A service data prediction model training method, device, equipment and medium

Abstract

The application discloses a service data prediction model training method, device and equipment and a medium, which can be applied to the financial field and the medical and pension field, and comprises the following steps: training a cross-domain feature extraction model according to training data corresponding to multiple service subfields; embedding the training data into a shared feature space through the cross-domain feature extraction model to obtain a cross-domain feature embedding result; performing feature modeling on the cross-domain feature embedding result through a cross-domain feature modeling network to obtain cross-domain label features; performing feature modeling on the cross-domain feature embedding result through a corresponding prediction feature modeling network according to the service subfield corresponding to the cross-domain feature embedding result to obtain prediction data features; and updating the prediction data feature modeling network according to the prediction data features and the corresponding cross-domain label features. Through the method, the training data of different service subfields is mapped into the shared feature space, cross-domain knowledge transfer and feature fusion are realized, and the generalization ability and training efficiency of the model are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence model training technology, and in particular to a method, apparatus, device and medium for training a business data prediction model. Background Technology

[0002] Currently, across various industries, the data systems of different product lines have long been independent due to significant differences in data structure, making it difficult to share semantic training. For example, in financial services, credit business focuses on structured data such as credit history and debt ratio, while wealth management business emphasizes time-series data such as portfolio details and transaction frequency. The data dimensions, structures, and core semantic concerns of the two are completely different, resulting in a huge semantic gap between the two businesses, making shared training impossible. In healthcare management, chronic disease management typically relies on structured monitoring data such as long-term continuous physical examination indicators and medication records, while rehabilitation and nursing care focuses on unstructured text data such as nursing logs and rehabilitation assessment scales. The semantics of the two businesses are completely separate, and there is no unified standard for data structure, making co-training and sharing difficult. In elderly care management, home-based elderly care focuses on daily life text and check-in data such as elderly people's daily routines and housekeeping service records, while community-based elderly care focuses on mixed data such as activity participation and health inspection records. The semantic differences between the two businesses are significant, and the data structures and annotation logics are different, making shared training difficult.

[0003] Meanwhile, when performing risk or pricing predictions based on data from different product lines, the models often suffer from insufficient generalization ability due to differences in data distribution. This necessitates independent training for each product line, leading to high development and maintenance costs for the prediction models. Furthermore, when new products emerge, the prediction models exhibit poor generalization ability, requiring the collection of new data and training of new models. This results in slow adaptation to new products, failing to meet the demands of rapid business iteration. Therefore, there is an urgent need for a model training method that enables knowledge transfer and sharing across business sub-domains. This invention constructs a cross-domain feature extraction model, mapping data from different business sub-domains to a unified shared feature space. Summary of the Invention This invention provides a method, apparatus, device, and medium for training a business data prediction model, in order to solve the problem in related technologies that it is difficult to achieve cross-business subdomain knowledge transfer and sharing, resulting in poor model generalization ability and high training costs.

[0004] Firstly, this disclosure provides a method for training a business data prediction model, including: Based on the training data corresponding to multiple business sub-domains, a cross-domain feature extraction model is trained. The cross-domain feature extraction model is used to map the training data corresponding to multiple business sub-domains to a shared feature space. By using a cross-domain feature extraction model, training data is embedded into a shared feature space to obtain cross-domain feature embedding results; By using a cross-domain feature modeling network, feature modeling is performed on the cross-domain feature embedding results to obtain cross-domain label features; Based on the business sub-domain corresponding to the cross-domain feature embedding result, feature modeling is performed on the cross-domain feature embedding result through the corresponding predictive feature modeling network to obtain predictive data features. The network complexity of the predictive feature modeling network is less than that of the cross-domain feature modeling network. Each business sub-domain corresponds to one predictive feature modeling network. The predictive data feature modeling network is updated based on the predicted data features and the corresponding cross-domain label features.

[0005] Secondly, this disclosure provides a business data prediction model training device, comprising: The cross-domain training module is used to train a cross-domain feature extraction model based on training data corresponding to multiple business sub-domains. The cross-domain feature extraction model is used to map the training data corresponding to multiple business sub-domains to a shared feature space. The shared embedding module is used to embed training data into a shared feature space through a cross-domain feature extraction model to obtain cross-domain feature embedding results. The cross-domain modeling module is used to perform feature modeling on the cross-domain feature embedding results through a cross-domain feature modeling network to obtain cross-domain label features; The single-domain modeling module is used to perform feature modeling on the cross-domain feature embedding results based on the business sub-domain corresponding to the cross-domain feature embedding results, and obtain the predicted data features through the corresponding predictive feature modeling network. The network complexity of the predictive feature modeling network is less than that of the cross-domain feature modeling network. Each business sub-domain corresponds to one predictive feature modeling network. The distillation update module is used to update the predictive data feature modeling network based on the predicted data features and the corresponding cross-domain label features.

