Information processing device, information processing method, and recording medium

Abstract

In this information processing device, an information gain calculation means acquires a sample set of target information and calculates an information gain of the sample set. A cost estimation means estimates the cost required for acquiring the sample set. A plan development means develops an optimal plan for obtaining a sample set that maximizes the information gain under cost-related constraint conditions. A utility function estimation means estimates a utility function for maximizing the optimum value obtained by achieving the optimal plan. A negotiation means executes negotiation so as to optimize a transaction condition of the sample on the basis of the utility function.

Description

Information processing device, information processing method, and recording medium

[0001] The present disclosure relates to trading data and learning models.

[0002] Previously, there were methods to estimate models without disclosing the original data through information consortia or federated learning, or to build learning models from confidential data, but these methods limited the opportunities for sharing and trading data and models.

[0003] Patent Document 1 describes a negotiation system that automatically negotiates terms between an ordering party and an order receiving party.

[0004] International Publication No. 2021 / 038626

[0005] One objective of the present disclosure is to propose a marketplace and trading method that enables the exchange and trading of data and learning models.

[0006] In one aspect of the present disclosure, an information processing device includes: an information gain calculation means for acquiring a sample set of target information and calculating the information gain of the sample set; a cost estimation means for estimating the cost required to acquire the sample set; a planning means for formulating an optimal plan for determining a sample set that maximizes the information gain under constraints related to the cost; a utility function estimation means for estimating a utility function for maximizing the optimal value obtained by achieving the optimal plan; and a negotiation means for conducting negotiations to optimize the trading conditions of the samples based on the utility function.

[0007] In another aspect of the present disclosure, an information processing method is executed by a computer, and includes acquiring a sample set of target information, calculating the information gain of the sample set, estimating the cost required to acquire the sample set, formulating an optimal plan to obtain a sample set that maximizes the information gain under constraints on the cost, estimating a utility function to maximize the optimal value obtained by achieving the optimal plan, and conducting negotiations to optimize trading terms for the samples based on the utility function.

[0008] In yet another aspect of the present disclosure, a recording medium records a program that causes a computer to execute the following processes: acquire a sample set of target information, calculate the information gain of the sample set, estimate the cost required to acquire the sample set, formulate an optimal plan to obtain a sample set that maximizes the information gain under constraints on the cost, estimate a utility function to maximize the optimal value obtained by achieving the optimal plan, and conduct negotiations to optimize the trading conditions of the samples based on the utility function.

[0009] According to the present disclosure, a marketplace and a trading method therefor that enable the exchange and trading of data and learning models can be proposed.

[0010] Fig. 1 shows the overall configuration of an information trading system to which an information processing device according to the present disclosure is applied. Fig. 2 is a block diagram showing the hardware configuration of the information trading device. Fig. 3 is a block diagram showing the functional configuration of the information trading device. Fig. 4 is a flowchart of information trading processing. Fig. 5 is a block diagram showing the functional configuration of an information processing device of a second embodiment. Fig. 6 is a flowchart of processing by the information processing device of the second embodiment.

[0011] <Prerequisites> Before describing the embodiments of the present disclosure, the following describes the underlying problems of conventional technologies. Current methods, including information consortia and federated learning, enable model estimation and the construction of learning models using confidential information while keeping the original data private. These methods make it possible to promote knowledge sharing while protecting data privacy.

[0012] However, while these methods have made progress in distributed learning and data sharing, they have not yet achieved the design of markets and rules that support data sharing, pricing, and the sharing of learning models under desired conditions between companies.To fill this gap, this embodiment proposes an automated marketplace and methodology for efficiently trading data and learning models.

[0013] Specifically, the proposed system calculates information gains from data samples, estimates trading costs, develops trading plans, estimates utility functions, and automatically negotiates with trading partners. This system is designed to automate the entire data trading process and facilitate agreement on terms, pricing, and sharing of learning models between companies. This system enables a market for data and learning models where participants can reach agreements based on their own terms.

[0014] The method of this embodiment can be applied to the trading of data and learning models in a wide range of fields, such as finance, medicine, marketing intelligence (MI), logistics, manufacturing, and space (especially data acquisition from satellite images).

