Table data modeling method and device, equipment, medium and product

Abstract

The embodiment of the invention provides a table data modeling method and device, equipment, a medium and a product. The method comprises the steps of obtaining a table data set of a downstream financial task by constructing a hybrid prior model comprising a router model and a plurality of prior learners, and inputting the table data set into a pre-trained hybrid prior model; dynamically determining at least one target prior learner from the plurality of prior learners through a router model therein, and performing weight distribution; and based on a Bayesian inference mechanism, performing parameter calibration on the target prior learner by using the training sample in the table data set to output a prediction result of the test sample. The method is used for achieving the effects of improving the generalization ability and prediction performance of the model and improving the modeling efficiency of the table data.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, and in particular to a table data modeling method, device, equipment, medium and product. BACKGROUND

[0002] As a basic and extensive data form, table data has core application value in the fields of finance, medical treatment, e-commerce, etc. With the development of artificial intelligence technology, table data modeling methods based on machine learning have become a key tool for improving business decision-making efficiency.

[0003] At present, existing table data modeling methods mainly rely on traditional machine learning algorithms (such as random forest, gradient boosting tree) and pre-training methods based on meta-learning. Among them, the traditional machine learning algorithm needs to be trained from scratch for each task, and the pre-training method based on meta-learning is pre-trained on synthetic table data to achieve rapid prediction in a small sample scenario.

[0004] However, due to the highly imbalanced nature of financial data, and the fact that the existing technical methods do not fully consider the causal relationship between features and target variables, it is easy to cause deviation between synthetic data and real distribution, reducing the generalization ability and predictive ability of the model in the imbalanced table data scenario, further reducing the modeling efficiency of table data. SUMMARY

[0005] The embodiments of the present application provide a table data modeling method, device, equipment, medium and product to improve the generalization ability and prediction performance of the model, and improve the modeling efficiency of table data.

[0006] In a first aspect, the embodiments of the present application provide a table data modeling method, comprising:

[0007] obtaining a table data set of a downstream financial task;

[0008] inputting the table data set into a pre-trained hybrid prior model, the hybrid prior model comprising a router model and a plurality of prior learners;

[0009] dynamically determining at least one target prior learner from the plurality of prior learners and performing weight allocation through the router model;

[0010] based on a Bayesian inference mechanism, calibrating parameters of the target prior learner using training samples in the table data set to output a prediction result for a test sample, the prediction result being a modeling result of the downstream financial task.

[0011] Optionally, dynamically determining at least one target prior learner from the plurality of prior learners and performing weight allocation through the router model, specifically comprising:

[0012] analyzing the table dataset through a router model to determine a feature distribution of the table dataset;

[0013] selecting a target prior learner from the mixed prior model according to the feature distribution;

[0014] assigning a sparse weight based on a prediction performance of each target prior learner.

[0015] Optionally, assigning a sparse weight based on a prediction performance of each target prior learner, specifically includes:

[0016] determining a prediction probability distribution of the target prior learner for any one target prior learner;

[0017] assigning a corresponding sparse weight to the target prior learner according to the prediction probability distribution.

[0018] Optionally, before constructing the mixed prior model, generating a diversified dataset containing causal relationships through a structural causal model.

[0019] Optionally, after parameter calibration of the target prior learner, performing operator fusion and memory optimization processing on the calibrated target prior learner to adapt to a target NPU hardware architecture.

[0020] Optionally, analyzing the table dataset through a router model to determine a feature distribution of the table dataset, specifically includes:

[0021] determining a feature dimension and a missing value ratio of the table dataset, and determining the feature distribution of the table dataset based on the feature dimension and the missing value ratio; and / or,

[0022] introducing a multi-granularity attention mechanism to simultaneously calculate local feature associations and global feature patterns of the table dataset, and determining the feature distribution of the table dataset based on the local feature associations and the global feature patterns.

[0023] In a second aspect, an embodiment of the present application provides a table data modeling device, comprising:

[0024] an acquisition module configured to acquire a table dataset of a downstream financial task;

[0025] a processing module configured to input the table dataset into a pre-trained mixed prior model, the mixed prior model comprising a router model and a plurality of prior learners;

[0026] the processing module is further configured to dynamically determine at least one target prior learner from the plurality of prior learners through the router model and perform weight assignment;

[0027] The processing module is further configured to calibrate parameters of the target prior learner based on a Bayesian inference mechanism and using training samples in the table data set, to output a prediction result for a test sample, the prediction result being a modeling result for a downstream financial task.

