure 5, time is dimensionless because only the relative comparison at different times is important. Furthermore, by plotting time cross sections of the spatiotemporal persona field ψ(x,t) at different times on a single graph, it is possible to confirm the time progression.

[0123] In this specification, a persona map refers to a diagram in which the norm ∥ψ(x,t)∥ or the gradient norm ∥∇ψ(x,t)∥ for at least a portion of the spatiotemporal persona field ψ(x,t) is plotted in a two- or more-dimensional coordinate space that includes the state coordinate x as a coordinate axis. The processing device 900 of the information processing system 100 according to this embodiment further includes a persona map generation unit 110 that generates map data for displaying the persona map on the input / output device 910. The persona map generation unit 110 may input data indicating at least a portion of the spatiotemporal persona field ψ(x,t) and calculate the norm ∥ψ(x,t)∥ or the gradient norm ∥∇ψ(x,t)∥ of the spatiotemporal persona field ψ(x,t) under predetermined conditions, or may input a collection of data indicating coordinate information (position vectors) to be plotted.

[0124] The input / output device 910 of the information processing system 100 according to this embodiment displays a predetermined persona map based on the map data generated by the map generation unit. In Fig. 2, the data flow related to the persona map generation unit 110 is indicated by a dashed arrow. In Fig. 2, the persona snapshot ψ generated by the persona field storage unit 103 is t For (x), the norm ∥ψ t (x)∥ and gradient norm∥∇ψ t The information of (x)∥ is obtained and the map data of the persona map is generated. t Since (x) is the time cross section of the space-time persona field ψ(x,t), the norm ∥ψ t (x)∥ is included in the concept of norm ∥ψ(x,t)∥, and the gradient norm ∥∇ψ t (x)∥ is included in the concept of the gradient norm∥∇ψ(x,t)∥. Figures 5, 6 and 9 show the norm∥ψ t (x)∥, and Figures 7 and 8 show the gradient norm∥∇ψ t This is a map of (x)∥.

[0125] Furthermore, the persona map generator 110 may generate map data of a map other than the norm ∥ψ(x,t)∥ or the gradient norm ∥∇ψ(x,t)∥. Examples of such maps include a map of the intervention effect τ generated by the causal inference unit 104 and a map of the minimum eigenvalue λ detected by the intervention candidate detector 105. min 18 and 19, the map of the minimum eigenvalue λ, the map of the critical region, the map of the intervention effect τ, and the map of the intervention candidate (the critical region where the intervention effect τ is positive) may be used. The dashed arrows from the causal inference unit 104 and the intervention candidate detection unit 105 to the persona map generation unit 110 indicate the data flow related to these maps. min A map of the critical region, a map of intervention effects, and a map of potential interventions are shown.

[0126] Let us now consider Figure 5 in more detail. At the beginning (time t = 0), as shown by the black dotted line in Figure 5, the intensity distribution of the one-dimensional Persona field ψ was a normal distribution (σ = 1.5) centered at x = 5. At time t = 3, as shown by the long-dashed gray line in Figure 5, the intensity distribution of the one-dimensional Persona field ψ is a normal distribution (σ = 1.5) centered at x = 5. 2 Due to the influence of ψ and the noise term η(x,t), natural diffusion occurs, flattening the peak at x=5 and expanding σ to 2.2.

[0127] Subsequently, at time t = 6, as an intervention, a temporary boost (forced input u(t)) was applied to the target layer for x > 7, which resulted in a step-like amplification on the right side of the intensity distribution of the one-dimensional persona field ψ, specifically bending upward at x = 7, as shown by the gray dashed line in Figure 5. Further time passed, at time t = 10, as shown by the gray solid line in Figure 5, and the combined effects of the diffusion of the initial distribution and the intervention converged, with the peak position of the intensity distribution of the one-dimensional persona field ψ, which was at x = 5 before the intervention, shifting to x = 6.5, and the overall shape of the distribution also becoming gradually smoother.

[0128] The above simulations can be summarized as follows: i) The spatiotemporal persona field ψ(x,t) is a simple approximation of a one-dimensional reaction-diffusion equation, and we were able to confirm the diffusion process in which the bandwidth expands as the variance σ increases over time. ii) When a forced input u(t) was applied to a limited region (x>7) at time t=6, the peak height increased by up to 30%. The effect of the intervention, which can boost the ROI in a short period of time, was visualized. iii) After the intervention, the diffusion continued, and the persona mass (the center of the intensity distribution of the one-dimensional persona field ψ) shifted to the right. This indicates that the "behavioral intention" shifted in the direction of the intervention. iv) Even with a single intervention, the peak position of the intensity distribution of the one-dimensional Persona field ψ was able to move by 1.5 units on the state coordinate x axis. When implementing interventions continuously, it was suggested that it would be effective to set the intervention around t = 10, when the effect of the first intervention converges.

[0129] Figure 6 is a contour plot showing the time progression of the spatiotemporal persona field ψ(x,y,t) with a two-dimensional persona element axis. Initially (time t = 0), as shown in the upper left diagram of Figure 6, the intensity distribution of the two-dimensional persona field ψ(x,y) was a single peak centered at (x,y) = (5,5) and consisting of Gaussian peaks with σx of 3 and σy of 2.5. Figure 6 makes it possible to confirm the time progression by displaying multiple contour plots obtained by individually plotting time cross sections of the two-dimensional spatiotemporal persona field ψ(x,y,t) at different times in a comparative manner.

[0130] After a certain amount of time (t=0.6), as shown in the upper right diagram of Fig. 6, the intensity distribution of the two-dimensional Persona field ψ(x,y) diffuses in both the x and y directions due to the influence of natural diffusion, and has a spread of more than 10 in the x direction and about 8 in the y direction. In addition, a tendency for the gradient near the center of the peak to decrease was confirmed.

[0131] When intervention was performed at time t = 1.0, a steep rise in the intensity of the 2D persona field ψ(x,y) due to the intervention occurred in the hatched rectangular region in the upper right ((x,y) = (7-10, 7-10)), as shown in the lower left diagram of Figure 6. Specifically, this region is the spatial section where the intervention measure u(t), the control input, was applied, and visually represents the set of coordinates where the response term f(ψ,u,x,t) is nonzero. Note that the degree of diffusion of the peak before intervention was equivalent to that at time t = 0.6, and this peak was essentially in equilibrium.

[0132] At time t=1.6, a predetermined amount of time after the intervention, the high intensity region formed by the intervention measure u(t) became a distribution with a peak centered at (x,y)=(8,8) due to diffusion, and this distribution was connected to the distribution with a peak centered at (x,y)=(5,5) that had existed from the beginning, and as a result, the intensity of the two-dimensional persona field ψ became a distribution with a bimodal structure, with the peak that had been near the center and a newly formed peak in the upper right corner.

[0133] Figure 7 shows the gradient vector ∇ψ for the time progression of the two-dimensional space-time persona field ψ(x, y, t) shown in Figure 6. Note that the gradient vector ∇ψ is plotted with 30 plots displayed on each axis, and the direction is set to "low to high." This allows you to intuitively understand where the latent desires of customers and other target individuals are heading from the direction of the arrow.

[0134] Initially (t=0), the arrows indicating the vectors converged to a central peak (an area where vectors with small norms centered around (x,y)=(5,5) gather). This initial peak can be seen as the focus of interest for customers and other subjects. At t=0.6, in comparison to the initial state (t=0), the arrows have spread out overall and the norm has become smaller. This indicates that the interest of the subjects is fading due to natural diffusion.

[0135] Immediately after the intervention (t=1.0), a steep inflow vector appeared in the upper right region along the periphery of the rectangular area based on the intervention measure u(t), confirming that the intervention measure u(t) was inducing all subjects simultaneously. By integrating the norm of the inflow vectors generated by this intervention measure u(t), it is possible to quantitatively evaluate the degree to which the intervention measure u(t) has induced subjects.

[0136] At the final stage (t = 1.6), a new peak, a region where vectors with small norms converged, was formed within the rectangular region. With the formation of this new peak, the center of the initial peak, which was centered at (x, y) = (5, 5), was confirmed to have been stretched to the upper right. This shift in the center of the two-dimensional persona field ψ(x, y) suggests that the point maximizing the ROI has been updated by the intervention measure u(t).

[0137] Figure 8 is a map in which the gradient intensity (norm ∥∇ψ∥ of the gradient vector ∇ψ) is shown in grayscale and overlaid on the diagram showing the gradient vector ∇ψ in Figure 7. In Figure 8, the length of the arrow showing the gradient vector ∇ψ is constant (not corresponding to the gradient intensity), making it easier to understand the direction of the gradient vector ∇ψ. Note that the grayscale range is optimized independently for each of the four individual diagrams, so the color tones can only be compared within each individual diagram.

[0138] At the initial stage (t=0), there is an apex consisting of a region of low gradient strength at the tip of the central peak centered at (x,y)=(5,5), and around that there is a region of high gradient strength, i.e., a high gradient degree, and further around that there is a region of low gradient strength, making it possible to visually confirm that the central peak is a Gaussian peak. At t=0.6, in comparison with the initial stage (t=0), the apex (region of low gradient strength) at the central peak has spread, making it easy to visually confirm the attenuation of interest due to natural diffusion.

[0139] Immediately after the intervention (t=1.0), the gradient strength of the rectangular area in the upper right corner based on the intervention measure u(t) is high, so the gradient strength of the other areas is displayed as the lowest level (black) on the grayscale. In addition, there is a part on the periphery of the rectangular area that is displayed as the highest level (white) on the grayscale, making it easy to visually confirm that the gradient strength is particularly high in this part.

[0140] At the final stage (t = 1.6), a gradient distribution has developed within the rectangular region, making it easy to visually confirm the formation of a new peak with a peak consisting of a region of low gradient intensity. Furthermore, because the maximum gradient intensity within the rectangular region has decreased, the gradient intensity distribution in the region outside the rectangular region, including the central peak, can be visually confirmed. Specifically, the shape of the dark region representing the peak of the central peak is now deviated from the elliptical shape of the original peak. This indicates that with the formation of the new peak in the rectangular region, the center of the initial peak, which was centered at (x, y) = (5, 5), has shifted to the upper right where the new peak is located. This figure also confirms that attention continues to be drawn from the surrounding areas (left and bottom) toward the new peak in the upper right.

