Big data automatic modeling method and device and electronic equipment

Abstract

The invention discloses a big data automatic modeling method and device and electronic equipment, relates to the technical field of IT, and aims to solve the problem of low automatic modeling efficiency. The method comprises the steps that original data are preprocessed to obtain target data, and the target data comprise a training data set, a historical model and data flow logic information; training an initial model based on the training data set and the data flow logic information to obtain a first model, the initial model being constructed based on a historical model; and performing fine tuning on the first model based on fine tuning training data generated by the self-game fine tuning mechanism to obtain a target model. According to the embodiment of the invention, the efficiency of automatic modeling can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information technology (IT), and in particular to a big data automatic modeling method and device and electronic equipment. BACKGROUND

[0002] Big data automatic modeling is a process of processing massive data, selecting an optimal model and optimizing parameters through automatic technology. At present, in big data automatic modeling, a single algorithm is usually taken as the core, and modeling is performed in an artificial design manner.

[0003] Since the automatic modeling process in the prior art basically relies on manual judgment and adjustment, efficient automatic verification and model iteration are difficult to achieve, resulting in low automatic modeling efficiency. SUMMARY

[0004] Embodiments of the present application provide a big data automatic modeling method and device and electronic equipment to solve the problem of low automatic modeling efficiency in the prior art.

[0005] In a first aspect, embodiments of the present application provide a big data automatic modeling method, comprising:

[0006] Preprocessing the original data to obtain target data, the target data comprising a training data set, a historical model and data flow direction logic information;

[0007] Training an initial model based on the training data set and the data flow direction logic information to obtain a first model, the initial model being constructed based on the historical model;

[0008] Fine-tuning the first model based on fine-tuning training data generated by a self-game fine-tuning mechanism to obtain a target model.

[0009] Optionally, fine-tuning the first model based on fine-tuning training data generated by a self-game fine-tuning mechanism to obtain a target model comprises:

[0010] Constructing a first intermediate model based on the first model, the first intermediate model comprising a knowledge fusion layer and a knowledge internalization layer, the knowledge fusion layer being configured to perform weighted fusion on multiple knowledge representations based on a cross-attention mechanism to obtain a first knowledge representation, and the knowledge internalization layer being configured to perform internalization processing on the first knowledge representation to obtain a second knowledge representation;

[0011] Fusing and internalizing knowledge representations of multiple external large models into the first intermediate model to obtain a second model;

[0012] Fine-tuning the second model based on fine-tuning training data generated by a self-game fine-tuning mechanism to obtain the target model.

[0013] Optionally, the fusing and internalizing of the knowledge representations of the plurality of external large models into the first intermediate model to obtain a second model comprises:

[0014] inputting the training data set into a plurality of the external large models for processing to obtain a plurality of original knowledge representations of the external large models;

[0015] continuously training the first intermediate model based on the plurality of original knowledge representations until a convergence condition is met to obtain the second model;

[0016] wherein the i-th training comprises:

[0017] extracting an i-th batch of training data from the plurality of original knowledge representations, i being a positive integer;

[0018] inputting the i-th batch of training data into the knowledge fusion layer, and performing weighted processing on the i-th batch of training data based on a cross-attention mechanism to obtain a first knowledge representation;

[0019] inputting the first knowledge representation into the knowledge internalization layer for internalization processing to obtain a second knowledge representation;

[0020] detecting the second knowledge representation using a loss function to obtain a loss value;

[0021] adjusting parameters of the first intermediate model based on the loss value.

[0022] Optionally, the fine-tuning training data generated based on the self-game fine-tuning mechanism is used to fine-tune the second model to obtain a target model, comprising:

[0023] adding a cross-domain fusion module to the second model to obtain a second intermediate model;

[0024] generating fine-tuning training data based on a self-game fine-tuning mechanism to fine-tune the second intermediate model to obtain a target model;

[0025] wherein the cross-domain fusion module comprises:

[0026] a plurality of parallel sub-networks, each of the sub-networks being used to process a knowledge representation of one domain, and the plurality of sub-networks performing information transmission based on a cross-attention mechanism;

[0027] a convergence layer used to integrate outputs of the plurality of sub-networks to obtain a multi-domain fusion representation.

[0028] Optionally, after the fine-tuning training data generated by the self-game fine-tuning mechanism is used to fine-tune the second model to obtain a target model, the method further comprises:

[0029] defining a hierarchy according to the type and characteristics of the first knowledge representation;

[0030] decomposing the first knowledge representation into a plurality of first data based on the hierarchy;

[0031] performing embedding processing on the plurality of first data to obtain a plurality of first embedding vectors;

[0032] performing fusion processing on the plurality of first embedding vectors based on a cross-attention mechanism to obtain a second embedding vector;

[0033] inputting the second embedding vector into a policy optimization process for self-game training based on a soft policy optimization mechanism to obtain a policy model.

[0034] Optionally, the training of the initial model based on the training data set and the data flow logical information to obtain the first model comprises:

[0035] performing meta-learning training on the initial model to obtain a third intermediate model;

[0036] adding a virtual branch to the third intermediate model to obtain a fourth intermediate model, the virtual branch being used to generate a virtual class;

[0037] performing incremental learning on the fourth intermediate model based on the training data set to obtain the first model.

[0038] Optionally, the third intermediate model comprises a backbone network and a multi-scale feature self-distillation module, the backbone network comprising an input layer, a plurality of intermediate layers and an output layer connected in sequence, and at least one intermediate layer of the plurality of intermediate layers being connected with the multi-scale feature self-distillation module;

[0039] wherein the multi-scale feature self-distillation module comprises:

[0040] a plurality of scale-different feature extraction branches, the feature extraction branches being used to perform feature extraction on features output by a first intermediate layer to obtain scale features, the first intermediate layer being an intermediate layer of the plurality of intermediate layers connected with the multi-scale feature self-distillation module;

[0041] a projection branch, the projection branch being used to uniformly project a plurality of the scale features to obtain projection features, the projection features being used to determine a consistency loss;

[0042] The fusion branch is used for inputting the plurality of scale features after fusion into a second intermediate layer, and the second intermediate layer is an intermediate layer located at a layer next to the first intermediate layer among the plurality of intermediate layers.

[0043] Optionally, the incrementally learning the fourth intermediate model based on the training data set to obtain the first model comprises:

[0044] The fourth intermediate model is incrementally learned based on the training data set for N rounds to obtain the first model.

[0045] The rth round of incrementally learning comprises:

[0046] The sample data of the rth category is extracted from the training data set, the rth category is a new category of the rth round of incrementally learning, N is a positive integer, and r is a positive integer less than or equal to N.

[0047] The sample data of the rth category and part of the old category sample data constitute a training subset, the old category sample data is sample data of the first r-1 categories, and the first r-1 categories are old categories of the first r-1 rounds of incrementally learning.

[0048] The model of the r-1th round of training is incrementally learned based on the training subset to obtain the model of the rth round of training, and the model obtained by the Nth round of training is the first model.

[0049] In a second aspect, an embodiment of the present application further provides a big data automatic modeling device, comprising:

[0050] A preprocessing module is configured to preprocess original data to obtain target data, wherein the target data comprises a training data set, a historical model and data flow direction logic information.

[0051] A training module is configured to train an initial model based on the training data set and the data flow direction logic information to obtain a first model, wherein the initial model is constructed based on the historical model.

[0052] A fine-tuning module is configured to fine-tune the first model based on fine-tuning training data generated by a self-game fine-tuning mechanism to obtain a target model.

[0053] In a third aspect, an embodiment of the present application further provides an electronic device, comprising a memory, a processor and a program stored in the memory and executable on the processor.

[0054] The processor is configured to read the program in the memory to implement the steps in the big data automatic modeling method of the first aspect.

