Prediction model training method, text data prediction method, computing device and storage medium

Abstract

Embodiments of the present specification provide a prediction model training method, a text data prediction method, a computing device and a storage medium, the prediction model training method comprising: obtaining first sample data from a training node corresponding to a target object, the first sample data comprising a specified field, the specified field representing task features of a target task; generating second sample data based on historical data accumulated by the training platform; and calling a training node to perform feature matching on the second sample data based on the task features to obtain third sample data, and training a prediction model corresponding to the target task based on the first sample data and the third sample data. According to the method, the non-real second sample data is generated based on the statistical information of the first sample data, then feature matching is performed in the training node to obtain the third sample data, and the prediction model is trained in combination with the first sample data and the third sample data, so that the data security is guaranteed and the sample diversity is enhanced on the premise of not exposing real data.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of artificial intelligence technology, and in particular to a predictive model training, text data prediction method, computing device, and storage medium. Background Technology

[0002] In the era of data-driven intelligent decision-making, it has become common practice for multiple stakeholders to jointly model and extract data value. This is especially true for the vast number of mid-to-long-tail customers, whose task growth is highly dependent on accurate risk assessment and user insight models.

[0003] However, these clients' modeling samples often suffer from inherent problems such as sparsity and label imbalance. Relying solely on their own data makes it difficult to train robust and reliable models. While external data can be introduced to expand the training set through sample transfer techniques, this process is highly susceptible to the risk of sensitive information leakage, resulting in low levels of data privacy compliance and severely hindering the deepening of joint modeling services and the expansion of commercial cooperation.

[0004] Therefore, there is an urgent need for a predictive model training method that can adequately protect data privacy and security while enriching the diversity of sample sets. Summary of the Invention

[0005] In view of this, embodiments of this specification provide a method for training a prediction model. One or more embodiments of this specification also relate to a text data prediction method, a prediction model training device, a text data prediction device, a training platform, a computing device, a computer-readable storage medium, and a computer program product, to address the technical deficiencies existing in the prior art.

[0006] According to a first aspect of the embodiments of this specification, a prediction model training method is provided, applied to a training platform, the training platform including training nodes corresponding to each object, and the training nodes being isolated from each other; the method includes: Obtain the first sample data from the training node corresponding to the target object. The first sample data includes a specified field, which represents the task characteristics of the target task. Based on the historical data accumulated on the training platform, second sample data is generated; The training node is invoked to perform feature matching on the second sample data based on task features to obtain the third sample data. Based on the first sample data and the third sample data, the prediction model corresponding to the target task is trained.

[0007] According to a second aspect of the embodiments of this specification, a text data prediction method is provided, comprising: In response to a text data prediction request, retrieve the text to be predicted; The text to be predicted is input into the prediction model to obtain the predicted text data for the target task. The prediction model is trained based on the prediction model training method in the first aspect.

[0008] According to a third aspect of the embodiments of this specification, a training platform is provided, including a request interface, a response unit, and training nodes corresponding to each object, wherein the training nodes are isolated from each other. The request interface is used to receive modeling requests sent by the target object; The response unit is used to obtain the first sample data from the training node corresponding to the target object based on the modeling request, and generate the second sample data based on the historical data accumulated by the training platform. The first sample data includes a specified field, which represents the task characteristics of the target task. The training node is used to perform feature matching on the second sample data based on the task features to obtain the third sample data, and to train the prediction model corresponding to the target task based on the first sample data and the third sample data.

[0009] According to a fourth aspect of the embodiments of this specification, a prediction model training apparatus is provided, applied to a training platform, the training platform including training nodes corresponding to each object, and the training nodes being isolated from each other; the apparatus includes: The first acquisition module is configured to acquire first sample data from the training node corresponding to the target object. The first sample data includes a specified field, which represents the task features of the target task. The generation module is configured to generate second sample data based on historical data accumulated on the training platform. The matching training module is configured to call the training node to perform feature matching on the second sample data based on the task features to obtain the third sample data, and to train the prediction model corresponding to the target task based on the first sample data and the third sample data.

[0010] According to a fifth aspect of the embodiments of this specification, a text data prediction apparatus is provided, comprising: The second acquisition module is configured to acquire the text to be predicted in response to a text data prediction request; The prediction module is configured to input the text to be predicted into the prediction model to obtain the predicted text data of the target task, wherein the prediction model is trained based on the prediction model training method of the first aspect.

[0011] According to a sixth aspect of the embodiments of this specification, a computing device is provided, comprising: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions. When the computer programs / instructions are executed by the processor, they implement the steps of the prediction model training method of the first aspect or the text data prediction method of the second aspect described above.

[0012] According to a seventh aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores a computer program / instructions, which, when executed by a processor, implement the steps of the prediction model training method of the first aspect or the text data prediction method of the second aspect.

[0013] According to an eighth aspect of the embodiments of this specification, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the steps of the prediction model training method of the first aspect or the text data prediction method of the second aspect described above.

[0014] One embodiment of this specification implements the acquisition of first sample data from the training node corresponding to the target object. The first sample data includes specified fields that characterize the task features of the target task. Second sample data is generated based on historical data accumulated by the training platform. The training node is then invoked to perform feature matching on the second sample data based on the task features to obtain third sample data. A prediction model corresponding to the target task is trained based on the first and third sample data. By acquiring first sample data containing task features and generating non-realistic second sample data based on its statistical information, the direct use of original sensitive data is avoided. Subsequently, feature matching is performed within the training node based on task features to obtain the third sample data. The prediction model is then trained by combining the first and third sample data, thereby ensuring data security without exposing real data. This allows the model to enhance sample diversity while maintaining the original task feature distribution, improving its generalization ability to the target task. Attached Figure Description

[0015] Figure 1 This is a flowchart of a prediction model training method provided in one embodiment of this specification; Figure 2 This is a flowchart illustrating a text data prediction method provided in one embodiment of this specification; Figure 3 This is a flowchart illustrating the processing steps of a prediction model training method provided in one embodiment of this specification. Figure 4 This is a timing diagram of a prediction model training method provided in one embodiment of this specification; Figure 5 This is a schematic diagram of the structure of a training platform provided in one embodiment of this specification; Figure 6This is a schematic diagram of the structure of a prediction model training device provided in one embodiment of this specification; Figure 7 This is a schematic diagram of the structure of a text data prediction device provided in one embodiment of this specification; Figure 8 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0016] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0017] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0018] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0019] Furthermore, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0020] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0021] Multi-party decision-making collaboration platform: A system that supports multiple participants to jointly conduct analysis, comparison and judgment under a unified framework. By establishing a standardized data interaction mechanism and a consistent rule foundation, it enables all parties to conduct joint evaluation, result verification and strategy negotiation based on commonly recognized features or indicators without directly sharing raw data, thereby achieving an efficient, reliable and compliant collaborative decision-making process.

[0022] Cloud desktop collaborative modeling space: A collaborative modeling space, which is essentially a sandbox environment based on ECS, providing a relatively neutral modeling environment among collaborating parties; in the embodiments of this specification, it is also referred to as "cloud space".

[0023] Synthetic data: Data generated by sampling from real data distributions through statistical models or deep generative networks (including but not limited to GAN, VAE, and DiffusionModel), which does not contain any real records but retains statistical characteristics; the synthetic model in this proposal is trained on the full sample on the collaborative modeling platform side, and the output synthetic data is similar to the collaborative modeling platform side in statistical characteristics, but there is no real matching data.

[0024] Sample transfer: The process of mapping synthetic samples generated in the source domain to the target domain through a certain transformation, so that the target can use them as supplementary training data to improve model performance.

[0025] In highly sensitive and data-intensive fields such as finance and risk control, collaborative modeling has become an important means to improve model performance and task decision-making capabilities. As the industry's demand for data intelligence applications continues to deepen, more and more institutions hope to introduce external data capabilities through cross-domain collaboration to make up for the lack of their own data dimensions or scale, while meeting increasingly stringent privacy protection and compliance regulatory requirements.

[0026] Against this backdrop, a type of "mid-to-long-tail" collaborator has gradually emerged. These collaborators are characterized by sparse numbers of their own samples and a severely imbalanced ratio of positive to negative samples, making it difficult for them to independently train stable and effective prediction models. Despite having a clear intention to model and specific task scenarios, their weak data foundation often prevents their model performance from meeting expectations, thus limiting the breadth of coverage and the release of value from collaborative modeling services.

[0027] To address this issue, existing solutions typically involve data providers directly importing their plaintext samples into the collaborative modeling space and mixing them with the partners' original data for model training. This approach achieves sample expansion through simple concatenation, has a low operational threshold, a short implementation path, and can quickly improve model performance in the short term.

[0028] However, this approach has significant drawbacks: the migrated samples are exposed in plaintext during the modeling process, allowing collaborating parties to directly observe data content they do not own, thus exposing the data provider's core assets to the risk of leakage. This model lacks an effective mechanism for controlling data visibility, making it difficult to meet the current data security regulations' requirements for principles such as "minimum necessary" and "usable but not visible," resulting in a very low level of compliance. In the long run, it is detrimental to establishing a sustainable and trustworthy multi-party collaborative ecosystem.