[0015] Preferred embodiments of the present disclosure will now be described with reference to the drawings. First Embodiment Overall Configuration FIG. 1 shows the overall configuration of an information trading system to which an information processing device according to the present disclosure is applied. The information trading system is a system for buying and selling information such as data and learning models between companies and individuals. Specifically, the information trading system includes an information trading device 100 used by the company or individual. The information trading device 100 may be used by a seller or a buyer. The information trading device 100 used by the seller and the information trading device 100 used by the buyer mutually exchange data and learning models to be bought and sold and their prices, and negotiate the buying and selling. Hereinafter, the data and learning models to be bought and sold will also be referred to as "trading information."

[0016] [Hardware Configuration] (Trading Device) Fig. 2 is a block diagram showing the hardware configuration of the information trading device 100. As shown in the figure, the information trading device 100 includes a processor 11, an interface (IF) 12, a ROM (Read Only Memory) 13, a RAM (Random Access Memory) 14, a database (DB) 15, and a storage medium 16. Each component is connected to the other via, for example, a bus 18.

[0017] The processor 11 is a computer such as a CPU (Central Processing Unit), and executes a prepared program to control the entire information trading device 100. Specifically, the processor 11 may be a CPU, a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), an MPU (Micro Processing Unit), an FPU (Floating Point number Processing Unit), a PPU (Physics Processing Unit), a TPU (Tensor Processing Unit), a quantum processor, a microcontroller, or a combination thereof.

[0018] The processor 11 also loads a program stored in the ROM 13 or the storage medium 16 into the RAM 14 and executes each process coded in the program. The processor 11 functions as a part or the whole of the information transaction device 100. The processor 11 executes the information transaction process described below.

[0019] The IF 12 transmits and receives data to and from external devices. Specifically, the information trading device 100 transmits and receives trading information and prices to and from other information trading devices 100 via the IF 12.

[0020] The ROM 13 stores various programs executed by the processor 11. The RAM 14 is used as a working memory while the processor 11 is executing various processes.

[0021] The DB 15 stores transaction information that is the subject of a transaction. The DB 15 may also store various algorithms, data, machine learning models, and the like that are used when the information trading device 100 executes the processes described below.

[0022] The storage medium 16 is a non-volatile, non-transitory storage medium such as a disk-shaped recording medium or a semiconductor memory. The storage medium 16 may be configured to be detachable from the information transaction device 100. The storage medium 16 stores various programs executed by the processor 11.

[0023] In addition to the above, the information transaction device 100 may also include a display device such as a liquid crystal display and an input device such as a keyboard and a mouse. These display device and input device are used by, for example, an operator of the information transaction device 100.

[0024] 3 is a block diagram showing the functional configuration of the information trading device 100. Functionally, the information trading device 100 includes an automatic negotiation unit 21, a utility function estimation unit 22, an information gain calculation unit 23, a plan formulation unit 24, and a cost estimation unit 25.

[0025] The information gain calculation unit 23 acquires a sample set of target transaction information and calculates the information gain of the sample set. The cost estimation unit 25 tally up the total costs of human labor, materials, effort, and time required to acquire the sample set. Through this process, the economic feasibility of data acquisition can be evaluated.

[0026] The planner 24 determines a sample set that maximizes information gain while taking into account budget and other cost constraints. Specifically, the planner 24 uses a greedy algorithm to select samples that offer the highest gain per additional sample cost, thereby constructing an optimal sample set. At this stage, an optimal plan is formulated that optimizes the balance between cost efficiency and information value.

[0027] The utility function estimator 22 constructs a function that evaluates the effect of perturbations to maximize the optimal value obtained by achieving the optimal plan. The utility function estimator 22 is designed to exhibit high utility when samples available in the market exceed the information gain of the existing sample set or have the same information gain but at a lower cost. This lays the foundation for formulating an optimal trading strategy.

[0028] The automatic negotiation unit 21 executes negotiations using the utility function constructed by the utility function estimation unit 22. The automatic negotiation unit 21 optimizes the trading conditions of the sample based on the utility function, aiming for the most advantageous result in negotiations in the market.

