Training meta-learning models based on causal transportability between datasets
By training meta-learning models using feature-dependent and difference information, the system addresses bias and ethical concerns in imbalanced datasets, ensuring accurate and ethical decision-making across domains without direct access to individual datasets.
Patent Information
- Application Number
- JP2023035801
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2022-03-15
- Filing Date
- 2023-03-08
- Publication Date
- 2026-11-11
- Estimated Expiration
- 2043-03-08
AI Technical Summary
Meta-learning models can become biased when trained on datasets that are imbalanced, leading to inaccurate results and unethical conclusions, and conventional systems require direct access to individual datasets, making them susceptible to distribution shifts and ethical concerns.
A system that trains meta-learning models based on causal transportability between datasets using feature dependency and difference information, determining structural causal models and ethical coefficients, and reliability scores without direct access to individual datasets, incorporating expert knowledge to address bias and ethical concerns.
The system reduces susceptibility to bias and ethical issues by training meta-learning models based on reliability scores and expert input, ensuring accurate and ethical decision-making across multiple domains.
Smart Images

Figure 0007948037000004 
Figure 0007948037000005 
Figure 0007948037000006
Abstract
Description
[Technical Field]
[0001] The embodiments discussed in this disclosure relate to training meta-learning models based on causal transportability between datasets. [Background technology]
[0002] Advances in the fields of machine learning and artificial intelligence (AI) have led to the development of meta-learning models that can be used for a variety of prediction tasks. Typically, such meta-learning models can be trained on one type of prediction task and then applied to other types of prediction tasks. For example, a meta-learning model may be trained on a binary classification task that distinguishes cat images from dog images, and even flower images from motorcycle images. Furthermore, unrelated datasets can be used to test meta-learning models. For example, a meta-learning model may be tested on a task that separates dog images from otter images. However, in certain cases, one or more training datasets may be biased towards one class (e.g., the dog class). For example, one or more training datasets may contain a significant number of dog images compared to cat images. In such cases, the meta-learning model may become biased in an unethical way. Such a biased meta-learning model may produce inaccurate results, which could lead to undesirable conclusions.
[0003] The scope claimed in this disclosure is not limited to embodiments that operate only in the environments described above or embodiments that resolve any defects. Rather, this background is provided solely to illustrate an example of the technical scope in which some of the embodiments described in this disclosure may be implemented. [Overview of the project]
[0004] In one embodiment, the method may include a set of actions that may include receiving a plurality of datasets relating to a plurality of application domains. The set of actions may further include determining feature dependency information related to a first dataset among the plurality of received datasets, based on a first user input. The set of actions may further include determining feature difference information related to the first dataset and a second dataset among the plurality of received datasets, based on a second user input and a set of ethical requirements, where the second dataset may be different from the first dataset. The set of actions may further include determining a set of structural causal models (SCMs) related to the first dataset, based on the determined feature dependency information and the determined feature difference information. The set of actions may further include determining a set of ethical coefficients related to the set of ethical requirements, based on the application of a causal transportability model to the determined set of SCMs. The set of actions may further include determining a reliability score related to the first dataset, based on the determined set of ethical coefficients. The set of actions may further include training a meta-learning model related to multiple application domains based on multiple confidence scores associated with multiple datasets, where the multiple confidence scores may include determined confidence scores associated with a first dataset.
[0005] The objectives and advantages of the embodiments are realized and achieved by at least the elements, features, and combinations specifically indicated in the claims.
[0006] Both the above summary and the following detailed description are given as examples and are illustrative, and not limitations on the claimed invention.
[0007] Exemplary embodiments are described and explained with further identification and detail through the use of the accompanying drawings. [Brief explanation of the drawing]
[0008] [Figure 1] This diagram illustrates an example of an environment relevant to training meta-learning models based on causal transportability between datasets. [Figure 2] This block diagram represents an exemplary electronic device for training a meta-learning model based on causal transportability between datasets. [Figure 3] This diagram illustrates an exemplary structural causal model (SCM) related to pairs of datasets corresponding to application domains. [Figure 4] This diagram illustrates a flowchart of an exemplary method for training a meta-learning model based on causal transportability between datasets. [Figure 5] This diagram illustrates an exemplary scenario for training a meta-learning model based on causal transportability between datasets. [Figure 6] This diagram illustrates an exemplary scenario for training a meta-learning model based on causal transportability between datasets. [Figure 7] This diagram illustrates an exemplary structural causal model (SCM) associated with pairs of datasets corresponding to a banking transaction application domain. [Modes for carrying out the invention]
[0009] All figures are in accordance with at least one embodiment described herein.
[0010] Some embodiments described in this disclosure relate to methods and systems for training meta-learning models based on causal transportability between datasets. In this disclosure, multiple datasets relating to multiple application domains may be received. Based on a first user input, feature dependency information related to the first dataset among the received multiple datasets may be determined. Furthermore, based on a second user input and a set of ethical requirements, feature difference information related to the first dataset and a second dataset among the received multiple datasets may be determined, where the second dataset may be different from the first dataset. Examples of sets of ethical requirements may include, but are not limited to, sets of fairness requirements, sets of privacy requirements, or sets of security requirements. Subsequently, a set of structural causal models (SCMs) related to the first dataset may be determined based on the determined feature dependency information and the determined feature difference information. Furthermore, a set of ethical coefficients related to the set of ethical requirements may be determined based on the application of a causal transportability model to the determined set of SCMs. A reliability score associated with the first dataset may be determined based on a set of determined ethical coefficients. Furthermore, a meta-learning model associated with multiple application domains may be trained based on multiple reliability scores associated with multiple datasets, where these multiple reliability scores may include the determined reliability score associated with the first dataset.
[0011] In accordance with one or more embodiments of this disclosure, the field of artificial intelligence (AI) models can be improved by configuring a computing system to have the ability to train meta-learning models based on causal transportability between datasets. The computing system may receive a plurality of datasets relating to a plurality of application domains. Based on a first user input, the computing system may determine feature dependency information relating to a first dataset among the plurality of received datasets. Furthermore, based on a second user input and a set of ethical requirements, the computing system may determine feature difference information relating to the first dataset and a second dataset among the plurality of received datasets, where the second dataset may be different from the first dataset. Examples of sets of ethical requirements may include, but are not limited to, sets of fairness requirements, sets of privacy requirements, or sets of security requirements. The computing system may determine a set of structural causal models (SCMs) relating to the first dataset based on the determined feature dependency information and the determined feature difference information. Furthermore, the computing system may determine a set of ethical coefficients wp associated with a set of ethical requirements, based on the application of a causal transportability model to the determined set of SCMs. In this way, causal effects related to multiple datasets and differences between various attributes across multiple datasets can be captured using SCMs based on feature dependency information and feature difference information. The computing system may determine a reliability score related to the first dataset based on the determined set of ethical coefficients. Furthermore, the computing system may train a meta-learning model related to multiple application domains based on multiple reliability scores related to multiple datasets, where the multiple reliability scores may include the determined reliability score related to the first dataset. In this way, the contribution of each dataset directed towards training the meta-learning model can be adapted based on the reliability score determined for that dataset.Since reliability scores may take expert knowledge into account, meta-learning models trained on multiple reliability scores can incorporate ethical requirements. Furthermore, undesirable consequences of dataset distribution shifts (e.g., bias) and ethical concerns (e.g., privacy and safety concerns) can be addressed by applying causal transportability models to SCM sets.
[0012] Conventional systems typically require access to individual datasets to train meta-learning models. The disclosed system, on the other hand, does not require access to individual datasets and can train meta-learning models by relying on feature-dependent and feature-diff information associated with one or more datasets. This feature-dependent and feature-diff information may be provided by expert users. The disclosed system can determine a set of SCMs based on the feature-dependent and feature-diff information. Furthermore, a causal transportability model may be applied to the SCM set to determine a confidence score associated with the datasets. The meta-learning model can then be trained based on the determined confidence score. Because individual datasets do not need to be directly accessed, meta-learning models trained using the disclosed system are not affected by distribution shifts associated with individual datasets. Moreover, unlike conventional systems, meta-learning models trained using the disclosed system are less susceptible to bias and ethical concerns because the disclosed system can include humans in its loops and rely on domain-specific knowledge from expert users.
[0013] Embodiments of this disclosure will be described with reference to the accompanying drawings.
[0014] Figure 1 is a diagram illustrating an example environment related to training a meta-learning model based on causal transportability between datasets, arranged in at least one embodiment described in the present disclosure. Referring to Figure 1, an environment 100 is depicted. The environment 100 may include an electronic device 102, a database 104, a user end device 106, and a communication network 112. The electronic device, the database 104, and the user end device 106 may be communicatively coupled to each other via the communication network 112. The electronic device 102 may include a set 108 of structural causal models (SCMs) and a meta-learning model 110. Figure 1 further illustrates a user 116 that is associated with or capable of operating the electronic device 102 or the user end device 106. A plurality of datasets 114 that may be associated with a plurality of application domains is further illustrated. The plurality of application domains may include a first application domain, a second application domain, ..., and an N-th application domain. The plurality of datasets 114 may include a first dataset 114A associated with the first application domain, a second dataset 114B associated with the second application domain, ..., and an N-th dataset 114N associated with the N-th application domain. The plurality of datasets 114 may be stored in the database 104.
[0015] The number N of datasets shown in Figure 1 is presented merely as an example. The plurality of datasets 114 may include only two datasets or more than N datasets without departing from the scope of the present disclosure. For the sake of brevity, only N datasets are shown in Figure 1. However, in some embodiments, there may be more than N datasets without limiting the scope of the present disclosure.
[0016] Although the set of SCMs 108 and the meta-learning model 110 are illustrated as being included in or stored in the electronic device 102, one or more of the set of SCMs 108 and the meta-learning model 110 may be disposed in a separate device, for example, a server (not shown in FIG. 1), the database 104, or the user end device 106, without departing from the scope of the present disclosure.