[0028] Optionally, the processing module is further configured to analyze the table data set by using a router model to determine a feature distribution of the table data set.

[0029] selecting the target prior learner from the mixed prior model according to the feature distribution;

[0030] assigning a sparse weight to each target prior learner based on a prediction performance of the target prior learner.

[0031] Optionally, the processing module is further configured to, for any one target prior learner, determine a prediction probability distribution of the target prior learner.

[0032] assigning a corresponding sparse weight to the target prior learner according to the prediction probability distribution.

[0033] Optionally, the processing module is further configured to, before constructing the mixed prior model, generate a diversified data set containing causal relationships by using a structural causal model.

[0034] Optionally, the processing module is further configured to, after calibrating parameters of the target prior learner, perform operator fusion and memory optimization processing on the calibrated target prior learner to adapt to a target NPU hardware architecture.

[0035] Optionally, the processing module is further configured to determine a feature dimension and a missing value ratio of the table data set, and determine the feature distribution of the table data set based on the feature dimension and the missing value ratio; and / or,

[0036] by introducing a multi-granularity attention mechanism to simultaneously calculate local feature associations and global feature patterns of the table data set, and determining the feature distribution of the table data set based on the local feature associations and the global feature patterns.

[0037] In a third aspect, an embodiment of the present application provides an electronic device, comprising: a memory and a processor.

[0038] The memory stores computer execution instructions.

[0039] The processor executes the computer execution instructions stored in the memory, so that the processor executes the first aspect and / or various possible implementation manners of the first aspect.

[0040] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, the computer readable storage medium stores computer execution instructions, and the computer execution instructions are executed by a processor to implement the first aspect and / or various possible implementation manners of the first aspect.

[0041] In a fifth aspect, an embodiment of the present application provides a computer program product, comprising a computer program which, when executed by a processor, implements the first aspect and / or various possible implementation manners of the first aspect.

[0042] The table data modeling method, device, equipment, medium and product provided by the embodiments of the present application can achieve the effects of improving the generalization ability and prediction performance of the model, and improving the modeling efficiency of the table data by constructing a hybrid prior model including a router model and multiple prior learners, obtaining a table data set of a downstream financial task, and inputting the table data set into the pre-trained hybrid prior model; dynamically determining at least one target prior learner from the multiple prior learners and performing weight distribution through the router model in the hybrid prior model; and performing parameter calibration on the target prior learner based on a Bayesian inference mechanism, using training samples in the table data set, to output a prediction result for a test sample. BRIEF DESCRIPTION OF DRAWINGS

[0043] The accompanying drawings, which are incorporated herein and form part of the specification, illustrate embodiments consistent with the present application and, together with the description, further serve to explain the principles of the application.

[0044] Figure 1 Flowchart of the table data modeling method provided by the present application Figure One ;

[0045] Figure 2 Flowchart of the table data modeling method provided by the present application Figure Two ;

[0046] Figure 3 Structure diagram of the table data modeling device provided by the present application;

[0047] Figure 4 Structure diagram of the electronic device provided by the present application.

[0048] The specific embodiments of the present application have been shown by the above drawings, and will be described in more detail hereinafter. These drawings and textual descriptions are not intended to limit the scope of the concept of the present application by any means, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0049] The exemplary embodiments will be described in detail herein with reference to the attached drawings. In the following description, like reference numerals refer to like elements throughout the description. The following exemplary embodiments described herein represent illustrations only. They are not intended to represent all aspects of devices and methods consistent with the present disclosure. Rather, they are merely examples consistent with some aspects of the present disclosure as detailed in the appended claims.

[0050] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of related data comply with relevant laws, regulations and standards in relevant regions, necessary security measures are taken, public order and good customs are not violated, and appropriate operation portals are provided for users to choose authorization or refusal.

[0051] First, the terms involved in the present application are explained:

[0052] Meta-learning: refers to a pre-training technology that enables artificial intelligence models to efficiently generalize to new task scenarios;

[0053] Bayesian inference: refers to a statistical learning technique that relies on pre-trained model priors to calibrate and update on downstream datasets, enabling rapid posterior prediction with small samples;

[0054] Context learning mechanism: refers to a machine learning technique that uses a small number of examples of training datasets containing target variable labels to infer the target variable labels of the test set.