[0029] [Specific Example] Next, specific examples of each element of the information trading device 100 will be described in detail. (Information Gain Calculation Unit) The information gain calculation unit 23 uses a Bayesian approach to quantify how valuable the number of samples obtained from a specific sample object under certain sampling conditions is in terms of information theory. This process is based on a prior distribution P0(θ) for an unknown parameter θ and a data model family P(X,Y|θ) that depends on the sampling condition X and the sample object Y.

[0030] Specifically, the information gain calculation unit 23 samples the sampling condition X from the sampling condition C, and samples the sample target Y from the distribution P(Y|X = x_i, θ) based on the given parameter θ and the sampling condition X, thereby generating N pieces of sample data (x_i, y_i). This method takes into account, for example, a function f(θ) that represents the information that a buyer of transaction information wants to know. The function f(θ) is generally a vector-valued function and provides information related to a specific problem, such as the parameter θ itself or the sampling condition X optimized based on the parameter θ. The information gain calculation unit 23 updates the posterior distribution of the parameter θ from the sample data (x_i, y_i) and the prior distribution P0(θ), and calculates the information gain from the resulting distribution of the function f(θ). The information gain G(C, Y, N) is defined as the inverse of the variance (matrix norm) of the distribution of the function f and measures the value of the information in terms of "how clearly it can be understood."

[0031] This approach avoids complex calculations and calculates the information gain through multiple simulations. Specifically, the information gain calculation unit 23 first samples the parameter θ from the prior distribution P0(θ), and then samples the sampling condition X according to the sampling condition C. Next, the information gain calculation unit 23 samples the sample object Y using the sampled condition X and the parameter θ, and repeats this process N times to construct a sample data set. Finally, the information gain calculation unit 23 calculates the posterior distribution of the parameter θ using the sample data, calculates the distribution of the function f(θ), and outputs the inverse of its variance as the information gain.

[0032] This information gain calculation unit 23 can be used to evaluate the value of data in the information market. When purchasing specific experimental data or a training dataset, buyers in the information market can automatically evaluate how many samples are required and how much information-theoretically valuable they are. This technology streamlines the decision-making process for information trading, enabling more efficient information trading.

[0033] Although Bayesian optimization has been described here as an example of a method for calculating information gain, it is not limited to this. Batch Bayesian optimization, which can simultaneously evaluate multiple samples, may also be used.

[0034] Now, consider a specific example in which this algorithm is applied to the learning data purchases in demand forecasting for convenience stores. In this case, the unknown parameter θ is a parameter that models factors (weather, season, day of the week, local events, etc.) that affect the daily sales volume of convenience store products. This parameter θ may be, for example, a prediction model trained using a machine learning algorithm from past data.

[0035] The prior distribution P0(θ) is set based on existing knowledge and past data about the influence of these factors. The data model family P(X, Y|θ) describes the distribution of sales of product Y (e.g., a particular beverage or snack) under a particular sampling condition X (e.g., a particular season or day of the week).

[0036] Sampling conditions C are a set of conditions that are of particular interest to buyers of the training data, such as "summer weekends" or "rainy weekdays." The function f(θ) represents the information that convenience store operators are most interested in, which could be the optimal inventory level for a product under specific conditions based on the parameter θ, or the magnitude of the impact that a particular promotional activity has on sales.

[0037] To calculate the information gain G(C, Y, N), the information gain calculation unit 23 first samples the parameter θ from the prior distribution and samples the sampling condition X (e.g., summer weekends) based on the selected sampling condition C. Next, the information gain calculation unit 23 samples the sales volume of product Y from the distribution P(Y|X = x_i, θ) using the sampling condition X and the parameter θ, and repeats this process N times to generate a sample data set. The information gain calculation unit 23 uses this data set to update the posterior distribution of the parameter θ and calculates the distribution of the function f(θ). Finally, the information gain calculation unit 23 calculates the information gain of the sample obtained under the specific sampling conditions by calculating the inverse of the variance (matrix norm) of the distribution of the function f(θ).

[0038] Based on this information gain, convenience store operators can evaluate the degree of improvement in accuracy they can expect in demand forecasts under specific conditions and use this information to help them make decisions when considering purchasing training data. Comparing the costs associated with purchasing data with the information gain enables them to select the most cost-effective dataset, contributing to improved demand forecast accuracy.