[0006] Thirdly, this disclosure provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned business data prediction model training method. Fourthly, this disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described business data prediction model training method.

[0007] In the above-mentioned scheme implemented by the training method, device, equipment, and medium for the business data prediction model, a cross-domain feature extraction model is trained based on training data corresponding to multiple business sub-domains. The cross-domain feature extraction model is used to map the training data corresponding to multiple business sub-domains to a shared feature space. The training data is embedded into the shared feature space through the cross-domain feature extraction model to obtain cross-domain feature embedding results. The cross-domain feature modeling network is used to perform feature modeling on the cross-domain feature embedding results to obtain cross-domain label features. Based on the business sub-domains corresponding to the cross-domain feature embedding results, the corresponding predictive feature modeling network is used to perform feature modeling on the cross-domain feature embedding results to obtain predicted data features. The network complexity of the predictive feature modeling network is less than that of the cross-domain feature modeling network. Each business sub-domain corresponds to one predictive feature modeling network. The predictive data feature modeling network is updated based on the predicted data features and the corresponding cross-domain label features. This method achieves knowledge transfer between multiple business subdomains by sharing a feature space, avoiding the repeated construction of highly complex models when training the predictive feature modeling networks corresponding to each business subdomain, effectively reducing training costs and resource consumption. At the same time, it uses cross-domain label features to guide each predictive feature modeling network to perform knowledge distillation, improving the model's generalization ability and prediction accuracy, ensuring that each predictive feature modeling network fully absorbs cross-domain common features while maintaining its specificity, thereby improving the overall modeling efficiency and stability. Attached Figure Description

[0008] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart of a business data prediction model training method in one embodiment of the present invention; Figure 2 This is another flowchart of a business data prediction model training method in one embodiment of the present invention; Figure 3 This is another flowchart of a business data prediction model training method in one embodiment of the present invention; Figure 4 This is another flowchart of a business data prediction model training method in one embodiment of the present invention; Figure 5 This is another flowchart of a business data prediction model training method in one embodiment of the present invention; Figure 6 This is a schematic diagram of a business data prediction model training device in one embodiment of the present invention; Figure 7This is a schematic block diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0010] In one embodiment, such as Figure 1 As shown, a method for training a business data prediction model is provided, including the following steps: S101, Train a cross-domain feature extraction model based on training data corresponding to multiple business sub-domains; S102, through the cross-domain feature extraction model, the training data is embedded into the shared feature space to obtain the cross-domain feature embedding result; S103 uses a cross-domain feature modeling network to perform feature modeling on the cross-domain feature embedding results to obtain cross-domain label features; S104. Based on the business sub-domain corresponding to the cross-domain feature embedding result, feature modeling is performed on the cross-domain feature embedding result through the corresponding predictive feature modeling network to obtain predictive data features. S105, update the prediction data feature modeling network based on the prediction data features and the corresponding cross-domain label features.

[0011] Among them, the cross-domain feature extraction model is used to map training data corresponding to multiple business sub-domains to a shared feature space. In other words, the cross-domain feature extraction model can uniformly represent training data from different business sub-domains, eliminate semantic differences between domains, and realize cross-domain knowledge sharing.

[0012] Among them, the network complexity of the predictive feature modeling network is less than that of the cross-domain feature modeling network. Each business subdomain corresponds to one predictive feature modeling network. In other words, the cross-domain feature modeling network is the teacher network of the predictive feature modeling network, which is used to guide the predictive feature modeling network to learn features. This enables knowledge distillation of multiple predictive feature modeling networks through one cross-domain feature modeling network, thereby improving the generalization ability and training efficiency of the prediction models in each subdomain. When applied, this predictive feature modeling network can be used to achieve prediction tasks such as risk prediction, pricing prediction, and claims prediction for specific business subdomains.

[0013] As an example, in step S101, the cross-domain feature extraction model can be used to encode the training data of multiple business sub-domains to obtain the encoding results of the training data. Since the business sub-domains corresponding to each training data are different, the encoding results obtained at this time still have inter-domain differences. Therefore, the cross-domain feature extraction model can be trained according to the inter-domain differences to construct a mapping space that can eliminate the inter-domain differences, so that the cross-domain feature extraction model can map the encoding results of different business sub-domains to a unified shared feature space and achieve semantic alignment.

[0014] The different business sub-fields may include financial businesses such as credit business and wealth management business, medical and health management businesses such as chronic disease management business, medical aesthetics business, and rehabilitation nursing business, and elderly care management businesses such as home-based elderly care business and community-based elderly care business. Each sub-field may have significant differences in data distribution, feature structure and business rules, but by constructing a shared feature space, the extraction and transfer of common features between different fields can be realized. However, it is not limited to this, and this disclosure does not limit the business sub-fields.