[0017] The electronic device 102 may include suitable logic, circuitry, interfaces, and / or code configured to train the meta-learning model 110 based on application of a causal transportability model to the set of SCMs 108, as described herein. In certain embodiments, the electronic device 102 may be configured to store the set of SCMs 108 and the meta-learning model 110, as shown in FIG. 1.
[0018] The electronic device 102 may be configured to receive a plurality of datasets 114 related to a plurality of application domains. The electronic device 102 may receive a first user input and a second user input from one or more users, for example, a user 116. In one example, the user 116 may be an expert who may have domain knowledge of the plurality of application domains. The electronic device 102 may be configured to determine feature dependency information associated with a first dataset (for example, the first dataset 114A) among the plurality of received datasets 114, based on the first user input. The determined feature dependency information may include, but is not limited to, at least one of a set of input variables associated with the first dataset 114A, a set of output variables associated with the first dataset 114A, dependency information associated with the set of input variables and the set of output variables, or a confidence level associated with the dependency information. The determination of the feature dependency information is further described, for example, in FIG. 4.
[0019] The electronic device 102 may be configured to determine feature difference information related to the first dataset 114A and the second dataset (e.g., the second dataset 114B) among the multiple datasets received, based on a set of second user input and ethical requirements. The determined feature difference information may correspond to inter-population differences between the first set of features related to the first dataset 114A and the second set of features related to the second dataset 114B. Here, the first set of features may correspond to the second set of features, and the second dataset 114B may be different from the first dataset 114A. For example, the first dataset 114A may include loan approval information related to a first set of individuals in a specific demographic and location, and the second dataset 114B may include loan approval information related to a second set of individuals in other demographic information and other locations. The determination of feature difference information is further described, for example, in Figure 4.
[0020] The electronic device 102 may be configured to determine a set of SCMs (e.g., a set of SCMs 108) associated with the first dataset 114A based on determined feature dependency information and determined feature difference information. The electronic device 102 may be further configured to determine a set of ethical coefficients associated with a set of ethical requirements based on the application of a causal transportability model to the determined set of SCMs 108. Examples of ethical requirements include, but are not limited to, a set of fairness requirements, a set of privacy requirements, or a set of safety requirements. The determination of the set of SCMs 108 and the set of ethical coefficients are further described, for example, in Figure 4. Examples of SCMs are shown in Figures 3 and 7.
[0021] The electronic device 102 may be configured to determine a reliability score associated with a first dataset 114A based on a determined set of ethical coefficients. The electronic device 102 may be further configured to train a meta-learning model (e.g., meta-learning model 110) associated with multiple application domains. Training of the meta-learning model 110 may be based on multiple reliability scores associated with multiple datasets 114, where the multiple reliability scores may include the determined reliability score associated with the first dataset 114A. The meta-learning model 110 may correspond to a generative learning model that includes, but is not limited to, at least one of a metric-based learning model, an optimization-based learning model, or a model-based learning model. The determination of the reliability score and the training of the meta-learning model 110 are further described, for example, in Figure 4.
[0022] Examples of electronic devices 102 may include, but are not limited to, causal analysis machines, meta-learning machines, mobile devices, desktop computers, laptops, computer workstations, computing devices, mainframe machines, servers, such as cloud servers, and groups of servers. In one or more embodiments, electronic device 102 may include user end terminal devices and servers coupled to the user end terminal devices for communication. Electronic device 102 may be implemented by hardware including a processor, a microprocessor (e.g., one that performs or controls one or more operations), a field-programmable gate array (FPGA), or an application-specific integrated circuit (ASIC). In some other cases, electronic device 102 may be implemented by a combination of hardware and software.
[0023] The database 104 may include appropriate logic, interfaces, and / or code that can be configured to store multiple datasets 114. In certain embodiments, a set of SCMs 108 and a meta-learning model 110 may be stored in the database 104. In other embodiments, datasets associated with the meta-learning model 110 (e.g., training datasets or test datasets) may be stored in the database 104. The database 104 may be derived from data from relational or non-relational databases, or from a set of comma-separated values (CSV) in conventional or big data storage.
[0024] The database 104 may be stored or cached in a device, such as a server or electronic device 102. The device storing the database 104 may be configured to receive queries from the electronic device 102 for a particular dataset (e.g., a first dataset 114A) among a plurality of datasets 114. In response, the database 104 device may be configured to read the queried database (e.g., the first dataset 114A) based on the received query and supply it to the electronic device 102. In some embodiments, the database 104 may be hosted on multiple servers stored in the same or different locations. The operation of the database 104 may be performed by hardware, including a processor, a microprocessor (e.g., one that performs or controls one or more operations), a field-programmable gate array (FPGA), or an application-specific integrated circuit (ASIC). In some other cases, the database 104 may be implemented in software.
[0025] The user end device 106 may include appropriate logic, circuitry, interfaces, and / or code that can be configured to render decision-making related to the context of an application domain to a display device (accompanying the user end device 106). The decision-making may be determined (e.g., by an electronic device 102) based on the application of a trained meta-learning model 110 to context-related inputs. For example, the context may be a loan approval scenario in a financial services application, and the inputs may include demographic data and personal financial data. The decision-making may correspond to automated decision-making indicating whether or not an individual's loan should be approved. For example, the user end device 106 may include web client software or email software, through which the user end device 106 may receive multiple datasets 114. For example, each of the multiple datasets 114 may be a comma-separated values (CSV) data file. Additionally or alternatively, the user end device 106 may include spreadsheet software that can be used to generate or edit the multiple datasets 114. The user end device 106 may upload multiple datasets 114 that have been generated, edited, or received to the electronic device 102. Furthermore, the user end device 106 may upload multiple datasets 114 to the database 104 for storage. In one embodiment, the user end device 106 may receive a query from the user 116 to train a meta-learning model 110 based on the multiple datasets 114. The user end device 106 may further send the query to the electronic device 102 to start training the meta-learning model 110. The user end device 106 may receive a first user input related to feature dependency information and a second user input related to feature difference information (for example, from the user 116). The user end device 106 may send the received first and second user inputs to the electronic device 102.Examples of user end devices 106 include, but are not limited to, mobile devices, desktop computers, laptops, computer workstations, computing devices, mainframe machines, servers, such as cloud servers, and groups of servers. In Figure 1, the user end device 106 is separated from the electronic device 102, however, in some embodiments, the user end device 106 may be incorporated into the electronic device 102 without exceeding the scope of this disclosure.
[0026] Each of the SCM sets 108 may include appropriate logic, circuits, interfaces, and / or code configured to represent causal relationships between variables associated with a particular dataset among a plurality of datasets 114. For example, each of the SCM sets 108 may correspond to a directed acyclic graph that can represent variables associated with a first dataset 114A. The directed acyclic graph may be augmented based on a set of selection variables that represent determined feature difference information. In an embodiment, the population associated with the first dataset 114A may be switched to a population associated with a second dataset 114B based on the application of a set of conditions to at least one different value in the set of selection variables. The set of selection variables may be determined based on user input (e.g., second user input) that may be received from a user, e.g., user 116. The user may be a domain knowledge expert associated with one or more of a plurality of application domains. The directed acyclic graph may show the causal effect of a set of input variables associated with the first dataset 114A on a set of output variables associated with the first dataset 114A. Furthermore, the directed acyclic graph may also include a set of unobserved variables that influence a set of input variables associated with the first dataset. Examples of SCMs are shown in Figures 3 and 7.
[0027] The meta-learning model 110 may be a meta-learning model, such as a neural network model, that can be configured to make decisions related to a particular application domain. In embodiments, the meta-learning model 110 may be trained on a dataset related to one application domain and may be tested or validated on other datasets that may be related to other application domains. Thus, the meta-learning model 110 can be used for predictions on data from various application domains. In other words, the training of the meta-learning model 110 may be transferable from a first application domain (where the meta-learning model 110 may be trained) to a second application domain (where the meta-learning model 110 may be used for predictions).
[0028] A neural network model (which may correspond to the meta-learning model 110) may be a system or computational network of artificial neurons arranged as nodes in multiple layers. The multiple layers of each neural network model may include an input layer, one or more hidden layers, and an output layer. Each of the multiple layers may include one or more nodes (or artificial neurons, e.g., represented by circles). The outputs of all nodes in the input layer may be coupled to at least one node in the hidden layer. Similarly, the inputs of each hidden layer may be coupled to the outputs of at least one node in the other layers of the neural network model. The outputs of each hidden layer may be coupled to the inputs of at least one node in the other layers of the neural network model. The nodes of the final layer may receive input from at least one hidden layer and output a result. The number of layers and the number of nodes in each layer may be determined from the hyperparameters of the neural network model. Such hyperparameters may be set before or during training the neural network model with training data.
[0029] Each node in a neural network model may correspond to a mathematical function (e.g., a sigmoid function or a normalized linear unit) that includes a set of parameters that can be tuned during training of the neural network model. The set of parameters may include, for example, weight parameters, regularization parameters, etc. Each node may use the above mathematical function to compute an output based on one or more inputs from nodes in other layers of the neural network model (e.g., previous layers). All or some of the nodes in a neural network model may correspond to the same or different mathematical functions.
[0030] In training a neural network model, one or more parameters of each node in the neural network model may be updated based on whether the output of the final layer for a given input from the training dataset matches the exact result based on the neural network model's loss function. The above process may be repeated for the same or different inputs until the minimum value of the loss function can be achieved, and training errors can be minimized. Several training methods, such as gradient descent, stochastic gradient descent, batch gradient descent, gradient boosting, and metaheuristics, are known in this technique.