[0141] FIG. 9 is a wireframe diagram showing the time progression of the two-dimensional space-time persona field ψ(x, y, t) shown in FIG.

[0142] At the initial stage (t=0), a single-peak structure is drawn, showing a central peak (a peak centered at (x,y)=(5,5)). As mentioned above, this indicates that the subject's attention is focused on one point.

[0143] At t=0.6, natural diffusion is visualized as the top of the central peak lowering and broadening, and the peripheral region of the central peak also broadening.

[0144] Immediately after the intervention (t=1.0), a new columnar peak appears in the upper right rectangular area based on the intervention.

[0145] In the final stage (t=1.6), the bimodal structure of a central peak and an upper right peak stabilized, demonstrating three-dimensionally that the intervention effect had been spatially established as a shift in the persona's center of gravity.

[0146] 3. Estimation of causal effects Next, the process of estimating the causal effect performed by the causal inference unit 104 will be described.

[0147] Conventional techniques, such as those disclosed in Patent Document 5, rely on correlation analysis to evaluate intervention effects, making it impossible to immediately quantify the true causal effects of intervention measures. This has led to concerns that incorrect measures may continue. The information processing system according to this embodiment uses a persona field dynamic causal model (PF-DCM) incorporating Pearl's do-calculus to estimate the intervention effect τ online via the spatiotemporal persona field ψ(x,t), which is a continuous vector field, thereby significantly reducing reliance on empirical rules.

[0148] In the information processing system according to this embodiment, correlation and causation are separated to calculate "how much the outcome Y changes when intervention A is performed," and the difference in outcome Y between when intervention A is performed and when it is not performed is calculated as the intervention effect τ. Quantifying the pure effect of the intervention measure makes it possible to avoid unnecessary intervention and maximize ROI. In a preferred example of this embodiment, the calculation of the intervention effect τ is performed using persona snapshot ψ t This is done in real time in accordance with the input timing of (x).

[0149] Specifically, in the information processing system according to this embodiment, the pure causal effect of the policy effect is estimated, and the obtained intervention effect τ is input to the intervention candidate detection unit 105. This makes it possible to correct noise and confounding, ensuring the reliability of the ROI simulation. It is also possible to receive model parameters γ (regularization, particle number, etc.) from the update learning unit 109 and optimize them online.

[0150] (1) Theoretical structure The structural equation model is as follows:

number

[0151] where Z is a covariate, A is an intervention variable, Y is an outcome, and U * is exogenous noise. Covariate Z is an intermediate variable that uses the spatiotemporal persona field ψ(x,t) and exogenous noise as input, and affects the intervention variable A and outcome Y. Exogenous noise Uz is involved in determining covariate Z but is an exogenous factor that is not directly observed within the model. Exogenous noise Ua is randomness or unobservable background factors involved in determining the intervention. Exogenous noise Uy is a random element or unobserved factor that affects the outcome but is unrelated to covariate Z or intervention variable A.

[0152] We define the probability distribution after the intervention as follows:

number

[0153] The intervention effect τ(x,t) is obtained by subtracting the expected value when no intervention is performed (do(A=0)) (expected value when no intervention is performed E0) from the expected value when intervention is performed (do(A=1)) in the space-time persona field ψ(x,t), as shown in the formula below.

number

[0154] (2) Estimation algorithm framework The inference of the causal effect executed by the causal inference unit 104 successively approximates the calculation formula for the intervention effect τ, and therefore, for example, the following three-stage estimator can be used.

[0155] i) Representation learning: The spatiotemporal persona field ψ(x,t) and the covariates Z are integrated to extract and compress the "intrinsic features and structures" that cannot be directly observed from the observed data, and a latent representation φ is obtained. tIn this way, by compressing the high-dimensional elements of the spatiotemporal persona field ψ(x,t), which is a continuous vector field, into low dimensions, the computational load can be reduced. In the outcome model that estimates the expected value of each outcome when intervention is performed (do(A=1)) and when no intervention is performed (do(A=0)), this latent representation φ t is used.

[0156] ii) Propensity score estimation: latent expression φ at time t t The propensity score e, which is the probability of choosing the intervention, conditional on t =Pr[A=1|φ t ] is estimated sequentially.

[0157] iii) Doubly Robust Estimation (also referred to as "DR Estimation" in this specification): The estimated value of the intervention effect at time t (estimated intervention effect τ hat t) is the propensity score e t The doubly robust estimator is calculated by combining the propensity score model and outcome model using the propensity score e t The estimated intervention effect τ is robust against both model misspecification and estimation bias because the estimates are consistent even if either the intervention effect or the outcome model is correct. hat When calculating t, a Bayesian hierarchical structure may be adopted, which introduces a prior distribution into the hyperparameters to perform overall control.

[0158] Specific examples of the estimator include a regression algorithm and a likelihood estimation method. The estimator can be selected appropriately depending on the data characteristics. For example, t can be trained with an autoencoder, and the propensity score e t may be estimated by stepwise logistic regression.

[0159] In the above explanation, the intervention effect τ is estimated using Doubly Robust (DR) estimation, but this is not limited to this. For example, propensity score reweighting (IPW, IPTW), propensity score et By designing a general-purpose estimation framework that can combine techniques such as neighborhood matching based on the method of causal correlation (CQC), instrumental variable (IV) techniques suitable for causal estimation under confounding conditions, causal forests that are flexible to nonlinearities and structural changes, and meta-learner sequences (T-Learner, X-Learner, etc.), and by applying a multi-armed bandit to select an estimator, it is possible to select the best model online using estimation error (least squared error) or log-likelihood as metrics. Note that estimation methods that directly consider the difference in observed mean outcomes between the intervention and non-intervention groups as the causal effect (naive difference) have low predictive accuracy for the intervention effect τ because they do not take confounding relationships into account.

[0160] (3) Data flow The causal inference unit 104 compares the latest spatiotemporal persona field ψ(x, t) acquired from the persona field storage unit 103 with the triplet stream <covariate vector Z t ,Intervention variable A t ,Outcome variable Y t > and are input, and the intervention effect τ(x,t) is sequentially estimated. Specifically, the estimated intervention effect τ hat (x,t) is a keyed stream < user ID,τ hat (x, t), timestamp> is sent to the intervention candidate detection unit 105. This forms a real-time pipeline that combines the continuous field (space-time persona field ψ(x, t)) and the intervention effect τ(x, t).

[0161] (4)Mathematical soundness To ensure the mathematical soundness of the intervention effect τ(x,t), the following measures may be taken:

[0162] i) Propensity score boundary: To reduce the estimation error, the propensity score e t In the range where is too small, there is a high possibility that problems such as large estimation errors will occur in weighted estimation using the propensity score value. t is clipped to [ε,1-ε] (0<ε<0.5) to obtain the propensity score et should not be less than ε or exceed 1-ε.

[0163] ii) Asymptotic convergence: latent representation φ t The number of dimensions (latent dimensions) of t ) to O(logN) order, and adopt online adaptive learning rate to apply learning rate η t η0 t -0.5 By suppressing the amount of weight updates as the learning progresses, convergence stability is improved.

[0164] iii) Delay Compensation: When there is a delay in the input stream (space-time persona field ψ(x,t)), a consistent time series is maintained by buffering.

[0165] (5) Diversity of implementation forms The processing performed by the causal inference unit 104 can be applied to batch processing, sequential processing, or micro-batch processing, and can operate on a single computing device or a distributed network. The specific algorithms, numerical optimization techniques, and latent representation learning methods used in the causal inference unit 104 can be changed as desired without departing from the spirit of this disclosure.

[0166] (6) Effects By executing the causal inference unit 104, the following effects can be obtained.

[0167] i) Early detection of erroneous measures: Because the effects of intervention are evaluated in real time, if the intervention effect τ(x,t)<0 is continuously detected for a certain user group, the measures for that user group can be immediately stopped, thereby minimizing the negative effects of the intervention.

[0168] ii) ROI optimization: For multiple interventions under consideration, the intervention effect τ(x,t) / intervention cost can be calculated before implementation to obtain the cost-effectiveness ratio (CER), and interventions can be prioritized based on the cost-effectiveness ratio (CER). This makes it possible to concentrate management resources on customers for whom the intervention will be most effective, which is expected to improve ROI.

[0169] iii) Critical point collaboration: Two-stage judgment with the intervention candidate detection unit 105 (minimum eigenvalue λmin≦threshold value ε c Furthermore, the intervention effect τ(x,t)>0) realizes the timing of intervention that achieves maximum effect with minimum stimulation.

[0170] As described above, the real-time estimation of causal effects executed by the causal inference unit 104 combines the continuous vector field model and do-calculus to achieve estimation of causal effects on the order of seconds, which was difficult with conventional technology, and thereby makes it possible to eliminate excessive, wasteful, and counterproductive interventions.

[0171] (7) Structural equation graph 10 is a structural equation graph used in the information processing system according to this embodiment. The causal graph referred to by the causal inference unit 104 when calculating the intervention effect τ has, as essential causal paths, a causal path (Z→A) from a covariate Z to an intervention variable A, a causal path (Z→Y) from the covariate Z to an outcome variable Y, and a causal path (A→Y) from the intervention variable A to the outcome variable Y, and further has, in this embodiment, a causal path (ψ→A) from the spatiotemporal persona field ψ(x, t) to the intervention variable A.

[0172] The central continuous vector field (space-time persona field ψ(x,t)) forms the ground state, and arrows extend from the space-time persona field ψ(x,t) to the covariate Z, intervention variable A, and outcome Y, which are made up of observed values, indicating that the space-time persona field ψ(x,t) is a latent confounder that determines the background state of all variables.

[0173] The path from intervention variable A to outcome Y (A → Y) is the main causal path that is the subject of estimating intervention effect τ. The path from the space-time persona field ψ(x,t) to intervention variable A (ψ(x,t) → A) is the main causal path when determining policy selection using the real-time potential demand estimation algorithm, and the influence of this path is emphasized when estimating intervention effect τ. For this reason, in Figure 10, this path is highlighted with a thick solid arrow.

[0174] The path (ψ(x,t) → Z) from the spatiotemporal persona field ψ(x,t) to the covariate Z indicates that the latent persona field (spatiotemporal persona field ψ(x,t)) influences the generation of the covariate Z (a specific example would be a behavioral log). By making the latent confounder explicit, the backdoor path ψ(x,t) ← Z → A is clearly expressed. A specific example of the causal path Z → A that constitutes a backdoor path is when the implementation of an intervention measure such as direct mail delivery is decided based on purchase history. This Z → A path represents an allocation bias in which intervention A (such as advertising delivery) is assigned based on Z.