[0015] The training data can be of the same or different modalities. For example, the training data may include text data, image data, structured tabular data, etc. This disclosure does not limit this.

[0016] As an example, in step S102, after the cross-domain feature extraction model constructs the shared feature space, it extracts features from the training data of each business sub-domain using the cross-domain feature extraction model, and embeds the extracted features into the shared feature space to form a cross-domain feature embedding result. This embedding result retains the original semantic information of each sub-domain while eliminating semantic differences between domains, providing a unified representation basis for subsequent feature modeling.

[0017] In other words, after the cross-domain feature extraction model has built a shared feature space, the training data can be used to extract features (i.e., the training data can be encoded), and the encoded features can be mapped to the shared feature space to form a unified cross-domain feature representation. This enables the training data to be embedded into the shared feature space, and the mapping result in the shared feature space is determined as the cross-domain feature embedding result.

[0018] As an example, in step S103, the cross-domain feature modeling network can utilize the common features and potential association structures between various business sub-domains to perform deep modeling of the cross-domain feature embedding results, generating cross-domain label features with cross-domain generalization capabilities.

[0019] Among them, cross-domain label features not only cover the specific semantics of each subdomain, but also integrate cross-domain shared semantic information, thereby assisting the cross-domain feature modeling network as a teacher network to guide multiple predictive feature modeling networks to complete knowledge distillation.

[0020] In other words, cross-domain label features are cross-domain semantic features generated by the cross-domain feature modeling network based on the cross-domain feature embedding results. For example, they are soft labels generated based on the cross-domain feature embedding results, used to characterize the potential relationships and common semantic patterns between different business sub-domains.

[0021] For example, in the financial sector, cross-domain labeling features can be used to characterize the potential relationship between credit default risk and financial product preferences. By mining common patterns in users' credit behavior and investment choices, these features reveal the intrinsic connection between their risk preferences and financial planning, thereby providing transferable knowledge support for personalized financial product business.

[0022] For example, taking healthcare management as an example, cross-domain tagging features can characterize the potential relationship between chronic disease management and health intervention measures. By analyzing common patterns in patients' medical records, medication behavior and lifestyle data, the inherent laws between disease progression and intervention response can be revealed, thereby providing transferable knowledge support for personalized health management.

[0023] For example, taking the field of elderly care management as an example, cross-domain label features can characterize the potential relationship between elderly health management business and elderly care service business. By integrating multi-source data of the elderly in health monitoring, daily care and social activities, we can explore the common patterns of their physiological state, psychological needs and service preferences, reveal the intrinsic relationship between the use of elderly care services and changes in health status, and thus provide transferable semantic feature representations for cross-scenario elderly care decision support.

[0024] As an example, in step S104, a predictive feature modeling network corresponding to the business sub-domain corresponding to the cross-domain feature embedding result can be selected, and the predictive feature modeling network can be used to perform domain-specific modeling on the cross-domain feature embedding result to generate predictive label features.

[0025] Since each predictive feature modeling network is only used to model the cross-domain feature embedding results in a single business subdomain, the network complexity of the predictive feature modeling network is less than that of the cross-domain feature modeling network, thereby reducing the computational cost of the model while ensuring domain-specific modeling accuracy.

[0026] It should be understood that after completing the feature modeling of the domain feature embedding results, the predictive feature modeling network can use the generated predictive data features to achieve corresponding tasks such as risk prediction and pricing prediction.

[0027] For example, taking financial business as an example, for the cross-domain feature embedding results of the credit sub-domain, feature modeling is performed by aligning with the corresponding predictive feature modeling network to obtain targeted predictive features such as credit risk features.

[0028] For example, taking healthcare management as an example, for the cross-domain feature embedding results of the chronic disease subdomain, feature modeling is performed through the corresponding predictive feature modeling network to obtain disease progression prediction features or intervention response prediction features, etc.

[0029] For example, taking the field of elderly care management as an example, based on the feature embedding results of the community elderly care sub-field, feature modeling is performed through the corresponding predictive feature modeling network to generate features for predicting elderly care service demand or health risk warning features, etc.

[0030] As an example, in step S105, the semantic loss of the predictive feature modeling network can be determined by comparing the differences between the predicted data features and the corresponding cross-domain label features. Then, the parameters of the predictive feature modeling network can be optimized based on the semantic loss, so that it gradually approaches the semantic representation ability of the cross-domain feature modeling network while retaining the domain specificity, thereby realizing the update of the predictive data feature modeling network.

[0031] In other words, cross-domain label features can be used as supervisory signals for knowledge distillation, guiding the predictive feature modeling network to optimize parameters. This reduces computational resource consumption while improving the model's generalization performance across various business subdomains. By minimizing the semantic differences between predictive label features and cross-domain label features, efficient knowledge transfer from the cross-domain feature modeling network to the predictive data feature modeling network is achieved, enhancing the model's understanding of cross-domain common semantics while retaining necessary domain-specific features.