[0029] To address the aforementioned problems, this specification provides a method for training a prediction model. One or more embodiments of this specification relate to a text data prediction method, a prediction model training device, a text data prediction device, a training platform, a computing device, a computer-readable storage medium, and a computer program product, which will be described in detail in the following embodiments.

[0030] See Figure 1 , Figure 1 This is a flowchart of a prediction model training method provided in one embodiment of this specification. The prediction model training method is applied to a training platform, which includes training nodes corresponding to each object. The training nodes are isolated from each other. Specifically, it includes the following steps 102-106.

[0031] Step 102: Obtain the first sample data from the training node corresponding to the target object. The first sample data includes a specified field, which represents the task characteristics of the target task.

[0032] Among them, the specified fields refer to a set of feature variable names and their structural specifications that are explicitly defined when the target object initiates the modeling task, which are used to define the input dimensions required by the target task; the training node is a closed virtual computing space allocated to each object on the training platform, which has independent storage and execution capabilities, and is completely isolated from the training nodes of other objects in terms of logic and permissions.

[0033] In practical applications, the system first receives a modeling request submitted by the target object. This request includes a task identifier, a backtracking time window, and a list of specified fields. Subsequently, the system locates the training node corresponding to the target object within the training platform and verifies its access permissions. After confirming permissions, the system selects matching sample records from the local historical records stored on the training node based on the specified field list and the backtracking time window, forming the first sample data. This process ensures that only the necessary subset of data for the task is extracted, avoiding the exposure of irrelevant information. During the extraction process, the system does not read the training node content of other objects, nor does it export the first sample data to a public area; all operations are completed in an isolated environment under platform control.

[0034] In one specific embodiment of this specification, the target is a licensed consumer finance company that initiates an anti-fraud modeling task on a training platform. The company specifies three fields in its request and sets the lookback period to the past 180 days. Upon receiving the request, the system locates the assigned training node, which already stores the company's historical lending behavior logs. Based on the field names and time windows, the system retrieves all user records that meet the criteria within the node. Each record contains the values ​​of the three fields and the corresponding identity identifier. Finally, the system assembles these records into first sample data, structured as a multi-row, three-column table, where each row represents a user sample and each column corresponds to a specified field.

[0035] In multi-party collaborative modeling systems, if multiple objects share the same computing environment, it can easily lead to security risks such as cross-access of data, permission confusion, and leakage of residual information. By dynamically allocating dedicated virtual computing resources to the target object and constructing an isolated computing environment as a training node in response to its modeling request before acquiring the first sample data, the system can ensure that the data processing of each object is completely independent at the logical and permission levels, providing a basic operational guarantee for subsequent secure modeling. Furthermore, before acquiring the first sample data from the training node corresponding to the target object, the system also includes: allocating virtual computing resources to the target object in response to its modeling request; and constructing a computing environment as a training node based on the virtual computing resources, wherein the computing environment allows the target object to access it with authorization.

[0036] Among them, the modeling request refers to the initialization instruction initiated by the target object, which includes task identifier and resource requirements; the virtual computing resource refers to the virtualized resource unit provided by the cloud platform, which has independent CPU, memory and storage quotas; the computing environment refers to the closed execution space built on the virtual computing resource, whose network policy, file system and process permissions are all controlled and configured, and only authorized target objects are allowed to access and operate through designated channels.

[0037] In practical applications, the system first receives a modeling request submitted by the target object, typically triggered through the task platform's front-end interface or API. After verifying the request's legitimacy, the system requests a set of virtual computing resources from the underlying resource scheduling layer, including virtual machine instances or container instances, and binds them to independent storage volumes and network namespaces. Subsequently, the system deploys a basic runtime environment on these resources, including an operating system image, dependency libraries, security agents, and remote access services (such as VNC or SSH), forming a fully functional but closed computing environment. This environment prohibits external inbound connections by default, opening only one access channel that has undergone authentication and permission verification, ensuring that only the target object can upload data, start programs, or obtain the trained model files through the collaborative modeling platform. After the environment is created, the system registers it as a dedicated training node for the target object and records its resource mapping relationship for subsequent steps. The entire construction process is automated, requiring no manual intervention, and the environment's lifecycle is bound to the task, avoiding security risks associated with long-term residency.

[0038] For the step "Build a computing environment as a training node based on virtual computing resources", a complete operating system instance can be launched using a virtual machine snapshot template, and network access rules can be strictly restricted through security group policies. This is suitable for scenarios with high requirements for the integrity of the running environment.

[0039] In one specific embodiment of this specification, a consumer finance company initiates an anti-fraud model training task through a multi-party decision-making collaboration platform. Upon receiving the modeling request, the system invokes the cloud resource management module to apply for a virtual machine instance and allocate 10GB of temporary storage space. Subsequently, the system loads a pre-configured security sandbox image, which includes a built-in Python runtime, model training framework, and sample migration program. Simultaneously, the system configures network policies, allowing only the company's administrator account to connect to the container via the platform's embedded VNC client, prohibiting access from all other IP addresses. After the container starts, it automatically mounts an empty local directory as a data workspace and starts a daemon process to listen for task commands. This container instance is then marked as the company's dedicated training node, and all subsequent data extraction, synthetic sample injection, model training, and other operations are performed within this environment, completely isolated from other clients.

[0040] Based on this, the system creates dedicated computing environments on demand, achieving task-level resource isolation and access control, eliminating the risk of cross-object data leakage at the infrastructure level, and providing a trusted execution foundation for privacy-secure collaborative modeling.

[0041] In collaborative modeling tasks, directly extracting all historical data from the target object's training nodes not only introduces irrelevant features and increases computational overhead, but may also lead to label invalidation or sample bias due to inappropriate time ranges. By responding to the data retrieval requests submitted by the target object, parsing the explicitly specified field and time information, and accurately extracting the first sample data accordingly, the system can minimize the data exposure range while ensuring task relevance, thus improving modeling efficiency and compliance. Furthermore, obtaining the first sample data from the training nodes corresponding to the target object includes: responding to the data retrieval requests submitted by the target object, parsing the field and time information contained in the data retrieval requests, where the field information represents the feature dimensions required by the target task, and the time information represents the historical time window required by the target task; and extracting the first sample data from the training nodes based on the field and time information.

[0042] Among them, field information refers to a set of feature variable names and their structural specifications that are explicitly defined when the target object initiates the task, which is used to define the dimensions required for model input; time information refers to the historical data time window on which the target task depends, which is used to limit the effective observation period of the samples; the first sample data is a set of structured records selected from the local storage of the training node under the constraints of the above two dimensions.

[0043] In practical applications, the system first receives a data retrieval request submitted by the target object through the platform interface or API. This request includes a list of fields (such as "user registration days", "number of transactions in the last 30 days", and "device fingerprint hash value") and a time range (such as "January 1, 2024 to June 30, 2024"). After parsing the request, the system verifies whether the fields belong to the preset feature directory and whether the time window conforms to the task rules (such as not crossing the unobserved period). Once confirmed, the system performs a data query operation within the training node corresponding to the target object: it iterates through the historical behavior logs or structured database stored locally on that node, filters out all records whose timestamps fall within the specified window, and retains only the columns specified in the field list. During the extraction process, the system does not read other fields or export the full text of the original logs; it only returns a subset that meets both conditions. The final generated first sample data is organized in tabular form, with rows representing sample instances and columns corresponding to specified fields. Each record also carries an implicit identifier for subsequent label association.

[0044] For the step "Extract the first sample data from the training node based on the field information and time information", one option is: the training node has a built-in SQL engine, and the system converts the field and time conditions into SELECT statements to directly query the relational table; another option is: the training node uses log files for storage, and the system uses streaming scanning combined with regular expression matching and time filters to filter valid records row by row, which is suitable for unstructured or semi-structured data sources.

[0045] In one specific embodiment of this specification, a regional bank initiates a user churn prediction task on a training platform. Its submitted data acquisition request includes three features: "recent login interval (days)," "current product holdings," and "customer service consultation frequency (last 7 days)"; the time information is set to "past 180 days." Upon receiving the request, the system locates the bank's user behavior log database on its dedicated training node. The log database is stored in daily partitions, and the system only loads partition data from January to June 2024. Subsequently, the system iterates through each log record, checking if the timestamp is within the window and extracting the values ​​of the three fields mentioned above. For logs with missing fields, the system retains the null value instead of discarding the entire record to maintain sample integrity. Finally, the system outputs a first sample of data containing approximately 5,000 user records, each with three columns, a uniform structure, and no unauthorized fields (such as user identification numbers, mobile phone numbers, or other sensitive information).

[0046] Based on this, the system achieves accurate data extraction through dual constraints of fields and time, strictly adhering to the principle of data minimization while meeting task requirements, effectively reducing the risk of privacy leakage and improving subsequent processing efficiency.

[0047] Step 104: Generate second sample data based on the historical data accumulated by the training platform.

[0048] Among them, the historical data accumulated by the training platform refers to the set of desensitized behavior records from multiple non-target objects that the platform has aggregated over a long period of time. All fields have been abstracted, and the identity markers have been removed and are irreversible. The second sample data refers to the data records newly constructed by the generation mechanism. Its field structure is compatible with the target task, and its content is randomly generated or transformed by the model. It does not correspond to any specific user instance in the historical data, nor does it retain any identifiable information.