[0055] In a fourth aspect, the embodiments of the present application further provide a readable storage medium for storing a program, which is executed by a processor to implement the steps in the big data automatic modeling method according to the first aspect.

[0056] In the embodiments of the present application, the original data is preprocessed to obtain target data, the target data including a training data set, a historical model and data flow direction logic information; an initial model is trained based on the training data set and the data flow direction logic information to obtain a first model, the initial model being constructed based on the historical model; the first model is fine-tuned based on fine-tuning training data generated by a self-game fine-tuning mechanism to obtain a target model. Through the above method, the training effect of the model can be improved based on the self-game fine-tuning mechanism, the model performance is optimized by self-generated data, the dependence on artificial labels is reduced, and the ability to handle complex tasks is enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the description of the embodiments of the present application will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0058] Figure 1 is a flow chart of the big data automatic modeling method provided by an embodiment of the present application;

[0059] Figure 2 is a flow chart of the big data automatic modeling method provided by another embodiment of the present application;

[0060] Figure 3 is a structural diagram of the big data automatic modeling device provided by an embodiment of the present application;

[0061] Figure 4 is a structural diagram of the electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0063] For the convenience of understanding, the following will first introduce some technical terms involved in the present application:

[0064] Multiple Mixing Self-Distillation (M2SD) is a model optimization method that combines multi-scale feature fusion and self-distillation learning, aiming to improve the feature representation ability and generalization performance of deep neural networks. The core idea is to realize efficient distillation without additional teacher models through multi-level feature interaction and knowledge transfer within the model.

[0065] Cross-Domain Few-Shot Learning is an important branch of Meta-Learning and Few-Shot Learning, aiming to solve the problem of how models quickly adapt to new tasks with very few samples (such as 1-5 samples per class) when there are significant distribution differences between the source domain and the target domain.

[0066] Model-Agnostic Meta-Learning (MAML) is a meta-learning method that aims to train a model to quickly adapt to new tasks with very few gradient updates. MAML does not directly learn the model parameters of each task, but learns a starting point from which it can quickly adapt to any task.

[0067] Self-Play Fine-Tuning (SPIN) is a training strategy that generates data and iteratively optimizes itself. The core idea is to let the model constantly confront or cooperate with its historical versions during training, generating diverse training samples and gradually improving performance.

[0068] PHI-3 model is a high-efficiency and lightweight language model, focusing on achieving high performance with small parameter quantity, especially suitable for edge devices, mobile terminals and resource-constrained scenarios.

[0069] Soft Policy Optimization (SPO) is a method of improving policy optimization in Reinforcement Learning (RL), which introduces an entropy regularization term in the policy optimization process to encourage the randomness of the policy, thereby improving the exploration ability and robustness.

[0070] Based on this, the embodiment of the present application provides a big data automatic modeling method, device and electronic equipment, aiming to improve the efficiency of big data automatic modeling. The present application can be applied to different fields, based on the big data automatic modeling in the field, which is not limited here. For the convenience of understanding, the present application will be applied to the operation and maintenance field, and the operation and maintenance data automatic modeling will be taken as an example for illustration.

[0071] As shown in Figure 1 , the big data automatic modeling method specifically includes the following steps:

[0072] Step 101, preprocessing the original data to obtain target data, the target data including training data set, historical model and data flow logic information.

[0073] When the operation and maintenance big data automatic modeling is based on the present method, the original data is operation and maintenance related data. In order to improve the efficiency of automatic modeling and the quality of model, the original data not only includes the training data set composed of actual operation and maintenance data, but also includes the historical model and data flow logic information previously constructed. In the case of applying the present method to the operation and maintenance field, the historical model can be the previously constructed and / or used discipline inspection model for operation and maintenance big data modeling. The data flow logic information specifically includes data flow and data logic. Exemplarily, the training data set includes actual data of each business system (such as procurement, finance, human resources, etc.); the historical discipline inspection model includes model architecture, model parameters and feature selection of historical discipline inspection model, etc.

[0074] As shown in Figure 2 , in some embodiments, the preprocessing of the original data specifically includes:

[0075] 1. Data source identification and classification. Specifically, the source of the original data is identified and classified, for example, the source of the actual operation and maintenance data in the original data includes provincial companies, different business systems (such as procurement, finance, human resources, etc.) or historical records (such as complaint records, audit reports, etc.).

[0076] 2. Data collection and integration. Specifically, the actual operation and maintenance data is extracted and converted using ETL (Extract, Transform, Load) tool, and integrated into a unified format. The data flow logic information is integrated into the central database through data mapping and conversion tool, so as to be accessed and used subsequently.

[0077] 3. Data cleaning and processing. Specifically, the integrated data is cleaned, for example, for actual operation and maintenance data, invalid data, duplicate data and outliers in the data are removed, and missing data is filled. For data flow logical information, these data are audited to ensure the integrity and accuracy of the information. Finally, the cleaned data is standardized to ensure consistency in format and dimension, providing high-quality data support for subsequent modeling.

[0078] 4. Data security and compliance check. Specifically, sensitive information in the original data is protected, and the transmission of the original data is encrypted. At the same time, it is checked whether the data processing and use comply with relevant regulations to ensure the legal use of data.

[0079] In some embodiments, in step 101, after preprocessing the original data to obtain target data, the target data is stored in an efficient and easily accessible data warehouse, and a data management mechanism is established to realize version control of historical models, metadata management and regular backup, to ensure the long-term availability and traceability of the target data. Through the above operations, it can be ensured that the target data required in the automatic modeling process can be easily accessed and used, and the latest state of the target data can be maintained.

[0080] Step 102, training an initial model based on the training data set and the data flow logical information to obtain a first model, the initial model being constructed based on the historical model.

[0081] It should be understood that the initial model is constructed based on the historical model, and in specific implementation, the initial model is a meta representation form of the historical model. In some embodiments, the number of historical models is multiple, and the initial model is obtained based on the meta representation forms of the multiple historical models.

[0082] Optionally, in some embodiments, the step 102 includes:

[0083] meta-learning training the initial model to obtain a third intermediate model;

[0084] adding a virtual branch to the third intermediate model to obtain a fourth intermediate model, the virtual branch being used to generate a virtual class;

[0085] incremental learning the fourth intermediate model based on the training data set to obtain the first model.

[0086] In some embodiments, the initial model is meta-learning trained through the MAML mechanism, and the trained initial model is the third intermediate model. In this embodiment, the third intermediate model trained through the meta-learning mechanism of MAML can quickly adapt to small sample data in a new field.

[0087] After the meta-learning training, a virtual branch is added on the backbone of the third intermediate model to obtain a fourth intermediate model. The number of virtual branches is multiple, and each virtual branch is used to generate a virtual class. The fourth intermediate model can generate virtual instances through the virtual branch to realize a multi-branch virtual class mechanism, thereby enhancing the expansibility and adaptability of the feature space.

[0088] In some embodiments, the fourth intermediate model is incrementally learned based on a training data set through a cross-domain few-shot learning mechanism to obtain a first model. Through the above cross-domain few-shot learning method, incremental learning can be performed using a small number of new class samples, while avoiding the problem of catastrophic forgetting.

[0089] Optionally, in some embodiments, the third intermediate model includes a backbone network and a multi-scale feature self-distillation module, the backbone network includes an input layer, a plurality of intermediate layers and an output layer connected in sequence, and at least one intermediate layer of the plurality of intermediate layers is connected with the multi-scale feature self-distillation module;

[0090] The multi-scale feature self-distillation module includes:

[0091] a plurality of feature extraction branches of different scales, the feature extraction branches being used for feature extraction on features output by a first intermediate layer to obtain scale features, the first intermediate layer being an intermediate layer of the plurality of intermediate layers connected with the multi-scale feature self-distillation module;

[0092] a projection branch, the projection branch being used for unified projection of a plurality of the scale features to obtain projection features, the projection features being used for determining a consistency loss;

[0093] a fusion branch, the fusion branch being used for inputting a plurality of the scale features after fusion into a second intermediate layer, the second intermediate layer being an intermediate layer of the plurality of intermediate layers located at a next layer of the first intermediate layer.