[0032] In summary, the business data prediction model training method proposed in this disclosure includes: training a cross-domain feature extraction model based on training data corresponding to multiple business sub-domains, the cross-domain feature extraction model being used to map the training data corresponding to multiple business sub-domains to a shared feature space; embedding the training data into the shared feature space through the cross-domain feature extraction model to obtain cross-domain feature embedding results; performing feature modeling on the cross-domain feature embedding results through a cross-domain feature modeling network to obtain cross-domain label features; performing feature modeling on the cross-domain feature embedding results based on the business sub-domains corresponding to the cross-domain feature embedding results through the corresponding predictive feature modeling network to obtain predicted data features, the network complexity of the predictive feature modeling network being less than that of the cross-domain feature modeling network, and one predictive feature modeling network corresponding to each business sub-domain; and updating the predictive data feature modeling network based on the predicted data features and the corresponding cross-domain label features. This method achieves knowledge transfer between multiple business subdomains by sharing a feature space, avoiding the repeated construction of highly complex models when training the predictive feature modeling networks corresponding to each business subdomain, effectively reducing training costs and resource consumption. At the same time, it uses cross-domain label features to guide each predictive feature modeling network to perform knowledge distillation, improving the model's generalization ability and prediction accuracy, ensuring that each predictive feature modeling network fully absorbs cross-domain common features while maintaining its specificity, thereby improving the overall modeling efficiency and stability.

[0033] In one embodiment, such as Figure 2As shown, step S101, which involves training a cross-domain feature extraction model based on training data corresponding to multiple business sub-domains, includes: S201, Multimodal encoding is performed on the training data to obtain the sub-domain feature vector corresponding to the training data; S202, based on subdomain feature vectors, performs semantic constraint training on cross-domain feature extraction models to obtain a shared feature space.

[0034] As an example, in step S201, training data of different modalities can be encoded separately. For instance, text data can be encoded using a Transformer encoder, image data can be encoded using an image encoder, and structured data can be encoded using a Multi-Layer Perceptron (MLP). In other words, different encoding methods can be used to encode training data of different modalities to obtain corresponding sub-domain feature vectors.

[0035] As an example, in step S202, by introducing a contrastive learning mechanism, the cross-domain feature extraction model is trained with semantic constraints to jointly train the subdomain feature vectors of multiple business subdomains, thereby achieving semantic alignment of features of different subdomains, and thus determining the space in which the semantically aligned subdomain feature vectors are located as a shared feature space.

[0036] The cross-domain feature extraction model can be trained with semantic constraints using the following formula: in, This represents the total loss function of the cross-domain feature extraction model. This represents the contrastive learning loss function. Represents the reconstruction loss function. Indicates the preset coefficient. This represents the subdomain feature vector of the i-th business subdomain. This represents the subdomain feature vector of the j-th business subdomain. This represents the training data corresponding to the subdomain feature vector of the i-th business subdomain. This represents the training data corresponding to the subdomain feature vector of the j-th business subdomain.

[0037] When the loss value calculated by the total loss function tends to converge (or is less than the preset loss threshold), it can be determined that the cross-domain feature extraction model has completed semantic constraint training. At this time, the model can effectively capture the common semantic information between different business sub-domains and achieve multimodal and cross-domain feature alignment in the shared feature space.

[0038] As an example, such as Figure 3 As shown, step S105, which involves updating the prediction data feature modeling network based on the prediction data features and the corresponding cross-domain label features, includes: S301, obtain the divergence parameter between the predicted data features and the corresponding cross-domain label features, the first feature mean of the predicted data features and the second feature mean of the corresponding cross-domain label features; S302, Based on the mean of the first feature and the mean of the second feature, determine the semantic deviation parameter between the predicted data features and the corresponding cross-domain label features; S303 updates the parameters of the predictive data feature modeling network based on divergence parameters and semantic bias parameters.

[0039] As an example, in step S301, the divergence parameter between the predicted data features and the cross-domain label features is calculated using a divergence calculation formula, for example, the divergence parameter is determined using the following formula: in, This indicates the preset divergence calculation formula. This represents the embedding result of the k-th cross-domain feature. This represents a predictive feature modeling network. This represents the cross-domain label feature corresponding to the k-th cross-domain feature embedding result. The predicted data features corresponding to the k-th cross-domain feature embedding result.

[0040] Furthermore, the second feature mean of the cross-domain label features can be determined by averaging all cross-domain label features; and the first feature mean can be obtained by averaging the predicted data features in each business sub-domain.

[0041] As an example, in step S302, the difference between the mean of the first feature and the mean of the second feature can be calculated by Euclidean distance or cosine similarity, thereby determining the semantic bias parameter, which is used to measure the degree of distribution shift of the predicted data features from the cross-domain label features in the shared feature space.