[0031] A neural network model may include electronic data, such as software programs, code, libraries, applications, scripts, or other logic or instructions executed by a processing device (e.g., the processor 204 of electronic device 102 in Figure 2). A neural network model may also include code and routines configured to enable a computing device, including a processor, to perform one or more tasks, such as making decisions related to a specific application domain (e.g., a banking domain) (e.g., loan approval decisions). Additionally, or alternatively, a neural network model may be implemented using hardware including a processor, a microprocessor (e.g., one that performs or controls one or more operations), a field-programmable gate array (FPGA), or an application-specific integrated circuit (AISC). Alternatively, in some embodiments, a neural network model may be implemented using a combination of hardware and software.
[0032] Examples of neural network models (e.g., meta-learning model 110) may include, but are not limited to, deep neural networks (DNNs), convolutional neural networks (CNNs), artificial neural networks (ANNs), fully connected neural networks, deep Bayesian neural networks, and / or combinations of such networks. In some embodiments, the neural network model may include numerical computation techniques that use a dataflow graph. In certain embodiments, the neural network model may be based on a hybrid architecture of multiple deep neural networks (DNNs).
[0033] The communication network 112 may include a communication medium through which the electronic device 102 can communicate with a server capable of storing a database 104 and a user end device 106. Examples of the communication network 112 may include, but are not limited to, the Internet, a cloud network, a Wireless Fidelity (Wi-Fi) network, a Personal Area Network (PAN), a Local Area Network (LAN), and / or a Metropolitan Area Network (MAN). Various devices within the environment 100 may be configured to connect to the communication network 112 according to various wired or wireless communication protocols. Examples of such wired and wireless communication protocols may include, but are not limited to, the Transmission Control Protocol and Internet Protocol (TCP / IP), User Data Graph Protocol (UDP), Hypertext Transfer Protocol (HTTP), File Transfer Protocol (FTP), ZigBee, EDGE, IEEE 802.11, Light Fidelity (Li-Fi), 802.16, IEEE 802.11s, IEEE 802.11g, Multihop Communication, Wireless Access Point (AP), Device-to-Device (D2D) Communication, Cellular Communication Protocol, and / or Bluetooth (BT) Communication Protocol, or at least one of these combinations.
[0034] Without exceeding the scope of this disclosure, modifications, additions, or deletions may be made to Figure 1. For example, the environment 100 may include more or fewer elements than those illustrated and described in this disclosure. For example, in some embodiments, the environment 100 may include the electronic device 102 but not the database 104 and the user end device 106. Furthermore, in some embodiments, the respective functions of the database 104 and the user end device 106 may be incorporated into the electronic device 102 without exceeding the scope of this disclosure.
[0035] Figure 2 is a block diagram representing an exemplary electronic device for training a meta-learning model based on causal transportability between databases, arranged according to at least one embodiment described herein. Figure 2 is described in relation to elements from Figure 1. Referring to Figure 2, a block diagram 200 of a system 202 including an electronic device 102 is shown. The electronic device 102 may include a processor 204, memory 206, persistent data storage 208, input / output (I / O) devices 210, a display screen 212, a network interface 214, a set of SCMs 108, and a meta-learning model 110.
[0036] The processor 204 may include appropriate logic, circuits, and / or interfaces that can be configured to execute program instructions related to various operations performed by the electronic device 102. For example, some operations may include receiving multiple datasets 114, determining feature dependency information, determining feature difference information, determining a set of SCMs 108, determining a set of ethical coefficients, determining a reliability score, and training a meta-learning model 110. The processor 204 may also include any appropriate dedicated or general-purpose computer, computing entity, or processing device, including various computer hardware or software modules, which may be configured to execute instructions stored in any applicable computer-readable storage medium. For example, the processor 204 may include a microprocessor, microcontroller, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or any other digital or analog circuit configured to interpret and / or execute program instructions and / or process data.
[0037] Although represented as a single processor in Figure 2, processor 204 may include any number of processors configured to individually or collectively perform or direct any number of operations of electronic device 102, as described in this disclosure. Furthermore, one or more processors may reside in one or more different electronic devices, such as different servers. In some embodiments, processor 204 may be configured to interpret and / or execute program instructions stored in memory 206 and / or persistent data storage 208 and / or to process stored data. In some embodiments, processor 204 may fetch program instructions from persistent data storage 208 and load the program instructions into memory 206. After the program instructions are loaded into memory 206, processor 204 may execute the program instructions. Some examples of processor 204 may be graphics processing units (GPUs), central processing units (CPUs), reduced instruction set computer (RISC) processors, ASIC processors, multiple instruction set computer (CISC) processors, coprocessors, and / or combinations thereof.
[0038] Memory 206 may include appropriate logic, circuitry, interfaces, and / or code that can be configured to store program instructions executable by the processor 204. In certain embodiments, memory 206 may be configured to store operating system and associated application-specific information. Memory 206 may include a computer-readable storage medium that carries or stores computer-executable instructions or data structures. Such a computer-readable storage medium may include any available medium that can be accessed by a general-purpose or dedicated computer, such as the processor 204. For example, and not as an limitation, such computer-readable storage media may include tangible or non-temporary computer-readable storage media, including random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), compact disk-read-only memory (CD-ROM) or other optical disk storage, magnetic disk storage or other magnetic storage devices, flash memory devices (e.g., solid-state memory devices), or any other storage media that may be used to carry or store specific program code in the form of computer-executable instructions or data structures and that can be accessed by a general-purpose or dedicated computer. Combinations of the above may also be included within the scope of computer-readable storage media. Computer-executable instructions may include, for example, instructions and data configured to cause a processor 204 to perform a specific operation or group of operations related to an electronic device 102.
[0039] The persistent data storage 208 may include appropriate logic, circuitry, interfaces, and / or code that can be configured to store program instructions, operating systems, and / or application-specific information, such as logs and application-specific databases, that can be executed by the processor 204. The persistent data storage 208 may include a computer-readable storage medium that carries or stores computer-executable instructions or data structures. Such a computer-readable storage medium may include any available medium that can be accessed by a general-purpose or dedicated computer, such as the processor 204.
[0040] For example, and not as an limitation, such computer-readable storage media may include tangible or non-temporary computer-readable storage media, including compact disk-read-only memory (CD-ROM) or other optical disk storage, magnetic disk storage or other magnetic storage devices (e.g., hard disk drives (HDDs)), flash memory devices (e.g., solid-state drives (SSDs), secure digital (SD) cards, other solid-state memory devices), or any other storage media that may be used to carry or store specific program code in the form of computer-executable instructions or data structures and that can be accessed by a general-purpose or dedicated computer. Combinations of the above may also be included within the scope of computer-readable storage media. Computer-executable instructions may include, for example, instructions and data configured to cause a processor 204 to perform a specific operation or group of operations related to an electronic device 102.
[0041] In some embodiments, either the memory 206 or the persistent data storage 208, or a combination thereof, may store the received datasets 114, determined feature dependency information, determined feature difference information, determined ethical coefficient sets, and determined reliability scores. Either the memory 206 or the persistent data storage 208, or a combination thereof, may further store the SCM sets 108 and the meta-learning model 110.
[0042] The I / O device 210 may include appropriate logic, circuitry, interfaces, and / or code that can be configured to receive user input. For example, the I / O device 210 may receive user input indicating a query to train a meta-learning model 110. In another example, the I / O device 210 may receive a first user input corresponding to feature-dependent information. The I / O device 210 may further receive a second user input corresponding to feature-difficulty information. The I / O device 210 may further receive user input including a query that determines a decision related to a particular application domain based on the application of the meta-learning model 110 to specific input data. The I / O device 210 may further be configured to supply an output in response to the received user input. For example, the I / O device 210 may render an automated decision output (which may be determined by the trained meta-learning model 110 of the electronic device 102) on a display screen 212. The I / O device 210 may include various input and output devices that can be configured to communicate with the processor 204 and other components, such as the network interface 214. Examples of input devices may include, but are not limited to, a touchscreen, keyboard, mouse, joystick, and / or microphone. Examples of output devices may include, but are not limited to, a display (e.g., display screen 212) and a speaker.
[0043] The display screen 212 may have appropriate logic, circuitry, interfaces, and / or code that can be configured to display the generated first image. The display screen 212 may be configured to receive user input from user 116 (e.g., queries to train the meta-learning model 110, first user input, second user input, and / or queries to determine decisions relating to a particular application domain). In such cases, the display screen 212 may be a touchscreen that receives user input. The display screen 212 may be implemented through several known technologies, including but not limited to liquid crystal displays (LCDs), light-emitting diode (LEDs), plasma displays, and / or organic LED (OLED) display technologies, and / or other display technologies.
[0044] The network interface 214 may include appropriate logic, circuitry, interfaces, and / or code that can be configured to establish communication between the electronic device 102 via the communication network 112, the server / device of the database 104, and the user end device 106. The network interface 214 may be implemented by the use of various known techniques to support wired or wireless communication of the electronic device 102 via the communication network 112. The network interface 214 may include, but is not limited to, an antenna, a radio frequency (RF) transceiver, one or more amplifiers, a tuner, one or more oscillators, a digital signal processor, a coder-decoder (CODEC) chipset, a subscriber identification module (SIM) card, and / or a local buffer.
[0045] Without departing from the scope of this disclosure, the exemplary electronic device 102 may be modified, added to, or deleted. For example, in some embodiments, the exemplary electronic device 102 may include any number of other components, which for the sake of brevity may not be expressly illustrated or described.
[0046] Figure 3 is a diagram representing an exemplary structural causal model (SCM) related to a pair of datasets corresponding to an application domain, according to embodiments of the present disclosure. Figure 3 is described in relation to elements from Figures 1 and 2. Referring to Figure 3, an exemplary structural causal model (SCM) 300 is shown. The SCM 300 may correspond to an application domain, for example, an online product recommendation application that may involve individuals with disabilities.