[0094] The backbone network is the core skeleton of the third intermediate model, carrying the basic representation ability and the classification task structure, and its specific structure is not limited here. The multi-scale feature self-distillation module is an expansion module set on the basis of the backbone network. The receptive field convolution kernels of the plurality of feature extraction branches of different scales are different, thereby realizing feature extraction of different scales and enhancing the perception ability of the model to complex data. The projection branch is used for unified projection of a plurality of scale features, and then realizes feature self-distillation, aligns and optimizes the semantics between the backbone network and the expansion features, and improves the semantic consistency and generalization performance of the entire model.

[0095] Feature self-distillation is a logical process structure, and is not repeated training of the entire network, but hierarchical guidance and attention distillation of fused scale features. By way of example, assuming that the number of feature extraction branches is 3, and the convolution kernel sizes thereof are 1x1, 3x3, and 5x5, respectively, for capturing semantic information of different granularities. If a multi-scale feature self-distillation module is connected to the third intermediate layer of the backbone network, then the features output by the third intermediate layer are first input into multiple feature extraction branches with different scales to extract features, to obtain scale features output by each feature extraction branch. Then, all scale features are uniformly projected through a projection branch to obtain corresponding projected features, and a consistency loss is determined by KL divergence or L2 loss distillation with the final representation of the main branch, to perform feature consistency optimization, wherein the main branch is a main path output channel of the model for a main task (such as classification / discrimination). Specifically, the main branch here refers to a structure in the backbone network, that is, a main feature expression path after fusion of the model. In this embodiment, by comparing the outputs of the multiple feature extraction branches after alignment with the output of the backbone network, the consistency and expression ability of the main features can be optimized.

[0096] In this embodiment, by combining feature enhancement technology with the self-distillation process, the classification performance of the model is significantly improved when learning new classes across domains. The focus of feature enhancement is to pass the diverse features of virtual classes to the backbone network and strengthen these features through attention mechanisms to improve the discrimination ability of the model.

[0097] It should be understood that in some embodiments, after meta-learning training of the initial model to obtain the third intermediate model, an adaptive feature space is pre-constructed using the M2SD method, reserving space for possible addition of new classes in the future. In the cross-domain small sample learning stage, the feature space pre-constructed using M2SD provides support for the introduction of new classes. The self-distillation mechanism enhances the adaptability of the model to features of different domains, and the KL divergence maintains the feature consistency between new and old classes.

[0098] Optionally, in some embodiments, the incremental learning of the fourth intermediate model based on the training data set to obtain the first model comprises:

[0099] N rounds of incremental learning of the fourth intermediate model based on the training data set to obtain the first model;

[0100] wherein the rth round of incremental learning comprises:

[0101] extracting sample data of an rth class from the training data set, the rth class being a new class of the rth round of incremental learning, N being a positive integer, and r being a positive integer less than or equal to N;

[0102] The sample data of the rth category and part of the old category sample data constitute a training subset, the old category sample data is the sample data of the first r-1 categories, and the first r-1 categories are the old categories of the first r-1 rounds of incremental learning.

[0103] Incremental learning is performed on the model of the r-1th round of training based on the training subset, and a model of the rth round of training is obtained, wherein the model obtained by the Nth round of training is the first model.

[0104] In each incremental learning stage, sample data of new categories is dynamically extracted from the training data set, combined with part of the old category sample data, to construct a training subset of the current stage. In some embodiments, MAML meta-learning algorithm is introduced in each incremental learning stage to optimize the parameters of the model to quickly adapt to the feature distribution of the new category. In other embodiments, in each incremental learning stage, the M2SD method is applied to feature remodeling and self-distillation of new categories to ensure that the model can maintain efficient knowledge transfer in new tasks and minimize interference with old category knowledge.

[0105] It should be understood that in some embodiments, after obtaining the first model in step 102, the model logic corresponding to the first model is generated based on the first model. In subsequent steps, instead of directly using the model structure of the first model for all subsequent tasks, the model logic corresponding to the first model is further automatically generated based on the trained model feature space, distilled semantic structure and virtual class representation to guide the training of subsequent models.

[0106] In specific implementation, after the model logic is generated, it needs to be verified to ensure that the model can effectively maintain the coherence and classification accuracy of knowledge when facing new data of different domains and categories.

[0107] As a specific embodiment, multi-scale feature self-distillation is combined with cross-domain small sample learning, and MAML is introduced as a meta-learning framework to improve the generalization ability of the model in cross-domain scenarios. Through the meta-learning mechanism of MAML, the model can quickly adapt to small sample data in new domains and enhance feature consistency through self-distillation, thereby effectively solving the catastrophic forgetting problem in cross-domain learning. This integration not only improves the adaptability and robustness of the model, but also provides a more flexible and efficient solution for cross-domain data modeling. In this embodiment, the update method of the model parameters is as follows:

[0108]

[0109] wherein θ represents the model parameters, α is a preset parameter for controlling the external learning rate, β is a preset parameter for controlling the internal learning rate, and α and β determine the speed of task generalization, representing the task a loss function on the current model, KL() is used to represent the Kullback-Leibler (KL) divergence, the KL divergence is used to measure the difference between the current model and the output of the previous round model, thereby reducing the feature shift, and λ is a coefficient for balancing the task loss and the self-distillation loss, is the output of the current model, is the output of the previous round model, as a distillation target, wherein, is the task distribution.

[0110] Step 103, fine-tuning the first model based on the fine-tuning training data generated by the self-game fine-tuning mechanism to obtain a target model.

[0111] The core of SPIN fine-tuning is to generate and optimize its own training data through self-game, thereby improving the performance of the model. The first model will generate simulated data scenarios in this stage, and identify its own logical weaknesses through self-antagonistic way. The fine-tuning training data generated by each round of self-game will be used to further fine-tune the model to continuously improve its decision-making ability in actual application, until the target model is obtained.

[0112] In this embodiment, the first model is further optimized using the SPIN mechanism, and the optimized first model is the target model. Optionally, in some embodiments, the step 103 comprises:

[0113] constructing a first intermediate model based on the first model, the first intermediate model comprising a knowledge fusion layer and a knowledge internalization layer, the knowledge fusion layer being configured to perform weighted fusion on multiple knowledge representations based on a cross-attention mechanism to obtain a first knowledge representation, and the knowledge internalization layer being configured to perform internalization processing on the first knowledge representation to obtain a second knowledge representation;

[0114] fusing and internalizing the knowledge representations of multiple external large models into the first intermediate model to obtain a second model;

[0115] fine-tuning the second model based on the fine-tuning training data generated by the self-game fine-tuning mechanism to obtain the target model.

[0116] It should be understood that since the first intermediate model is further constructed based on the first model, the first intermediate model also inherits and extends the network structure (such as the backbone network and the multi-scale feature self-distillation module) of the first model, and embeds new structures (such as the knowledge fusion layer and the knowledge internalization layer) related to knowledge transfer, fusion and internalization on this basis, thereby realizing the complete upgrade from feature space modeling to multi-source knowledge integration.

[0117] It should be understood that in some embodiments, the plurality of external large models are all PHI-3 models, and the first intermediate model is a PHI-3 model. Instead of directly adding a knowledge fusion layer and a knowledge internalization layer on the basis of the first model, the first intermediate model is generated based on the first model to further generate a classification logic path for a new field (i.e., a model logic corresponding to the first model), thereby realizing adaptive discrimination. Therefore, the first intermediate model is a model obtained based on an interpretable reasoning rule or decision path derived from the first model.