[0042] For example, the semantic deviation parameter can be determined using the following formula: in, This represents the mean of the second characteristic. Let ∥⋅∥2 represent the first characteristic mean, and let ∥⋅∥2 represent the Euclidean distance. The semantic deviation parameter refers to the semantic deviation between predicted data features and cross-domain label features in a certain business subdomain. The semantic deviation parameters determined in different business subdomains may be the same or different, and this disclosure does not limit them.

[0043] As an example, in step S303, the predictive data feature modeling network can be jointly iteratively trained using semantic bias parameters and divergence parameters until the obtained semantic bias parameters and divergence parameters both meet the corresponding preset parameter threshold conditions, thereby realizing the parameter update of the predictive data feature modeling network.

[0044] In one embodiment, such as Figure 4 As shown, after step S105, which updates the prediction data feature modeling network based on the prediction data features and the corresponding cross-domain label features, the following steps are also included: S401, When new business subdomains are added to the training data, calculate the distribution difference parameter between the updated data and the training data; S402 updates the cross-domain feature extraction model and / or prediction feature network based on the distribution difference parameter.

[0045] As an example, in step S401, when the training data is updated and the business subdomain corresponding to the updated data has not appeared in the training data, the distribution difference parameter between the updated data and the existing training data can be calculated to determine whether the cross-domain feature extraction model needs to be adjusted to adapt to the feature distribution of the new subdomain.

[0046] As an example, in step S402, the cross-domain feature extraction model and / or prediction feature network are dynamically updated according to the determined distribution difference parameters, so that the updated model and network can effectively integrate the feature information of the new business sub-domain, while maintaining the stability of the prediction performance of the existing sub-domain.

[0047] Specifically, if the distribution difference parameter exceeds the preset difference threshold, a lightweight adaptation component is added to the cross-domain feature extraction model, the shared semantic layer parameters of the cross-domain feature extraction model are frozen, and the lightweight adaptation component is trained with loss until the loss parameter corresponding to the lightweight adaptation component is less than or equal to the preset loss threshold. If the distribution difference parameter does not exceed the preset threshold, a corresponding predictive feature modeling network is constructed for the updated data, and based on the updated data, a cross-domain feature modeling network is used to perform knowledge distillation for the corresponding predictive feature modeling network.

[0048] In other words, when the distribution difference parameter exceeds the preset difference threshold, it indicates that the cross-domain feature extraction model is difficult to directly adapt to the updated data of the new business sub-domain. A lightweight adaptation component needs to be introduced to enhance the domain adaptability of the cross-domain feature extraction model, and a dedicated predictive feature modeling network should be added for the new business sub-domain. When the distribution difference is small, it indicates that the feature distribution of the updated data of the new business sub-domain is close to the feature distribution of the training data. There is no need to adjust the cross-domain feature extraction model. A dedicated predictive feature modeling network can be directly added for the new business sub-domain, and the knowledge in the cross-domain feature modeling network can be transferred to the predictive feature modeling network corresponding to the new business sub-domain through knowledge distillation. This will quickly improve the modeling accuracy and generalization ability of the new business sub-domain while retaining the original domain knowledge.

[0049] In this process, when fine-tuning the lightweight adaptation component, the training process can be supervised by a pre-set loss function. When the loss parameter calculated by the pre-set loss function is less than or equal to the preset loss threshold, the training of the lightweight adaptation component is stopped. This ensures that the original model is not compromised while adapting to features in the new sub-domain, thereby achieving efficient incremental updates and multi-domain compatibility of the cross-domain feature extraction model.

[0050] It should be understood that when adding lightweight adaptation components, the fine-tuned lightweight adaptation components need to be verified to ensure their effectiveness in feature extraction in the new business subdomain, and to avoid negative impacts on the feature extraction performance of existing business subdomains after adding lightweight adaptation components to cross-domain feature extraction models.

[0051] In one embodiment, such as Figure 5 As shown, after step S105, which updates the prediction data feature modeling network based on the prediction data features and the corresponding cross-domain label features, the following steps are also included: S501, Based on the updated prediction data feature modeling network, feature modeling is performed on the cross-domain feature embedding results to obtain updated prediction features; S502, based on the updated prediction features, the predicted data features and cross-domain label features, determine the information gain corresponding to the cross-domain feature embedding results; S503, the cross-domain feature embedding results with information gain greater than or equal to the preset gain threshold are identified as high information gain samples; S504 updates parameters of a cross-domain feature modeling network based on high information gain samples.

[0052] As an example, in step S501, after the prediction data feature modeling network is updated, the cross-domain feature modeling network can be iteratively optimized periodically to improve the accuracy and generalization performance of the cross-domain feature modeling network.