[0175] Furthermore, the causal path (Z → A → Y) from covariate Z through intervention variable A to outcome Y shows the causal structure in which intervention variable A is determined based on covariate Z, which in turn produces outcome Y. This type of causal path corresponds to the framework for policy decision-making and effectiveness measurement in the real world.

[0176] By making such unobserved common causes apparent in a causal graph, the backdoor path (ψ(x,t)←Z→A→Y) that is the source of bias is visualized. When estimating the causal discrimination formula do(A), by conditioning on the covariate Z and closing the causal paths Z→A and Z→ψ, the backdoor path is closed and the bias in the intervention effect τ is eliminated.

[0177] When covariate Z affects both intervention variable A and outcome Y (A ← Z → Y), this causal pathway also becomes a backdoor path and causes confounding. A specific example of a causal pathway represented by Z → Y is when purchase history directly influences purchases. Therefore, to accurately estimate the causal effect, adjustments such as conditioning on Z and closing the Z → Y pathway are necessary. Note that when designing a causal graph, if it can be clearly determined that Z does not have a direct effect on Y, it is not necessary to include an arrow from Z to Y.

[0178] The path (ψ(x,t) → Y) from the spatiotemporal persona field ψ(x,t) to outcome Y indicates the latent confounding factors that directly affect the latent persona field (spatiotemporal persona field ψ(x,t)) on outcome Y, such as purchases or health indicators, and is a typical expression of an unobserved confounding factor. Because the influence of this path is relatively minor, it is shown with a dotted line in Figure 10.

[0179] Uz, Ua, and Uy, located in the upper part of Figure 10, represent exogenous noise (independent and identically distributed error terms) for the covariate Z, the intervention variable A, and the outcome Y, respectively. In Figure 10, the influence of such exogenous noise is shown by dotted lines, similar to the path ψ(x,t) → Y, to distinguish it from random effects in the structural equation model.

[0180] The dashed arrow represents Pearl's operational intervention (forced intervention) do(A=a), indicating that the intervention is forcibly set from outside the structural equation. By visualizing the forced intervention, when estimating the intervention effect τ(x,t), the causal edge (causal arrow) between the intervention variable A and the parent node (covariate Z, etc.) is cut, making it possible to estimate the pure intervention effect τ(x,t) of the intervention variable A on the outcome Y.

[0181] (8) Details of the causal inference unit 104 Fig. 11 is a functional block diagram illustrating an example of a specific operation of the causal inference unit 104. As shown in Fig. 11, the causal inference unit 104 includes, as functional blocks, a triplet event acquisition unit 141, a persona snap acquisition unit 142, an update parameter acquisition unit 143, a PF-DCM inference unit 144, and an intervention effect map generation unit 145.

[0182] The triple event acquisition unit 141 included in the causal inference unit 104 acquires the triple stream generated by the data collection unit 101. <Z t ,A t ,Y t The data may be acquired from the data collection unit 101 or from the recording medium 920. In this embodiment, the data collection unit 101 acquires data including the triplet stream <Z t ,A t,Y t The triplet event creation unit performs processing including evaluating the information acquired via the input / output device 910 from the perspective of an explanatory variable describing the state before the intervention to set the covariate Z, evaluating it from the perspective of whether or not there is an intervention to set the intervention variable A, and evaluating it from the perspective of whether or not there is an outcome (for example, whether a purchase is made / not made) to set the outcome variable Y.

[0183] Triad Stream <Z t ,A t ,Y t > the covariate vector Z t are a group of explanatory variables that describe the state before the intervention is implemented, and have numerical information that is a summary statistic of the characteristics of the subjects and context, and provide the basic information for the causal inference model to perform a "homogeneous comparison." As mentioned above, the data collection unit 101 collects the triplet stream <Z t ,A t ,Y t >Perform the process to generate it.

[0184] Intervention variable A t In one example, the outcome variable Y is binary data indicating whether or not an intervention was performed at time t. t is the outcome indicator observed at time t after the intervention. In one example, it is a variable directly linked to a business KPI such as a purchase or click. This variable may also be expressed as binary data (purchase completed: 1 / not purchased: 0).

[0185] Covariate vector Z t Subjects with high similarity in the above are set as specific contrast targets, and for these, the intervention variable A t indicates intervention and intervention variable A t The outcome variable Y in the intervention group was calculated by dividing the intervention group into two groups: t The mean value of the outcome variable Y in the reference group tBased on this result, the PF-DCM inference unit 144 (described later) performs counterfactual inference, asking, "If the intervention had not been implemented, would the subject not have responded as expected (for example, made a purchase)?", and calculates a quantitative intervention effect τ.

[0186] The persona snap acquisition unit 142 acquires the persona snap ψ generated by the persona snap generation unit 134. t Data including (x) is acquired (step S142). The data may be acquired from the persona snap generation unit 134 or from the recording medium 920.

[0187] The update parameter acquisition unit 143 acquires update data of parameters used for causal inference performed in the PF-DCM inference unit 144 from the update learning unit 109 (step S143). Examples of parameters to be updated include the number of particles, parameters related to regularization strength such as the regularization coefficient λ, and prior distribution in Haze estimation. Note that the update learning unit 109 performs re-learning by reinforcement learning using the effect of intervention as a reward based on the reaction log obtained from the distribution / reaction measurement unit 108.

[0188] The PF-DCM inference unit 144 calculates the triplet stream <Z t ,A t ,Y t > and Persona Snap ψ tUsing (x) as input, the backdoor path is adjusted (i.e., the covariate Z is adjusted), and state estimation is performed using an extended Kalman filter (EKF) or a sequential particle filter. The intervention effect τ is calculated from the difference in expected value between intervention (do(A=1)) and no intervention (do(A=0)) (step S144). Calculating the intervention effect τ requires forward integration along the time axis toward the future to predict the outcome when intervention occurs, and backward integration along the time axis toward the past to construct a scenario in which no intervention occurs. This time calculation process can be calculated using the time evolution of a reaction-diffusion equation describing the spatiotemporal persona field ψ(x,t). When calculating the intervention effect τ, confidence intervals (e.g., 95% confidence intervals) are also calculated.

[0189] Specifically, the PF-DCM inference unit 144 calculates the persona snap ψ t The intervention effect τ is calculated in cell units set based on the resolution Δx of (x). As described above, in this embodiment, the persona snap ψ t Since (x) is composed of a collection of discrete data, the intervention effect τ obtained by the processing of the PF-DCM inference unit 144 is also a collection of multiple calculation results. However, the present disclosure is not limited to this, and the intervention effect τ may be set as an equation describing a continuum. This method is arbitrary, and for example, the continuum may be obtained from a collection of multiple data that constitute the intervention effect τ.

[0190] The intervention effect map generation unit 145 outputs data including the intervention effect τ calculated on a cell-by-cell basis and information on the confidence interval to the intervention candidate detection unit 105 (step S145). A recording medium 920 may be interposed in the data flow from the intervention effect map generation unit 145 to the intervention candidate detection unit 105. Note that, from the intervention effect τ calculated on a cell-by-cell basis, it is possible to obtain the intensity distribution (intervention effect map) of the intervention effect τ in the coordinate system indicated by x.

[0191] (9) Numerical implementation example Figure 12A is a scatter plot showing the results of a simulation of the relationship between the latent expression φ1 and the true propensity score e. In Figure 12A, the gray band-like area indicates the 95% confidence interval (confidence band), and the gray square marks indicate the local mean (the average value of each bin). Note that in Figure 12A, the standard error SE is shown as a bar above and below each mark, but because the number of samples is sufficiently large, the standard error SE is extremely small and overlaps with the marker, making it so short that it is not visible. The solid black line is the curve approximation result using a logistic function.

[0192] As shown in Figure 12A, it was confirmed that as the latent expression φ1 increases from -3 to +3, the true propensity score e increases in an S-shaped pattern from 0 to 1. The larger the latent expression φ1, that is, the more pronounced the influence of the latent expression φ1, the more the propensity score e increases in a sigmoidal pattern. This trend allows visual confirmation that the latent feature is a sufficient statistic for the allocation probability. The width of the confidence band (gray band in Figure 12A) is narrowest in the central region where the latent expression φ1 is near 0. This indicates that estimation accuracy improves in areas with high data density.

[0193] Figure 12B is a plot of the difference (residual) between the plotted values ​​of the propensity score e shown in Figure 12A and the fitting curve. The residuals were randomly distributed around 0, and no significant trend was observed in the residuals. Therefore, the relationship between the latent expression φ1 and the true propensity score e can be adequately explained by a logistic model.

[0194] 12A and 12B visualize that a monotonic correspondence between "latent space and intervention probability" is established at the stage of representation learning (generation of latent representation φ1) that reduces the dimensionality of the spatiotemporal persona field ψ(x,t) and the stage of estimating the propensity score e. A monotonic relationship is established between the position in the latent space obtained by representation learning and the corresponding intervention probability, and this structure enables stable and semantic estimation of the propensity score.

[0195] Figure 13A is a graph showing the results of estimating the intervention effect using the online double correction estimation formula. In Figure 13A, the gray dotted line represents the instantaneous estimated intervention effect τ hat inst(t), and the solid black line represents the instantaneous estimated intervention effect τ hat The cumulative average estimated intervention effect τ is the cumulative value of inst(t). hat cum(t).

[0196] Here, the instantaneous estimated intervention effect τ hat inst(t) and cumulative average estimated intervention effect τ hat cum(t) is the following formula

number

number

[0197] In Figure 13A, the gray band represents the cumulative average estimated intervention effect τ hat Theoretical confidence interval (τ hat cum(t)±1.96·σ / √n), and the black dashed line indicates the true value of the intervention effect τ (τ=1.5). Figure 13B shows the cumulative average estimated intervention effect τ in Figure 13A. hat This is a graph plotting the absolute value of the difference between cum(t) and the true value of the intervention effect τ.