[0039] (Planning Unit) The planning unit 24 uses an optimization algorithm to formulate a strategy for efficiently acquiring data samples. As an example, a case will be described in which the planning unit 24 creates a plan using a greedy algorithm. The planning unit 24 aims to select and acquire the sample with the highest utility within the available budget. Here, "utility" is a numerical representation of various factors, such as the value of the information provided by the sample and its impact on improving the performance of the learning model. In this embodiment, utility is a value calculated from the information gain and cost described above.

[0040] After the automated negotiation unit 21 successfully negotiates and obtains data samples from the seller, the planning unit 24 uses the samples to update the posterior distribution of the model parameter θ using Bayesian optimization techniques. That is, the posterior distribution is updated based on the principles of Bayesian statistics to reflect new information obtained from the obtained samples. Through this process, the uncertainty of the model is reduced, enabling more accurate predictions and analysis.

[0041] The planner 24 calculates this posterior distribution using newly acquired samples and uses the result as a new prior distribution. This allows for more informed decision-making when acquiring the next sample. The planner 24 can incrementally acquire samples that provide the maximum utility within a budget, efficiently improving the accuracy of the learning model.

[0042] For example, the process of purchasing learning data for demand forecasting at convenience stores will be described below using a specific example. Using the planning unit 24, a strategy is adopted for gradually purchasing the most cost-effective data samples.

[0043] First, when a sample containing information that is predicted to have a significant impact on the demand for a particular product at the convenience store, as calculated by the information gain calculation unit 23 described above, is purchased, the plan formulation unit 24 updates the convenience store's demand forecasting model. Specifically, the newly acquired data is used to calculate the posterior distribution of the model, which is then used as a new prior distribution. This update process allows the demand forecasting model to reflect the latest market conditions and environmental changes, improving forecast accuracy. Through this iterative process, the convenience store can efficiently obtain data with the greatest utility within its limited budget and gradually improve the accuracy of its demand forecasting.

[0044] (Utility Function Estimation Unit) The utility function estimation unit 22 constructs a model for quantifying the utility of agreed conditions in data transactions, and provides a standard for the automatic negotiation unit 21, which will be described later. This model evaluates both conditions that result in high information gain and conditions that result in low price as levels of utility. Specifically, the following utility function is used to express the relationship between the information gain of data and the transaction price. Utility function U(C, Y, N, price) = G(C, Y, N) - α × price

[0045] Here, G(C, Y, N) is a function that calculates the information gain, and C, Y, and N are parameters that affect the calculation of the information gain. The price refers to the price of the traded data or model, and α is a parameter that represents the weight of the price utility.

[0046] The design of the utility function is based on the principle that the higher the information gain and the lower the price, the higher the utility. This function calculates the final utility by subtracting α times the price from the information gain G(C, Y, N). The value of α is set based on the criterion that "if the cost of conducting the experiment yourself is the same as the price in the market, the utility will be zero." This makes it possible to evaluate the terms of agreement taking into account the balance between price and information gain.

[0047] The utility function estimator 22 refines the transaction decision-making process. Trading participants can use this utility function to evaluate the utility of proposed transaction conditions and select the transaction conditions with the highest utility. Through this process, participants can pursue the optimal balance between cost and information gain, enabling more valuable data transactions. The utility function estimator 22 improves the accuracy of negotiations and decision-making in the data market, contributing to the realization of fairer and more efficient transactions.

[0048] (Automated Negotiation Unit) Next, a method for the automated negotiation unit 21 to automatically negotiate using the above-described utility function will be described. The automated negotiation unit 21 executes an automated negotiation process based on the utility function, proposal evaluation and response, and generation of a negotiation strategy. The utility function calculated by the above-described process quantitatively evaluates the value of a proposal in the negotiation process. The utility function quantifies the importance and preference of each negotiation element and indicates the degree to which a proposal matches the negotiator's goals. This function reflects the priorities and values ​​related to the negotiation issues and provides a basis for making optimal decisions during the negotiation process.