[0053] Specifically, the updated prediction data feature modeling network is used to perform secondary feature extraction on the cross-domain feature embedding results to obtain more discriminative updated prediction features, laying the foundation for subsequent calculation of information gain values ​​of features in each dimension by combining prediction data features and cross-domain label features.

[0054] As an example, in step S502, the information gain of a sample can be determined by comparing the degree of change in error of the predictive data feature modeling network before and after learning the sample.

[0055] Specifically, the absolute value of the difference between the cross-domain label features and the predicted data features is determined as the first semantic difference value; the absolute value of the difference between the cross-domain label features and the updated predicted features is determined as the second semantic difference value; the second semantic difference parameter is subtracted based on the first semantic difference value, and the result of the subtraction is determined as the information gain.

[0056] The first semantic difference value reflects the deviation between the cross-domain label features and the predicted data features before updating the predicted data feature modeling network, i.e., the prediction deviation of the predicted data feature modeling network before learning from the samples. The second semantic difference value reflects the deviation between the cross-domain label features and the updated predicted features after updating the predicted data feature modeling network, i.e., the residual deviation of the predicted data feature modeling network after learning from the samples. By subtracting the above two difference values, the degree of improvement in the ability of the predicted data feature modeling network before and after distillation learning can be determined, and the result of the subtraction can be determined as the information gain corresponding to the domain feature embedding result (i.e., the above-mentioned learned samples).

[0057] As an example, in step S503, cross-domain feature embedding results with information gain greater than or equal to a preset gain threshold can be identified as high information gain samples.

[0058] It should be understood that the greater the information gain, the more significant the improvement of the semantic extraction capabilities of the cross-domain feature embedding result on the cross-domain feature modeling network and the predictive feature modeling network, and the more it helps to enhance the representation capabilities of the predictive feature modeling network in the corresponding business sub-domain.

[0059] By setting a preset gain threshold to filter high information gain samples, and using these high information gain samples to update the parameters of the cross-domain feature modeling network, we can effectively retain the feature expressions that contribute significantly to the improvement of model performance and suppress the interference of inefficient or redundant features.

[0060] As an example, in step S504, the parameters of the cross-domain feature modeling network can be updated based on high information gain samples.

[0061] In other words, the cross-domain feature embedding results corresponding to high information gain samples can be backpropagated to the cross-domain feature modeling network. Using the cross-domain feature embedding results and the corresponding training samples, the cross-domain feature modeling network can be retrained to further enhance its ability to express cross-domain features.

[0062] In summary, this disclosure has the following beneficial effects: 1. By sharing the feature space, knowledge transfer between multiple business sub-domains can be achieved, avoiding the repeated construction of highly complex models when training the predictive feature modeling network corresponding to each business sub-domain, thus effectively reducing training costs and resource consumption; 2. By utilizing cross-domain label features to guide the knowledge distillation of each prediction feature modeling network, the generalization ability and prediction accuracy of the model are improved, ensuring that each prediction feature modeling network fully absorbs cross-domain common features while maintaining its specificity, thereby improving the overall modeling efficiency and stability. 3. By adding updated data corresponding to new business sub-domains to the training data, the distribution difference parameter between the updated data and the training data is calculated. When the distribution difference parameter exceeds a preset threshold, a lightweight adaptation component is added to enhance the domain adaptability of the cross-domain feature extraction model, avoiding the computational overhead caused by global reconstruction of the cross-domain feature extraction model. When the distribution difference parameter does not exceed the preset threshold, the predictive feature modeling network corresponding to the updated data is added directly, maintaining the stability of the model structure and achieving incremental learning. This avoids full training of the predictive feature modeling network, significantly improving model iteration efficiency and reducing maintenance costs. 4. By introducing a high information gain sample screening mechanism, the cross-domain feature modeling network is retrained using high information gain samples, which effectively improves the model's learning efficiency and representation ability of key features. While ensuring semantic integrity, it significantly enhances the adaptability and accuracy of the predictive feature modeling network in the target business sub-domain.

[0063] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0064] In one embodiment, a business data prediction model training apparatus is provided, which corresponds one-to-one with the business data prediction model training method described in the above embodiments. For example... Figure 6 As shown, the business data prediction model training device includes a cross-domain data training module 601, a shared embedding module 602, a cross-domain modeling module 603, a single-domain modeling module 604, and a distillation update module 605. Detailed descriptions of each functional module are as follows: The cross-domain training module 601 is used to train a cross-domain feature extraction model based on training data corresponding to multiple business sub-domains. The cross-domain feature extraction model is used to map the training data corresponding to multiple business sub-domains to a shared feature space. The shared embedding module 602 is used to embed training data into a shared feature space through a cross-domain feature extraction model to obtain cross-domain feature embedding results. The cross-domain modeling module 603 is used to perform feature modeling on the cross-domain feature embedding results through a cross-domain feature modeling network to obtain cross-domain label features; The single-domain modeling module 604 is used to perform feature modeling on the cross-domain feature embedding results based on the business sub-domain corresponding to the cross-domain feature embedding results, and obtain the predicted data features through the corresponding predictive feature modeling network. The network complexity of the predictive feature modeling network is less than that of the cross-domain feature modeling network. Each business sub-domain corresponds to one predictive feature modeling network. The distillation update module 605 is used to update the prediction data feature modeling network based on the prediction data features and the corresponding cross-domain label features.