[0118] In some embodiments, the model logic corresponding to the first model serves as a structural prior or semantic guide to participate in the training and feature design of the current first intermediate model. Instead of being directly calculated in the network, it affects feature construction, distillation path, task division, and other strategies, so that the first intermediate model is more directional and interpretable in the training process.

[0119] It should be understood that "fusion" is based on two processes including "externalization" and "transfer". First, through the externalization process, the probability distribution of the external large model is obtained, and the probability distribution is taken as the original knowledge representation. Then, through the transfer process, the original knowledge representation is transferred to the first intermediate model.

[0120] It should be understood that the core mechanism of "internalization" is to obtain the first knowledge representation by weight fusion of each plurality of original knowledge representations through a hierarchical cross-attention module (Hierarchical Cross-Attention Block), while introducing a residual feedback structure to maintain the stability of the original model. Subsequently, these first knowledge representations are mapped to the representation of the current hidden state space of the first intermediate model through a knowledge reconstruction multilayer perceptron (Multilayer Perceptron, MLP), and are used to drive the parameter update of the first intermediate model, thereby completing the knowledge structure adjustment. This process can be called internalization.

[0121] It should be understood that the external large model is a pre-trained large model, and the fusion and internalization of the knowledge representation of the plurality of external large models to the first intermediate model can make the obtained second model have strong cross-domain knowledge processing capability. In specific implementation, first, the knowledge representation of the plurality of large models is extracted and fused, retaining the unique advantages of each large model. Then, through the hierarchical cross-attention mechanism and the continuous training method, these knowledge is transferred and internalized to the first intermediate model, and finally the second model is obtained. In some embodiments, after obtaining the second model, the model logic corresponding to the second model can be generated based on the second model, and the model logic corresponding to the second model can guide the subsequent training of the second model.

[0122] Optionally, in some embodiments, the fusing and internalizing the knowledge representations of the plurality of external large models into the first intermediate model to obtain a second model comprises:

[0123] inputting the training data set into a plurality of the external large models for processing to obtain a plurality of original knowledge representations of the external large models;

[0124] continuously training the first intermediate model based on the plurality of original knowledge representations until a convergence condition is met to obtain the second model;

[0125] wherein the i-th training comprises:

[0126] extracting an i-th batch of training data from the plurality of original knowledge representations, i being a positive integer;

[0127] inputting the i-th batch of training data into the knowledge fusion layer, and performing weighted processing on the i-th batch of training data based on a cross-attention mechanism to obtain the first knowledge representation;

[0128] inputting the first knowledge representation into the knowledge internalization layer for internalization processing to obtain the second knowledge representation;

[0129] detecting the second knowledge representation using a loss function to obtain a loss value;

[0130] adjusting parameters of the first intermediate model based on the loss value.

[0131] To obtain the knowledge representations of the plurality of external large models, first, the training data set is input into the plurality of external large models for processing. For each external large model, the training data set is input into the external large model to obtain the probability distribution of the next token prediction output by the external large model based on the input semantics. Each external large model generates a probability distribution of the next token based on the knowledge representation obtained by training, and these probability distributions reflect the understanding and domain knowledge of each external large model for the input data. Therefore, the probability distribution output by the external large model is regarded as the original knowledge representation of the external large model for a specific problem. By analyzing the output of each external large model, the advantages and tendencies of the external large model in a specific task can be identified.

[0132] The original knowledge representations output by different external large models are weighted processed by the hierarchical cross-attention mechanism in the knowledge fusion layer to obtain the first knowledge representation corresponding to each original knowledge representation. Subsequently, the plurality of first knowledge representations are input into the knowledge internalization layer for internalization processing to obtain the second knowledge representation. In some embodiments, the knowledge internalization layer is composed of a residual connection, a Multilayer Perceptron (MLP) knowledge reconstructor, and an attention selection unit.

[0133] In the above manner, the parameters of the first intermediate model are adjusted through continued training to gradually absorb the externalized knowledge and internalize it as its own knowledge representation. The goal of the training is to make the probability distribution generated by the first intermediate model as close as possible to the output of the external large model, so that the knowledge characteristics and advantages of multiple external large models can be retained to the greatest extent. In specific implementation, a KL divergence or other measurement method is used to minimize the probability distribution difference between the first intermediate model and the external large model, thereby achieving effective knowledge transfer.

[0134] Optionally, in some embodiments, the fine-tuning training data generated based on the self-game fine-tuning mechanism is used to fine-tune the second model to obtain a target model, including:

[0135] A cross-domain fusion module is added to the second model to obtain a second intermediate model;

[0136] The fine-tuning training data generated based on the self-game fine-tuning mechanism is used to fine-tune the second intermediate model to obtain a target model;

[0137] The cross-domain fusion module includes:

[0138] A plurality of parallel sub-networks, each of which is used to process the knowledge representation of one domain, and the plurality of sub-networks perform information transmission based on a cross-attention mechanism;

[0139] A convergence layer that integrates the outputs of the plurality of sub-networks to obtain a multi-domain fusion representation.

[0140] In specific implementation, after obtaining the second intermediate model, the second intermediate model needs to be trained again based on this to obtain a trained second intermediate model, and then fine-tune based on the trained second intermediate model to obtain a target model.

[0141] Specifically, in the present embodiment, a hierarchical cross-attention mechanism is adopted to process knowledge representations of different domains, through which multiple parallel sub-networks are established within the model, each of which focuses on processing knowledge of different domains and interacts information between them, and the sub-networks realize cross information transmission through shared attention guiding units. While each sub-network encodes the knowledge in the field, it perceives the semantic representation of other sub-networks through the cross-attention mechanism, and then dynamically adjusts its own representation weight. Finally, the fused representation is integrated by a unified aggregation layer for further fine-tuning of the main model parameters. Through the above setting of the cross-domain fusion module, it is ensured that the second intermediate model can mine potential connections from multiple knowledge perspectives when processing data, thereby improving the cross-domain understanding, reasoning and decision-making ability.

[0142] In the present embodiment, after completing the "externalization, transfer, fusion and internalization" of knowledge, cross-domain knowledge fusion is also needed, so a cross-domain fusion module is added based on the second model to obtain a second intermediate model. Externalization and transfer are knowledge acquisition and migration, while cross-domain fusion in the present embodiment is to realize unified representation generation of multi-source knowledge, internalization is absorption at the model parameter level, and cross-domain fusion is deep collaborative modeling of different domain knowledge within the model.

[0143] In the present embodiment, the SPIN mechanism is used to further optimize the second model, and the optimized second model is the target model. In some embodiments, after constructing the second intermediate model based on the second model, the SPIN mechanism is used to further optimize the second intermediate model, and the optimized second intermediate model is the target model.

[0144] In one specific embodiment, during the externalization and transfer of the knowledge of the plurality of external PHI-3 models, the SPIN mechanism is introduced to further optimize the performance of the second model. Through cross-domain knowledge fusion and self-game fine-tuning, the second model not only successfully absorbs the knowledge advantages of the plurality of external PHI-3 models, but also enhances the reasoning ability and adaptability of the model in data analysis by self-generating and optimizing training data, significantly improving the analysis efficiency and accuracy of the model. In the present embodiment, the updating method of the model parameters is as follows:

[0145]

[0146] wherein, θ t represents the current model parameter, θ t+1 represents the updated model parameter, η is a preset learning rate, ∈ is a coefficient for balancing knowledge transfer and strategy optimization, is a strategy parameter, R(τ) is the return on trajectory τ, which is used to reflect the effectiveness of the strategy, a first knowledge representation for representing the first intermediate model output, a second knowledge representation for representing the first intermediate model output, a strategy reward generated for self-game.