[0055] Tabular data is a basic and widely used form of data, especially in the fields of finance, medicine, e-commerce, etc. For example, in the financial field, tabular data is widely used in customer behavior analysis (such as potential customer mining, loss prevention prediction), financial product preference prediction (such as large-denomination deposit, financial product, and fund purchase willingness), risk control modeling (such as credit risk assessment, fraud detection), and time series prediction (such as yield volatility analysis) scenarios. The tasks in the above scenarios usually have small sample modeling requirements and data imbalance characteristics. For example, in the marketing scenario, target customers (such as high net worth customers) may only account for 0.01% of the total samples, while in the risk control scenario, the proportion of fraud samples or default samples may be lower than 0.01%.

[0056] Currently, the existing table data modeling methods can be divided into two categories: traditional machine learning algorithms (such as random forest, gradient boosting tree) and pre-training methods based on meta-learning. Among them, the traditional machine learning algorithm needs to be trained using a large number of samples, and needs to be re-modeled for each new task, which easily increases the training time, and the model under this method has poor reusability. And in the small sample scene, the model has poor prediction accuracy for unbalanced data (such as data sets with fewer positive samples). For the pre-training method based on meta-learning, the model is pre-trained based on synthetic table data to enable the model to master the general data correlation pattern, and then in the downstream task, it can directly predict without the need for secondary training. However, there is a gap between the distribution of the synthetic data and the real data using this method, which limits the generalization ability of the model, and the prediction result has poor accuracy when analyzing and predicting unbalanced data, and has poor adaptability to different devices (such as difficult to use GPU (Graphics Processing Unit, graphics processor) / NPU (Neural Processing Unit, neural processing unit) and other high-performance hardware acceleration), thereby reducing the inference efficiency.

[0057] In combination with the above scenarios, it can be seen that the existing technology has poor table data modeling accuracy in real scenarios, low inference efficiency, and low hardware utilization.

[0058] The table data modeling method provided by the present application first constructs a hybrid prior model including a router model and a plurality of prior learners, synthesizes diversified training data sets conforming to the financial data distribution using a structural causal model, and then performs multi-granularity analysis on the feature distribution of the input table data set through the router model to dynamically select the target prior learner and assign sparse weights to the target prior learner. Through the combination of the Bayesian inference mechanism, rapid parameter calibration in the small sample scene is realized, and operator fusion, memory optimization, and parallel scheduling are performed for the target NPU hardware for deep optimization processing, thereby solving the technical problems of low modeling efficiency, insufficient prediction performance, weak generalization ability, and poor hardware adaptability of the existing technology in the small sample and unbalanced financial table data scene, and improving the efficiency and performance of the financial table data modeling.

[0059] The technical solutions of the present application and how the technical solutions of the present application solve the above technical problems will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present application will be described below with reference to the accompanying drawings.

[0060] Figure 1 Flowchart of the table data modeling method provided by the present application Figure One For example,Figure 1 The method comprises the following steps:

[0061] S101, acquiring a table dataset of a downstream financial task.

[0062] More specifically, the table dataset comprises a plurality of feature columns and a target column, and is small sample data with a sample size less than a preset threshold. The downstream financial task includes but is not limited to potential customer mining, customer loss prevention, financial product preference prediction, credit risk assessment, fraud detection, etc.

[0063] Optionally, before acquiring the table dataset of the downstream financial task, a hybrid prior model is constructed.

[0064] For example, a plurality of prior learners (e.g., feature-level encoders) are integrated into a unified architecture, and each prior learner learns differentiated knowledge independently.

[0065] Optionally, the hybrid prior model is an ensemble model composed of a plurality of prior learners, and a part of the learners are selected to participate in prediction through a sparse activation mechanism. For example, only 3 prior learners are selected from 8 prior learners to participate in prediction.

[0066] Optionally, the hybrid prior model is pre-trained, and the pre-training step comprises: synthesizing a diversified synthetic table dataset through a structural causal model; taking table prediction as a meta-learning task, and pre-training the hybrid prior model using the synthetic table dataset to obtain a pre-trained hybrid prior model adapted to general table data rules after training.

[0067] Optionally, before constructing the hybrid prior model, a diversified dataset containing causal relationships is generated through a structural causal model (SCM). The structural causal model is used to simulate data generation mechanisms through a directed acyclic graph to reflect the causal relationships between variables, for example, to define the causal relationship between customer income and financial product purchase behavior through a neural network.