[0047] SCM300 may relate to a pair of datasets associated with an online product recommendation application domain. For example, SCM300 may relate to a first dataset 114A and a second dataset 114B. SCM300 may correspond to a directed acyclic graph representing variables associated with a dataset (e.g., the first dataset 114A). The directed acyclic graph of SCM300 may include multiple nodes, e.g., a set of first nodes, a set of second nodes, a set of third nodes, and a set of selection nodes. The directed acyclic graph may show the causal effect of a set of input variables associated with the first dataset 114A on a set of output variables associated with the first dataset 114A. The set of first nodes may include a first node 302A that may represent a user click variable (represented as "uc") and a second node 302B that may represent a product recommendation variable (represented as "pr"). The user click variable (represented by "uc") may be an input variable, and the product recommendation variable (represented by "pr") may be an output variable. The second set of nodes may include a first node 304A that may represent a physical condition variable (represented by "pc") and a second node 304B that may represent a user history variable (represented by "uh"). Both the physical condition variable (represented by "pc") and the user history variable (represented by "uh") may be input variables. The directed acyclic graph of SCM300 may further include a set of unobserved variables that may influence the set of input variables associated with the first dataset 114A. The third set of nodes may include a first node 306A that may represent a first unobserved variable (represented by "Ux") that may influence the user click variable (represented by "uc"). For example, the first unobserved variable may correspond to a user's purchase intention, which does not need to be directly observed. The third set of nodes may further include a second node 306B which may represent a second unobserved variable (represented as "Uc") corresponding to a product characteristic that may influence the physical condition variable (represented as "pc").
[0048] The directed acyclic graph of SCM300 can be augmented based on a set of selection variables that indicate determined feature difference information between a first dataset 114A and a second dataset 114B. Here, a population associated with the first dataset 114A can be switched to a population associated with the second dataset 114B based on the application of a set of conditions to at least one different value in the set of selection variables. For example, the set of selection nodes corresponding to SCM300 may include node 308, which may indicate that a physical condition variable (denoted as "pc") can be a selection variable that can switch a population in the first dataset 114A to a population in the second dataset 114B. For example, the first dataset 114A and the second dataset 114B may contain differences based on an individual's physical condition. In other words, the physical condition variable may differ across datasets such as the first dataset 114A and the second dataset 114B. Furthermore, as shown in Figure 3, a particular node in SCM300 may be a rectangular node. Rectangular nodes can indicate that the variables represented by such nodes may differ among different individuals within the same dataset. For example, the user history variable (represented as "uh") and the physical condition variable (represented as "pc") may differ among different individuals within the first dataset 114A. An example of an SCM related to a banking transaction application domain is further described, for example, in Figure 7.
[0049] It should be noted that the SCM300 shown in Figure 3 is for illustrative purposes only and should not be intended to limit the scope of this disclosure.
[0050] Figure 4 is a flowchart illustrating an example of a method for training a meta-learning model based on causal transportability between datasets, according to embodiments of the present disclosure. Figure 4 is described in relation to elements from Figures 1, 2, and 3. Referring to Figure 4, flowchart 400 is shown. The method represented in flowchart 400 may begin at 402 and may be performed by any suitable system, apparatus, or device, for example, by the electronic device 102 in the example of Figure 1 or the processor 204 in Figure 2. Steps and operations related to one or more of the blocks of flowchart 400, although represented in separate blocks, may be divided into additional blocks, combined into fewer blocks, or deleted, depending on the particular implementation.
[0051] In block 402, multiple datasets 114 relating to multiple application domains may be received. In an embodiment, the processor 204 may receive multiple datasets 114 relating to multiple application domains. For example, the multiple datasets 114 may include a first dataset 114A that may correspond to a banking transaction domain associated with a first location. Furthermore, the multiple datasets 114 may include a second dataset 114B that may correspond to a banking transaction domain associated with a second location. In another example, the first dataset 114A may correspond to a banking transaction domain associated with a first location, and the second dataset 114B may correspond to an insurance domain associated with a first location. In a particular scenario, the processor 204 may receive multiple datasets 114 from the database 104. Upon receiving the multiple datasets 114, the processor 204 may store the received multiple datasets 114 in memory 206. In other scenarios, multiple datasets 114 may be pre-stored in memory 206, and the processor 204 may read the pre-stored datasets 114 from memory 206.
[0052] In block 404, feature dependency information related to the first dataset 114A among the multiple datasets 114 received may be determined based on the first user input. In an embodiment, the processor 204 may be configured to determine feature dependency information related to the first dataset 114A based on the first user input. In an embodiment, the feature dependency information may include, but is not limited to, at least one of the following: a set of input variables related to the first dataset 114A, a set of output variables related to the first dataset 114A, dependency information related to the set of input variables and the set of output variables, or a confidence level related to the dependency information. For example, the processor 204 may receive a first user input which may include a set of input variables and a set of output variables related to the first dataset 114A. As an example, if the first dataset 114A relates to a banking transaction domain, the set of input variables may include an income variable, an age variable, a gender variable, and a credit score variable. Furthermore, in such a case, the set of output variables may include a loan decision variable (which may be a binary variable representing the predictive output of the meta-learning model 110). The first user input received may further include dependency information related to the input variable set and the output variable set. This dependency information may indicate whether a particular input variable in the input variable set influences a particular output variable in the output variable set. The first user input may also include confidence levels related to the dependency information. These confidence levels may indicate the level of certainty that user 116 may have regarding the dependency information. In other words, a higher confidence level indicates a greater certainty that user 116 has that a particular input variable influences a particular output variable in the first dataset 114A. For example, user 116 may have a 90% confidence level that an individual's income and credit score (two input variables) influence the loan decision output (output variable) associated with the first dataset 114A corresponding to the banking transaction domain. Therefore, in a particular case, the dependency information may indicate that the income variable and credit score variable may influence the loan decision variable, and that a 90% confidence level may exist associated with such dependency information.
[0053] The first user input may be received from an expert user (e.g., user 116) who may possess domain knowledge related to one or more of the application domains. For example, the expert user may be a financial consultant with knowledge of the banking transaction domain. Furthermore, since the first user input may include confidence levels related to the dependency information, the degree of reliability of the dependency information is quantifiable. Therefore, the first user input (from the expert user) can be accurate and useful for determining the feature dependency information.
[0054] In block 406, feature difference information related to the first dataset 114A and the second dataset 114B among the multiple datasets 114 can be determined based on a set of second user input and ethical requirements. Here, the first dataset 114A may be different from the second dataset 114B. For example, the first dataset 114A may correspond to a banking transaction domain associated with a first location, and the second dataset 114B may correspond to a banking transaction domain associated with a second location different from the first location. As an example, the second location may be in a different region from the first location. In an embodiment, the processor 204 may be configured to determine feature difference information related to the first dataset 114A and the second dataset 114B based on a set of second user input and ethical requirements. The feature difference information may correspond to differences between sets of first features associated with the first dataset 114A and sets of second features associated with the second dataset 114B. Furthermore, the set of first features may correspond to the set of second features. For example, the second user input may include one or more input variables that may differ across the first dataset 114A and the second dataset 114B. Consider a scenario where the first dataset 114A corresponds to a bank transaction domain dataset associated with rural locations, and the second dataset 114B corresponds to a bank transaction domain dataset associated with urban locations. As an example, the first set of features may include a gender variable associated with the first dataset 114A for rural individuals, and the second set of features may include a gender variable associated with the second dataset 114B for urban individuals. The gender variable may have different values across the first dataset 114A and the second dataset 114B. Typically, rural men are considerably more likely to obtain loan approval than rural women of a particular age group (e.g., 25-40 years old). However, the loan approval probabilities for urban men and urban women of the same age group (i.e., 25-40 years old) show a smaller difference. In such cases, the gender bias (more male-oriented) in loan approval decisions is greater in rural groups compared to urban groups. Therefore, the gender variable may have different values between datasets for rural locations and datasets for urban locations.
[0055] In the embodiment, the second user input may further include a confidence level associated with the feature difference information. The confidence level associated with the feature difference information may indicate the level of certainty that user 116 may have about the feature difference information, which may include one or more input variables that may have different values across the first dataset 114A and the second dataset 114B. In other words, the higher the confidence level, the more certain user 116 is that a particular input variable may have different values across the first dataset 114A and the second dataset 114B. For example, user 116 may have a 90% confidence level that the gender variable may have different values across the first dataset 114A and the second dataset 114B. Thus, in the present case, the second user input may indicate that user 116 may have a 90% confidence level that the gender variable may have different values across the first dataset 114A and the second dataset 114B.
[0056] The determination of feature difference information may further be based on a set of ethical requirements that may be relevant to the application domain on which the meta-learning model 110 can be trained. Examples of sets of ethical requirements may include, but are not limited to, a set of fairness requirements, a set of privacy requirements, or a set of security requirements. For example, if a set of requirements related to a banking transaction domain includes a privacy requirement, the feature difference information may be determined without the need to access the underlying demographic or financial data of individuals in the first dataset 114A and the second dataset 114B. In such a case, the feature difference information may be determined from a second user input that is based on the domain knowledge of an expert user related to the banking transaction domain, but does not have to be based on the underlying data or statistics related to the first dataset 114A and the second dataset 114B. Thus, privacy is ensured because the underlying demographic or financial data does not need to be shared with anyone. In another example, when a set of requirements related to the banking transaction domain includes a set of fairness requirements, the gender variable may be selected as a feature difference because there may be differences in value or bias with respect to the gender variable across the first dataset 114A and the second dataset 114B.
[0057] In block 408, a set of SCMs 108 associated with the first dataset 114A may be determined based on the determined feature dependency information and the determined feature difference information. In the embodiment, the processor 204 may be configured to determine the set of SCMs 108 associated with the first dataset 114A based on the determined feature dependency information and the determined feature difference information. An example of the set of SCMs 108 is further described, for example, in Figure 5. An example of an SCM is further described, for example, in Figures 3 and 7.