[0049] In practical applications, the system uses accumulated historical data from the training platform's sample pool as the basis for generation. Subsequently, the system invokes a pre-trained generative model (such as a generative adversarial network, diffusion model, or autoregressive model), which undergoes unsupervised training on the historical data, learning the manifold structure of its high-dimensional feature space. During the generation phase, the model takes random noise as input and outputs new samples through an internal mapping mechanism. The generation process does not rely on the first sample data of the target object, nor does it impose statistical distribution constraints; it only ensures that the output field dimensions are consistent with the target task. The final second sample data is entirely synthesized by the model, verified not to duplicate any records in the historical data (e.g., through hash fingerprint comparison or nearest neighbor detection), and cannot be used to infer the original user behavior.

[0050] In one specific embodiment of the present invention, the historical data accumulated by the training platform contains 2 million abstracted records, each with 64 feature fields, all of which are outputs or encoding results of the platform's internal sub-models. The system loads a StyleGAN2 model pre-trained on this dataset, disables the conditional input module, and uses only standard normal distribution noise as the starting point for generation. The model outputs 10,000 new records. Each record is independently hashed and compared with the hash library of the entire historical data set to confirm that there are no collisions (Hamming distances are all greater than the threshold). The output second sample data is saved as a structured file, with field names consistent with the fields specified in the target task, but the content is newly generated and has no user traceability.

[0051] In modeling scenarios targeting long-tail customers, the target object's own samples are sparse and skewed in distribution, making it difficult to support stable model training. Directly using the original historical data of other objects violates the principles of data isolation and privacy protection. By acquiring fourth sample data accumulated on the platform from other objects as a distribution reference source, and using the statistical information of the target object's first sample data as a constraint to generate second sample data, the system can synthesize non-sensitive samples that retain the diversity of platform data while matching the distribution characteristics of the target task without touching any real individual records, achieving safe and effective data augmentation. Furthermore, generating second sample data based on historical data accumulated on the training platform includes: acquiring the fourth sample data accumulated on the training platform, where the fourth sample data consists of historical data from objects other than the target object; and generating second sample data that conforms to the statistical information, using the fourth sample data as a reference sample and the statistical information of the first sample data as a constraint.

[0052] The fourth sample data refers to a collection of historical behavior records from multiple non-target objects, aggregated by the training platform under compliance conditions. Its content has been anonymized and is not associated with specific identity identifiers. The statistical information of the first sample data refers to the distribution feature parameters extracted from the specified fields of the target object, including the mean, variance, and quantile of numerical fields and the frequency distribution of categorical fields. The second sample data is a dataset newly synthesized by the generative model based on the above references and constraints.

[0053] In practical applications, the system first retrieves historical data with task isomorphism to the target task from the platform's sample database, forming the fourth sample data. This dataset is typically large-scale and covers diverse scenarios, but identifiable information has been removed and the field structure has been standardized. Subsequently, the system performs statistical analysis on the first sample data, calculates the distribution parameters of each specified field, and encodes these parameters as constraints for the generative model. Next, the system calls a pre-trained generative model, which has been unsupervised trained on the fourth sample data and possesses the ability to learn complex joint distributions. During the generation phase, the model takes random noise as input, combines iterative optimization with statistical constraints, and outputs new samples that meet the distribution requirements. During generation, the model only utilizes the distribution pattern of the fourth sample data, without copying its specific records, ensuring that the output data is "usable but not visible." The resulting second sample data is aligned with the first sample data in terms of field dimensions, value range, and overall distribution, and can be safely used for subsequent transfer processing.

[0054] For the step "using the fourth sample data as a reference sample and the statistical information of the first sample data as a constraint, generate the second sample data that conforms to the statistical information", one possible approach is to use a conditional generative adversarial network, inputting the statistical information as a conditional vector into the generator to guide its output to match the distribution; another possible approach is to use a flow-based generative model, embedding the statistical moments of the first sample into the loss function of the generation process through an invertible transformation, forcing the output distribution to approximate the target statistical characteristics.

[0055] In one specific embodiment of the present invention, the target object is a small loan assistance platform. Its first sample data contains three specified words, with a sample size of 600 records, of which positive examples (overdue) account for 5%. The system calls anonymized historical data contributed by dozens of partner institutions from the sample library accumulated by the training platform as the basis for generation. The total amount of this historical data exceeds one million records, and all fields have been processed by hashing, sub-model scoring, or irreversible encoding, without original scene meaning and cannot be associated with specific users. The system loads a generative adversarial network (GAN) model pre-trained on this historical data. During the training phase, the model only learns the distribution structure of the overall feature space and does not introduce any statistical constraints of the target task. In the generation phase, the model takes a random noise vector as input and independently generates 10,000 new records. Each record contains three fields with the same structure as the first sample data. After generation, the system performs deduplication verification on all output records: calculates the hash fingerprint of each record and compares it with the hash library of the entire historical data set and the first sample data to confirm that there are no duplicates or nearest neighbor matches (the minimum Hamming distance is greater than a preset threshold). The final second sample data is entirely synthesized by the model. Its content cannot be traced back to any real user behavior, nor does it reflect the specific distribution characteristics of the first sample data. It is only used as a structurally compatible augmentation sample for subsequent processing.

[0056] Based on this, the system integrates the distribution breadth of external data with the statistical precision of the target task, and produces high-fidelity synthetic samples under the premise of strictly isolating the original data, effectively alleviating the model bias problem caused by sample sparsity for long-tail customers.

[0057] When using historical data from the platform to assist in generating synthetic samples, blindly using all accumulated data may introduce distribution patterns that are irrelevant to or even conflict with the target task, causing the generated samples to deviate from the target domain and reducing model effectiveness. By identifying historical data sources that are task-related to the target task and filtering historical data that match specified fields, the system can ensure that the fourth sample data is highly semantically and structurally related to the target task, providing a high-quality, highly adaptable distribution reference basis for subsequent generation. Furthermore, obtaining the fourth sample data accumulated on the training platform includes: identifying historical data sources that are task-related to the target task within the training platform; and filtering historical data that match specified fields from the historical data sources to obtain the fourth sample data.

[0058] Among them, task relevance refers to the semantic consistency or transferability between the behavior type, task scenario or prediction target recorded by the historical data source and the current target task; specified fields refer to the set of feature variables explicitly defined by the target object in the modeling request; the fourth sample data is a set of de-identified historical records extracted from historical data sources with task relevance and whose field structure is aligned with the specified fields.

[0059] In practical applications, the system first analyzes the metadata of the target task, including task type (such as credit assessment, churn prediction), target tag definition, and semantic description of specified fields, using this as the basis for correlation judgment. Subsequently, the system traverses the historical data source directory registered on the platform, comparing their task tags, field semantics, and data collection context to filter out data sources that are logically isomorphic to the current task. For example, if the target task is "predicting repayment risk based on user device behavior," data sources that also involve the correlation between device behavior and credit results are prioritized, rather than pure marketing click logs. After determining candidate data sources, the system further checks whether their field sets contain all or most of the specified fields. For data sources missing some fields, the system can perform compatibility judgments based on field semantic mapping rules; those that are completely mismatched are excluded. Finally, the system extracts all records from the retained data sources and retains only columns that are equivalent to the specified field names or semantics, forming a structurally unified fourth sample data. The entire process is completed within the platform; the original data does not leave its security domain, and only a subset of anonymized fields is output.

[0060] For the step "Filter historical data that matches the specified field from historical data sources", one option is to perform semantic alignment based on a predefined field ontology library, automatically mapping fields with different names but the same meaning to a unified identifier; another option is to use vector embedding technology to calculate the similarity of field description text, and consider it a match when the similarity exceeds a threshold, which is suitable for heterogeneous data sources that lack standardized naming.

[0061] In a specific embodiment of this invention, the target task is to predict loan delinquency risk based on user behavior characteristics. The specified fields are "feat_087" (anonymized activity representation feature automatically generated by the platform), "subscore_txn_v2" (the sub-model output score calculated based on transaction patterns, without original frequency or amount information), and "overdue_risk_enc" (risk indication feature generated by irreversibly encoding historical performance). The system first analyzes that this task belongs to the credit risk control category, with the target label being "whether it is overdue." Subsequently, the system retrieves all historical data sources within the platform, excluding non-financial data sources such as e-commerce recommendations and advertising clicks, and retaining historical credit behavior databases from three regional banks and two consumer finance companies. The features in these data sources have all been uniformly processed by the platform and transformed into abstract representations. For example, the original login activity, after being scored by the sub-model, is output as "weekly_active_score"; the transaction frequency, after normalization and binning, is encoded as "txn_pattern_code"; and the overdue history is irreversibly hashed as "risk_hist_hash". The system examines the field lists of each data source and identifies features compatible with the specified field structure: two institutions provide fields “feat_login_7d”, “txn_subscore_30”, and “max_od_enc”, whose dimensions, types, and missing patterns are consistent with “feat_087”, “subscore_txn_v2”, and “overdue_risk_enc”; another institution, although its field name is “active_score_w”, its metadata indicates that it is an equivalent representation under the same abstract feature space. The system extracts all records from these three data sources and maps the corresponding fields uniformly to the specified field names of the target task, ultimately forming a fourth sample dataset containing 850,000 records. Each record contains three target fields, and all values ​​are anonymized abstract numerical values ​​or codes, making it impossible to infer the original user behavior.