[0068] Optionally, the diversified dataset is used to provide diversified training data for the pre-training step, so that the hybrid prior model learns the context through the diversified dataset to enhance the rapid modeling and small sample learning ability on the downstream task, and enhance the generalization ability of the hybrid prior model.

[0069] Optionally, the method of generating a diversified dataset containing causal correlations through a structural causal model can be at least one of: dynamically adjusting the generation parameters of the structural causal model according to the data distribution of the downstream financial task, and then generating synthetic data through data initialization, forward propagation and post-processing steps to obtain a diversified dataset; or, based on the feature templates of multiple preset financial scenarios, preferentially generating a diversified dataset corresponding to the feature template of the current financial scenario.

[0070] Optionally, the feature template is a data distribution feature preset for a specific financial scenario (such as marketing, risk control). For example, a customer behavior data template with a high imbalance ratio and a long-tail distribution.

[0071] Optionally, by dynamically adjusting the generation parameters (such as topology complexity and causal strength) of the structural causal model or presetting the feature template, a diversified dataset that better fits the actual business needs is generated, and the matching degree of the synthetic data and the actual business scenario is improved through dynamic parameter adjustment and template matching.

[0072] For example, in the task of predicting stock fund preferences, according to the characteristics of a high imbalance ratio in this scenario, the sample imbalance ratio parameter of the structural causal model is dynamically adjusted, and the inter-layer connection probability is increased to simulate complex customer behavior correlations, thereby synthesizing a high-imbalance dataset that fits the actual situation.

[0073] The diversified dataset generated by the structural causal model in this embodiment enhances the quality of the pre-training data, so that the pre-training model not only learns statistical correlations but also reliable causal laws, thereby improving the generalization ability and decision-making reliability of the hybrid prior model in real financial scenarios. For example, in the task of predicting customer behavior, the hybrid prior model can more accurately capture causal correlations and avoid the prediction distortion caused by distribution bias in traditional techniques.

[0074] S102, input the table dataset into the pre-trained hybrid prior model, and the hybrid prior model includes a router model and multiple prior learners.

[0075] More specifically, the prior learner is a feature-level encoder with attention mechanisms in the feature dimension and the sample dimension, and the small sample learning based on the context learning mechanism is realized.

[0076] S103, dynamically determine at least one target prior learner from the multiple prior learners through the router model and perform weight allocation.

[0077] More specifically, the router model is used to dynamically determine at least one target prior learner from a plurality of prior learners and assign weights, specifically including: analyzing the table data set by the router model to determine the feature distribution of the table data set; selecting a target prior learner from the mixed prior model according to the feature distribution; and assigning sparse weights based on the prediction performance of each target prior learner.

[0078] Optionally, the feature distribution refers to the statistical characteristics of the input data, including but not limited to feature dimension, missing value ratio, category distribution, etc. For example, the proportion of continuous features in financial customer data, the proportion of positive samples, etc. The Bayesian inference mechanism refers to a parameter updating method based on a probability model, which calculates the posterior distribution through the prior distribution and the likelihood function.

[0079] Optionally, the router model is used to analyze the feature distribution (such as the proportion of positive samples and feature correlation) of the input table data set, dynamically select a matching prior learner, and assign sparse weights.

[0080] In one possible embodiment, the router model is a fully connected neural network. The router model analyzes the positive-negative sample imbalance ratio and feature dimension of the input table data set to obtain the feature distribution. Then, the router model predicts the adaptation degree of each prior learner to the feature distribution, and finally selects a preset number of prior learners in descending order of the adaptation degree, and assigns initial weights to the selected prior learners.

[0081] The embodiments of the present application improve the adaptability of the model to financial data sets with different characteristics (such as high-dimensional feature sets or credit scoring data containing a large number of missing values), and improve the generalization ability and robustness of the model, thereby enhancing the prediction performance of the model when facing diversified downstream tasks.

[0082] The table data modeling method provided by the embodiments of the present application dynamically determines the target prior learner through the router model in the pre-trained mixed prior model, and directly outputs the prediction result by performing rapid parameter calibration. This method does not need to start from zero to train the model, and only a small amount of samples and a short time are needed to complete modeling, thereby improving the prediction efficiency and modeling efficiency, and meeting the small sample and high timeliness requirements (such as quickly online marketing activities or real-time fraud detection) in the financial field.