[0058] In block 410, a set of ethical coefficients associated with a set of ethical requirements may be determined based on the application of a causal transportability model to the determined set of SCMs 108. In embodiments, the processor 204 may be configured to determine a set of ethical coefficients associated with a set of ethical requirements based on the application of a causal transportability model to the determined set of SCMs. The set of ethical coefficients may include an ethical coefficient for each ethical requirement in the set of ethical requirements. For example, the set of ethical coefficients may include a fairness coefficient associated with a fairness requirement, a privacy coefficient associated with a privacy requirement, and a safety coefficient associated with a set of safety requirements. In embodiments, each in the set of ethical coefficients may be determined based on a confidence level associated with feature difference information and causal transportability coefficients associated with the first dataset 114A and the second dataset 114B. Furthermore, the causal transportability coefficients may be determined by the use of a causal transportability model. The causal transportability model may correspond to a causal fusion tool. The determination of the ethical coefficient set is further described, for example, in Figures 5 and 6.
[0059] In block 412, a reliability score associated with the first dataset 114A may be determined based on a set of determined ethical coefficients. In an embodiment, the processor 204 may be configured to determine a reliability score associated with the first dataset 114A based on a set of determined ethical coefficients. In an embodiment, the reliability score may be determined based on a weighted average of the set of determined ethical coefficients. The determination of the reliability score is further described, for example, in Figures 5 and 6.
[0060] In block 414, a meta-learning model 110 related to multiple application domains may be trained based on multiple confidence scores related to multiple datasets 114. The multiple confidence scores may include determined confidence scores related to a first dataset 114A. In embodiments, the processor 204 may be configured to train the meta-learning model 110 related to multiple application domains based on multiple confidence scores related to multiple datasets 114. The determination of the multiple confidence scores may be similar to the determination of the confidence scores related to the first dataset 114A. The meta-learning model 110 may correspond to a generative learning model that includes, but is not limited to, at least one of a metric-based learning model, an optimization-based learning model, or a model-based learning model. Once the meta-learning model 110 is trained, it may be used for prediction tasks related to various application domains, such as an insurance domain related to a first location or a banking domain related to a third location, etc. The training of the meta-learning model 110 is further described, for example, in Figures 5 and 6. Control moves to termination.
[0061] Flowchart 400 is represented as separate operations such as 402, 404, 406, 408, 410, 412, and 414. However, in certain embodiments, such separate operations may be further divided into additional operations, combined into fewer operations, or deleted, depending on the particular implementation, without departing from the essence of the disclosed embodiments.
[0062] Figure 5 is a diagram illustrating an exemplary scenario for training a meta-learning model based on causal transportability between datasets, according to embodiments of the present disclosure. Figure 5 is described in relation to elements from Figures 1, 2, 3, and 4. Referring to Figure 5, an exemplary scenario 500 is shown. Scenario 500 may represent the training of a meta-learning model 110 based on causal transportability between datasets. Scenario 500 may include operations for individual datasets 502, confidence score calculation 504, and meta-learner training 506.
[0063] For example, an individual dataset 502 may contain multiple datasets that may be related to multiple application domains. Multiple application domains may include Domain-1, Domain-2, ..., and Domain-N. A domain expert or expert user (e.g., user 116) may be related to each of the multiple application domains. A domain expert (or expert user) related to a particular application domain may possess domain knowledge related to that particular application domain. Figure 5 shows a set of blocks that may represent domain knowledge related to each of the multiple application domains. For example, Domain-1 expert knowledge 508A may be related to Domain 1, Domain-2 expert knowledge 508B may be related to Domain-2, ... Domain-N expert knowledge 508N may be related to Domain-N.
[0064] The number N of application domains related to the individual dataset 502 in Figure 5 is shown merely as an example. Multiple application domains may include as few as two or more than N application domains without exceeding the scope of this disclosure. For brevity, only N application domains are shown. However, in some embodiments, there may be more than N application domains without exceeding the scope of this disclosure.
[0065] Based on domain knowledge of a specific application domain, the processor 204 may receive a first user input (corresponding to feature dependency information) and a second user input (corresponding to feature difference information) from a domain expert associated with that specific application domain. For example, each domain expert may supply a first user input that may include a set of input variables and a set of output variables related to the dataset of the domain expert's application domain, the dependency between the set of input variables and the set of output variables, and the confidence level associated with the dependency. Furthermore, each domain expert may identify features where differences may exist between the dataset associated with the domain expert's application domain and the datasets associated with the remaining applications among the multiple application domains. The domain expert may also identify the confidence level associated with the identified feature difference. The domain expert may supply the identified feature difference and the confidence level associated with the feature difference as a second user input.
[0066] The determination of feature difference information may further be based on sets of ethical requirements, such as sets of privacy requirements, sets of fairness requirements, and sets of security requirements. For example, for security requirements, a domain expert may encode differences in variables (across the dataset) that may differ in terms of vulnerability to attack. For instance, some features, such as credit scores, may be vulnerable to attack because they may be easily manipulated. However, features such as addresses are less likely to be easily manipulated. Therefore, the second user input may include an indication that the credit score feature may differ across the dataset and may be a security requirement relevant to SCM decisions. In another example, for privacy requirements, a domain expert may encode differences in variables (across the dataset) that may be personal based on local regulations. For example, a candidate's age may be a personal feature for job searching. Therefore, the second user input may include an indication that the age feature may differ across the dataset and may be a personal requirement relevant to SCM decisions. In yet another example, for fairness requirements, a domain expert may indicate variables that may be biased across the dataset. For example, the second user input may include a note stating that the gender variable may differ across the dataset and may be a fairness requirement related to the SCM decision.
[0067] The processor 204 may determine a set of SCMs for each application domain based on feature dependency information and feature difference information. For a particular dataset (e.g., the first dataset 114A related to domain-1), the processor 204 may determine (N-1) SCMs (in the case of N datasets), and each such SCM may consider a pair of datasets containing the first dataset 114A and other datasets (which may be related to other domains). Furthermore, since the feature difference information can vary with respect to various ethical requirements, the processor 204 may determine a separate SCM for each ethical requirement. For example, in the case of three ethical requirements (e.g., privacy requirements, fairness requirements, and security requirements), the processor 204 may determine three separate SCMs for each dataset. Thus, a total of 3 × (N-1) SCMs may be determined as a set of SCMs related to domain-1 (corresponding to the first dataset 114A) (e.g., the first set of SCMs 510A). Similarly, the processor 204 may determine 3 × (N-1) SCMs as a second set of SCMs 510B associated with domain-2 (corresponding to the second dataset 114B). Each SCM may model a causal relationship (e.g., causal effect 502A) related to the application domain associated with that SCM.
[0068] Each of the determined SCM sets may correspond to a directed acyclic graph representing a variable associated with a particular dataset. In an embodiment, the first user input may indicate a node in the SCM associated with a dataset (e.g., first dataset 114A). The first user input may further include dependency information between the nodes of the SCM. This dependency information may be represented in the SCM by directed edges between the nodes. In an embodiment, the second user input may indicate a node in the SCM that may correspond to a feature difference between a pair of datasets corresponding to the SCM (e.g., first dataset 114A and second dataset 114B). Such indicated nodes may be represented as quadrilateral nodes in the SCM. One or more features that may differ between the pair of datasets may also be selected as a set of selection variables. The directed acyclic graph of the SCM may be augmented based on the set of selection variables (which may represent determined feature difference information). The second user input may include a set of selection nodes that may be represented in the SCM by concentric circular nodes. An example of SCM related to the banking transaction domain is further described, for example, in Figure 7.
[0069] The reliability score calculation 504 may include operations such as causal transportability estimation 512, reliability score determination 514 (for each dataset along different dimensions based on various ethical requirements), and score aggregation 516 (for each dataset). The processor 204 may perform operations related to the reliability score calculation 504 by using a causal effect evaluator 504A, such as a causal transportability model. In embodiments, the causal transportability model may correspond to a causal fusion tool. Further details regarding the reliability score calculation 504 are described, for example, in Figure 6.
[0070] Based on the output of the reliability score calculation 504, the processor 204 may train the meta-learning model 110 by using operations such as meta-learner training 506. For example, the total reliability score output may be determined based on the score aggregation 516. Based on the total reliability score output, the processor 204 may perform meta-learner training 506 on the meta-learning model 110 to adjust the weights of the meta-learning model 110, thereby training the meta-learning model 110. Meta-learner training 506 is further described, for example, in Figure 6.
[0071] It should be noted that Scenario 500 in Figure 5 is for illustrative purposes only and should not be intended to limit the scope of this disclosure.
[0072] Figure 6 is a diagram illustrating an exemplary scenario for training a meta-learning model based on causal transportability between datasets, according to embodiments of the present disclosure. Figure 6 is described in relation to elements from Figures 1, 2, 3, 4, and 5. Referring to Figure 6, an exemplary scenario 600 is shown. Scenario 600 may represent the training of a meta-learning model 110 based on causal transportability between datasets. Scenario 600 may include multiple sets of SCM, operations for confidence score calculation 504, and operations for meta-learner training 506. The decision 602 of the meta-learning model 110 is also shown.
[0073] Multiple sets of SCMs may include a first set of SCMs 510A associated with a first dataset 114A (corresponding to domain-1), a second set of SCMs 510B associated with a second dataset 114B (corresponding to domain-2), ... and a nth set of SCMs 510N associated with a dataset N (corresponding to domain-N). The determination of multiple sets of SCMs is further illustrated, for example, in Figure 5. An example of an SCM associated with a banking transaction domain is further illustrated, for example, in Figure 7.