[0049] The automated negotiation unit 21 evaluates the received proposals based on a utility function, for example, and either accepts the proposal depending on its value or makes an improved counterproposal including more favorable terms. This step is repeated until a mutually satisfactory agreement is reached, with each proposal being continuously adjusted as the negotiation progresses.

[0050] Negotiation strategy generation includes time-dependent strategies and strategies using deep reinforcement learning. The Boulware strategy, an example of a time-dependent strategy, increases the probability of reaching an agreement by making more concessions as time passes. On the other hand, strategic models based on deep reinforcement learning learn from past negotiation data and results to generate optimal proposed strategies. These strategies are important for adapting to negotiation dynamics and finding a final agreement.

[0051] [Information Transaction Processing] Next, a description will be given of the flow of information transaction processing by the information transaction device 100. Fig. 4 is a flowchart of the information transaction processing. This processing is realized by the processor 11 shown in Fig. 2 executing a prepared program and operating as the elements shown in Fig. 3.

[0052] First, the information gain calculation unit 23 acquires a set of samples and calculates an optimality criterion for each sample (step S11). This step quantifies the information value of each sample in the data set and evaluates its gain.

[0053] Next, the cost estimation unit 25 tally up the costs of manpower, materials, labor, and time required for each sample set (step S12). At this stage, the data acquisition cost is calculated and an evaluation of economic feasibility is performed.

[0054] Next, the planner 24 operates to find a sample set that maximizes information gain under budget and other cost constraints. Specifically, the planner 24 uses a greedy algorithm to select samples with an optimal ratio of additional cost to gain, and develops an optimal plan for cost-effective data collection (step S13).

[0055] Next, the utility function estimator 22 constructs a function for evaluating perturbations to maximize the optimal value achieved by the optimal plan (step S14). This function serves as a criterion for optimizing the sample procurement strategy in the market.

[0056] Finally, the automatic negotiation unit 21 negotiates sample trading conditions in the market based on the above-mentioned utility function (step S15). The automatic negotiation unit 21 automatically advances the negotiation process to achieve an optimal balance between information gain and cost.

[0057] Through the above series of steps, the information trading system automates the trading process of data and learning models, and pursues optimal trading conditions efficiently and economically.

[0058] As described above, the information trading device 100 of this embodiment calculates information gain from a data sample, estimates the cost of the data, formulates an optimal plan for trading, estimates a utility function, formulates a negotiation strategy, and automatically executes negotiations. This allows the user to more accurately evaluate the true value of the data and models and conduct efficient and effective trading.

[0059] 4 is a block diagram showing the functional configuration of an information processing apparatus according to Embodiment 2. The information processing apparatus 70 includes an information gain calculation unit 71, a cost estimation unit 72, a plan formulation unit 73, a utility function estimation unit 74, and a negotiation unit 75.

[0060] 5 is a flowchart of processing by the information processing device of the second embodiment. The information gain calculation means 71 acquires a sample set of target information and calculates the information gain of the sample set (step S71). The cost estimation means 72 estimates the cost required to acquire the sample set (step S72). The planning means 73 develops an optimal plan to find a sample set that maximizes the information gain under cost constraints (step S73). The utility function estimation means 74 estimates a utility function to maximize the optimal value obtained by achieving the optimal plan (step S74). The negotiation means 75 negotiates to optimize the trading conditions of the samples based on the utility function (step S75).

[0061] The information processing device 70 of the second embodiment proposes a marketplace that enables the exchange and trading of data and learning models, and a trading method thereof.

[0062] A part or all of the above-described embodiments can be described as, but not limited to, the following supplementary notes.

[0063] (Supplementary Note 1) An information processing device comprising: an information gain calculation means for acquiring a sample set of target information and calculating the information gain of the sample set; a cost estimation means for estimating the cost required to acquire the sample set; a planning means for formulating an optimal plan for determining a sample set that maximizes the information gain under constraints on the cost; a utility function estimation means for estimating a utility function for maximizing an optimal value obtained by achieving the optimal plan; and a negotiation means for negotiating to optimize trading conditions for the samples based on the utility function.

[0064] (Supplementary Note 2) The information processing device according to Supplementary Note 1, wherein the information gain calculation means calculates the informational value of the number of samples obtained from the sample subject using a Bayesian approach.