[0083] Optionally, the table dataset is analyzed by the router model to determine the feature distribution of the table dataset, specifically including: determining the feature dimension and the missing value ratio of the table dataset, and determining the feature distribution of the table dataset based on the feature dimension and the missing value ratio; and / or introducing a multi-granularity attention mechanism to simultaneously calculate the local feature association and the global feature mode of the table dataset, and determining the feature distribution of the table dataset based on the local feature association and the global feature mode.

[0084] Optionally, the multi-granularity attention mechanism refers to an attention mechanism that simultaneously focuses on local features (such as the causal relationship between a single feature and a target) and global features (such as the interaction between features).

[0085] Optionally, by analyzing the feature dimension and the missing value ratio of the input table dataset, or introducing a multi-granularity attention mechanism, the perception ability of the router model for the feature distribution is improved.

[0086] In a possible embodiment, through the multi-granularity attention mechanism in the router model, bidirectional self-attention is calculated on the feature dimension to capture the association mode between different features such as “age” and “number of financial product holdings”, and bidirectional self-attention is calculated on the sample dimension to identify customer groups with similar behavior patterns such as “high net worth middle-aged customers”, thereby generating deep data feature publishing.

[0087] The embodiment provides a selection of different complexities and performances for the analysis method of the feature distribution: based on the basic statistical characteristics (i.e., the feature dimension and the missing value ratio), and introducing a multi-granularity attention mechanism to analyze local and global feature modes. Through the method based on the basic statistical characteristics, the macro features of the dataset are quickly grasped at a lower computational cost, and through the method of introducing a multi-granularity attention mechanism, deep mining of complex and deep semantic relationships in the data is realized, thereby improving the accuracy of dynamic selection when the router model processes financial data (such as user behavior sequences) with complex feature interactions.

[0088] Optionally, the sparse weight is assigned based on the prediction performance of each target prior learner, specifically including: for any one target prior learner, determining the prediction probability distribution of the target prior learner; and assigning the corresponding sparse weight to the target prior learner according to the prediction probability distribution.

[0089] In a possible embodiment, for the selected prior learners (for example, 3 prior learners are selected), the confidence of the prediction probability distribution output by each prior learner on the historical financial task is calculated respectively. Then, the prior learner with high confidence is assigned a higher weight.

[0090] The embodiment optimizes the sparse weight distribution by analyzing the prediction probability distribution, for example, in a financial product preference prediction task, the mixed prior model can dynamically adjust the weight according to the confidence of the prediction probability, thereby avoiding the interference of the low-confidence prior learner, and improving the prediction accuracy of the mixed prior model.

[0091] In S104, based on the Bayesian inference mechanism, the training samples in the table data set are used to calibrate the parameters of the target prior learner to output the prediction result of the test sample.

[0092] More specifically, based on the Bayesian inference mechanism, the training samples (a small number of samples) in the table data set are used to calibrate the parameters (or context learning) of the target prior learner to generate a final prediction model, and the prediction result of the test sample is output through the generated final prediction model, which is the modeling result of the downstream financial task.

[0093] Optionally, after calibrating the parameters of the target prior learner, the calibrated target prior learner is subjected to operator fusion and memory optimization processing to adapt to the target NPU hardware architecture.

[0094] In one possible embodiment, on the NPU of the target device, the bidirectional self-attention of the feature-level encoder in the feature dimension and the bidirectional self-attention in the sample dimension are fused into a single GEMM (General Matrix Multiply, general matrix multiplication) operation; at the same time, by sharing LayerNorm (Layer Normalization, layer normalization) and position embedding, weight sharing is realized, and the training set KV (Key-Value Cache, key-value cache) cache is stored in the NPU to avoid repeated calculation, further improving the inference speed.

[0095] The embodiment introduces the operator fusion and memory optimization of the specific hardware (target NPU) oriented backend processing step after calibrating the model parameters, realizes the deep optimization of the hardware level, enhances the adaptability to the device, and improves the inference speed and throughput of the mixed prior model, while reducing the power consumption. The modeling method provided in the present application can meet the demand for high concurrency and low delay real-time prediction in financial business.