[0074] The processor 204 may perform operations related to the reliability score calculation 504 based on multiple sets of SCMs. The processor 204 may determine a set of ethical coefficients related to a set of ethical requirements based on the application of a causal transportability model to the determined set of SCMs. In an embodiment, each of the ethical coefficients in the set (for a certain dataset, e.g., a first dataset 114) may be determined based on a confidence level related to feature difference information and causal transportability coefficients (which may be related to the first dataset 114A and the second dataset 114B). Referring to Figure 5, the reliability score calculation 504 may include operations such as causal transportability estimation 512, reliability score determination 514, and score aggregation 516. In causal transportability estimation 512, the processor 204 may determine causal transportability coefficients related to a pair of datasets that may be related to each of the multiple sets of SCMs. For example, the first SCM in the first set of SCMs 510A may be associated with a first dataset 114A and a second dataset 114B. For the first SCM, the processor 204 may determine a causal transportability coefficient associated with the first dataset 114A and the second dataset 114B. In an embodiment, the causal transportability coefficient may be determined by using a causal transportability model. The causal transportability model may correspond to a causal fusion tool. Based on one or more feature differences between the two datasets (indicated by feature difference information), the causal fusion tool may automatically determine whether a causal effect from one dataset can be carried to the other dataset. In an embodiment, the causal transportability coefficient may be a binary value, where a value of "0" indicates that the causal effect cannot be carried, while a value of "1" indicates that the causal effect can be carried from one dataset to the other.
[0075] In reliability score determination 514, the processor 204 may determine ethical coefficients associated with each of the sets of ethical requirements for each of the multiple sets of SCMs. For example, for the first dataset 114A (corresponding to domain-1), the first set of SCMs 510A may include a first subset of SCMs corresponding to fairness requirements, a second subset of SCMs corresponding to privacy requirements, and a third subset of SCMs corresponding to security requirements. Each of the three subsets of SCMs may contain (N-1) SCMs corresponding to a particular ethical requirement. The (N-1) SCMs within a subset of SCMs may be determined based on feature dependency information of the first dataset 114A and feature difference information related to the first dataset 114A and other datasets in the multiple datasets 114. For example, the first SCM among the (N-1) SCMs may capture the feature difference between the first dataset 114A and the second dataset 114B, the second SCM among the (N-1) SCMs may capture the feature difference between the first dataset 114A and the third dataset 114C, and so on. In one example, the processor 204 calculates the ethical coefficient (c) for each of the three subsets of the SCM for the first dataset 114A by using equation (1) as follows: ij It is represented by ).) You may decide:
number
[0076] For example, as shown in FIG. 6, for the first data set 114A, the processor 204 may calculate the ethical coefficient "c 11 " associated with the fairness requirement, the ethical coefficient "c 12 " associated with the privacy requirement, and the ethical coefficient "c 13 " associated with the security requirement. Similarly, for the second data set 114B, the processor 204 may determine a set of ethical coefficients (for example, "c 21 ", "c 22 " and "c 23 ") associated with the set of ethical requirements (for example, the fairness requirement, the privacy requirement, and the security requirement).
[0077] In the score aggregation 516, the processor 204 may determine the reliability score (represented as c i ) associated with each data set (for example, the first data set 114A) based on the determined set of ethical coefficients associated with that data set. In embodiments, the reliability score may be determined based on a weighted average of the determined set of ethical coefficients. In one example, the processor 204 may determine the reliability score associated with a data set (for example, the first data set 114A) using Equation (2) as follows:
Equation
[0078] In a embodiments, weights related to the ethical coefficient (i.e., weight j The sum of ) may be equal to 1, and the reliability score (i.e., c i The value of ) may be between 0 and 1. Therefore, the processor 204 calculates the reliability score (i.e., c) for each dataset in the multiple datasets 114. i ) can be determined. For example, multiple reliability scores c1, c2, ...c n This can be determined for multiple datasets 114 (including the first dataset 114A, the second dataset 114B, ..., and the Nth dataset 114N).
[0079] In meta-learner training 506, the processor 204 may train the meta-learning model 110 based on multiple confidence scores associated with multiple datasets 114. As described, the multiple confidence scores may include determined confidence scores associated with the first dataset 114A. In embodiments, the meta-learning model 110 may correspond to a generative learning model that includes, but is not limited to, at least one of a metric-based learning model, an optimization-based learning model, or a model-based learning model. For example, the meta-learning model 110 may be associated with a particular application domain and may be pre-trained based on a training dataset (e.g., historical data) associated with that application domain. Based on the pre-training of the meta-learning model 110, a set of weights associated with the meta-learning model 110 may be determined. The set of weights may correspond to various ethical requirements associated with the application domain of the meta-learning model 110. Thus, the previous ethical behavior of the meta-learning model 110 may be known based on historical data in the training dataset. For example, based on historical data of the banking transaction domain, the meta-learning model 110 may be trained so that security requirements may have a higher priority than fairness requirements. The processor 204 may adjust the determined set of weights of the meta-learning model 110 to train the meta-learning model 110. The adjustment of the determined set of weights of the meta-learning model 110 involves multiple confidence scores associated with multiple datasets 114 (i.e., c1, c2, ..., and c n The test dataset may include ) as follows: The processor 204 may determine a decision 602 (e.g., a loan approval decision) related to the application domain (e.g., a banking transaction domain) corresponding to the meta-learning model 110 by using equation (3):
number
[0080] It should be noted that Scenario 600 in Figure 6 is for illustrative purposes only and should not be intended to limit the scope of this disclosure.
[0081] Figure 7 is a diagram representing an exemplary structural causal model (SCM) related to a pair of datasets corresponding to a banking transaction application domain, according to embodiments of the present disclosure. Figure 76 is described in relation to elements from Figures 1, 2, 3, 4, 5, and 6. Referring to Figure 7, an exemplary structural causal model (SCM) 700 is shown. For example, SCM 700 may be related to a banking transaction application domain. SCM 700 may be determined based on a pair of datasets related to the banking transaction domain (e.g., a first dataset 114A and a second dataset 114B). In one example, the first dataset 114A may be related to a first location or territory (e.g., a first country), while the second dataset 114B may be related to a second location or territory (e.g., a second country).
[0082] SCM700 may be a directed acyclic graph representing variables related to the first dataset 114A. The directed acyclic graph of SCM700 may show the causal effect of a set of input variables related to the first dataset 114A on a set of output variables related to the first dataset 114A. For example, SCM700 may include a first node 702A, a second node 702B, and a third node 702C. The first node 702A (represented by "X") may represent input variables, such as credit scores, income, and age related to individuals in the first dataset 114A related to a banking transaction domain. The second node 702B (represented by "Y") may represent output (or result) variables, such as loan decisions for individuals in the first dataset 114A. As shown in Figure 7, an input variable (represented by "X") may causally influence an output variable (represented by "Y"), which can be represented in SCM700 by a directed edge from the first node 702A to the second node 702B. In one example, the causal effect of X on Y (where X is considered an input binary random variable and Y is considered an outcome variable) can be determined as E(Y|do(X=1))-E(X|do(X=0)), where E() can represent the expected value and do() can represent an intervention (also called a "do-action") that changes the value of the variable.
[0083] The directed acyclic graph of SCM700 may be augmented based on a set of selection variables (represented by "S") that show determined feature difference information related to the first dataset 114A and the second dataset 114B. The third node 702C (represented by "Z") may represent a variable that may correspond to the feature difference between the first dataset 114A and the second dataset 114B. For example, the feature difference (i.e., the set of selection variables) may include a gender variable. The directed acyclic graph of SCM700, which can be augmented based on a set of selection variables (represented by "S"), may also be called a selection diagram. The set of selection variables (represented by "S") may be represented as nodes shown as concentric circles in the selection diagram. The population related to the first dataset 114A may be switched to the population related to the second dataset 114B based on the application of a set of conditions to at least one different value in the set of selection variables.
[0084] For example, consider "D" as a choice diagram that can characterize two datasets or groups, such as "π" and "π*", and "S" may be a set of choice variables in "D". A hierarchy-specific causal effect P*(y|do(x),z) can be carried from π to π* if Z “d-separates” Y from S within a version of D manipulated by X, i.e., Z satisfies (Y⊥S|Z,X)Dx', where DX' may correspond to a post-intervention graph associated with the choice diagram, where all incoming arrows to X can be removed. The “d-separation” between the S node (i.e., a set of choice nodes) and the resulting variable Y can be achieved based on a conditional set T (i.e., a set of conditions). A conditional set of variables that satisfies (Y⊥S|T) in Dx' may also be called s-acceptable (with respect to the causal effect of X on Y), where ⊥ represents conditional independence. The conditional set T may be selected such that the outcome variable Y can be conditionally independent from the set of selection variables S based on the conditional set T. Thus, the set of conditions may correspond to the s-admissibility (or conditional independence) of the selection variables with respect to the causal effect of X on Y.
[0085] It should be noted that the SCM700 in Figure 7 is for illustrative purposes only and should not be intended to limit the scope of this disclosure.
[0086] It should be noted that the loan approval decision scenario related to the banking transaction domain is presented merely as an illustrative scenario for the application of this disclosure and should not be considered as limiting the scope of this disclosure. This disclosure may be applicable to a variety of other scenarios without exceeding the scope of this disclosure. For example, this disclosure may be applicable to use cases for autonomous vehicle safety assistance, credit evaluation, recommendation tools / platforms, or language modeling. For example, in the use case for autonomous vehicle safety assistance, the disclosed electronic device 102 may be used to determine whether a meta-learning model related to an autonomous vehicle system, which may be developed based on a dataset from a first location, can be safely used in a second location. In another example, in the use case for credit evaluation, the electronic device may be used to determine whether a meta-learning model related to the banking transaction domain, which may be trained using a dataset within one region, can be applied to other regions. In another example, in a recommendation tool / platform use case, the electronic device 102 may be used to determine whether a meta-learning model trained to predict product recommendations for users of one demographic or class of people may be applicable across other demographics or classes of people (e.g., minority groups of people). In yet another example, in a language modeling use case, the electronic device 102 may be used to relax the effects of a large-scale language model trained using one language (e.g., a primary language such as English) to use cases of other languages (e.g., a secondary language such as Spanish).