[0147] In specific implementation, the second model can be further fine-tuned based on the fine-tuning training data generated by the SPIN mechanism, and through continuous self-game training, the second model can continuously optimize its reasoning path and logical structure. Through the above-mentioned manner, not only can the final target model adapt to different types of data, but also can dynamically adjust its analysis framework according to the scene, ensuring high precision and consistency in actual application.

[0148] Taking the discipline inspection data scene as an example, when the target model is applied to the discipline inspection data scene, the reasoning ability of the target model can be used to automatically analyze and model the structured and unstructured data from the discipline inspection system to automatically generate data model logic. In this process, the target model uses the fused cross-domain knowledge and the reasoning ability optimized through the SPIN mechanism to automatically process and analyze complex discipline inspection data.

[0149] In specific implementation, the adaptability of the target model in different scenes needs to be verified through multiple applications and tests, and the target model needs to be further iteratively optimized according to feedback. At the same time, as the application scene expands and the data environment changes, the target model continues to be optimized through the feedback loop, and regular SPIN fine-tuning training ensures that the target model can continuously absorb knowledge in new data and scenes, and continuously improve its analysis ability and decision accuracy. Through this continuous iteration, the model can maintain foresight and innovation in data analysis.

[0150] Optionally, in some embodiments, after the fine-tuning training data generated based on the self-game fine-tuning mechanism is used to fine-tune the second model to obtain the target model, the method further includes:

[0151] defining a hierarchical structure according to the type and characteristics of the first knowledge representation;

[0152] decomposing the first knowledge representation into a plurality of first data based on the hierarchical structure;

[0153] performing embedding processing on the plurality of first data to obtain a plurality of first embedding vectors;

[0154] performing fusion processing on the plurality of first embedding vectors based on a cross-attention mechanism to obtain a second embedding vector;

[0155] inputting the second embedding vector into a strategy optimization process for self-game training based on a soft strategy optimization mechanism to obtain a strategy model.

[0156] In this embodiment, the model effectively integrates information from different levels through a hierarchical cross-attention mechanism, providing richer feature input for policy optimization. Furthermore, the soft policy optimization mechanism, through a self-game and zero-sum game framework, eliminates reliance on reward models and adversarial training, improving the efficiency and stability of policy optimization.

[0157] Specifically, the hierarchical structure of the data is first defined, and its hierarchical structure is defined according to the type and characteristics of the input data (the first knowledge representation obtained in the previous step, specifically including information of multiple modalities such as text, images, and structured disciplinary inspection data). For example, for multimodal data (such as text, images, and sensor data), it can be divided into multiple levels, each level representing a specific feature or source of information. The definition of this hierarchical structure ensures that the model can process information at different levels of abstraction, laying the foundation for the subsequent cross-attention mechanism. Based on the hierarchical structure defined above, the first knowledge representation is decomposed into multiple first data, and the multiple first data are data of different hierarchical structures. Then, the first data is converted into a representation vector using an embedding method (such as a Transformer encoder or a convolutional neural network) to obtain a first embedding vector. The first embedding vector will serve as the input of the subsequent hierarchical cross-attention mechanism to ensure that the information at each level is properly represented and transmitted.

[0158] The following is an explanation of the specific implementation of the cross attention mechanism:

[0159] A cross-attention mechanism is introduced between adjacent or related layers, and the cross-attention weight is calculated using the following formula to generate the second embedding vector:

[0160]

[0161] Among them, d k The dimension used to represent the key vector to ensure numerical stability; Q i Used to represent the query vector, which represents the information that the model is currently “focusing on”, K j Used to represent the key vector, representing the information that can be queried in the hierarchy, V j Used to represent the value vector, which contains the actual information or features in the hierarchy.

[0162] The representations generated by the cross-attention mechanism are hierarchically combined, and a new comprehensive representation, namely the second embedding vector, is formed by weighted summation of the attention weights of each layer.

[0163] In this embodiment, the SPO self-game training mechanism is introduced, and the second embedding vector generated by the hierarchical cross-attention is input into the policy optimization process in the SPO framework. SPO is trained through self-game, compares these trajectories using a preference model, evaluates the pros and cons of each pair of trajectories, and records the win rate. Based on the comparison results of the trajectories, the policy update is guided by calculating the win rate of the trajectories. Unlike traditional RLHF methods, SPO does not rely on a reward model, but ensures the convergence of the policy through a zero-sum game. The specific policy update formula is as follows:

[0164]

[0165] wherein, π t is the current policy, π t+1 is the updated policy, WinRate(τ) represents the win rate of the current trajectory τ, is the learning rate.

[0166] In this embodiment, through multiple rounds of self-game training, the SPO framework gradually optimizes the policy and makes it converge to the globally optimal policy. Since SPO avoids adversarial training, the entire training process is more stable and efficient.

[0167] As a specific embodiment, the combination of hierarchical cross-attention mechanism and SPO reinforcement learning creates an efficient policy optimization method without adversarial training. This embodiment provides more rich feature input for policy optimization through multi-level information fusion, and realizes efficient convergence and global optimization of the policy using the self-game framework of SPO. This combination not only improves the efficiency of policy optimization of the model, but also enhances the stability and adaptability of the system. In this embodiment, the specific formula for policy update is as follows:

[0168]

[0169] wherein, π t is the current policy, π t+1 is the updated policy, ε is the learning rate of policy update, Q is the query vector, K is the key vector, V is the value vector, softmax(·) represents multi-level attention calculation combined with cross-layer information, WinRate(τ) represents the win rate of the trajectory τ, reflecting the effect of policy optimization.

[0170] In some embodiments, after obtaining the target model, the model logic corresponding to the target model is automatically generated based on the target model. The model logic represents the reasoning structure and knowledge path of the model when processing data, and can guide the training of the model in the subsequent model optimization process.

[0171] After obtaining the target model and the strategy model, a comprehensive evaluation and optimization of the target model and the strategy model are needed to ensure the effectiveness and practicality of the newly built model logic. Optionally, in some embodiments, the performance of the target model and the strategy model is evaluated and necessary adjustments and improvements are made by introducing reinforcement learning and adversarial generative model techniques.

[0172] First, evaluation indicators are defined for evaluating the performance of the model logic corresponding to the target model and the strategy model. The evaluation indicators are used to evaluate the prediction accuracy, processing efficiency, adaptability to actual data, and consistency with expected results of the model. For example, accuracy, recall rate, F1 score, and other evaluation indicators can be used to measure the performance of the model. At the same time, optimization objectives are defined to ensure that the output of the model meets the actual business requirements and compliance requirements.

[0173] In some embodiments, reinforcement learning is used for preliminary evaluation, and the generated model is preliminarily evaluated by the SPO mechanism method in reinforcement learning. SPO can evaluate the performance of the model through a self-optimization mechanism and adjust the model strategy according to the actual results. This method uses environmental feedback and self-improvement mechanisms to help identify the strengths and weaknesses of the model in specific scenarios and provide improvement suggestions.

[0174] In some embodiments, a generative adversarial network (GAN) is introduced to compare and evaluate the target model. The input of the evaluation includes actual business data, abnormal input generated by GAN, and trajectory execution record of the strategy model. The output includes model prediction results, strategy win rate, and other multi-dimensional evaluation indicators. Through these indicators, the performance of the model under different input conditions is judged, and the potential defects of the model in the logic structure or strategy mechanism are located. The adversarial generative model can simulate various potential challenge scenarios through GAN technology to test the robustness of the model in complex environments. This method can reveal the potential vulnerabilities of the model and provide improvement directions to ensure that the model performs stably in the face of abnormal situations in actual applications.