[0198] Figure 13A shows that the initial variability of the confidence interval converged from 0.3 to 0.05 within 50 steps, even in a high-noise environment with σ = 3. Figure 13B also shows that the cumulative average estimated intervention effect τ hat The error between cum(t) and the true value of the intervention effect τ (τ = 1.5) was less than 0.01 at 200 steps, confirming that the bias had essentially disappeared. Also, as shown in Figure 13B, the cumulative average estimated intervention effect τ hat The absolute value of the difference from the true value of cum(t) is 10 -1 10 from order -3 We were able to confirm the asymptotic unbiasedness of the online DR estimation, which decays exponentially to the order of magnitude.

[0199] 13A and 13B show the causal inference unit 104 receiving the spatiotemporal persona field ψ(x, t) sequentially transmitted as a stream from the persona field storage unit 103, and the estimated intervention effect τ hat The graphs also show that the DR estimation algorithm converges even when it is online (receiving data sequentially). t Because it combines both estimation of the model and estimation of the outcome model, the two estimation errors compensate for each other, and it is thought that bias can be significantly reduced if at least one of the models is correctly specified.

[0200] Figure 14 shows the DR estimates (300-step instantaneous estimated intervention effect τ hat inst(t)) (number of bins: 20). In Figure 14, the dashed line extending vertically indicates the true value of the intervention effect (τ = 1.5). As shown in Figure 14, the histogram shows a frequency distribution with a high frequency in the range of 1.4 to 1.6, concentrated around the true value (τ = 1.5). In addition, the symmetry of the distribution was high around the true value. From these characteristics, the estimated intervention effect τ hat The distribution characteristics of t suggest that the bias is small. When the variance is checked, the standard deviation is approximately 0.28, which is the theoretical value based on the population standard deviation σ≒4.7 and the sample size N=300.

number

[0201] Figure 15 shows the simulation results showing the relationship between the inference method of causal effects for calculating the intervention effect τ and the estimation error (root mean square error, RMSE). As shown in Figure 15, the estimation error was the largest because it was the naive difference between the intervention group (when the intervention was implemented) and the control group (when the intervention was not implemented). This is thought to be because this inference method does not correct for confounding relationships. Hereafter, we will refer to the inverse probability weighted estimation (IPTW), propensity score e t The estimation error decreased in the following order: Matching (P-score) based on neighborhood matching, Doubly Robust DR (which combines IPTW and regression correction), Causal Forest, and Bayes-DR, which averages the Bayesian posterior distribution for the parameters of DR estimation. This suggests that when there are cost constraints, it is preferable to adopt the DR method.

[0202] 4. Intervention timing decision process (1) Details of the intervention candidate detection unit 105 The following describes the process of determining the timing of intervention based on critical points, which is performed by the intervention candidate detection unit 105. It is known that the behavior of a subject changes suddenly near critical points where the tendency of the subject's psychological state changes (e.g., social criticality phenomenon, impulse buying, etc.). Conventionally, intervention times have been determined based on empirical rules, which makes it easy to miss the optimal timing. In this disclosure, the criticality theory of statistical physics is applied to detect phase changes in the spatiotemporal persona field ψ(x,t) to trigger intervention, thereby achieving maximum effect with minimum stimulation.

[0203] Here, since the spatiotemporal persona field ψ(x,t) indicating the psychological state is a continuum, the critical state, which is one form of the psychological state, can also be expressed as a set of multiple points on the continuum. Therefore, in this specification, the term "critical region" may be used instead of "critical point" to express the critical state. Also, in the implementation stage, when the spatiotemporal persona field ψ(x,t) is latticed and expressed as a set of multiple cells, it is in a critical state (in this embodiment, "minimum eigenvalue λ min ≦threshold ε c " relationship is satisfied.) A lattice cell is defined as a critical cell, and a set formed by a series of adjacent critical cells is specifically called a critical cell region. In this specification, the term "critical region" is used to mean a region indicating a critical state on a continuum, and when referring to a set approximated on a discrete lattice of the space-time Persona field ψ(x,t), the term "critical cell region" is used as a sub-term of the critical region.

[0204] 16 is a flowchart illustrating the processing executed by the intervention candidate detection unit 105. The intervention candidate detection unit 105 detects the latest persona snapshot ψ generated by the persona field storage unit 103. t (x) is acquired (step S151). The latest persona snapshot ψ t (x) may be received directly from the persona field storage unit 103 or may be the latest persona snapshot ψ stored in the recording medium 920. t (x) may also be read.

[0205] The intervention candidate detection unit 105 detects critical cells using the free energy functional F[ψ] of the spatiotemporal persona field ψ(x, t), expressed by the following equation, and further extracts a set in which critical cells are continuously distributed as a critical cell region.

number

[0206] Here, L is a Laplacian linear operator, and U(ψ) is a potential. As a specific example, the potential U(ψ) is assumed to be downward convex, and the minimum eigenvalue λmin has a finite lower bound. One example of such a potential U(ψ) is -αψ 2 +βψ 4 (β>0). The intervention candidate detection unit 105 uses the latest persona snapshot ψ obtained from the persona field storage unit 103 or the recording medium 920. t For (x), the free energy functional F[ψ] is calculated (step S152).

[0207] When detecting a critical cell from the free energy functional F[ψ] obtained in step S152, the following Hessian matrix H(x,t), which is a second-order differential matrix, is used.

number

[0208] The critical cell is the minimum eigenvalue (minimum eigenvalue λ) of the Hessian matrix H(x,t) defined by the following equation: min ) is used. As mentioned above, when critical cells are spatially adjacent to each other to form a continuous region, the set is called a critical cell region.

number

[0209] In this embodiment, the persona snap ψ t (x) contains not only the persona vector for each cell but also the Laplacian (spatial second derivative, ∇ 2 ψ t (x)) is also stored, so in the region where the Hessian matrix is ​​isotropic and each eigenvalue can be considered to be approximately equal,

number

[0210] In step S153, the minimum eigenvalue λ min The minimum eigenvalue λ is calculated. min To calculate the eigenvalues, the free energy functional F[ψ] is approximated by finite difference on a grid, and the Hessian matrix H(x,t) is constructed using sparse diagonal blocks through sparse finite difference discretization. This reduces the computational complexity even when the space-time persona field ψ(x,t) is high-dimensional. To calculate the eigenvalues, it is preferable to use a power iteration method with linear scaling computational complexity O(Kn) with respect to the state dimension n, without performing a full eigendecomposition. This allows for real-time approximation of the curvature index in high-dimensional space, thereby reducing processing delays and memory consumption and enabling immediate intervention decisions and stability estimation. In addition to the power approximation method described above, other well-known Krylov subspace algorithms that can efficiently extract the smallest eigenvalues, such as the shift-and-invert method, Rayleigh quotient iteration, Lanczos method, Arnoldi method, and subspace iteration, can also be used.

[0211] Thus, the smallest eigenvalue λ min Once obtained, the smallest eigenvalue λ min Threshold ε compared to c (Step S154). c may be a constant value, but the threshold ε c is the time-varying dynamic threshold ε t Specifically, the minimum eigenvalue λ of the Hessian matrix observed in the most recent ΔT time width min Based on the distribution of t By adopting the percentile method, which sets the eigenvalue as λ, it is possible to track temperature fluctuations in the system and time drifts in data statistics (such as mean and variance). It also prevents overreactions in determining the local flatness of the model while ensuring consistency in intervention decisions. The (1-q) quantile is the minimum eigenvalue λ in the most recent specified period (ΔT, for example, several seconds). minIt may be preferable to determine it from the distribution of

[0212] Threshold ε c When the threshold ε is dynamically changing c The method for setting the minimum eigenvalue λ is not limited to the percentile method. min The distribution of is modeled using a prior distribution (such as a gamma distribution), and the posterior mean is set to a threshold ε c By Bayesian estimation, the threshold ε c Alternatively, the smallest eigenvalue λ may be determined. min The threshold value ε is set by sequential change detection when the cumulative sum (CUSUM) of c In either case, the optimum method is selected depending on the system characteristics and response requirements.

[0213] The intervention candidate detection unit 105 detects the minimum eigenvalue λ min is a predetermined numerical range, specifically, a threshold value ε c (threshold ε c is a small value near 0, and the positive or negative sign is designed according to the detection target) when the condition based on min ≦ε c or λ min ≧ε c When the persona snap ψ t The local curvature of the space-time persona field ψ(x,t) corresponding to (x) is judged to be critically small, and it is considered as a candidate for a critical cell region (a set of cells that are latticed in the critical region on the continuum). For example, λ min If you want to detect just before ε turns negative, c <0 (e.g., ε c =-0.3), and λ min If you want to detect the amplitude where ε approaches 0 from the positive side, c Just set it to 0.

[0214] The threshold value ε set in this way c and the calculated minimum eigenvalue λ min Compare with λ min ≦ε cIt is determined whether or not there is a cell that satisfies the above condition (step S155). If the determination result is "no", that is, if λ is not satisfied for all cells, min >ε c If this is satisfied, the persona snapshot ψ t (x) is determined not to have a critical cell region, and the processing of the intervention candidate detection unit 105 is terminated. c as the dynamic threshold ε t When using λ in the most recent time interval ΔT, min Due to the distribution of ε t is updated sequentially and the same judgment is made.

[0215] In step S155, λ min ≦ε c If there is a cell determined to be such, it is determined that the cell belongs to the critical cell region (step S156), and the cell is designated as a critical cell. The functional block that performs the processes from step S152 to step S156 above is the critical region detection unit 105A. Note that the specific information of the critical cell, which is the output result of the critical region detection unit 105A, can be used for purposes other than determining the next intervention candidate. For example, it is also possible to perform cluster classification of the spatiotemporal persona field ψ(x, t) using information on the distribution of critical cells (examples include area, center of gravity position, and their time fluctuations).

[0216] When specific information of the critical cell region (aggregate of critical cells) is obtained as an output of the critical region detection unit 105A, the latest intervention effect τ(x, t) is obtained from the causal inference unit 104 (step S157). min ≦ε c For critical cells that are determined to belong to the critical cell region, it is determined whether τ(x, t)>0 (positive intervention effect) is satisfied (step S158), and critical cells that satisfy τ(x, t)>0 are determined to be intervention candidates (step S159). On the other hand, if it is determined in step S158 that there are no critical cells that satisfy τ(x, t)>0, it is determined that the critical cell is not suitable for intervention timing, and the processing of the intervention candidate detection unit 105 is terminated. The functional block that performs the processing from step S157 to step S159 above is the intervention candidate determination unit 105B.