[0087] The disclosed electronic device 102 may be configured to train a meta-learning model based on causal transportability between datasets. The electronic device 102 may receive multiple datasets 114 relating to multiple application domains. Based on a first user input, the electronic device 102 may determine feature dependency information related to the first dataset 114A among the multiple received datasets 114. Furthermore, based on a second user input and a set of ethical requirements, the electronic device 102 may determine feature difference information related to the first dataset 114A and the second dataset 114B among the multiple received datasets 114, where the second dataset 114B may be different from the first dataset 114A. Examples of sets of ethical requirements may include, but are not limited to, sets of fairness requirements, sets of privacy requirements, or sets of security requirements. Based on the determined feature dependency information and determined feature difference information, the electronic device 102 may determine a set of SCMs 108 related to the first dataset 114A. Furthermore, the electronic device 102 may determine a set of ethical coefficients associated with a set of ethical requirements based on the application of a causal transportability model to the determined set of SCMs. Thus, causal effects related to multiple datasets 114 and differences between various attributes between the multiple datasets can be captured using SCMs based on feature dependency information and feature difference information. The electronic device 102 may determine a reliability score associated with the first dataset 114A based on the determined set of ethical coefficients. Furthermore, the electronic device 102 may train a meta-learning model 110 associated with multiple application domains based on multiple reliability scores associated with multiple datasets. Here, the multiple reliability scores may include the determined reliability score associated with the first dataset 114A. In this way, the contribution of each dataset directed towards training the meta-learning model can be adapted based on the reliability score determined for that dataset. Since the reliability score may take expert knowledge into account, a meta-learning model trained on multiple reliability scores may incorporate ethical requirements.Furthermore, undesirable consequences such as dataset distribution shifts (e.g., bias) and ethical concerns (e.g., privacy and security concerns) can be addressed by applying causal transportability models to the SCM set.
[0088] Conventional systems typically require access to individual datasets to train a meta-learning model. However, with the disclosed electronic device 102, access to individual datasets is not required; the meta-learning model 110 can be trained relying on feature-dependent information and feature-diff information associated with one or more datasets. The feature-dependent information and feature-diff information may be provided by an expert user. Based on the feature-dependent information and feature-diff information, a set of SCMs 108 can be determined using the disclosed electronic device 102. Furthermore, a causal transportability model may be applied to the set of SCMs 108 to determine a reliability score associated with the dataset. The meta-learning model 110 can then be trained based on the determined reliability score. Since individual datasets do not need to be directly accessed, the meta-learning model 110 trained using the disclosed electronic device 102 is not affected by distribution shifts associated with individual datasets. Furthermore, unlike conventional systems, the meta-learning model 110, which can be trained using the disclosed electronic device 102, is less prone to bias and ethical concerns because the disclosed electronic device 102 can include humans in the loop and rely on domain-specific knowledge from expert users. The disclosed electronic device 102 also enables expert users to provide confidence levels related to feature differences, thereby receiving a bias tendency of features rather than simply classifying features as biased or unbiased features. The disclosed electronic device 102 may further use confidence levels to determine confidence scores. Moreover, the meta-learning model 110, which can be trained by the electronic device 102, may be data- and model-independent (agnostic) and applicable across various use cases to mitigate the effects of dataset distribution shifts.
[0089] Various embodiments of this disclosure may provide one or more non-temporary computer-readable storage media configured to store instructions causing a system (e.g., an example electronic device 102) to perform an action in response to being executed. The action may include receiving a plurality of datasets relating to a plurality of application domains. The action may further include determining feature dependency information related to a first dataset among the plurality of received datasets, based on a first user input. The action may further include determining feature difference information related to the first dataset and a second dataset among the plurality of received datasets, based on a second user input and a set of ethical requirements, where the second dataset and the first dataset may be different. The action may further include determining a set of structural causal models (SCMs) related to the first dataset, based on the determined feature dependency information and the determined feature difference information. The action may further include determining a set of ethical coefficients related to the set of ethical requirements, based on the application of a causal transportability model to the determined set of SCMs. The operation may further include determining a reliability score associated with the first dataset based on the determined set of ethical coefficients. The operation may further include training a meta-learning model associated with multiple application domains based on multiple reliability scores associated with multiple datasets, where the multiple reliability scores may include the determined reliability scores associated with the first dataset.
[0090] As used in this disclosure, the terms “module” or “component” may refer to a specific hardware implementation configured to perform the operation of a module or component, and / or a software object or software routine stored in and / or executed by general-purpose hardware of the electronic device 102 (e.g., computer-readable media, processing devices, etc.). In some embodiments, different components, modules, engines, and services described in this disclosure may be implemented as objects or processes that run on the electronic device 102 (e.g., as separate threads). While some of the systems and methods described in this disclosure are generally described as being implemented in software (stored in and / or executed by general-purpose hardware), specific hardware implementations, or combinations of software and specific hardware implementations, are also possible and conceivable. As used herein, “computation entity” may be any electronic device 102 as previously defined in this disclosure, or any module or combination of modules that runs on the electronic device 102.
[0091] In this disclosure, the terms used in particular in the attached claims (e.g., the text of the attached claims) are generally intended to be “open” terms (for example, the word “including” should be interpreted as “including but not limited to,” the word “having” should be interpreted as “having at least,” and the word “includes” should be interpreted as “including but not limited to,” etc.).
[0092] Furthermore, if a specific number is intended in an introduced claim recitation, that intention must be clearly stated in the claim; if there is no such statement, then no such intention exists. For example, to facilitate understanding, subsequent appended claims may use introductory phrases such as "at least one" and "one or more" to introduce a claim. However, the use of such phrases should not be interpreted as suggesting that a particular claim containing the introduced claim recitation is limited to cases that include only one instance of that item, even if the claim includes both an introductory phrase such as "one or more" or "at least one" and an indefinite article such as "a" or "an" (for example, "a" and / or "an" should be interpreted as meaning "at least one" or "one or more"). The same applies when introducing a claim recitation using a definite article.
[0093] Furthermore, even if a specific number is explicitly stated in the introduced claim description, it will be understood by those skilled in the art that such description should generally be interpreted to mean at least the number stated (for example, if there is a description of only "two descriptions" without any other modifiers, this description means at least two descriptions, or two or more descriptions). Furthermore, when a notation similar to "at least one of A, B, and C, etc." or "one or more of A, B, and C, etc." is used, such a structure is generally intended to include A only, B only, C only, both A and B, both A and C, both B and C, and / or all of A, B, and C, etc.
[0094] Furthermore, any disjunctions and / or disjunctions representing two or more selectable terms, whether in the specification, claims, or drawings, should be understood as intended to include the possibility of including one of those terms, either of those terms, or both of those terms. For example, the phrase "A or B" should be understood to include the possibility of "A or B" or "A and B".
[0095] All examples and conditional language cited herein are intended for educational purposes to help readers understand the concepts and inventions to which the inventors have contributed to the advancement of the art, and should be construed as not being limited to such specifically cited examples and conditions. While embodiments of this disclosure have been described in detail, various modifications, substitutions, and alternatives may be made without departing from the spirit and scope of this disclosure.