[0175] As a specific embodiment, the input of the evaluation stage mainly includes: actual business data samples, adversarial data generated by GAN, and execution trajectories generated by the strategy model. The output includes the prediction results of the model, the SPO trajectory win rate, and key evaluation indicators. GAN generates these extreme or irregular input data to test the robustness of the model. The evaluation indicators are explicitly used for multi-dimensional performance evaluation and optimization direction guidance. For example, the accuracy and SPO trajectory win rate indicators reveal the performance differences of the strategy in the adversarial environment. These indicators will serve as the quantitative basis for whether the model needs to be adjusted, and through fine-grained diagnosis of different types of failure points, support subsequent hyperparameter adjustment, module enhancement, and strategy iteration optimization.

[0176] In some embodiments, the target model is compared and analyzed with previously constructed historical models. By comparing the performance of the new model with the historical models on the same data set, it is evaluated whether the target model has improved in terms of prediction accuracy, processing speed and adaptability, and the analysis results can help identify the strengths and weaknesses of the model and provide the basis for further optimization.

[0177] In some embodiments, the generated target model and strategy model are adjusted and optimized according to the evaluation results. Based on the feedback of reinforcement learning and adversarial generative model, parameter tuning, structure optimization or algorithm improvement are carried out. In specific implementation, hyperparameter optimization techniques can be used to further improve the performance of the model, ensuring that it can achieve the expected effect in actual application.

[0178] In some embodiments, after the optimized model is adjusted, it is tested and verified in the field. Representative actual data sets are selected and applied to the model for testing to verify the performance of the model in real scenarios. This process ensures that the model is not only effective in theory, but also meets business needs in actual operation.

[0179] As a specific embodiment, taking discipline inspection data as an example, the method provided by the present application is used to automatically model the discipline inspection data, and the final output includes the overall model (including structure and parameters) that can be deployed; the model logic used by the overall model when running; the strategy model optimized through the SPO mechanism. The final model refers to the target model generated after multiple optimization stages such as reinforcement learning, self-game, and adversarial evaluation. In the optimization process, the model logic represents the reasoning structure and knowledge path of the model when processing discipline inspection data. The strategy model is a decision-making module trained through the SPO mechanism. The three are not independent, but are integrated and optimized as a whole in the optimization process, and finally form a complete solution that has both efficient reasoning logic and flexible decision-making. In actual application, the input of the final model is discipline inspection business data, and the output is analysis logic and strategy suggestion; the strategy model is responsible for decision guidance during the generation process, and its input is feature representation and output is optimized strategy path or instruction.

[0180] In this embodiment, the cross-domain self-distillation mechanism improves the generalization ability and performance of the model between different data sources, optimizes the extraction of data features and the adaptability of the model. Through unsupervised self-optimization, the training efficiency and effect of the model are improved, and the ability of the model in handling complex tasks is enhanced. In reinforcement learning, the hierarchical attention mechanism and adaptive strategy optimization are used to improve the decision-making ability and strategy adjustment efficiency of the model in dynamic environment. Through the above methods, efficient feature extraction, accurate model optimization and flexible strategy adjustment are realized in data automatic modeling, which greatly improves the automation degree of the modeling process and the overall performance of the model.

[0181] Please refer to Figure 3 The embodiment of the present application also provides a big data automatic modeling device 300. Figure 3 Figure 1 is a structural diagram of the big data automatic modeling device 300 provided by the embodiment of the present application. Since the principle of solving problems of the big data automatic modeling device 300 is similar to the big data automatic modeling method shown in the embodiment of the present application, the implementation of the big data automatic modeling device 300 can be referred to the implementation of the method, and the repeated parts will not be described here. Figure 1

[0182] As shown in Figure 3 , the big data automatic modeling device 300 comprises:

[0183] A preprocessing module 301 is configured to preprocess original data to obtain target data, wherein the target data comprises a training data set, a historical model and data flow direction logic information.

[0184] A training module 302 is configured to train an initial model based on the training data set and the data flow direction logic information to obtain a first model, wherein the initial model is constructed based on the historical model.

[0185] A fine-tuning module 303 is configured to fine-tune the first model based on fine-tuning training data generated by a self-game fine-tuning mechanism to obtain a target model.

[0186] Optionally, the fine-tuning module 303 comprises:

[0187] A construction unit is configured to construct a first intermediate model based on the first model, wherein the first intermediate model comprises a knowledge fusion layer and a knowledge internalization layer, the knowledge fusion layer is configured to weight and fuse a plurality of knowledge representations based on a cross-attention mechanism to obtain a first knowledge representation, and the knowledge internalization layer is configured to internalize the first knowledge representation to obtain a second knowledge representation.

[0188] A fusion internalization unit is configured to fuse and internalize knowledge representations of a plurality of external large models to the first intermediate model to obtain a second model.

[0189] A fine-tuning unit is configured to fine-tune the second model based on fine-tuning training data generated by a self-game fine-tuning mechanism to obtain the target model.

[0190] Optionally, the fusion internalization unit is specifically configured to:

[0191] Input the training data set into a plurality of external large models for processing to obtain original knowledge representations of the plurality of external large models.

[0192] ​continuously train the first intermediate model based on the plurality of original knowledge representations until a convergence condition is met, to obtain the second model;

[0193] wherein the i-th training comprises:

[0194] extracting an i-th batch of training data from the plurality of original knowledge representations, i being a positive integer;

[0195] inputting the i-th batch of training data into the knowledge fusion layer, and performing weighted processing on the i-th batch of training data based on a cross-attention mechanism, to obtain the first knowledge representation;

[0196] inputting the first knowledge representation into the knowledge internalization layer for internalization processing, to obtain the second knowledge representation;

[0197] detecting the second knowledge representation using a loss function, to obtain a loss value;

[0198] adjusting parameters of the first intermediate model based on the loss value.

[0199] Optionally, the fine-tuning unit is specifically configured to:

[0200] adding a cross-domain fusion module to the second model, to obtain a second intermediate model;

[0201] fine-tuning the second intermediate model using fine-tuning training data generated based on a self-game fine-tuning mechanism, to obtain a target model;

[0202] wherein the cross-domain fusion module comprises:

[0203] a plurality of parallel sub-networks, each of the sub-networks being configured to process knowledge representation of one domain, and the plurality of sub-networks being configured to perform information transmission based on a cross-attention mechanism;

[0204] a convergence layer configured to integrate outputs of the plurality of sub-networks, to obtain a multi-domain fusion representation.

[0205] Optionally, the big data automatic modeling apparatus 300 further comprises:

[0206] a definition module configured to define a hierarchical structure according to types and features of the first knowledge representation;

[0207] a decomposition module configured to decompose the first knowledge representation into a plurality of first data based on the hierarchical structure;

[0208] an embedding processing module configured to perform embedding processing on the plurality of first data, to obtain a plurality of first embedding vectors;

[0209] a fusion processing module configured to perform fusion processing on the first embedding vectors based on a cross-attention mechanism to obtain second embedding vectors;

[0210] a self-game training module configured to perform self-game training on the second embedding vectors in a policy optimization process based on a soft policy optimization mechanism to obtain a policy model.

[0211] Optionally, the training module 302 includes:

[0212] a meta-learning training unit configured to perform meta-learning training on the initial model to obtain a third intermediate model;

[0213] a setting unit configured to add a virtual branch to the third intermediate model to obtain a fourth intermediate model, the virtual branch being configured to generate a virtual class;

[0214] an incremental learning unit configured to perform incremental learning on the fourth intermediate model based on the training data set to obtain the first model.