[0065] In one embodiment, the cross-domain training module 601 is further configured to train a cross-domain feature extraction model, including: Multimodal encoding is performed on the training data to obtain the sub-domain feature vectors corresponding to the training data; Based on subdomain feature vectors, the cross-domain feature extraction model is trained with semantic constraints to obtain a shared feature space.

[0066] In one embodiment, the distillation update module 605 is further configured to obtain the divergence parameter between the predicted data features and the corresponding cross-domain label features, the first feature mean of the predicted data features, and the second feature mean of the corresponding cross-domain label features; Based on the mean of the first feature and the mean of the second feature, the semantic deviation parameter between the predicted data features and the corresponding cross-domain label features is determined; Based on the divergence parameter and semantic bias parameter, the parameters of the predictive data feature modeling network are updated.

[0067] In one embodiment, the distillation update module 605 is further configured to calculate the distribution difference parameter between the updated data and the training data when new updated data corresponding to a new business subdomain is added to the training data; Based on the distribution difference parameter, the cross-domain feature extraction model and / or prediction feature network are updated.

[0068] In one embodiment, the distillation update module 605 is further configured to, if the distribution difference parameter exceeds a preset difference threshold, add a lightweight adaptation component to the cross-domain feature extraction model, freeze the shared semantic layer parameters of the cross-domain feature extraction model, and perform loss training on the lightweight adaptation component until the loss parameter corresponding to the lightweight adaptation component is less than or equal to the preset loss threshold. If the distribution difference parameter does not exceed the preset threshold, a corresponding predictive feature modeling network is constructed for the updated data, and based on the updated data, knowledge distillation is performed on the corresponding predictive feature modeling network using the cross-domain feature modeling network.

[0069] In one embodiment, the distillation update module 605 is further configured to model the cross-domain feature embedding results based on the updated prediction data feature modeling network to obtain updated prediction features; Based on the updated prediction features, the predicted data features and cross-domain label features are used to determine the information gain corresponding to the cross-domain feature embedding results. Cross-domain feature embedding results with information gain greater than or equal to a preset gain threshold are identified as high information gain samples. Based on high information gain samples, parameters of the cross-domain feature modeling network are updated.

[0070] In one embodiment, the distillation update module 605 is further configured to determine the absolute value of the difference between the cross-domain label features and the predicted data features as the first semantic difference value; The absolute value of the difference between the cross-domain label features and the updated predicted features is determined as the second semantic difference value; The second semantic difference parameter is subtracted from the first semantic difference value, and the result of the subtraction is determined as the information gain.

[0071] This invention provides a business data prediction model training device, comprising: a cross-domain training module for training a cross-domain feature extraction model based on training data corresponding to multiple business sub-domains, wherein the cross-domain feature extraction model maps the training data corresponding to multiple business sub-domains to a shared feature space; a shared embedding module for embedding the training data into the shared feature space through the cross-domain feature extraction model to obtain cross-domain feature embedding results; a cross-domain modeling module for performing feature modeling on the cross-domain feature embedding results through a cross-domain feature modeling network to obtain cross-domain label features; a single-domain modeling module for performing feature modeling on the cross-domain feature embedding results based on the business sub-domains corresponding to the cross-domain feature embedding results through a corresponding predictive feature modeling network to obtain predicted data features, wherein the network complexity of the predictive feature modeling network is less than that of the cross-domain feature modeling network, and each business sub-domain corresponds to one predictive feature modeling network; and a distillation update module for updating the predictive data feature modeling network based on the predicted data features and the corresponding cross-domain label features. This device achieves knowledge transfer between multiple business subdomains by sharing a feature space, avoiding the repeated construction of highly complex models when training the predictive feature modeling networks corresponding to each business subdomain, effectively reducing training costs and resource consumption. At the same time, it uses cross-domain label features to guide each predictive feature modeling network to perform knowledge distillation, improving the model's generalization ability and prediction accuracy, ensuring that each predictive feature modeling network fully absorbs cross-domain common features while maintaining its specificity, thereby improving the overall modeling efficiency and stability.

[0072] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database is used for data employed in a business data prediction model training method. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a business data prediction model training method.

[0073] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described business data prediction method.

[0074] In one embodiment, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described business data prediction method.