[0217] Persona snap ψ to be processed t If there is a critical cell in (x) where τ(x, t)>0, intervention candidate data including information identifying the cell (cell ID) and τ(x, t) is generated, and is sent to the intervention queue control unit 106 or temporarily stored in the recording medium 920, and the processing of the intervention candidate detection unit 105 is terminated. Note that if the processing of the intervention candidate detection unit 105 is terminated due to a negative determination in step S155 or step S158, information to that effect may be stored in the recording medium 920 as a log.

[0218] In a non-limiting example, the intervention candidate detection unit 105 performs steps S151 to S159 at 0.5 second intervals.

[0219] In the above process flow, the detection of the critical point is performed by the minimum eigenvalue λ min However, the present invention is not limited to this. For example, detection may be performed using different types of indicators, such as the norm ∥ψ∥ or the gradient norm ∥∇ψ∥ of the latest spatiotemporal persona field ψ(x,t) being equal to or greater than a predetermined value (e.g., ∥ψ∥≧δ1, ∥∇ψ∥≧δ2), or the difference ΔF in variational free energy being equal to or greater than a predetermined value, or multiple detection methods may be used in combination. A specific example of using multiple detection methods in combination is to identify a primary region by performing processing using a detection method with a relatively low computational load, and then perform processing on the identified primary region using a detection method with a relatively high computational load. When performing stepwise processing, the resolution Δx may be reduced in steps.

[0220] The above-mentioned processing (steps S151 to S159) executed by the intervention candidate detection unit 105 can be expected to have a high effect with minimal stimulation. Since intervention is performed only near the critical point, the response coefficient is easily maximized. Also, since intervention is performed only on critical zone users from all users (customers), the overall number of interventions is reduced, which reduces the cost of measures and improves ROI. Furthermore, since intervention is performed according to the individual circumstances of each user, the timing of intervention is naturally distributed, reducing the need for high-load processing such as simultaneous distribution. This reduces the load on the distribution server.

[0221] The above-described processing performed by the intervention candidate detection unit 105 detects phase changes in the vector field and automatically extracts "moments when intervention is likely to be effective," and by combining this with the causal effect (intervention effect τ) obtained by the causal inference unit 104, it is possible to trigger measures only for users who are "positively oriented and most susceptible." This makes it possible to simultaneously improve ROI and minimize the cost of measures.

[0222] In the above explanation, Persona Snap ψ t Since (x) is a collection of discrete data corresponding to a plurality of cells, the critical points are also found as critical cells in cell units, but this is not limiting.

[0223] (2) Numerical implementation example 17 is a 2-row×2-column block diagram showing the processing flow performed by the intervention candidate detection unit 105. In FIG. 17, the upper row (block 1 → block 2) calculates the minimum eigenvalue λ from the space-time persona field ψ(x, t). min The bottom line (Block 3 → Block 4) shows the logical product line of the criticality judgment and the intervention effect τ(x,t). The functions of each block are as follows:

[0224] Block 1: Persona snap ψ based on the latest spatiotemporal persona field ψ(x, t) from the persona field storage unit 103 t Receive (x).

[0225] Block 2: Persona Snap Psi tThe Hessian matrix H of the free energy of (x) is approximated by difference, and the locally minimum eigenvalue λ is found by power iteration. min Calculate.

[0226] Block 3: Adaptive Threshold (Dynamic Threshold ε t ) c The critical cells are extracted by the following formula to generate a threshold mask C.

[0227] Block 4: The positive effect cell (intervention effect τ>0) obtained from the causal inference unit 104 is logically ANDed with the threshold mask C, and the obtained target user is registered in the intervention queue control unit 106 as an intervention candidate.

[0228] In Figure 17, the two dashed arrows (the ψ stream and the τ stream) indicate the external input paths for persona field data and causal effect data, respectively. The horizontal arrows indicate serial processing between blocks, and the vertical arrows indicate the propagation of the analysis results in the upper level to the decision block in the lower level, visualizing both data flow and control flow at the same time.

[0229] Fig. 18 is a diagram showing the output results of each function block shown in Fig. 17. The upper left of Fig. 18 shows a persona snap ψ of a two-dimensional time-space persona field ψ(x, y, t) based on two state coordinates x, y, which is generated sequentially (online) in the persona field storage unit 103 as a result of the processing of block 1. t (x,y) are plotted. In this figure, the persona snap ψ t The stronger the intensity of (x, y) (positive and with a larger absolute value), the darker the color is displayed. t It can be seen that (x,y) consists of a dark central Gaussian peak centered at (x,y)=(5,5) and small noise. This central peak represents a group of increasing state quantities, such as purchasing desire, and is surrounded by a gentle lowland. Criticality has not yet been determined at the processing stage of Block 1.

[0230] The top right of Figure 18 shows the minimum eigenvalue λ of the Hessian matrix H(x,y,t) of the free energy F[ψ] of the space-time persona field ψ(x,y,t) snapshotted by the processing of Block 1 as a result of the processing of Block 2. min The spatial distribution of λ is shown. min The smaller the value (negative and with a larger absolute value), the brighter the color. min =ε c ), i.e., λ min The deep valleys with low σ are unstable regions of the system and form potential critical points.

[0231] 18 shows a threshold mask C as a result of the processing of block 3. Specifically, the minimum eigenvalue λ obtained by the processing of block 2 min Among the cells that make up the spatial distribution of min is the threshold ε c Below (λ min ≦ε c ) cells are set to "1", otherwise (λ min >ε c ) is shown as a binary map. In the figure, "1" is shown in black and "0" is shown in white, so the distribution of the critical region can be visually grasped. When a slight stimulus is applied to the critical region shown in black in this figure (implementation of intervention), the state of the spatiotemporal persona field ψ(x,y,t) can change significantly.

[0232] The processing result of block 4 is shown in the lower right of Figure 18. Specifically, the distribution obtained by taking a snapshot of the causal effect (intervention effect τ) supplied as a stream from the causal inference unit 104 is shown in grayscale. Also, the criticality (λ min ≦ε cThe area where cells where both a positive effect (τ>0) and a positive effect are present is indicated by cross-hatching. This allows the distribution of users (customers) who are candidates for intervention to be visually grasped. By implementing measures (implementing intervention) on subjects corresponding to cells located in the cross-hatched area, high effectiveness can be expected with minimal stimulation. In this way, by executing the processing of Block 4, it is possible to intuitively recognize that the "moments when the intervention is likely to be effective" have been automatically extracted.

[0233] Figure 19 is a graph showing three-dimensionally the distribution of each block shown in Figure 18. The diagram on the bottom right shows the distribution of the intervention effect τ.

[0234] (3) Effects of parameters The influence of the parameters used in the process is explained. Figure 20 shows the relationship between the calculation time T and the critical cell number N crit The results of a simulation were shown in Table 1 and Figure 20.

[0235] [Table 1]

[0236] As shown in Figure 20, when the resolution Δx was changed in the range of 0.1 to 1.6, it was found that the calculation time T (solid white circle line) was exponentially reduced as the resolution Δx was made coarser (larger). crit It was confirmed that the lower the resolution Δx, the more rapidly the value (white triangle dashed line) decreases, increasing the risk of under-extraction. crit The decreasing trend of was exponential with the calculation time T.

[0237] In the information processing system according to this embodiment, the resolution Δx and the threshold value ε cThe estimation method for the intervention effect τ can be dynamically adjusted. This makes it possible to balance system load and extraction accuracy in real time. It also makes it possible to maintain a high ROI while preventing an excess or shortage of critical cells. Furthermore, it is possible to achieve both estimation accuracy and computational resources. Note that the specific numerical values ​​shown in the above explanation are merely examples and may vary depending on data distribution, computational resources, and business requirements, but do not in any way limit the technical scope of the present invention.

[0238] 5. Details of other functional blocks (1) Measure generation department Fig. 21 is a functional block diagram illustrating an example of a specific operation of the measure generation unit 107. As shown in Fig. 21, the measure generation unit 107 includes, as functional blocks, a queue acquisition unit 171, a template pool acquisition unit 172, a history / policy acquisition unit 173, a machine learning unit 174, a personalization unit 175, and an action packet generation unit 176.

[0239] The queue acquisition unit 171 acquires intervention candidate data from the intervention queue control unit 106 (step S171). As described above, the intervention queue control unit 106 controls the input timing of the intervention candidate data to the measure generation unit 107, thereby ensuring the operational stability of the measure generation unit 107.

[0240] The template pool acquisition unit 172 acquires data (template data) that will serve as a template for an intervention measure from a template pool formed, for example, by the recording medium 920 (step S172). For example, if the intervention measure to be implemented is an email, text-based data in which the subject and email body are written in an appropriate language is acquired as the email template. If the intervention measure to be implemented is a communication by a sales representative, data indicating the content of the email to be sent by the sales representative becomes the template data.

[0241] The history and policy acquisition unit 173 acquires access history information and policy information related to the subject who is a candidate for intervention, for example, from the recording medium 920 (step S173). A specific example of access history information is a click history. A specific example of policy information is a prohibited medium for access (for example, limited to SNS, not email), a prohibited access time period (for example, no notifications allowed from 1:00 AM to noon), etc.

[0242] The machine learning unit 174 generates specific intervention measure data based on the template data through machine learning (step S174). The algorithm used for generation is not particularly limited. Examples include contextual Thompson sampling, which is widely used in online advertising, and policy gradient, which directly optimizes an agent's policy in reinforcement learning. This machine learning unit 174 may involve multiple AI agents. Examples include a channel selection AI agent that determines the medium for implementing the measure from email, push notification, advertisement, etc. and outputs the selected channel, and a frequency control AI agent that receives response logs from the target person as input, estimates the customer's fatigue score, adjusts the transmission interval, and outputs the next delivery time.

[0243] The personalization unit 175 optimizes the intervention measure data generated by the machine learning unit 174 for the target person for whom the intervention measure is to be taken (step S175). Specific examples of optimization include inserting the target person's name into an email that serves as the intervention measure, and switching (translating) the language.

[0244] In the action packet generation unit 176, an action packet is generated that simultaneously optimizes the items to be communicated to the target person, the medium, and the timing of delivery through the processes executed in the machine learning unit 174 and the personalization unit 175 (including the processes performed by the AI ​​agent described above), and this is converted into a format that can be received by the delivery and response measurement unit 108 and output to the delivery and response measurement unit 108 (step S176). A recording medium 920 may be interposed in this data flow.