[0215] Optionally, the third intermediate model includes a backbone network and a multi-scale feature self-distillation module, the backbone network including an input layer, a plurality of intermediate layers and an output layer connected in sequence, and at least one of the plurality of intermediate layers being connected with the multi-scale feature self-distillation module;

[0216] The multi-scale feature self-distillation module includes:

[0217] a plurality of scale-different feature extraction branches configured to perform feature extraction on features output by a first intermediate layer to obtain scale features, the first intermediate layer being an intermediate layer of the plurality of intermediate layers connected with the multi-scale feature self-distillation module;

[0218] a projection branch configured to uniformly project a plurality of the scale features to obtain projection features, the projection features being configured to determine a consistency loss;

[0219] a fusion branch configured to input a plurality of the scale features after fusion into a second intermediate layer, the second intermediate layer being an intermediate layer of the plurality of intermediate layers located at a next layer of the first intermediate layer.

[0220] Optionally, the incremental learning unit is specifically configured to:

[0221] perform N rounds of incremental learning on the fourth intermediate model based on the training data set to obtain the first model;

[0222] The rth round of incremental learning includes:

[0223] Extracting sample data of the rth category from the training data set, where the rth category is a new category for the rth round of incremental learning, where N is a positive integer and r is a positive integer less than or equal to N;

[0224] The sample data of the rth category and part of the sample data of the old category constitute a training subset, where the sample data of the old category is the sample data of the first r-1 categories, and the first r-1 categories are the old categories of the first r-1 rounds of incremental learning;

[0225] Incremental learning is performed on the model trained in the r-1th round based on the training subset to obtain a model trained in the rth round, wherein the model obtained in the Nth round of training is the first model.

[0226] In the several embodiments provided in this application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection of some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0227] In addition, the functional units in various embodiments of the present invention may be integrated into a single processing unit, each unit may be physically included separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or hardware plus software functional units.

[0228] The above-mentioned integrated unit implemented in the form of a software functional unit can be stored in a computer-readable storage medium. The above-mentioned software functional unit is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to perform some steps of the sending and receiving methods described in various embodiments of the present invention. The aforementioned storage medium includes: a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc., various media that can store program code.

[0229] like Figure 4 As shown, an embodiment of the present invention further provides an electronic device 400, which includes a processor 401 configured to read a program in a memory 402 and execute the following steps:

[0230] Preprocess the original data to obtain target data, the target data including a training data set, a historical model and data flow direction logic information;

[0231] Train an initial model based on the training data set and the data flow direction logic information to obtain a first model, the initial model being constructed based on the historical model;

[0232] Fine-tune the first model based on fine-tuning training data generated by a self-game fine-tuning mechanism to obtain a target model.

[0233] Optionally, the processor 401 is further configured to read a program in the memory 402 and perform the following steps:

[0234] Construct a first intermediate model based on the first model, the first intermediate model including a knowledge fusion layer and a knowledge internalization layer, the knowledge fusion layer being configured to perform weighted fusion on multiple knowledge representations based on a cross-attention mechanism to obtain a first knowledge representation, and the knowledge internalization layer being configured to perform internalization processing on the first knowledge representation to obtain a second knowledge representation;

[0235] Fuse and internalize knowledge representations of multiple external large models to the first intermediate model to obtain a second model;

[0236] Fine-tune the second model based on fine-tuning training data generated by a self-game fine-tuning mechanism to obtain the target model.

[0237] Optionally, the processor 401 is further configured to read a program in the memory 402 and perform the following steps:

[0238] Input the training data set into multiple external large models for processing to obtain original knowledge representations of the multiple external large models;

[0239] Continuously train the first intermediate model based on multiple original knowledge representations until a convergence condition is met to obtain the second model;

[0240] Wherein, the i-th training includes:

[0241] Extract the i-th batch of training data from the multiple original knowledge representations, i being a positive integer;

[0242] Input the i-th batch of training data into the knowledge fusion layer, and perform weighted processing on the i-th batch of training data based on a cross-attention mechanism to obtain the first knowledge representation;

[0243] Input the first knowledge representation into the knowledge internalization layer for internalization processing to obtain the second knowledge representation;

[0244] detecting the second knowledge representation by using a loss function to obtain a loss value;

[0245] adjusting parameters of the first intermediate model based on the loss value.

[0246] Optionally, the processor 401 is further configured to read a program in the memory 402 and perform the following steps:

[0247] adding a cross-domain fusion module to the second model to obtain a second intermediate model;

[0248] fine-tuning the second intermediate model based on fine-tuning training data generated by a self-game fine-tuning mechanism to obtain a target model;

[0249] The cross-domain fusion module comprises:

[0250] a plurality of parallel sub-networks, each of which is used to process knowledge representation of one domain, and the plurality of sub-networks perform information transmission based on a cross-attention mechanism;

[0251] a convergence layer used to integrate outputs of the plurality of sub-networks to obtain multi-domain fusion representation.

[0252] Optionally, the processor 401 is further configured to read a program in the memory 402 and perform the following steps:

[0253] defining a hierarchical structure according to types and characteristics of the first knowledge representation;

[0254] decomposing the first knowledge representation into a plurality of first data based on the hierarchical structure;

[0255] performing embedding processing on the plurality of first data to obtain a plurality of first embedding vectors;

[0256] performing fusion processing on the plurality of first embedding vectors based on a cross-attention mechanism to obtain a second embedding vector;

[0257] inputting the second embedding vector into a policy optimization process for self-game training by using a soft policy optimization mechanism to obtain a policy model.

[0258] Optionally, the processor 401 is further configured to read a program in the memory 402 and perform the following steps:

[0259] performing meta-learning training on the initial model to obtain a third intermediate model;

[0260] adding a virtual branch to the third intermediate model to obtain a fourth intermediate model, the virtual branch being used to generate a virtual class;

[0261] performing incremental learning on the fourth intermediate model based on the training data set to obtain the first model.

[0262] Optionally, the third intermediate model comprises a backbone network and a multi-scale feature self-distillation module, the backbone network comprises an input layer, a plurality of intermediate layers and an output layer connected in sequence, and the multi-scale feature self-distillation module is connected to at least one intermediate layer of the plurality of intermediate layers.

[0263] The multi-scale feature self-distillation module comprises:

[0264] a plurality of scale-different feature extraction branches, the feature extraction branches being configured to perform feature extraction on features output by a first intermediate layer to obtain scale features, the first intermediate layer being an intermediate layer of the plurality of intermediate layers to which the multi-scale feature self-distillation module is connected.

[0265] a projection branch, the projection branch being configured to uniformly project a plurality of the scale features to obtain projection features, the projection features being configured to determine a consistency loss.

[0266] a fusion branch, the fusion branch being configured to input a plurality of the scale features after fusion into a second intermediate layer, the second intermediate layer being an intermediate layer of the plurality of intermediate layers located at a next layer of the first intermediate layer.

[0267] Optionally, the processor 401 is further configured to read a program in the memory 402 and perform the following steps:

[0268] performing N rounds of incremental learning on the fourth intermediate model based on the training data set to obtain the first model.

[0269] The rth round of incremental learning comprises:

[0270] extracting sample data of an rth category from the training data set, the rth category being a new category of the rth round of incremental learning, N being a positive integer, and r being a positive integer less than or equal to N;

[0271] constructing a training subset from the sample data of the rth category and part of old category sample data, the old category sample data being sample data of the first r-1 categories, and the first r-1 categories being old categories of the first r-1 rounds of incremental learning;

[0272] performing incremental learning on a model trained in the r-1th round based on the training subset to obtain a model trained in the rth round, wherein a model trained in the Nth round is the first model.

[0273] The electronic device 400 provided in the embodiment of the application can execute the above-mentioned big data automatic modeling method embodiment, and has similar implementation principles and technical effects, which will not be described here again.