[0075] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0076] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0077] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for training a business data prediction model, characterized in that, include: A cross-domain feature extraction model is trained based on training data corresponding to multiple business sub-domains. The cross-domain feature extraction model is used to map the training data corresponding to the multiple business sub-domains to a shared feature space. The training data is embedded into the shared feature space using the cross-domain feature extraction model to obtain the cross-domain feature embedding result. Cross-domain label features are obtained by performing feature modeling on the cross-domain feature embedding results through a cross-domain feature modeling network. Based on the business sub-domain corresponding to the cross-domain feature embedding result, feature modeling is performed on the cross-domain feature embedding result through the corresponding predictive feature modeling network to obtain predictive data features. The network complexity of the predictive feature modeling network is less than that of the cross-domain feature modeling network. Each business sub-domain corresponds to one predictive feature modeling network. The predicted data feature modeling network is updated based on the predicted data features and the corresponding cross-domain label features.

2. The method according to claim 1, characterized in that, The step of training a cross-domain feature extraction model based on training data corresponding to multiple business sub-domains includes: The training data is multimodal encoded to obtain the sub-domain feature vector corresponding to the training data; Based on the subdomain feature vectors, the cross-domain feature extraction model is trained with semantic constraints to obtain a shared feature space.

3. The method according to claim 1, characterized in that, The step of updating the predictive data feature modeling network based on the predicted data features and the corresponding cross-domain label features includes: Obtain the divergence parameter between the predicted data features and the corresponding cross-domain label features, the first feature mean of the predicted data features, and the second feature mean of the corresponding cross-domain label features; Based on the mean of the first feature and the mean of the second feature, a semantic deviation parameter between the predicted data feature and the corresponding cross-domain label feature is determined; Based on the divergence parameter and the semantic bias parameter, the parameters of the predicted data feature modeling network are updated.

4. The method according to claim 1, characterized in that, After updating the predicted data feature modeling network, the method further includes: When new business subdomains are added to the training data, the distribution difference parameter between the updated data and the training data is calculated. Based on the distribution difference parameters, the cross-domain feature extraction model and / or the prediction feature network are updated.

5. The method according to claim 4, characterized in that, The step of updating the cross-domain feature extraction model and / or the prediction feature network based on the distribution difference parameter includes: If the distribution difference parameter exceeds the preset difference threshold, a lightweight adaptation component is added to the cross-domain feature extraction model, the shared semantic layer parameters of the cross-domain feature extraction model are frozen, and the lightweight adaptation component is trained with loss until the loss parameter corresponding to the lightweight adaptation component is less than or equal to the preset loss threshold. If the distribution difference parameter does not exceed a preset threshold, a corresponding predictive feature modeling network is constructed for the updated data, and based on the updated data, the cross-domain feature modeling network is used to perform knowledge distillation on the corresponding predictive feature modeling network.

6. The method according to claim 1, characterized in that, After updating the predicted data feature modeling network, the method further includes: Based on the updated predicted data feature modeling network, feature modeling is performed on the cross-domain feature embedding results to obtain updated predicted features; Based on the updated prediction features, the predicted data features, and the cross-domain label features, the information gain corresponding to the cross-domain feature embedding result is determined; The cross-domain feature embedding results with information gain greater than or equal to a preset gain threshold are identified as high information gain samples. Based on the high information gain samples, the parameters of the cross-domain feature modeling network are updated.

7. The method according to claim 6, characterized in that, The step of determining the information gain corresponding to the cross-domain feature embedding result based on the updated prediction features, the prediction data features, and the cross-domain label features includes: The absolute value of the difference between the cross-domain label features and the predicted data features is determined as the first semantic difference value; The absolute value of the difference between the cross-domain label feature and the updated prediction feature is determined as the second semantic difference value; The second semantic difference parameter is subtracted based on the first semantic difference value, and the result of the subtraction is determined as the information gain.

8. A training device for a business data prediction model, characterized in that, include: The cross-domain training module is used to train a cross-domain feature extraction model based on training data corresponding to multiple business sub-domains. The cross-domain feature extraction model is used to map the training data corresponding to the multiple business sub-domains to a shared feature space. The shared embedding module is used to embed the training data into the shared feature space through the cross-domain feature extraction model to obtain the cross-domain feature embedding result; The cross-domain modeling module is used to perform feature modeling on the cross-domain feature embedding results through a cross-domain feature modeling network to obtain cross-domain label features; A single-domain modeling module is used to perform feature modeling on the cross-domain feature embedding result according to the business sub-domain corresponding to the cross-domain feature embedding result, and obtain predicted data features through the corresponding predictive feature modeling network. The network complexity of the predictive feature modeling network is less than that of the cross-domain feature modeling network. Each business sub-domain corresponds to one predictive feature modeling network. The distillation update module is used to update the prediction data feature modeling network based on the prediction data features and the corresponding cross-domain label features.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the business data prediction model training method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the business data prediction model training method as described in any one of claims 1 to 7.

Images