[0065] (Supplementary Note 3) The information processing device according to Supplementary Note 1, wherein the planning means sequentially selects samples with an optimal cost-to-benefit ratio using a greedy method.

[0066] (Supplementary Note 4) The information processing device according to Supplementary Note 1, wherein the utility function estimation means constructs a utility function that represents the relationship between the information gain and the transaction price.

[0067] (Supplementary Note 5) The information processing device according to Supplementary Note 1, wherein the negotiation means performs negotiation according to a time-dependent strategy or a strategy based on deep reinforcement learning.

[0068] (Supplementary Note 6) An information processing method that is executed by a computer, acquires a sample set of target information, calculates the information gain of the sample set, estimates the cost required to acquire the sample set, formulates an optimal plan to obtain a sample set that maximizes the information gain under constraints on the cost, estimates a utility function to maximize the optimal value obtained by achieving the optimal plan, and executes negotiations to optimize the trading conditions of the samples based on the utility function.

[0069] (Supplementary Note 7) A recording medium having recorded thereon a program that causes a computer to execute the following processes: acquire a sample set of target information, calculate the information gain of the sample set, estimate the cost required to acquire the sample set, formulate an optimal plan to obtain a sample set that maximizes the information gain under constraints on the cost, estimate a utility function to maximize the optimal value obtained by achieving the optimal plan, and execute negotiations to optimize the trading conditions of the samples based on the utility function.

[0070] Although the present disclosure has been described above with reference to embodiments and examples, the present disclosure is not limited to the above embodiments and examples. Various modifications that would be understood by those skilled in the art can be made to the configuration and details of the present disclosure within the scope of the present disclosure. In other words, the present invention naturally includes various modifications and alterations that would be possible for those skilled in the art based on the entire disclosure, including the claims, and the technical ideas. Furthermore, the disclosures of the above-cited patent documents and the like are incorporated herein by reference.

[0071] This application claims priority based on Japanese Patent Application No. 2014-27272, filed February 27, 2024, the disclosure of which is incorporated herein by reference in its entirety.

[0072] REFERENCE SIGNS LIST 11 Processor 21 Automatic negotiation unit 22 Utility function estimation unit 23 Information gain calculation unit 24 Planning unit 25 Cost estimation unit 100 Information trading device

Claims

1. An information processing device comprising: an information gain calculation means for acquiring a sample set of target information and calculating the information gain of the sample set; a cost estimation means for estimating the costs required to acquire the sample set; a planning means for formulating an optimal plan for determining a sample set that maximizes the information gain under constraints related to the costs; a utility function estimation means for estimating a utility function for maximizing the optimal value obtained by achieving the optimal plan; and a negotiation means for negotiating to optimize the trading conditions of the samples based on the utility function.

2. The information processing device according to claim 1, wherein said information gain calculation means calculates the informational value of the number of samples obtained from the sample subject using a Bayesian approach.

3. The information processing apparatus according to claim 1, wherein said planning means uses a greedy algorithm to sequentially select samples with the optimum cost-to-benefit ratio.

4. The information processing device according to claim 1, wherein said utility function estimation means constructs a utility function that represents the relationship between said information gain and transaction price.

5. The information processing device according to claim 1, wherein the negotiation means performs negotiation according to a time-dependent strategy or a strategy based on deep reinforcement learning.

6. An information processing method that is executed by a computer, acquires a sample set of target information, calculates the information gain of the sample set, estimates the cost required to acquire the sample set, formulates an optimal plan to obtain a sample set that maximizes the information gain under constraints related to the cost, estimates a utility function to maximize the optimal value obtained by achieving the optimal plan, and conducts negotiations to optimize the trading conditions of the samples based on the utility function.

7. A recording medium having recorded thereon a program that causes a computer to execute the following processes: acquire a sample set of target information, calculate the information gain of the sample set, estimate the cost required to acquire the sample set, formulate an optimal plan to obtain a sample set that maximizes the information gain under constraints related to the cost, estimate a utility function to maximize the optimal value obtained by achieving the optimal plan, and negotiate to optimize the trading conditions of the samples based on the utility function.

Images