[0245] (2) AI Agents Specific examples of AI agents provided in the information processing system 100 according to this embodiment are as follows.

[0246] i) Message generation AI agent: Used in the LLM encoder 112, it takes the spatiotemporal persona field ψ(x, t) and the attributes of the user (subject) as input, generates candidates for text, images, and videos, and outputs a set of candidate intervention measures. Specifically, an example is given of using an LLM and a template.

[0247] ii) Scoring AI agent: Used in the distribution and response measurement unit 108, it is a time-space persona field ψ that reflects the effect of intervention measure A. * (x, t), the subject's past response log (clicks, ratings, stay time, etc.), and the feature vector Φ combining the intervention effect τ(x, t) obtained by the causal inference unit 104 are input, and the expected reward Q(ψ * ,A)=E[R|Φ], where R is the final performance indicator (sales, CV, degree of health improvement, etc.), and the scoring AI agent sequentially updates the Q function using online reinforcement learning.

[0248] iii) Channel selection AI agent: Used in the machine learning unit 174, it takes the spatiotemporal persona field ψ(x, t) and distribution history as input, determines the medium for implementing the measures from email, push notification, advertisement, etc., and outputs the selected channel.

[0249] iv) Frequency control AI agent: Used in the machine learning unit 174, it takes the response log from the user (customer) as input, estimates the customer's fatigue score, adjusts the transmission interval, etc., and outputs the next delivery time.

[0250] v) Cost management AI agent: Used in the distribution and response measurement unit 108, it takes budget information, ROI indicators, etc. as input, monitors the upper cost limit of measures, allocates budgets, etc., and outputs a flag indicating whether or not to implement measures (interventions) under consideration.

[0251] FIG. 22 is a flowchart showing an example of processing using an AI agent executed in the information processing system according to this embodiment.

[0252] First, a message generation AI agent generates multiple (e.g., N) candidate measures (step S201). Next, a probabilistic evaluation of the multiple candidate measures is performed using methods such as the upper confidence bound (UCB), Thompson sampling, or the off-policy maximum entropy reinforcement learning method SAC (Soft Actor-Critic), and a predetermined number (e.g., k) of top candidate measures are selected (step S202). Next, a channel selection AI agent and a frequency control AI agent determine the medium and time, and deliver the selected intervention measures (step S203).

[0253] Once the intervention measures have been implemented in this manner, the learning means updates the policies of each AI using the user response log, including the response to the intervention measures, as a reward (step S204). Step S204 is executed by the update learning unit 109. The execution time for this series of processes is arbitrary, and may be, for example, a few seconds or a few minutes. Depending on the type and content of distribution, it may take some time to obtain a response, so after executing step S203, it is possible to wait to execute step S204 until the number of responses required for appropriate feedback learning has been obtained. The learning means is not limited, and in addition to Q-learning, specific examples include SAC (Soft Actor-Critic) and distributional reinforcement learning based on a distributional representation of Q values.

[0254] The mathematical points to note in the processing shown in FIG. 22 are as follows.

[0255] i) Exploration-exploitation balance: The evaluation algorithm may employ an upper-bounded search method that theoretically demonstrates convergence.

[0256] ii) Reward normalization: By normalizing revenue, cost, and fatigue penalty to the same scale, the excessive influence of negative rewards is prevented. Note that fatigue penalties can be designed using any metric, such as notification mute rate, viewing time, or survey index.

[0257] iii) Learning Stabilization: It may be desirable to heuristically limit the policy parameter update rate, thereby avoiding rapid policy fluctuations.

[0258] When executing the process shown in FIG. 22, cooperation between multiple AI agents is optional, and may be of the parameter sharing type, experience sharing type, or independent learning type.

[0259] By executing the process shown in FIG. 22, the following effects can be obtained.

[0260] i) Rapid response to environmental changes: Policy policies are updated autonomously in response to changes in external factors and user status, enabling immediate response to environmental changes.

[0261] ii) Maintaining a high ROI: Since measures are only delivered when both the critical point condition and the positive causal effect (intervention effect τ>0) are simultaneously met, the implementation of ineffective measures, i.e., wasted efforts, is suppressed, and ROI is increased.

[0262] iii) User fatigue reduction: The frequency control AI agent learns fatigue penalties, so over-delivery is automatically suppressed.

[0263] Therefore, by performing the above process, it is possible to autonomously derive the "optimal content x optimal medium x optimal timing" even in a dynamic environment, thereby simultaneously achieving continuous improvement in ROI and improved user experience.

[0264] 6.Summary The information processing system 100 according to this embodiment provides an integrated intervention optimization platform that integrates a qualitative data quantization mechanism based on LLM and a dynamic update mechanism of an AI multi-agent, with spatiotemporal persona field theory and random field dynamic causal model (PF-DCM) at its core. Based on theoretical suggestions from mathematical analysis and simulation, it is expected to achieve at least the following significant technical effects:

[0265] i) Dramatic improvement in intervention accuracy By capturing the state of the target customer at high resolution as a spatiotemporal persona field ψ(x, t), it becomes possible to grasp subtle state changes that were lost with conventional discrete clustering, enabling intervention at the optimal timing.

[0266] ii) Ex-ante quantitative prediction of causal effects By combining PF-DCM with Pearl's do-calculus, it is possible to estimate true causal effects without relying on correlation analysis, supporting the selection of measures based on scientific evidence.

[0267] iii) Limiting excessive intervention and optimizing ROI Critical zone detection based on critical phenomenon theory automatically identifies the phases where maximum effect can be expected with minimal stimulation, thereby increasing return on investment while avoiding increased costs and user fatigue due to excessive intervention.

[0268] iv) Integrating and utilizing qualitative and multimodal data By quantizing qualitative data using LLM, unstructured information such as text, images, and audio, which was previously difficult to incorporate into explanatory variables, can be incorporated into intervention design, significantly improving the rate at which information can be utilized.

[0269] v) Autonomous adaptation to environmental changes and scalability The architecture, in which multiple specialized AI agents work together to continuously learn and update models, enables rapid response to changes in the external environment and user behavior, maintaining practical response times even in large-scale data environments.

[0270] vi) Balancing explainability and privacy protection LLM allows the reasons for policies to be generated in natural language, making it possible to provide explanations without requiring specialized knowledge, and the introduction of differential privacy and federated learning technologies enables highly accurate inference while ensuring the protection of personal information.

[0271] Because the information processing system 100 according to this embodiment has these effects, it can serve as an integrated platform that fundamentally overcomes the limitations of conventional technology in a variety of fields that require intervention and optimization of human behavior, such as direct marketing, medicine, education, finance, public policy, etc. In other words, because the information processing system 100 according to this embodiment has high versatility, it can realize optimal intervention based on scientific evidence in any field involving human decision-making or behavioral change.

[0272] The effects that can be brought about by the information processing system 100 according to this embodiment can be summarized from another perspective as follows.

[0273] I) Dramatic expansion of the scope of data utilization By converting previously unutilized unstructured data such as text, images, audio, and behavior logs into explanatory variables without omission, it is possible to capture high-dimensional states with high resolution.In addition, since all information is embedded in the spatiotemporal persona field ψ(x,t) as a continuous vector, state transitions, gradients, and phase changes can be smoothly grasped.

[0274] II) Improving the accuracy of policy effect estimation The structural causal model (PF-DCM) explicitly separates correlation and causation. In addition, the online DR estimation sequentially updates the intervention effect τ, enabling ROI evaluation with less statistical bias.

[0275] III) Achieving minimum stimulus and maximum response timing The smallest eigenvalue λ of the free energy F[ψ] of the space-time persona field ψ(x,t) minBy detecting critical regions using this method, it is possible to automatically extract peak susceptibility times. Intervention candidates are limited to users who satisfy the relationship "critical cell ∧ positive causal effect," which reduces wasted interventions and adverse effects.

[0276] IV) Autonomous optimization by AI multi-agent By having multiple reinforcement learning agents, each responsible for content selection, channel selection, and delivery frequency, coordinate and update the system, it is possible to quickly respond to changes in the environment and user status. By incorporating a user (customer) fatigue index into the learning penalty, it is possible to prevent over-delivery while maintaining a high level of outcome.

[0277] V) Real-time closed loop Each method runs the sensor (qualitative data) → inference (ψ·τ) → control (policy delivery) loop in seconds, enabling decision-making without delay even in dynamic environments. Furthermore, because modules communicate using the same vector representation and stream transfer, system integration costs are reduced.

[0278] VI) Expansion of applicable areas It is expected to be applied not only in marketing but also in many other fields, including medicine (behavioral change intervention), education (learning promotion notifications), finance (investment behavior support), and public policy and politics (behavioral insights).Each method employs a variable parameter design, making it easy to customize according to the amount of data an organization has and legal and regulatory requirements.

[0279] The fields in which the information processing system 100 according to this embodiment can be involved are listed below. Specific applications in each application field can be organized, for example, by classifying them from the perspective of the purpose of the intervention. Specific examples of such perspectives of the purpose of the intervention include (Objective A) optimizing intervention for existing subjects, (Objective B) predicting the probability of success of a challenge such as developing new customers, and (Objective C) risk management such as crisis detection.