[0062] Based on this, the system uses a dual screening mechanism of task semantics and field structure to ensure that the fourth sample data is highly aligned with the target task in terms of task logic and feature dimensions, providing a reliable and relevant distribution prior for the subsequent generation of synthetic samples that conform to the target distribution.

[0063] In the process of sample augmentation, relying solely on the sparse data of the target object itself makes it difficult to learn a robust generation distribution, while directly using the original data of other objects poses a risk of privacy leakage. By training a generative model with the fourth sample data as a reference to capture its overall data characteristics, and inputting the statistical information of the first sample data as a conditional constraint into the model, the system can generate second sample data that inherits the diversity of platform data and strictly matches the statistical characteristics of the target task without reproducing any real records, thus achieving secure and high-fidelity data synthesis. Furthermore, using the fourth sample data as a reference sample and the statistical information of the first sample data as a constraint, generating second sample data that conforms to the statistical information includes: training a generative model based on the fourth sample data, wherein the generative model is used to learn the data characteristics of the fourth sample data; and inputting the statistical information of the first sample data into the generative model to obtain second sample data that conforms to the data characteristics and statistical information.

[0064] Here, the generative model refers to a machine learning model that can learn the joint distribution from the training data and generate new samples; the data characteristics of the fourth sample data refer to its overall distribution pattern on a specified field, including marginal distribution, inter-field correlation and higher-order statistical structure; the statistical information of the first sample data refers to the set of distribution parameters extracted from the real samples of the target task, which is used to guide the generation process; the second sample data is a brand-new synthetic record output by the generative model under statistical constraints, which does not contain any original individuals in the fourth sample data.

[0065] In practical applications, the system first uses the fourth sample data to perform unsupervised training on the generative model. During training, the model gradually approximates the high-dimensional joint distribution of the fourth sample data on specified fields through mechanisms such as adversarial learning, maximum likelihood estimation, or streaming transformation, mastering its inherent data characteristics, such as the skewness of numerical fields and the co-occurrence patterns of categorical fields. After training, the model has the ability to generate new samples with similar distributions from random noise. Subsequently, in the generation stage, the system encodes the statistical information of the first sample data (such as the mean, variance, quantiles, and class proportions of each field) into conditional vectors or constraint signals and inputs them into the trained generative model. The model dynamically adjusts its output during the generation process, so that the synthesized samples are aligned with the first sample data at the macro-statistical level, for example, by introducing a moment matching term through the loss function or embedding a conditional batch normalization layer in the generator network. The final output second sample data is consistent with the first sample in field structure, retains the diversity of the fourth sample data in distribution characteristics, satisfies the statistical constraints of the first sample data, and each record is newly generated by the model without corresponding to a real user.

[0066] For the step "inputting the statistical information of the first sample data into the generative model", one option is to concatenate the statistical information as a conditional vector into the input noise of the generator and implement distribution guidance through a conditional generative adversarial network architecture. Another option is to add a statistical moment matching term to the loss function of the generative model, such as minimizing the L2 distance between the generated sample and the first sample in terms of mean and variance, thereby implicitly imposing constraints during the optimization process.

[0067] In one specific embodiment of this specification, the fourth sample data contains a large-scale anonymized credit records from multiple financial institutions, with fields including "User Lifecycle Stage Score," "Economic Capacity Stratification Identifier," and "Number of Historical Delinquencies." The system first uses this dataset to train a Wasserstein GAN with Gradient Penalty (WGAN-GP) model, enabling it to generate new samples that conform to the overall credit population distribution. Subsequently, for the first sample data (only 800 records) of a specific target object, the system calculates its "User Lifecycle Stage Score" as having a mean of 36.5 and a standard deviation of 10.2, the distribution of the "Economic Capacity Stratification Identifier" as [low:40%, medium:50%, high:10%], and the 75th percentile of the "Number of Historical Delinquencies" as 1. These statistical parameters are encoded into a conditional vector and input into the trained generator. At each sampling, the generator adjusts its output based on this conditional vector, ensuring that the generated "User Lifecycle Stage Score" is concentrated between 25 and 50 years old, the proportion of the "Economic Capacity Stratification Identifier" matches the target distribution, and the "Number of Historical Delinquencies" is mostly 0 or 1. Finally, 20,000 second sample data were generated. After verification, the deviations of the statistical indicators of the second sample data from the first sample data in the specified fields were all less than the preset thresholds, and no records were duplicated with the fourth sample data.

[0068] Based on this, the system combines the prior distribution of external data with the statistical constraints of the target task to produce high-fidelity and highly adaptable synthetic samples while ensuring privacy and security, effectively supporting the model training needs in long-tail scenarios.

[0069] Step 106: Call the training node to perform feature matching on the second sample data based on the task features to obtain the third sample data, and train the prediction model corresponding to the target task based on the first sample data and the third sample data.

[0070] Among them, task features refer to the feature space structure defined by specified fields, including the data type, semantic meaning and effective value range of each field; feature matching refers to the field structure alignment and value range adaptation processing of the second sample data to make it compatible with the first sample data in terms of form and distribution; the third sample data is the synthetic data after the above processing, which can be seamlessly spliced ​​with the first sample data for supervised learning.

[0071] In practical applications, the system sends the second sample data along with a pre-configured transfer procedure to the training node of the target object. Upon entering this isolated environment, the transfer procedure first parses the field metadata and value range constraints of the first sample data to obtain the technical attributes and task rules of each specified field. Then, the program performs field structure alignment on the second sample data: if there are inconsistencies in field naming or type, it renames or converts the types based on the metadata; next, it performs value range adaptation processing, such as truncating values ​​exceeding a reasonable range, unifying the mapping of category labels, and filling in null values ​​according to the missing patterns of the first sample. After adaptation, the system saves the processing result as the third sample data within the training node. Subsequently, the training program loads the first sample data and its corresponding ground truth labels, merges them with the third sample data and its synthetic labels to construct a complete training set, and trains the prediction model corresponding to the target task using standard machine learning procedures (such as gradient descent optimization). The entire process is completed in a closed environment without any leakage of raw data.

[0072] In a specific embodiment of this specification, the first sample data includes the fields "feat_087" (device risk assessment score generated by the platform, integer, value 0–100) and "region_enc" (a string of standardized encoded registration region codes, in the format of a six-digit regional code). The corresponding fields in the second sample data are "device_risk_score" (floating-point, range 0.0–1.0) and "region_id" (integer, eight digits). After the migration program runs within the training node, it first identifies the semantic equivalence of the fields, multiplies "device_risk_score" by 100 and rounds it to an integer, ensuring it falls within the 0–100 range; for "region_id", it truncates the first six digits and converts it to a string, consistent with the encoding format of the first sample. Simultaneously, the program checks whether the "device risk score" has negative values ​​or outliers exceeding 100, and corrects them to boundary values. After processing, the field names, types, and value ranges of the third sample data are completely consistent with those of the first sample data. Subsequently, the system merges the first sample (containing real overdue labels) with the third sample (containing synthetic overdue labels), inputs them into the logistic regression model for end-to-end training, and iteratively optimizes until convergence.

[0073] In collaborative modeling, directly transmitting the second sample data to the training node in its raw form may lead to field incompatibility or misalignment due to a lack of supporting processing logic, and there is also execution uncertainty caused by the separation of program and data. By encapsulating the second sample data and a preset transfer procedure into a unified data package and sending it as a whole to the training node, the system can ensure that the feature matching logic and the synthesized data arrive synchronously and execute in the same environment, guaranteeing the consistency, integrity, and security of the processing. Furthermore, before calling the training node to perform feature matching on the second sample data based on task features to obtain the third sample data, the process includes: encapsulating the second sample data and the preset transfer procedure into a data package, wherein the transfer procedure is used to perform feature matching on the second sample data; and sending the data package to the training node.

[0074] The transfer procedure refers to a pre-built executable code module used to perform feature matching operations. Its functions include parsing task features, aligning field structures, and calibrating value range distributions. The data packet refers to a transmission unit with self-contained processing capabilities, which is composed of the second sample data and the transfer procedure. The training node is an isolated computing environment dedicated to the target object, which is only allowed to receive and execute authorized data packet content.

[0075] In practical applications, after generating the second sample data, the system retrieves a migration program matching the current task type from the program repository. This program has been pre-compiled into a platform-wide format and has undergone security auditing. The system packages the second sample data and the migration program into a single file or compressed archive, forming a data package. The data package is organized using a standardized structure, including a "data / " directory for sample files, a "program / " directory for executable logic, and a "manifest.json" file describing field mapping rules and execution entry points. Subsequently, the system pushes the data package to the training node corresponding to the target object through an encrypted channel. Upon receiving the data package, the training node first verifies its digital signature and integrity. After confirming the source is trustworthy, it automatically decompresses and starts the migration program in a sandbox environment. The entire transmission and loading process requires no manual intervention, and the data package does not expose plaintext content during transmission and leaves no intermediate state after execution, effectively preventing program tampering or data leakage.

[0076] For the step "encapsulate the second sample data with the preset migration program to form a data package", one option is to use container image technology to build the migration program and its dependent environment together with the second sample data into a lightweight image to achieve runtime consistency; another option is to use a self-extracting executable file format to embed the sample data as encrypted resources into the program binary, and dynamically decrypt and process it at runtime, which is suitable for lightweight nodes without container support.