[0096] Figure 2 The flowchart of the table data modeling method provided in the present application Figure Two As Figure 2As shown, the method comprises: first, synthesizing a diversified unbalanced dataset through a structural causal model. Specifically, a data synthesis algorithm based on a structural causal model is adopted, a directed acyclic graph simulating the logic of financial business is constructed, and recursive sampling and data generation are performed in the graph, so as to synthesize a tabular dataset including different feature dimensions, sample quantities and unbalanced proportions.

[0097] Then, the hybrid prior model is pre-trained based on the synthesized dataset generated in the previous step, that is, the synthesized diversified tabular dataset is input into the hybrid prior model for meta-learning. Specifically, the hybrid prior model in this embodiment is composed of 1 router and 8 prior learners based on feature-level encoders, and the pre-training task is a tabular prediction task, that is, the labeled training set and the unlabeled test set are input to make the model predict the labels of the test set. In the training process, the method of load balancing regularization term is adopted to make each prior learner learn differential knowledge, and after the training is completed, the pre-trained hybrid prior model parameters are output.

[0098] Finally, the inference optimization and NPU deployment step is performed, that is, the pre-trained model is adapted to the target NPU hardware. Specifically, through operator fusion, memory optimization, parallel scheduling and weight sharing technology, the calculation bottleneck in the hybrid prior model is optimized for hardware customization, so as to form a programmed product supporting user-defined input and evaluation, and realize the function of high throughput and low delay tabular data prediction.

[0099] Figure 3 The structural schematic diagram of the table data modeling device provided in the present application is shown in Figure 3 As shown, the table data modeling device 30 provided in the present embodiment comprises:

[0100] The acquisition module 301 is configured to acquire a tabular dataset of a downstream financial task.

[0101] The processing module 302 is configured to input the tabular dataset into a pre-trained hybrid prior model, and the hybrid prior model comprises a router model and a plurality of prior learners.

[0102] The processing module 302 is further configured to dynamically determine at least one target prior learner from the plurality of prior learners through the router model and perform weight distribution.

[0103] The processing module 302 is further configured to calibrate the parameters of the target prior learner based on a Bayesian inference mechanism and using the training samples in the tabular dataset, so as to output a prediction result of the test samples, and the prediction result is a modeling result of the downstream financial task.

[0104] Optionally, the processing module 302 is further configured to analyze the tabular dataset through the router model to determine the feature distribution of the tabular dataset.

[0105] selecting the target prior learner from the mixture prior model according to the feature distribution;

[0106] assigning the sparse weight based on the prediction performance of each target prior learner.

[0107] Optionally, the processing module 302 is further configured to determine, for any one target prior learner, a prediction probability distribution of the target prior learner.

[0108] assigning the corresponding sparse weight to the target prior learner according to the prediction probability distribution.

[0109] Optionally, the processing module 302 is further configured to generate a diversified dataset containing causal relationships through a structural causal model before constructing the mixture prior model.

[0110] Optionally, the processing module 302 is further configured to perform operator fusion and memory optimization processing on the calibrated target prior learner to adapt to the target NPU hardware architecture after parameter calibration of the target prior learner.

[0111] Optionally, the processing module 302 is further configured to determine a feature dimension and a missing value ratio of the table dataset, and determine the feature distribution of the table dataset based on the feature dimension and the missing value ratio; and / or,

[0112] The multi-granularity attention mechanism is introduced to simultaneously calculate the local feature association and the global feature pattern of the table dataset, and the feature distribution of the table dataset is determined based on the local feature association and the global feature pattern.

[0113] The table data modeling apparatus provided in this embodiment can perform the method provided in the method embodiments, and has similar implementation principles and technical effects. Details are not described herein again.

[0114] Figure 4 A structural schematic diagram of an electronic device provided in this application is shown in FIG. 4. As shown in FIG. 4, the electronic device 40 provided in this embodiment includes at least one processor 401 and a memory 402. Optionally, the device 40 further includes a communication component 403. The processor 401, the memory 402 and the communication component 403 are connected through a bus 404. Figure 4

[0115] In the specific implementation process, the at least one processor 401 executes the computer execution instructions stored in the memory 402, so that the at least one processor 401 performs the method described above.

[0116] The specific implementation process of the processor 401 can refer to the method embodiments described above, and has similar implementation principles and technical effects. Details are not described herein again. ​

[0117] In the above embodiments, it should be understood that the processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The steps of the disclosed method can be directly embodied as hardware processor execution, or a combination of hardware and software modules in the processor.

[0118] The memory can include a random access memory (RAM), and can also include a non-volatile memory (NVM), such as at least one disk memory.