[0280] i) Direct marketing field (major applications) E-commerce site: Purchasing promotion (Objective A), Cart abandonment prevention (Objective C), Cross-selling and up-selling (Objectives A and B), Maximizing customer lifetime value (Objective A) Email marketing: optimal delivery timing (Objective A), personalized content (Objective A), increased engagement (Objective A) Mobile app: Push notification optimization (Objective A), in-app messaging (Objective A), user retention (Objective A) Social media advertising: Improved targeting accuracy (Objective A), prevention of ad fatigue (Objective C), maximization of ROI (Objective A) Subscription services: Prevent cancellations (Objective C), encourage upgrades (Objectives A and B), and improve customer satisfaction (Objective A) Retail: Store visit promotion (Objective A), inventory-linked promotion (Objective B), omnichannel optimization (Objective A)

[0281] ii) Medical and healthcare field Preventive medicine: Prevention of lifestyle-related diseases (Objectives A and C), promotion of health checkups (Objectives A and B), support for early detection (Objective C) Chronic disease management: Improving medication adherence (Objectives A and C), preventing symptom exacerbation (Objective C), improving quality of life (Objective A) Mental health: stress management (Objective C), early intervention (Objective C), suicide prevention (Objective C) Rehabilitation: Optimizing the recovery process (Objective A), maintaining motivation (Objective C), improving function (Objectives A and B) Elderly care: monitoring services (purposes A and C), dementia prevention (purpose C), support for independence (purposes A and B) Telemedicine: Optimizing online medical care (Objective A), emergency response (Objective C), continuing care (Objective A)

[0282] iii) Education and Human Resources Development Online education: Learning support (Objective A), Dropout prevention (Objective C), Level-based instruction (Objectives A, B) Corporate training: Skill development (Objective A, B), Career support (Objective A), Talent management (Objective A, B) Lifelong learning: Individually optimized learning path (Objectives A, B), maintaining motivation (Objectives A, C), maximizing results (Objective A) Language learning: Proficiency-based curriculum (Objective A), pronunciation correction (Objective A), practical conversation skill improvement (Objective A) Qualification exam preparation: Overcome weaknesses (Objective A), optimize study plans (Objective A), improve pass rates (Objective A), and suggest optimal qualifications (Objective B) ·STEAM education: creativity development (purpose B), problem-solving ability improvement (purpose A), cooperative learning support (purpose A, B)

[0283] iv) Financial services sector Credit management: default prediction (objective C), limit optimization (objective A), risk minimization (objective C) Asset management: Portfolio optimization (Objectives A and B), investment advice (Objectives A and B), risk management (Objective C) Insurance: Risk assessment (Objective C), customized product development (Objectives A and B), claim processing optimization (Objective A) Payment services: fraud detection (objective C), promotion of usage (objective A), promotion of cashless payments (objectives A and B) Loans: Repayment assistance (purpose A, B), refinancing proposals (purpose A, B), interest rate optimization (purpose A) Fintech: Personal financial management (Objective A), savings support (Objective A, B), financial inclusion (Objective A, B)

[0284] v) Other application areas Manufacturing: predictive maintenance (objective C), quality control (objective A), supply chain optimization (objectives A and B) Energy: Demand forecasting (Objectives B and C), energy conservation support (Objective A), renewable energy utilization (Objective A) Transportation and logistics: Delivery optimization (Objective A), congestion prediction (Objective B, C), promotion of modal shift (Objective A, B) Real estate: Property recommendation (Objective A), price optimization (Objective A), tenant matching (Objective A, B) Human Resources Services: Job Seeker Matching (Objectives A and B), Career Support (Objective A), Skill Development (Objective B) Entertainment: Content recommendation (Objective A), continued viewing (Objectives A and C), fan engagement (Objective A)

[0285] As described above, this disclosure relates to a technology that utilizes a big data infrastructure related to the behavior of subjects, such as customers, and combines it with generative AI technology to estimate and reproduce the psychological state of individual subjects in real time, thereby maximizing the effectiveness of, for example, advertising delivery. Specifically, the technology disclosed herein involves applying generative AI to multidimensional data contained in big data, such as past purchase history, browsing logs, location information, and social data, to generate and update a spatiotemporal continuum, the spatiotemporal persona field ψ(x,t). The spatiotemporal persona field ψ(x,t) is a "digital twin of the psychological state" for each user, modeling the user's internal state, such as interests, desires, stress, and purchasing intent, over time. The spatiotemporal persona field ψ(x,t) is constructed with a structure that allows for causal analysis of the relationship between behavioral changes before and after interventions such as advertising exposure.

[0286] Furthermore, the technology disclosed herein realizes dynamic intervention control (e.g., optimization of ad delivery timing, automatic generation of appeal content) triggered by changes in psychological state. Therefore, when the technology disclosed herein is applied to the advertising field, it distinguishes itself from conventional advertising optimization technology in that it enables the development of advertising expression and appeal strategies that correspond to "the psychology of the moment," rather than simply delivering segments.

[0287] In one application example, the technology disclosed herein simultaneously achieves improved personalization accuracy and ROI in the marketing field, and is positioned as an advanced technology that can be implemented and has high applicability in the fusion field of big data, generative AI, behavioral science, and advertising optimization.

[0288] Although the present embodiment has been described above, the present invention is not limited to these examples. For example, those in which a person skilled in the art appropriately adds or deletes components or modifies the design of the above-described embodiments, or appropriately combines features of the configuration examples of the embodiments, are also included within the scope of the present invention as long as they include the gist of the present invention. [Example]

[0289] Below, specific simulation results using the information processing system 100 according to this embodiment will be shown as examples, but the present disclosure is not limited to these examples.

[0290] In this example, a simulation was performed using the information processing system 100 according to this embodiment to optimize the timing of direct mail distribution of childcare goods targeted at parents of babies aged 0 to 12 months.

[0291] The state coordinate x of the spatiotemporal persona field ψ(x,t) in this simulation was set to the category of childcare goods and the age of the subject's infant. Therefore, the persona snap ψ, which is a time cross-section of the spatiotemporal persona field ψ(x,t), t The intensity of (x) (norm ∥ψ t (x)∥) can be expressed in a two-dimensional coordinate system with an axis indicating the category of childcare goods (category axis x) and an axis indicating the age of the baby in months (age axis y). In this simulation, the product categories set on the category axis x were set as shown in Table 2 below.

[0292] [Table 2]

[0293] The state space is defined by assigning numbers indicating the age of the moon to coordinates on the y-axis.

[0294] The processing executed in the qualitative data conversion unit 102 was as follows. First, in the preprocessor 111, qualitative data such as the subject's reviews, SNS posts, and Q&As was converted in real time into a token sequence T using a conversion encoder, generating a multimodal token sequence T. Note that this multimodal token sequence T was tagged with information about the subject who provided the qualitative data (such as the infant's age in months, generation, and gender). In addition, the preprocessor 111 performed appropriate processing to protect personal information (hashing personal identification attributes and adding differential privacy noise), and only information from which personal information could not be identified was output from the preprocessor 111.

[0295] Next, the LLM encoder 112 performed inference using the multimodal token sequence T generated by the preprocessor 111 as input, and obtained information from the multimodal token sequence T including the product category (x1 to x8), an emotion score that scores satisfaction on a scale of 1 to 5, and a relevance rating that indicates purchase intention as a probability value ranging from 0 to 1. In other words, the LLM encoder 112 performed processing to obtain, from the multimodal token sequence T, a two-dimensional numerical vector consisting of coordinates on the category axis x, the emotion score, and the relevance rating.

[0296] Next, as shown in Table 3, the reduction block 114 collected two-dimensional numerical vectors with the same coordinate position in the state coordinates x and y, and reduced the dimension of each numerical vector. Furthermore, the initial persona generation unit 115 averaged the reduced-dimensional vectors with the same coordinate position in the state coordinates x and y, to obtain the initial persona vector φ at that coordinate position. Specifically, as shown in Table 3, two-dimensional numerical vectors based on posts where the category axis x was 3 (stroller) and the age axis y was 3 (3rd month) were collected, and these two-dimensional numerical vectors were reduced to one-dimensional vectors. Furthermore, the average norm of the one-dimensional vectors was calculated, and a one-dimensional vector with a norm of 0.57 was obtained as the initial persona vector φ.

[0297] [Table 3]

[0298] A similar process was performed to obtain the initial persona vector φ corresponding to a coordinate position in the state coordinates x and y where the coordinate values ​​of each axis are integers (for example, (x, y) is (1, 1), (3, 6), etc.). Figure 23 shows a graph of the age dependency in months of the intensity of the spatiotemporal persona field ψ, which is made up of a collection of the obtained initial persona vectors φ, for each item (childcare goods) that makes up the category axis x. In Figure 23, the vertical axis indicating the intensity of the spatiotemporal persona field ψ is displayed as "Latent demand ψ," and the horizontal axis y, the age axis in months, is displayed as "Baby age (months)."

[0299] As shown in Figure 23, the items that make up the category axis x tended to have higher intensities of the spatiotemporal persona field ψ at the ages corresponding to their respective appeal ages. The interpretation of these peaks is shown in Table 4.

[0300] [Table 4]

[0301] FIG. 24 shows a persona snap ψ according to the embodiment. t The intensity of (x) (norm ∥ψ t (x)∥) in two-dimensional state coordinates. Specifically, in the persona field storage unit 103, a spatiotemporal persona field ψ(x,t) consisting of a continuum described by a reaction-diffusion PDE is obtained using the multiple initial persona vectors φ obtained in this embodiment as initial values, and a time section discretized with predetermined resolutions Δx and Δy is obtained from the spatiotemporal persona field ψ(x,t), and a persona snap ψ t (x). This persona snapshot ψ t The intensity of (x) (norm ∥ψ tA persona map showing (x)∥) in two-dimensional state coordinates consisting of a category axis x and a lunar age axis y is shown in Fig. 24. Therefore, the persona map shown in Fig. 24 shows the spatiotemporal persona field ψ(x,t) more smoothly than the graphs of each category shown in Fig. 23.

[0302] In FIG. 24, the vertical axis labeled "Product category" is the category axis x, and the horizontal axis labeled "Baby age (months)" is the age axis y. Norm ∥ψ t (x)∥ is color-coded in eight grayscale levels, and the norm ∥ψ t The larger (x)∥ is, the darker the color is. t The area where (x)∥ is large is a distribution that spreads from the bottom left to the top right in a two-dimensional coordinate system consisting of the category axis x and the age axis y. If the category axis x is replaced with the axis of the age at which the target item is appealed, it can also be said that there is a positive correlation. In Figure 24, the area where the diagonal hatching overlaps is the area where the norm∥ψ t This is the region where (x)∥ exceeds 0.75, and it can be understood that this is the time to implement intervention measures.

[0303] FIG. 25 shows a persona snap ψ according to the embodiment. t Gradient norm of (x) ∇ψ t FIG. 10 is a diagram showing a persona map in which (x)∥ is shown in two-dimensional state coordinates.