[0077] In one specific embodiment of this specification, the second sample data is a CSV file containing 10,000 records, with fields "user_age_float" (user age field), "txn_freq_norm" (transaction frequency field), and "device_risk_score" (device risk assessment field). The system selects a general migration program from the model service repository, which supports automatic identification of field semantics and alignment based on the metadata of the first sample. The system packages the CSV file and the Python migration script (including the dependency list requirements.txt) into a ZIP format data package, and attaches a configuration file specifying the field definitions of the first sample: "User lifecycle stage score (integer, 0–120)", "Transaction frequency (integer, ≥0)", and "Device risk score (integer, 0–100)". The data package is uploaded to the training node of the target financial institution via HTTPS encryption. After the proxy service within the node verifies the signature, it decompresses the data package, creates a temporary virtual environment, installs dependencies, and executes the migration script. The script reads the configuration file, rounds "user_age_float" to an integer and truncates it to 0–120, rounds "txn_freq_norm" and corrects negative values ​​to zero, and scales "device_risk_score" to the 0–100 range, finally outputting the third sample data. The entire process is completed in a closed environment, and external parties cannot interfere with the program logic or access the original samples.

[0078] Based on this, the system integrates data and processing logic into a single package to ensure that feature matching operations are executed accurately, reliably, and securely on the target side, avoiding processing failures or result deviations caused by environmental differences or program deficiencies.

[0079] In cross-domain data augmentation scenarios, although the second sample data is statistically close to the target task, its underlying feature distribution (such as field value density, skewness direction, and class ratio) may still have a systematic offset from the first sample data. Directly mixing and training will introduce distribution noise and reduce model performance. By calling the training node, the system performs a domain adaptation transformation on the second sample data based on the feature distribution of the first sample data. The system can dynamically calibrate the distribution shape of the synthesized data and generate third sample data that is aligned with the real data in a high-dimensional feature space, thereby achieving safe and effective sample fusion. Further, the task is the feature distribution of the sample data; the training node performs feature matching on the second sample data based on the task features to obtain the third sample data, including: calling the training node, performing a domain adaptation transformation on the second sample data according to the feature distribution of the first sample data, and obtaining the third sample data. The domain adaptation transformation is used to align the feature distribution of the third sample data with the feature distribution of the first sample data.

[0080] Among them, the feature distribution of the task sample data refers to the joint probability distribution structure of the input features as the core modeling basis of the target task; the feature distribution of the first sample data refers to its empirical distribution on the specified field, including the marginal distribution of each dimension and the dependency relationship between fields; the domain adaptation transformation refers to a data transformation operation used to adjust the distribution of the source domain (second sample data) to approximate the distribution of the target domain (first sample data); the third sample data is the synthetic data output after the transformation, which is highly consistent with the first sample data in terms of feature distribution.

[0081] In practical applications, the system initiates the domain adaptation process within the training node. First, the transfer processing loads the first sample data and calculates its distribution characteristics across specified fields, such as kernel density estimates or quantile sequences for numerical fields, frequency vectors for categorical fields, and mutual information or correlation coefficient matrices between fields. Then, the program takes the second sample data as input and performs field-by-field transformations: for continuous variables, quantile matching maps the cumulative distribution function of the second sample to the cumulative distribution function of the first sample; for discrete variables, reweighting or label smoothing adjusts the class proportions; for multidimensional dependency structures, covariance alignment or maximum mean difference (MMD) minimization strategies can be introduced for joint correction. The entire transformation process is completed locally on the training node, without transmitting the original data back, using only the distribution summary information of the first sample to guide the adjustments. The final output third sample data maintains consistent feature distributions with the first sample data at both univariate and multivariate levels, allowing for seamless integration for subsequent model training.

[0082] For the step "perform domain adaptation transformation on the second sample data according to the feature distribution of the first sample data", one possible approach is to use non-parametric quantile-quantile mapping (QQ mapping) to stretch or compress the quantiles of the second sample to the corresponding quantiles of the first sample field by field; another possible approach is to construct a lightweight calibration network, use the distribution statistics of the first sample as the supervision signal, and fine-tune the network within the training nodes so that it projects the embedding representation of the second sample onto the target distribution manifold.

[0083] In a specific embodiment of this specification, the first sample data includes three fields: "feat_loan_amt" (application intensity score generated by the platform, right-skewed distribution, 75th percentile of 8,000), "feat_age_enc" (user lifecycle stage encoding, approximately normal, mean 34, standard deviation 9), and "channel_type_enc" (classification feature of registration channel after anonymization mapping, three categories, proportions A:60%, B:30%, C:10%). Although the second sample data is generated by the generative model, its "feat_loan_amt" is generally low (75th percentile of 5,000), and its "channel_type_enc" proportions are A:40%, B:40%, C:20%. In the training node, the transfer learning program first calculates the quantile vectors (e.g., 10%, 25%, 50%, 75%, 90%) and class frequencies of the first sample. Subsequently, quantile matching was performed on "feat_loan_amt": the value at position 5,000 in the second sample was mapped to the position corresponding to 8,000; "feat_age_enc" was linearly scaled to match the mean and variance; and "channel_type_enc" was resampled or soft-allocated according to the target proportion. After the transformation, the distribution histogram of the third sample data highly overlapped with that of the first sample, and the correlation between fields (such as high-scoring applicants mostly coming from channel A) was also preserved.

[0084] Based on this, the system performs a domain adaptation transformation based on distribution alignment in an isolated environment, enabling synthetic data to achieve deep compatibility with real data in the feature space, effectively eliminating cross-domain bias, and providing a high-quality sample foundation with consistent distribution for fusion training.

[0085] In cross-source data fusion modeling, although the second sample data is generated by the generative model, its field naming, data type, or value logic may be inconsistent with the real data of the target object. If it is directly used for training, it will lead to feature misalignment or model misjudgment. By parsing the feature space structure of the first sample data inside the training node, obtaining field metadata and value range constraints, and accordingly performing field structure alignment and value range adaptation processing on the second sample data, the system can achieve accurate alignment of the synthetic data at the formal and semantic levels without exposing the original data, ensuring the consistency and reliability of subsequent training inputs. Furthermore, in the training node, the second sample data is subjected to domain adaptation transformation based on the feature distribution of the first sample data to obtain the third sample data. This includes: in the training node, parsing the first sample data to obtain field definition information and field value range constraints, wherein the field definition information represents the field meta-information of each specified field, and the field value range constraints represent the effective value range of each specified field; based on the field meta-information, the second sample data is aligned with the field structure so that the second sample data contains a set of specified fields that match the first sample data; based on the effective value range, the second sample data after field structure alignment is subjected to value range adaptation processing to obtain the third sample data.

[0086] Among them, field meta-information refers to the structural definition and task meaning of each specified field, including field name, data type (such as integer, floating point, string) and semantic role (such as "user lifecycle stage score" representing age in years); field value range constraint refers to the valid range of values ​​allowed for each specified field in the target task, such as lower / upper limits of values, enumeration of category sets or format regularity rules; field structure alignment refers to adjusting the field set of the second sample data to make it completely match the first sample data in terms of name, quantity and type; value range adaptation processing refers to correcting the values ​​of the aligned fields to make them fall within the target value range, avoiding invalid or outlier values ​​from interfering with model learning.

[0087] In practical applications, after the system starts the migration process within the training node, it first reads the header and several sample rows of the first sample data, automatically inferring the field metadata and value range constraints of each specified field. For example, if a field is an integer between 0 and 120 in all records, it is determined to be a "User Lifecycle Stage Score" field, with an integer data type and a value range constraint of [0, 120]. Subsequently, the program checks the field list of the second sample data: if the field names are different but semantically mappable (e.g., "age_user" corresponds to "User Lifecycle Stage Score"), it renames the field according to a pre-set semantic dictionary or field description; if the data types are inconsistent (e.g., floating-point User Lifecycle Stage Score), it performs type conversion (rounding to integer); if a field is missing, it decides whether to fill in the default value or discard the sample according to the task rules. After aligning the field structures, the program iterates through each record and performs value range adaptation on each field: values ​​exceeding the upper limit are truncated to the maximum value, values ​​below the lower limit are set to the minimum value, illegal categories are replaced with "unknown" or the nearest valid category, and missing values ​​are filled according to the missing pattern of the first sample (such as the mode or a specific label). The entire process is completed locally on the training node, without relying on external services, and without outputting intermediate states, ensuring the closedness and security of data processing.

[0088] For the step "align the field structure of the second sample data according to the field meta information", one option is to perform automatic semantic matching based on the field ontology library pre-registered by the platform to achieve cross-naming system field mapping; another option is to attach a field mapping configuration file to the modeling request, and let the migration program perform alignment according to explicit rules, which is suitable for highly customized scenarios.