[0096] In addition to the embodiments described above, the following additional information is disclosed. (Note 1) A method that is executed by a processor, Receiving multiple datasets related to multiple application domains, Based on the first user input, the feature-dependent information related to the first dataset among the multiple datasets received is determined, Based on the second user input and ethical requirements, the feature difference information related to the first dataset and the second dataset among the multiple received datasets is determined, and the second dataset is different from the first dataset. Based on the determined feature dependency information and the determined feature difference information, a set of structural causal models (SCMs) related to the first dataset is determined. Based on the application of the causal transportability model to the determined set of SCMs, a set of ethical coefficients related to the set of ethical requirements is determined, Based on the set of ethical coefficients determined above, a reliability score associated with the first dataset is determined, The method involves training a meta-learning model related to the multiple application domains based on multiple reliability scores related to the multiple datasets, wherein the multiple reliability scores include the determined reliability score related to the first dataset. A method of having. (Note 2) The aforementioned feature-dependent information is A set of input variables related to the aforementioned first dataset, A set of output variables related to the aforementioned first dataset, Dependency information related to the set of input variables and the set of output variables, or Confidence level related to the aforementioned dependency information including at least one of the following: The method described in Appendix 1. (Note 3) The feature difference information corresponds to the difference between groups between the first set of features associated with the first dataset and the second set of features associated with the second dataset. The aforementioned set of the first feature corresponds to the aforementioned set of the second feature. The method described in Appendix 1. (Note 4) The aforementioned set of ethical requirements is, A set of fairness requirements, A set of privacy requirements, or Set of safety requirements including at least one of the following: The method described in Appendix 1. (Note 5) Each of the determined sets of SCMs corresponds to a directed acyclic graph that represents the variables associated with the first dataset, The directed acyclic graph is augmented based on the set of selection variables that represent the determined feature difference information. The method described in Appendix 1. (Note 6) The population associated with the first dataset is switched to the population associated with the second dataset based on the application of a set of conditions to at least one different value in the set of selection variables. The method described in Appendix 5. (Note 7) The directed acyclic graph shows the causal effect of a set of input variables related to the first dataset on a set of output variables related to the first dataset. The method described in Appendix 5. (Note 8) The directed acyclic graph further includes a set of unobserved variables that influence the set of input variables associated with the first dataset. The method described in Appendix 5. (Note 9) Each of the sets of ethical coefficients is determined based on the confidence level associated with the feature difference information and the causal transportability coefficient associated with the first and second datasets. The method described in Appendix 1. (Note 10) The causal transportability coefficient is determined by using the causal transportability model. The aforementioned causal transportability model corresponds to a causal fusion tool. The method described in Appendix 9. (Note 11) The reliability score is determined based on the weighted average of the determined set of ethical coefficients. The method described in Appendix 1. (Note 12) The meta-learning model corresponds to a generative learning model that includes at least one of a metric-based learning model, an optimization-based learning model, or a model-based learning model. The method described in Appendix 1. (Note 13) When executed, the electronic device will Receiving multiple datasets related to multiple application domains, Based on the first user input, the feature-dependent information related to the first dataset among the multiple datasets received is determined, Based on the second user input and ethical requirements, the feature difference information related to the first dataset and the second dataset among the multiple received datasets is determined, and the second dataset is different from the first dataset. Based on the determined feature dependency information and the determined feature difference information, a set of structural causal models (SCMs) related to the first dataset is determined. Based on the application of the causal transportability model to the determined set of SCMs, a set of ethical coefficients related to the set of ethical requirements is determined, Based on the set of ethical coefficients determined above, a reliability score associated with the first dataset is determined, The method involves training a meta-learning model related to the multiple application domains based on multiple reliability scores related to the multiple datasets, wherein the multiple reliability scores include the determined reliability score related to the first dataset. One or more non-temporary computer-readable storage media configured to store instructions for performing an action having [a certain characteristic]. (Note 14) The aforementioned feature-dependent information is A set of input variables related to the aforementioned first dataset, A set of output variables related to the aforementioned first dataset, Dependency information related to the set of input variables and the set of output variables, or Confidence level related to the aforementioned dependency information including at least one of the following: One or more non-temporary computer-readable storage media as described in Appendix 13. (Note 15) The feature difference information corresponds to the difference between groups between the first set of features associated with the first dataset and the second set of features associated with the second dataset. The aforementioned set of the first feature corresponds to the aforementioned set of the second feature. One or more non-temporary computer-readable storage media as described in Appendix 13. (Note 16) The aforementioned set of ethical requirements is, A set of fairness requirements, A set of privacy requirements, or Set of safety requirements including at least one of the following: One or more non-temporary computer-readable storage media as described in Appendix 13. (Note 17) Each of the determined sets of SCMs corresponds to a directed acyclic graph that represents the variables associated with the first dataset, The directed acyclic graph is augmented based on the set of selection variables that represent the determined feature difference information. One or more non-temporary computer-readable storage media as described in Appendix 13. (Note 18) The population associated with the first dataset is switched to the population associated with the second dataset based on the application of a set of conditions to at least one different value in the set of selection variables. One or more non-temporary computer-readable storage media as described in Appendix 17. (Note 19) Each of the sets of ethical coefficients is determined based on the confidence level associated with the feature difference information and the causal transportability coefficient associated with the first and second datasets. One or more non-temporary computer-readable storage media as described in Appendix 13. (Note 20) The memory that stores the instructions, It is coupled to memory and executes the instruction, Receiving multiple datasets related to multiple application domains, Based on the first user input, the feature-dependent information related to the first dataset among the multiple datasets received is determined, Based on the second user input and ethical requirements, the feature difference information related to the first dataset and the second dataset among the multiple received datasets is determined, and the second dataset is different from the first dataset. Based on the determined feature dependency information and the determined feature difference information, a set of structural causal models (SCMs) related to the first dataset is determined. Based on the application of the causal transportability model to the determined set of SCMs, a set of ethical coefficients related to the set of ethical requirements is determined, Based on the set of ethical coefficients determined above, a reliability score associated with the first dataset is determined, The method involves training a meta-learning model related to the multiple application domains based on multiple reliability scores related to the multiple datasets, wherein the multiple reliability scores include the determined reliability score related to the first dataset. A processor that executes a process having An electronic device having [Explanation of Symbols]
[0097] 100 Environment 102 Electronic Devices 104 Databases 106 User End Devices 108. Structural Causal Models (SCMs) 110 Meta-Learning Models 112 Communication Networks 114 datasets 116 users 202 System 204 Processors 206 memory 208 Persistent Data Storage 210 I / O devices 212 display screens 214 Network Interfaces 300,700 SCM
Claims
1. A method that is executed by a processor, Receiving multiple datasets related to multiple application domains, Based on the first user input, the feature-dependent information related to the first dataset among the multiple datasets received is determined, Based on the second user input and ethical requirements, the feature difference information related to the first dataset and the second dataset among the multiple received datasets is determined, and the second dataset is different from the first dataset. Based on the determined feature dependency information and the determined feature difference information, a set of structural causal models (SCMs) related to the first dataset is determined. Based on the application of the causal transportability model to the determined set of SCMs, a set of ethical coefficients related to the set of ethical requirements is determined, Based on the set of ethical coefficients determined above, a reliability score associated with the first dataset is determined, The method involves training a meta-learning model related to the multiple application domains based on a plurality of reliability scores related to the plurality of datasets, wherein the plurality of reliability scores include the determined reliability score related to the first dataset. A method of having.
2. The aforementioned feature-dependent information is A set of input variables related to the first dataset, A set of output variables related to the first dataset, Dependency information related to the set of input variables and the set of output variables, or Confidence level related to the aforementioned dependency information Including at least one of the following: The method according to claim 1.
3. The feature difference information corresponds to the difference between the sets of first features associated with the first dataset and the sets of second features associated with the second dataset. The set of the first features corresponds to the set of the second features. The method according to claim 1.
4. The aforementioned set of ethical requirements is, A set of fairness requirements, A set of privacy requirements, or Set of safety requirements Including at least one of the following: The method according to claim 1.
5. Each of the determined sets of SCMs corresponds to a directed acyclic graph that represents the variables associated with the first dataset, The directed acyclic graph is augmented based on the set of selection variables that represent the determined feature difference information. The method according to claim 1.
6. The group associated with the first dataset is switched to the group associated with the second dataset based on the application of a set of conditions to at least one different value in the set of selection variables. The method according to claim 5.
7. The directed acyclic graph shows the causal effect of a set of input variables related to the first dataset on a set of output variables related to the first dataset. The method according to claim 5.
8. The directed acyclic graph further includes a set of unobserved variables that influence the set of input variables associated with the first dataset. The method according to claim 5.
9. Each of the sets of ethical coefficients is determined based on the confidence level associated with the feature difference information and the causal transportability coefficient associated with the first and second datasets. The method according to claim 1.
10. The causal transportability coefficient is determined by using the causal transportability model. The aforementioned causal transportability model corresponds to a causal fusion tool. The method according to claim 9.
11. The reliability score is determined based on the weighted average of the determined set of ethical coefficients. The method according to claim 1.
12. The meta-learning model corresponds to a generative learning model that includes at least one of a metric-based learning model, an optimization-based learning model, or a model-based learning model. The method according to claim 1.
13. When executed, the electronic device will Receiving multiple datasets related to multiple application domains, Based on the first user input, the feature-dependent information related to the first dataset among the multiple datasets received is determined, Based on the second user input and ethical requirements, the feature difference information related to the first dataset and the second dataset among the multiple received datasets is determined, and the second dataset is different from the first dataset. Based on the determined feature dependency information and the determined feature difference information, a set of structural causal models (SCMs) related to the first dataset is determined. Based on the application of the causal transportability model to the determined set of SCMs, a set of ethical coefficients related to the set of ethical requirements is determined, Based on the set of ethical coefficients determined above, a reliability score associated with the first dataset is determined, The method involves training a meta-learning model related to the multiple application domains based on a plurality of reliability scores related to the plurality of datasets, wherein the plurality of reliability scores include the determined reliability score related to the first dataset. One or more non-temporary computer-readable storage media configured to store instructions for performing an action having the following characteristics.
14. The aforementioned feature-dependent information is A set of input variables related to the first dataset, A set of output variables related to the first dataset, Dependency information related to the set of input variables and the set of output variables, or Confidence level related to the aforementioned dependency information Including at least one of the following: One or more non-temporary computer-readable storage media according to claim 13.
15. The feature difference information corresponds to the difference between the sets of first features associated with the first dataset and the sets of second features associated with the second dataset. The set of the first features corresponds to the set of the second features. One or more non-temporary computer-readable storage media according to claim 13.
16. The aforementioned set of ethical requirements is, A set of fairness requirements, A set of privacy requirements, or Set of safety requirements Including at least one of the following: One or more non-temporary computer-readable storage media according to claim 13.
17. Each of the determined sets of SCMs corresponds to a directed acyclic graph that represents the variables associated with the first dataset, The directed acyclic graph is augmented based on the set of selection variables that represent the determined feature difference information. One or more non-temporary computer-readable storage media according to claim 13.
18. The group associated with the first dataset is switched to the group associated with the second dataset based on the application of a set of conditions to at least one different value in the set of selection variables. One or more non-temporary computer-readable storage media according to claim 17.
19. Each of the sets of ethical coefficients is determined based on the confidence level associated with the feature difference information and the causal transportability coefficient associated with the first and second datasets. One or more non-temporary computer-readable storage media according to claim 13.
20. The memory that stores the instructions, It is coupled to memory and executes the instruction, Receiving multiple datasets related to multiple application domains, Based on the first user input, the feature-dependent information related to the first dataset among the multiple datasets received is determined, Based on the second user input and ethical requirements, the feature difference information related to the first dataset and the second dataset among the multiple received datasets is determined, and the second dataset is different from the first dataset. Based on the determined feature dependency information and the determined feature difference information, a set of structural causal models (SCMs) related to the first dataset is determined. Based on the application of the causal transportability model to the determined set of SCMs, a set of ethical coefficients related to the set of ethical requirements is determined, Based on the set of ethical coefficients determined above, a reliability score associated with the first dataset is determined, The method involves training a meta-learning model related to the multiple application domains based on a plurality of reliability scores related to the plurality of datasets, wherein the plurality of reliability scores include the determined reliability score related to the first dataset. A processor that executes a process having An electronic device having
Citation Information
Patent Citations
Computer system
JP2020071694A
Method and apparatus for performing learning from demonstrations, particularly imitation learning
US20210192391A1
Method for detecting and mitigating bias and weakness in artificial intelligence training data and models
US20220012591A1