[0304] As described above, in this embodiment, the persona snap ψ t (x) is the gradient norm ∥∇ψ as information about the gradient of each cell. t (x)∥. This gradient norm∥∇ψ t Figure 25 shows a persona map in which (x)∥ is plotted on a two-dimensional state coordinate system consisting of the category axis x (vertical axis) and the age axis y (horizontal axis), similar to Figure 24. In Figure 25, the gradient norm∥∇ψ t (x)∥ is colored in seven grayscale levels, and the gradient norm ∥∇ψ t The smaller (x)∥ is, the darker the color is. Therefore, the norm ∥ψt In Figure 24, which shows the distribution of (x)∥, the norm ∥ψ t The dark areas corresponding to the peaks of (x)∥ are the gradient norm ∥∇ψ shown in Figure 25. t The distribution of (x)∥ is also shown in a relatively dark color, and the path of the demand peak movement can be seen as a band-like area extending from the bottom left to the top right on the map.

[0305] Figure 25 shows the Persona Snap ψ t Gradient ∇ψ in each cell of (x) t The direction of (x) is shown by a normalized arrow. Specifically, for each cell, the gradient norm ∥∇ψ is calculated for that cell (the center cell) and each of the eight cells surrounding the center cell (the surrounding cells). t The absolute value of the difference between (x)∥ was calculated, and the arrow of the center cell was set to point toward the surrounding cell where this value was greatest. Looking at the distribution of the arrows in Figure 25, we see that they point toward the upper right in the band-like area that indicates the path along which peaks of demand and interest move. From this trend, we can see that as the value of the age axis y (horizontal axis) increases, the interest of the subjects (parents of infants) shifts toward product categories with larger values ​​on the category axis x (vertical axis). In other words, we can see that the interest of the subjects shifts in accordance with the age of the infant in months, corresponding to the product category's appealing age in months.

[0306] Specifically, in the range of the age axis y (horizontal axis) from 0 to 3 (0 to 3 months), the arrow on the category axis x (vertical axis) points from swaddles / onesies to strollers / rattles. This indicates that during this period, as the infant begins to respond to external stimuli, the interest (demand) of the subjects (the infant's parents) shifts from clothing to other infant items. Around the age axis y (horizontal axis) of 6 (6 months), the category axis x (vertical axis) steeply slopes toward the weaning set & high chair area. This suggests that the infant has reached the weaning food stage and is beginning to demand feeding-related items. In the range of the age axis y (horizontal axis) from 9 to 12 (9 to 12 months), a strong gradient can be seen on the category axis x (vertical axis) toward the walker / edu. toys area. This gradient indicates that as infants grow older, their motor functions and brain development increases, and as a result, the interest (demand) of the target audience (parents of infants) is shifting towards walking aids and educational products.

[0307] In Figure 25, the gradient ∇ψ t The area where the direction of (x) is concentrated (the gradient ∇ψ t Since there is a high possibility that a critical region exists in the region where (x) is large), this region is used as an input and the critical region detection unit 105A is executed to find the minimum eigenvalue λ min The critical region may be found by calculating the gradient ∇ψ t Areas where (x) is large can be set as "bucket boundaries" where personas with different tendencies are adjacent, and direct mail delivery schedules can be set according to these divided buckets. [Industrial Applicability]

[0308] The information processing system 100 according to this embodiment provides an intervention design platform that integrates "continuous vector field x causal inference x criticality theory x AI autonomous update" in one stop. This makes it possible to simultaneously improve all four areas: data utilization rate, estimation accuracy, intervention effect, and operational efficiency. Therefore, the information processing system 100 according to this embodiment can significantly advance the technological level of behavioral analysis and intervention optimization in industry and the public sector. [Explanation of symbols]

[0309] 100: Information Processing Systems 101: Data Collection Department 102: Qualitative Data Conversion Department 103: Persona Field Memory Section 104:Causal inference part 105: Intervention candidate detection unit 105A: Critical region detection unit 105B: Intervention candidate determination unit 106: Intervention queue control unit 107: Policy generation department 108: Distribution and response measurement section 109: Update learning section 110: Persona map generation unit 111 : Preprocessor 112: LLM encoder 113: Quality Evaluation Department 114: Reduction block 115: Initial persona generation section 119: Update learning section 131: Initial Persona Acquisition Section 132: Update parameter acquisition unit 133: Persona Field Update Department 134: Persona Snap Generation Unit 141: Triple event acquisition unit 142: Persona Snap Acquisition Department 143: Update parameter acquisition unit 144 :PF-DCM inference part 145: Intervention effect map generation section 171: Queue acquisition unit 172: Template pool acquisition unit 173: History and policy acquisition section 174: Machine Learning Department 175: Personalization Department 176: Action packet generation unit 900: Processing equipment 910: Input / output device 920: Recording media 930: External database 940, 960: Information terminal 950: External information processing device 1000: Information processing equipment NW: Network

Claims

1. An information processing system comprising a processing device capable of data processing, an input / output device, and a recording medium that is a non-transitory computer-readable medium, The processing device includes: a qualitative data conversion unit that performs processing including converting unstructured data including information about the subject input via the input / output device into an initial persona vector φ(x) including a numerical vector r that indicates the persona of the subject at a predetermined time point using a trained generative model; and a persona field storage unit that executes processing including generating a spatiotemporal persona field ψ(x,t), which is described using a reaction-diffusion partial differential equation in which the initial value is time-evolved by applying a weighted kernel integral using the initial persona vector ψ(x) as an initial value, or by applying the weighted kernel integral and an extended Kalman filter assimilation process, and in which the psychological state of the subject is expressed as a continuous vector field with state coordinate x and time t as parameters; To be prepared An information processing system characterized by:

2. The qualitative data conversion unit a preprocessor that uses the trained generative model to perform a normalization process to convert the unstructured data into a token string T; an LLM encoder that performs an embedding process to encode the token sequence into a d-dimensional embedding vector E corresponding to a space of interest; an initial persona generation unit that executes a process of generating an initial persona vector φ(x) including the numerical vector r based on the d-dimensional embedding vector E; The information processing system according to claim 1 , comprising:

3. An information processing system as described in claim 2, wherein the unstructured data includes data other than text, and the token sequence T is a multimodal token sequence.

4. The information processing system described in Claim 2, wherein the qualitative data conversion unit is provided with a reduction block that performs a compression process to generate the numerical vector r, including compressing the d-dimensional embedded vector E or a vector based on the d-dimensional embedded vector E using a projection matrix.

5. The information processing system of claim 4, wherein the qualitative data conversion unit receives the generation probability p 1 output by the trained generative model and the consistency probability p 2 output by an external fact-checking model for the d-dimensional embedding vector E, calculates the geometric mean, and performs masking on components for which the calculated p conf is lower than a predetermined threshold, thereby generating a masked embedding vector E' as a vector based on the d-dimensional embedding vector E.

6. The persona field storage unit includes a persona snap generation unit that executes processing including extracting a time cross section of the space-time persona field ψ(x, t) and generating a persona snap ψ t (x) that is an instantaneous vector; Persona Snap ψ t (x) includes a collection of discrete data in which information of the time cross section of the space-time persona field ψ(x,t) is discretized into each of a plurality of cells obtained by dividing the time cross section into a grid with a resolution Δx, The information processing system according to claim 1 , wherein the discrete data includes data indicating a result of a spatial differentiation operation of the time cross section corresponding to the cell to which the discrete data belongs.

7. The processing device includes: Data including a covariate Z, an intervention variable A, and an outcome variable Y generated based on information acquired via the input / output device; The persona snap ψ generated by the persona snap generating unit t (x) and With input, The information processing system according to claim 6 , further comprising a causal inference unit that executes processing including calculating an intervention effect τ defined by the following formula: τ=E 1 -E 0 Here, E 1 is the expected value when the intervention is implemented, and E 0 is the expected value when no intervention is implemented.

8. The processing device t (x) further comprising a critical region detection unit that executes a process including detecting a region in a critical state as a critical region; The information processing system according to claim 7 , further comprising an intervention candidate determination unit that executes a process including determining the critical region in which the intervention effect τ is positive as an intervention candidate.

9. The critical region detection unit detects the person snap ψ t 9. The information processing system according to claim 8, wherein the minimum eigenvalue is calculated from the Hessian matrix of (x), and the region in which the minimum eigenvalue is within a predetermined range of values ​​is determined to be the critical region.

10. The information processing system according to claim 9, wherein the processing device further includes a measure generation unit that executes processing including determining at least one of an intervention content, an intervention medium, and an intervention timing for the critical region determined to be an intervention candidate.

11. The processing device performs the following with respect to the information acquired via the input / output device: The covariate Z is set by evaluating it from the viewpoint of an explanatory variable describing the state before the intervention, Evaluate in terms of whether or not to intervene and set the intervention variable A; Evaluate in terms of the presence or absence of an outcome and set the outcome variable Y. The information processing system according to claim 7 , further comprising a triplet event creation unit that executes processing including:

12. The causal graph referred to when the causal inference unit calculates the intervention effect τ is A causal path from the covariate Z to the intervention variable A (Z → A), a causal path from the covariate Z to the outcome variable Y (Z → Y); and The causal path from the intervention variable A to the outcome variable Y (A → Y) The information processing system according to claim 7 , further comprising:

13. The causal graph further comprises: The causal path (ψ → A) from the space-time persona field ψ(x,t) to the intervention variable A The information processing system according to claim 12 , further comprising:

14. The information processing system according to claim 7 , wherein the causal inference unit estimates the intervention effect τ using a double correction estimation formula that combines a propensity score model and an outcome model.

15. The information processing system according to claim 10, wherein the processing device further comprises an update learning unit that executes processing including updating parameters of the reaction-diffusion partial differential equation describing the space-time persona field ψ(x,t) using information including a reaction to an intervention measure implemented based on the decision of the measure generation unit as input.

16. The persona field storage unit includes a persona field update unit that acquires updated parameters of the reaction-diffusion partial differential equation from the update learning unit and executes a process of updating the spatiotemporal persona field ψ(x, t), The information processing system according to claim 15 , wherein the persona snap generation unit executes a process of generating the persona snap ψ t (x) by determining a time cross section of the spatiotemporal persona field ψ(x, t) updated by the persona field update unit.

17. The information processing system according to claim 15 , wherein the update learning unit executes a process including updating parameters used in the causal inference unit.

18. The information processing system of any one of claims 1 to 5, wherein the processing device further comprises a persona map generation unit that generates map data for displaying on the input / output device a persona map in which the norm ∥ψ(x,t)∥ or the gradient norm ∥∇ψ(x,t)∥ for at least a portion of the space-time persona field ψ(x,t) is plotted in a two- or more-dimensional coordinate space that includes the state coordinate x as a coordinate axis.

Images