[0089] In a specific embodiment of this specification, the first sample data includes three specified fields: "feat_reg_days" (user lifecycle stage score, integer, value range [0, 3650]), "device_risk_enc" (the anonymized string of the device risk assessment result, with enumerated values ​​of "low", "medium", and "high"), and "login_active_7d" (recent activity sub-model output score, integer, value range ≥ 0). The corresponding fields in the second sample data are "reg_days" (floating-point), "device_risk" (integer, 0 / 1 / 2), and "login_cnt_7d" (integer, some values ​​are negative). After parsing the first sample in the training node, the transfer processing program determines the field metadata and value range constraints. Subsequently, "reg_days" is rounded to an integer and truncated to the 0–3650 range; the 0 / 1 / 2 values ​​of "device_risk" are mapped to "low", "medium", and "high" respectively; and negative values ​​in "login_cnt_7d" are corrected to 0. The final output of the third sample data is completely consistent with the first sample in terms of field names, types, and value logic. Each record can be seamlessly spliced ​​with the real sample for model training.

[0090] Based on this, the system ensures that the synthetic data is strictly compatible with the target task in terms of field definition and value specification by implementing a dual alignment mechanism based on feature space structure in an isolated environment. This effectively eliminates feature noise caused by format or semantic differences, laying a data foundation for high-quality fusion training.

[0091] Furthermore, the prediction model training method also includes: obtaining a label data file provided by the target object, wherein the label data file contains label information corresponding to the sample identifiers in the first sample data; inputting the label data file into the training node, and updating the first sample data through the label data file.

[0092] The label data file refers to a structured file uploaded by the target object that contains sample identifiers and corresponding label values; the sample identifier is an anonymized key value (such as a de-identified user ID or request serial number) used to uniquely associate feature records and label records; the label information refers to the target variable value that the target task wants to predict, such as the binary classification label "0 / 1" or continuous rating; updating the first sample data means that within the training node, the label information is matched according to the sample identifier and then appended to the corresponding feature record to form a complete training sample with labels.

[0093] In practical applications, the system first receives a label data file submitted by the target object through a secure channel. This file is typically in CSV or Parquet format and contains two columns: a sample identifier column and a label column. After verifying the file format and permissions, the system encrypts and transmits the label data file to the target object's dedicated training node. Upon entering the training node, the migration program or training agent loads the first sample data (containing only feature fields and sample identifiers) and the label data file. Based on the sample identifiers, it performs an inner join operation, appending the successfully matched label values ​​to the end of the corresponding feature rows. For sample identifiers present in the label file but missing in the feature data, the system can choose to ignore them or log them; for samples without corresponding labels in the feature data, they are marked as unsupervised samples or removed, with the specific strategy determined by the task configuration. The entire label association process is completed locally on the training node; the label data does not leave this isolated environment and is not shared with other objects, ensuring minimal exposure of sensitive result information.

[0094] In one specific embodiment of this specification, the first sample data uploaded by a consumer finance company contains 5,000 records, each with a sample identifier "req_id_2024xxxx" and three feature fields: "Transaction frequency in the last 30 days," "Number of device replacements," and "Registration channel." Subsequently, the company uploads a tag data file through the platform interface, containing two columns: "req_id" and "is_overdue" (0 or 1). The system pushes this tag file to its training node. The processing program within the node reads the first sample data and the tag file, constructs a hash table with "req_id" as the key, and matches each record. Ultimately, 4,850 records are successfully associated with tags, forming labeled training samples; the remaining 150 records are marked as unobserved samples and excluded due to missing tags. The updated first sample data becomes a four-column structure, which can be directly used for subsequent joint training with the third sample data.

[0095] Based on this, the system constructs a compliant, complete, and usable supervised learning dataset by completing the secure alignment of labels and features within closed training nodes, providing accurate learning signals for model training while ensuring label privacy.

[0096] Furthermore, based on the first sample data and the third sample data, a prediction model corresponding to the target task is trained, including: calling the training node to train a first prediction model based on the first sample data; calling the training node to train a second prediction model based on the combined sample data of the first and third sample data; and determining the prediction model corresponding to the target task from the first and second prediction models based on the performance evaluation rules corresponding to the target task.

[0097] The first prediction model refers to the baseline model trained using only the first sample data (i.e., the real labeled samples of the target object); the second prediction model refers to the augmented model trained using an extended training set composed of the first sample data and the third sample data (synthetic samples after feature matching); the performance evaluation rule refers to the model selection criteria predefined for the target task, including but not limited to cross-validation accuracy, AUC, KS value, precision under recall threshold, or a comprehensive scoring function under task constraints.

[0098] In practical applications, the system first loads the updated labeled first sample data within the training node and uses the built-in machine learning framework for training to generate the first prediction model. The training process follows a standard procedure, including feature encoding, default hyperparameter configuration or grid search, and k-fold cross-validation to prevent overfitting. Subsequently, the system concatenates the third sample data (containing synthetic labels) with the first sample data row-by-row to form combined sample data, ensuring complete consistency in field structure. On this extended dataset, the system retrains using the same algorithm and hyperparameter configuration to obtain the second prediction model. After training both models, the system evaluates them on the same validation set (usually partitioned from the first sample data) according to the performance evaluation rules declared in the modeling request for the target task. For example, if the task focuses on identifying high-risk users, the evaluation rule might be "choose the model with the highest precision when recall is ≥80%"; if the task emphasizes overall ranking ability, AUC is used as the primary metric. The system compares the evaluation scores of the two models, selects the one that meets the rules and has better performance as the final prediction model, and serializes and saves it within the training node for subsequent prediction calls.

[0099] For the step "Based on the performance evaluation rules corresponding to the target task, determine the prediction model corresponding to the target task from the first prediction model and the second prediction model", one possible approach is to use a single-index threshold judgment method. If the second model is better than the first model in the core index and the improvement exceeds the preset minimum gain (e.g., AUC improvement ≥ 0.02), then the second model is adopted. Another possible implementation approach is to construct a multi-objective weighted scoring function, comprehensively consider performance, stability and the proportion of synthetic data used, and select the one with the highest total score. This approach is suitable for complex scenarios that need to balance performance and compliance.

[0100] In a specific embodiment of this specification, the objective task is to predict loan defaults for micro and small enterprises, and the performance evaluation rule is set as "AUC ≥ 0.75 and KS value ≥ 0.45". The first sample data contains 1,200 real samples, with positive examples accounting for 6%. The system first trains a first prediction model (LightGBM), which, under five-fold cross-validation, yields an AUC of 0.72 and a KS of 0.40, failing to meet the requirements. Subsequently, the system merges 10,000 third sample data (with the synthetic positive example ratio adjusted to 6%) with the first sample to train a second prediction model. The evaluation results show that AUC = 0.78 and KS = 0.48, meeting the rule requirements. Based on this, the system selects the second prediction model as the final model and deploys it to the prediction service module. The entire selection process is automatically executed by the training nodes without manual intervention, ensuring the objectivity and reproducibility of the decision.

[0101] Based on this, the system introduces a controlled model comparison mechanism to improve performance using synthetic data, while using task-oriented evaluation rules as a "safety valve" to avoid introducing invalid or harmful biases due to data augmentation, ensuring that the final model is both effective and reliable.

[0102] If training nodes and their occupied virtual computing resources are not promptly cleared after the collaborative modeling task is completed, it will lead to sensitive data residue, resource waste, and potential security risks. By proactively clearing the training nodes corresponding to the target object and releasing their occupied virtual computing resources after model training, the system can achieve closed-loop management of the task lifecycle, ensuring that data is not left behind, resources are recyclable, and the environment is reusable, thereby enhancing privacy protection and platform operation and maintenance efficiency. Furthermore, after training the prediction model corresponding to the target task based on the first and third sample data, the process also includes: clearing the training nodes corresponding to the target object and releasing the virtual computing resources occupied by the training nodes.

[0103] Among them, a training node refers to an isolated computing environment temporarily allocated for the target object, which includes memory, storage, processes, and network context; virtual computing resources include underlying resources such as CPU quotas, memory capacity, disk space, and network bandwidth dynamically allocated by the cloud platform; clearing a training node means terminating all its running processes, deleting local storage data, revoking access permissions, and unbinding resources.

[0104] In practical applications, after completing the training and evaluation of the prediction model, the system triggers a resource reclamation process. First, the agent program within the training node stops all running training or migration tasks and clears temporary files, logs, and sample data (including first sample data, third sample data, label files, and intermediate artifacts) from the working directory. Then, the system sends a release command to the platform's resource scheduling layer, notifying it to reclaim the virtual machine instance or container resources occupied by the node. After the platform verifies that the task status is "completed," it forcibly destroys the running instance of the node, removes its identity binding from the target object, and returns the released CPU, memory, and storage resources to the resource pool for reuse by subsequent tasks. The entire process is driven by automated scripts, requiring no manual intervention, and multiple rounds of data erasure verification are performed before resource release to ensure that no user data remains in memory or disk cache.

[0105] The step "clear the training node corresponding to the target object and release the virtual computing resources occupied by the training node" can be initiated through the cloud platform API in a virtual machine scenario, and the associated persistent volume can be deleted in conjunction with the storage service. This is suitable for financial-grade tasks with higher isolation requirements.

[0106] In one specific embodiment of this specification, after a regional bank completes the training of its anti-fraud model, the system detects that the training task status has changed to "successful." Immediately, the cleanup module within the training node is activated: all CSV files (including real and synthetic samples) in the ` / data` directory, temporary checkpoints in the ` / model` directory, and log caches are deleted. Simultaneously, a Pod termination command is sent to the Kubernetes cluster, and the platform reclaims the temporary storage occupied by the Pod within 30 seconds. After the cloud monitoring system confirms the resource release, it updates the available quota in the resource pool and records the audit log: "Training node [bank_A_task_20250401] has been safely destroyed, with no data residue." After this, the bank can no longer access the node or recover any data from it.