[0274] The embodiment of the present application further provides a readable storage medium, and a program is stored on the readable storage medium. The program is executed by a processor to implement each process of the above-mentioned big data automatic modeling method embodiment and achieve the same technical effects. To avoid repetition, details are not described herein.

[0275] The readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to a magnetic storage (for example, a floppy disk, a hard disk, a magnetic tape, a magneto-optical disk (MO), etc.), an optical storage (for example, a compact disk (CD), a digital versatile disk (DVD), a Blu-ray disk (BD), a high-definition versatile disk (HVD), etc.), and a semiconductor storage (for example, a read-only memory (ROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a non-volatile memory (NAND FLASH), a solid state disk (SSD), etc.).

[0276] It should be noted that, in this document, the term "comprising" or "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such a process, method, article or device. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0277] Those skilled in the art can clearly understand the above-mentioned embodiment method can be realized by means of software and necessary general hardware platform, of course, also can be realized by hardware, but in many cases, the former is a better embodiment. According to such understanding, the technical scheme of the present application can be embodied in the form of software product, and the computer software product is stored in a storage medium (such as ROM / RAM, disk, optical disk), including a plurality of instructions to make a terminal (which can be a mobile phone, computer, server, air conditioner, or network equipment, etc.) execute the method described in various embodiments of the present application.

[0278] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the above-mentioned specific embodiments, and the above-mentioned specific embodiments are only illustrative, not limiting, and those skilled in the art can make many forms without departing from the purpose of the present application and the scope protected by the claims under the inspiration of the present application, which all belong to the protection of the present application.

Claims

1. A big data automatic modeling method, characterized in that: include: Preprocessing the original data to obtain target data, wherein the target data includes a training data set, a historical model, and data flow logic information; Training an initial model based on the training data set and the data flow logic information to obtain a first model, wherein the initial model is constructed based on the historical model; The first model is fine-tuned based on the fine-tuning training data generated by the self-game fine-tuning mechanism to obtain a target model.

2. The method according to claim 1, characterized in that Fine-tuning the first model using the fine-tuning training data generated based on the self-game fine-tuning mechanism to obtain a target model includes: constructing a first intermediate model based on the first model, the first intermediate model comprising a knowledge fusion layer and a knowledge internalization layer, the knowledge fusion layer being used to perform weighted fusion of multiple knowledge representations based on a cross-attention mechanism to obtain a first knowledge representation, and the knowledge internalization layer being used to internalize the first knowledge representation to obtain a second knowledge representation; fusing and internalizing the knowledge representations of multiple external large models into the first intermediate model to obtain a second model; The second model is fine-tuned based on the fine-tuning training data generated by the self-game fine-tuning mechanism to obtain the target model.

3. The method according to claim 2, characterized in that The step of fusing and internalizing the knowledge representations of multiple external large models into the first intermediate model to obtain the second model includes: Inputting the training data set into the plurality of the external large models for processing to obtain original knowledge representations of the plurality of the external large models; Continuously training the first intermediate model based on the multiple original knowledge representations until a convergence condition is met to obtain the second model; Among them, the i-th training includes: Extracting the i-th batch of training data from the plurality of original knowledge representations, where i is a positive integer; Inputting the i-th batch of training data into the knowledge fusion layer, and performing weighted processing on the i-th batch of training data based on a cross-attention mechanism to obtain the first knowledge representation; Inputting the first knowledge representation into the knowledge internalization layer for internalization processing to obtain the second knowledge representation; Using a loss function to detect the second knowledge representation to obtain a loss value; Parameters of the first intermediate model are adjusted based on the loss value.

4. The method according to claim 2, characterized in that Fine-tuning the second model using the fine-tuning training data generated based on the self-game fine-tuning mechanism to obtain a target model includes: Adding a cross-domain fusion module to the second model to obtain a second intermediate model; Fine-tune the second intermediate model based on the fine-tuning training data generated by the self-game fine-tuning mechanism to obtain a target model; The cross-domain fusion module includes: Multiple parallel sub-networks, each of which is used to process knowledge representation in a certain field, and multiple sub-networks transfer information based on a cross-attention mechanism; The aggregation layer is used to integrate the outputs of multiple sub-networks to obtain a multi-domain fusion representation.

5. The method according to claim 3, characterized in that After fine-tuning the second model using the fine-tuning training data generated based on the self-game fine-tuning mechanism to obtain a target model, the method further includes: defining a hierarchical structure according to the type and characteristics of the first knowledge representation; Decomposing the first knowledge representation into a plurality of first data based on the hierarchical structure; Performing embedding processing on the plurality of first data to obtain a plurality of first embedding vectors; Fusing the multiple first embedding vectors based on a cross-attention mechanism to obtain a second embedding vector; The second embedding vector is input into the strategy optimization process for self-game training using a soft strategy optimization mechanism to obtain a strategy model.

6. The method according to claim 1, characterized in that The step of training the initial model based on the training data set and the data flow logic information to obtain a first model includes: Performing meta-learning training on the initial model to obtain a third intermediate model; Adding a virtual branch to the third intermediate model to obtain a fourth intermediate model, wherein the virtual branch is used to generate a virtual class; Incremental learning is performed on the fourth intermediate model based on the training data set to obtain the first model.

7. The method according to claim 6, characterized in that The third intermediate model includes a backbone network and a multi-scale feature self-distillation module, wherein the backbone network includes an input layer, multiple intermediate layers, and an output layer connected in sequence, and at least one of the multiple intermediate layers is connected to the multi-scale feature self-distillation module; The multi-scale feature self-distillation module includes: Multiple feature extraction branches of different scales, each of which is used to extract features output by a first intermediate layer to obtain scale features, wherein the first intermediate layer is an intermediate layer among the multiple intermediate layers to which the multi-scale feature self-distillation module is connected; A projection branch, wherein the projection branch is used to uniformly project the multiple scale features to obtain a projection feature, and the projection feature is used to determine the consistency loss; The fusion branch is used to fuse the multiple scale features and input them into a second intermediate layer, where the second intermediate layer is an intermediate layer located below the first intermediate layer among the multiple intermediate layers.

8. The method according to any one of claims 6 or 7, characterized in that The performing incremental learning on the fourth intermediate model based on the training data set to obtain the first model includes: Performing N rounds of incremental learning on the fourth intermediate model based on the training data set to obtain the first model; Among them, the rth round of incremental learning includes: Extracting sample data of the rth category from the training data set, where the rth category is a new category for the rth round of incremental learning, where N is a positive integer and r is a positive integer less than or equal to N; The sample data of the rth category and part of the sample data of the old category constitute a training subset, where the sample data of the old category is the sample data of the first r-1 categories, and the first r-1 categories are the old categories of the first r-1 rounds of incremental learning; Incremental learning is performed on the model trained in the r-1th round based on the training subset to obtain a model trained in the rth round, wherein the model obtained in the Nth round of training is the first model.

9. A big data automatic modeling device, characterized in that: include: A preprocessing module is used to preprocess the original data to obtain target data, wherein the target data includes a training data set, a historical model, and data flow logic information; a training module, configured to train an initial model based on the training data set and the data flow logic information to obtain a first model, wherein the initial model is constructed based on the historical model; The fine-tuning module is used to fine-tune the first model based on the fine-tuning training data generated by the self-game fine-tuning mechanism to obtain a target model.

10. An electronic device comprising: A memory, a processor, and a program stored in the memory and executable on the processor; characterized in that: The processor is used to read the program in the memory to implement the steps in the big data automatic modeling method as described in any one of claims 1 to 8.

11. A readable storage medium for storing a program, characterized in that: When the program is executed by a processor, the steps of the big data automatic modeling method according to any one of claims 1 to 8 are implemented.

Images