[0119] The bus can be an industry standard architecture (ISA) bus, a peripheral component (PCI) bus, an extended industry standard architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus in the drawings of the present application does not limit to only one bus or one type of bus.

[0120] The present application also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the above method.

[0121] The present application also provides a computer readable storage medium, which stores computer execution instructions, and when a processor executes the computer execution instructions, the above method is implemented.

[0122] The above readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The readable storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0123] An example readable storage medium is coupled to the processor such that the processor can read information from the readable storage medium and can write information to the readable storage medium. Of course, the readable storage medium can also be a part of the processor. The processor and the readable storage medium can be located in an application specific integrated circuit (ASIC). Of course, the processor and the readable storage medium can also exist as discrete components in the device.

[0124] The division of units is only a logical functional division, and in actual implementation, there can be another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0125] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, can be located in one place, or can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0126] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit.

[0127] If the functions are realized in the form of software functional units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the part of the present application that essentially contributes to the prior art or the part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method of each embodiment of the present application. The foregoing storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0128] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction related hardware. The foregoing program can be stored in a computer readable storage medium. The program executes to perform the steps of the above-mentioned method embodiments; and the foregoing storage medium includes various media capable of storing program codes, such as ROM, RAM, magnetic disk, or optical disk.

[0129] Finally, it should be noted that other embodiments of the present application will readily occur to those skilled in the art upon consideration of the specification and practice of the present application disclosed herein. The present application is intended to include all such variations, uses, or adaptations of the application in which the general principles of the application are used to best advantage and encompassed within its scope. The present application is not limited to the precise structures described and shown in the accompanying drawings and figures, and can be practiced with variation of modifications and alterations without departing from the scope of the present application. The scope of the present application is limited only by the claims appended hereto.

Claims

1. A method for modeling tabular data, characterized in that, include: Obtain the tabular dataset for downstream financial tasks; The tabular dataset is input into a pre-trained hybrid prior model, which includes a router model and multiple prior learners; The router model dynamically determines at least one target prior learner from the plurality of prior learners and assigns weights accordingly. Based on the Bayesian inference mechanism, the target prior learner is calibrated using training samples in the tabular dataset to output prediction results for test samples. These prediction results are the modeling results for the downstream financial task.

2. The method according to claim 1, characterized in that, The router model dynamically determines at least one target prior learner from the plurality of prior learners and assigns weights accordingly, specifically including: The tabular dataset is analyzed using the router model to determine the feature distribution of the tabular dataset; Select the target prior learner from the hybrid prior model based on the feature distribution; Sparse weights are assigned based on the prediction performance of the prior learner for each target.

3. The method according to claim 2, characterized in that, Sparse weights are assigned based on the prediction performance of each target prior learner, specifically including: For any target prior learner, determine the predicted probability distribution of the target prior learner; The target prior learner is assigned corresponding sparse weights based on the predicted probability distribution.

4. The method according to claim 1, characterized in that, Also includes: Before constructing the hybrid prior model, a diverse dataset containing causal relationships is generated through a structural causal model.

5. The method according to claim 1, characterized in that, Also includes: After parameter calibration of the target prior learner, operator fusion and memory optimization are performed on the calibrated target prior learner to adapt it to the target NPU hardware architecture.

6. The method according to claim 2, characterized in that, The tabular dataset is analyzed using the router model to determine its feature distribution, specifically including: Determine the feature dimensions and missing value ratio of the tabular dataset, and determine the feature distribution of the tabular dataset based on the feature dimensions and missing value ratio; and / or, By introducing a multi-granularity attention mechanism, the local feature associations and global feature patterns of the tabular dataset are calculated simultaneously, and the feature distribution of the tabular dataset is determined based on the local feature associations and global feature patterns.

7. A tabular data modeling device, characterized in that, include: The acquisition module is used to acquire tabular datasets for downstream financial tasks; The processing module is used to input the tabular dataset into a pre-trained hybrid prior model, which includes a router model and multiple prior learners; The processing module is further configured to dynamically determine at least one target prior learner from the plurality of prior learners through the router model and perform weight allocation; The processing module is further configured to calibrate the parameters of the target prior learner using training samples in the tabular dataset based on a Bayesian inference mechanism, so as to output the prediction results for the test samples, wherein the prediction results are the modeling results of the downstream financial task.

8. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-6.

Images