[0107] Based on this, the system, through an automatic resource recycling mechanism after the task ends, ensures the modeling effect while completely eliminating the risk of data retention, realizing a secure computing paradigm of automatic destruction and no data retention upon task completion, effectively supporting privacy protection scenarios with stringent compliance requirements.

[0108] One embodiment of this specification provides a method for training a prediction model. The method involves obtaining first sample data from a training node corresponding to the target object. The first sample data includes specified fields that characterize the task features of the target task. Second sample data is generated based on historical data accumulated by the training platform. The training node is then invoked to perform feature matching on the second sample data based on the task features to obtain third sample data. A prediction model corresponding to the target task is trained based on the first and third sample data. By obtaining first sample data containing task features and generating non-realistic second sample data based on its statistical information, the direct use of original sensitive data is avoided. Subsequently, feature matching based on task features is performed within the training node to obtain the third sample data. The prediction model is then trained by combining the first and third sample data. This ensures data security without exposing real data, enhances sample diversity while maintaining the original task feature distribution, and improves the model's generalization ability to the target task.

[0109] See Figure 2 , Figure 2 This is a flowchart of a text data prediction method provided in one embodiment of this specification, specifically including the following steps 202-204.

[0110] Step 202: In response to the text data prediction request, obtain the text to be predicted; Step 204: Input the text to be predicted into the prediction model to obtain the predicted text data of the target task. The prediction model is trained based on the prediction model training method in steps 102-106.

[0111] Among them, the text to be predicted refers to the structured record data submitted by the target object, which represents the user's recent credit product usage behavior and contains multiple task-related fields; the predicted text data refers to the prediction results of the model on the target task (such as whether the loan can be repaid on time), which are usually classification labels or probability values; the prediction model is trained by fusing real samples and synthetic samples aligned with features, and has stronger generalization ability and robustness.

[0112] In practical applications, the system first receives a text data prediction request from a small or medium-sized financial institution, which includes the credit behavior records of the user to be predicted. After parsing the request, the system extracts the text to be predicted, which consists of one or more lines of structured data, each line corresponding to a user, and each column corresponding to a specified field (such as "login frequency in the last 30 days," "current loan balance," "historical overdue payments," etc.). The system then inputs this text into a deployed prediction model. The model internally standardizes the input fields to ensure consistency with the feature space during the training phase, and then calculates the prediction result through forward propagation. Finally, the system returns the predicted text data to the requester, including a judgment of each user's repayment ability (such as "on-time repayment" or "potential overdue") and the corresponding confidence score. The entire prediction process is executed in a controlled environment, without caching or recording the original input, ensuring the temporality and privacy of the user behavior data.

[0113] In one specific embodiment of this specification, a regional microfinance company needs to assess the repayment ability of a batch of newly applying users. The text submitted by the company contains 100 user records, each with three fields: "Number of active days on the APP in the last 7 days" (integer, value 0-7), "Current credit utilization rate" (floating-point, value 0.0-1.0), and "Number of historical defaults on associated devices" (integer, value 0-5). After receiving the text, the system inputs it into a gradient boosting tree model built based on the aforementioned training method. During the training phase, the model has integrated the company's own real samples (approximately 800 records, with 8% positive examples) and platform-generated synthetic samples (2000 records, after field alignment and distribution calibration). The model performs feature encoding and ensemble decision-making on each input record, outputting a binary prediction and probability value for "whether the loan can be repaid on time". For example, given a user's "6 active days on the app in the last 7 days", "current credit usage rate = 0.35", and "0 historical defaults on associated devices", the model outputs "on-time repayment" with a probability of 0.92. All prediction results are aggregated into predicted text data and returned to the company in structured table format for subsequent approval decisions.

[0114] One embodiment of this specification provides a text data prediction method. Since the prediction model is trained based on the rich sample data obtained by the aforementioned prediction model training method, which integrates real samples and synthetic samples aligned with features, the coverage and distribution diversity of the training data are effectively expanded, thereby significantly improving the model's generalization ability and robustness in complex, sparse or long-tail scenarios, making the prediction results of the text to be predicted more accurate and stable.

[0115] The following is in conjunction with the appendix Figure 3Taking the application of the prediction model training method provided in this specification in the assessment of user behavior by financial institutions as an example, the prediction model training method will be further explained. Figure 3 This is a flowchart illustrating the processing steps of a prediction model training method provided in one embodiment of this specification, specifically including the following steps.

[0116] Step 302: The customer logs into the multi-party decision-making collaboration platform, applies for cloud desktop modeling, and uploads sample ID1 + backtracking time.

[0117] Step 304: Initiate the selection of feature products to the feature data retrieval platform and create a backtracking task.

[0118] Step 306: The feature data acquisition platform returns the feature file (ID1, X1) to the multi-party decision-making collaboration platform.

[0119] Specifically, the client applies for cloud desktop modeling space (i.e., the aforementioned training node) through the multi-party decision-making collaboration platform (i.e., the aforementioned training platform). The client inputs the modeling sample [client ID1, feature field set X1, backtracking time], where the feature field set is specified by the client through the feature directory list provided by the collaborative modeling platform, and a collaborative modeling task is created. After the multi-party decision-making collaboration platform completes the approval process internally, it will automatically schedule the feature data retrieval platform to create a backtracking task and return the corresponding feature file [ID1, X1].

[0120] Step 308: The multi-party decision-making collaboration platform uploads feature files to the synthetic data service and initiates a data synthesis task.

[0121] Step 310: Data synthesis service item feature retrieval platform returns data synthesis file (ID2, X2, Y2).

[0122] Specifically, after the feature acquisition platform completes the feature file, the multi-party decision-making collaboration platform synchronously initiates a data synthesis task. The key information for initiating the task must include the feature file [ID1, X1], the amount of synthesized data [gen_cnt], the applied synthesis model [aigd_model], and the data columns to be synthesized [x0, x1...xn].

[0123] Step 312: The multi-party decision-making collaboration platform uploads the feature files (ID1, X1) and synthetic data files (ID2, X2, Y2) to the cloud desktop modeling space.

[0124] Step 314: The customer uploads sample label Y1 to the cloud desktop modeling space.

[0125] Step 316: The multi-party collaborative platform disconnects the network from the cloud to the desktop modeling space.

[0126] Specifically, the multi-party decision-making collaboration platform packages and installs the feature files [ID1, X1] (i.e., the aforementioned first sample data), the synthetic data files [ID2, X2, Y2] (i.e., the aforementioned second sample data), and the sample recommendation program Recommend.bin (i.e., the aforementioned migration program) into the cloud desktop modeling space. The customer uploads the sample label file [Y1] (i.e., the aforementioned label data file) through the multi-party decision-making collaboration platform. Finally, the external access network of the cloud space only retains a special VNC channel, and only the customer has unique access control permissions.

[0127] Step 318: Run the sample recommendation program in the cloud desktop modeling space. Based on real data (ID1, X1, Y1), use synthetic data (ID2, X2, Y2) to apply the sample transfer algorithm to make sample recommendations.

[0128] Step 320: The sample recommendation program returns the recommended synthetic data (ID3, X3, Y3) to the cloud desktop modeling space.

[0129] Specifically, the sample recommendation program is an executable program within the cloud space. When triggered by the client, it automatically executes the sample transfer algorithm and generates recommended synthetic data files [ID3, X3, Y3] (i.e., the aforementioned third sample data). Note that this stage is fundamentally different from using plaintext data for sample recommendation, because the result files returned by the sample transfer algorithm can be directly viewed by the client during modeling. Therefore, the synthetic data plays a crucial role at this stage by providing data that is usable but not "visible".

[0130] Step 322: In the cloud desktop modeling space, the user completes the modeling using real data (ID1, X1, Y1) + recommended synthetic data (ID3, X3, Y3).

[0131] Step 324: The cloud desktop modeling space returns the model to the multi-party decision-making collaboration platform and deploys it.

[0132] Specifically, customers log into the cloud modeling space and use real data [ID1, X1, Y1] and combined data [ID1+ID3, X1+X3, Y1+Y3] to complete modeling (model_true) and model_recommend respectively. They evaluate the performance metrics of the models, such as KS and AUC, and select the best model. If the final model_recommend performs better, it means that the sample transfer based on the synthetic data has achieved the expected results.

[0133] Through steps 302-304 above, the cloud-based collaborative modeling scheme requires users to access data in the sandbox environment during the modeling process. Because synthetic data is used, the displayed data cannot be directly or indirectly located to the real data, so the data is essentially invisible. The synthetic data generated based on the real data has similar statistical characteristics, so in most scenarios, combined with the sample transfer algorithm, it can bring additional model effect gains, that is, the data has usability.

[0134] See Figure 4 , Figure 4 This is a timeline diagram of a prediction model training method provided in one embodiment of this specification. Specifically, it includes four stages: feature acquisition, data synthesis, sample recommendation, and model training.

[0135] Feature Data Acquisition Phase: In this phase, the client first logs into the multi-party decision-making collaboration platform, submits a cloud desktop modeling application, and uploads the sample ID and backtracking time information. Upon receiving the request, the platform guides the client to select a feature product and create a backtracking task. Subsequently, it retrieves a feature file containing the user's identity ID (ID1) and a set of feature fields (X1) from the feature data acquisition platform. This process ensures a precise match between the client's specified task requirements and historical data, guaranteeing the accuracy of the feature dimensions used in subsequent modeling.

[0136] Data synthesis phase: The client uploads the feature files (ID1, X1) to the synthetic data service, initiating a data synthesis task. The synthetic data service generates a new synthetic dataset (ID2, X2, Y2) based on the platform's full historical data (non-client data) and constrained by the feature structure provided by the client. This dataset retains the overall distribution characteristics of the platform's data but does not contain any real user records, thus expanding the sample size while ensuring privacy and security, providing high-quality data support for subsequent modeling.

[0137] Sample Recommendation Phase: The client uploads the original feature file (ID1, X1) and the synthetic data file (ID2, X2, Y2) to the cloud desktop modeling space, and simultaneously uploads the real label file Y1. After disconnecting from the network, the system enters an isolated environment and executes a sample transfer algorithm through the sample recommendation program: based on the feature space structure of the real data (ID1, X1, Y1) and the synthetic data (ID2, X2, Y2), the synthetic data undergoes domain adaptation transformation to generate recommended synthetic data (ID3, X3, Y3) that is highly aligned with the real data in terms of field definitions, value ranges, and statistical distributions, achieving secure sample augmentation that is "usable but not visible".

[0138] Model training phase: Customers train their models using real data (ID1, X1, Y1) and recommended synthetic data (ID3, X3, Y3) in the cloud desktop modeling space. After training, the system returns the trained model to the multi-party decision-making collaboration platform and completes deployment. This process, by fusing real samples with adapted synthetic samples, significantly improves the model's generalization ability and stability in long-tail scenarios. Furthermore, the entire process is executed in an isolated environment, ensuring data security and compliance.

[0139] Corresponding to the above method embodiments, this specification also provides task platform embodiments. Figure 5 This is a schematic diagram of the structure of a training platform provided in one embodiment of this specification. Figure 5 As shown, the task platform 500 includes: a request interface 502, a response unit 504, and training nodes 506 for multiple objects; Request interface 502 is used to receive modeling requests sent by the target object; The response unit 504 is used to obtain first sample data from the training node corresponding to the target object based on the modeling request, and generate second sample data based on the historical data accumulated by the training platform. The first sample data includes a specified field, which represents the task characteristics of the target task. Training node 506 is used to perform feature matching on the second sample data based on task features to obtain the third sample data, and to train the prediction model corresponding to the target task based on the first sample data and the third sample data.

[0140] The above is an illustrative scheme of a training platform according to this embodiment. It should be noted that the technical solution of this training platform and the technical solution of the prediction model training method described above belong to the same concept. For details not described in detail in the technical solution of the training platform, please refer to the description of the technical solution of the prediction model training method described above.

[0141] Corresponding to the above method embodiments, this specification also provides embodiments of a prediction model training device. Figure 6 This is a schematic diagram of the structure of a prediction model training device provided in one embodiment of this specification. Figure 6 As shown, the device includes: The first acquisition module 602 is configured to acquire first sample data from the training node corresponding to the target object, wherein the first sample data includes a specified field, and the specified field represents the task features of the target task. The generation module 604 is configured to generate second sample data based on historical data accumulated on the training platform. The matching training module 606 is configured to call the training node to perform feature matching on the second sample data based on the task features to obtain the third sample data, and to train the prediction model corresponding to the target task based on the first sample data and the third sample data.

[0142] Optionally, the first acquisition module 602 is further configured to: in response to a data acquisition request submitted by the target object, parse the field information and time information contained in the data acquisition request, wherein the field information represents the feature dimensions required by the target task, and the time information represents the historical time window required by the target task; and extract the first sample data from the training node based on the field information and time information.

[0143] Optionally, the generation module 604 is further configured to: acquire the fourth sample data accumulated by the training platform, wherein the fourth sample data is historical data of other objects besides the target object; and generate second sample data that conforms to the statistical information by using the fourth sample data as a reference sample and the statistical information of the first sample data as a constraint.

[0144] Optionally, the generation module 604 is further configured to: determine historical data sources that are task-related to the target task in the training platform; and filter historical data that match a specified field from the historical data sources to obtain fourth sample data.

[0145] Optionally, the generation module 604 is further configured to: train a generation model based on the fourth sample data, wherein the generation model is used to learn the data features of the fourth sample data; input the statistical information of the first sample data into the generation model to obtain second sample data that conforms to the data features and statistical information.

[0146] Optionally, the predictive model training apparatus further includes a building module configured to allocate virtual computing resources to the target object in response to a modeling request from the target object; and to build a computing environment as a training node based on the virtual computing resources, wherein the computing environment allows the target object to access it through authorization.

[0147] Optionally, the prediction model training device further includes an encapsulation module configured to encapsulate the second sample data with a preset transfer procedure to form a data packet, wherein the transfer procedure is used to perform feature matching on the second sample data; and to send the data packet to the training node.

[0148] Optionally, the matching training module 606 is further configured to: invoke the training node to perform a domain adaptation transformation on the second sample data according to the feature distribution of the first sample data to obtain the third sample data, wherein the domain adaptation transformation is used to align the feature distribution of the third sample data with the feature distribution of the first sample data.

[0149] Optionally, the matching training module 606 is further configured to: in the training node, parse the first sample data to obtain field definition information and field value range constraints, wherein the field definition information represents the field meta-information of each specified field, and the field value range constraints represent the effective value range of each specified field; according to the field meta-information, perform field structure alignment on the second sample data so that the second sample data contains a set of specified fields that match the first sample data; according to the effective value range, perform value range adaptation processing on the field structure aligned second sample data to obtain the third sample data.

[0150] Optionally, the prediction model training device further includes an update module configured to acquire a label data file provided by the target object, wherein the label data file contains label information corresponding to the sample identifiers in the first sample data; input the label data file into the training node, and update the first sample data through the label data file.

[0151] Optionally, the matching training module 606 is further configured to: call the training node to train a first prediction model based on the first sample data; call the training node to train a second prediction model based on the combined sample data of the first sample data and the third sample data; and determine the prediction model corresponding to the target task from the first prediction model and the second prediction model based on the performance evaluation rules corresponding to the target task.

[0152] Optionally, the prediction model training device also includes a lifecycle management module, configured to clear the training node corresponding to the target object and release the virtual computing resources occupied by the training node.

[0153] Applied to this predictive model training device, the first acquisition module 602 obtains first sample data containing task features from the training node of the target object, ensuring accurate alignment with modeling requirements; the generation module 604 generates non-realistic second sample data based on its statistical information to avoid leakage of original sensitive data; the matching training module 606 calls a transfer procedure within the training node to perform feature matching on the second sample data based on task features to generate third sample data, and then merges the first and third sample data to complete model training. These modules work together to achieve "data usable but invisible," enhancing sample diversity and improving model generalization ability while protecting privacy.

[0154] The above is an illustrative scheme of a predictive model training device according to this embodiment. It should be noted that the technical solution of this predictive model training device and the technical solution of the predictive model training method described above belong to the same concept. For details not described in detail in the technical solution of the predictive model training device, please refer to the description of the technical solution of the predictive model training method described above.

[0155] Corresponding to the above method embodiments, this specification also provides embodiments of a text data prediction device. Figure 7 This is a schematic diagram of the structure of a text data prediction device provided in one embodiment of this specification. Figure 7 As shown, the device includes: The second acquisition module 702 is configured to acquire the text to be predicted in response to a text data prediction request; The prediction module 704 is configured to input the text to be predicted into the prediction model to obtain the predicted text data of the target task, wherein the prediction model is trained based on the prediction model training method.

[0156] Applied to this text data prediction device, the second acquisition module 702 responds to the prediction request and acquires the text to be predicted, which is then input into the prediction model trained by the aforementioned prediction model training method by the prediction module 704. Because this model integrates real samples and feature-aligned synthetic samples for training, it significantly improves the breadth of data coverage and the diversity of distribution. Therefore, the prediction module 704 possesses stronger generalization ability and robustness when processing text to be predicted in complex, sparse, or long-tail scenarios, and the output predicted text data is more accurate and stable.

[0157] The above is an illustrative scheme of a text data prediction device according to this embodiment. It should be noted that the technical solution of this text data prediction device and the technical solution of the above-described text data prediction method belong to the same concept. For details not described in detail in the technical solution of the text data prediction device, please refer to the description of the technical solution of the above-described text data prediction method.

[0158] Figure 8 A structural block diagram of a computing device 800 according to one embodiment of this specification is shown. The components of the computing device 800 include, but are not limited to, a memory 810 and a processor 820. The processor 820 is connected to the memory 810 via a bus 830, and a database 850 is used to store data.

[0159] The computing device 800 also includes an access device 840, which enables the computing device 800 to communicate via one or more networks 860. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 840 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0160] In one embodiment of this specification, the above-described components of the computing device 800 and Figure 8 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 8 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0161] The computing device 800 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 800 can also be a mobile or stationary server.

[0162] The processor 820 is used to execute the following computer program / instruction, which, when executed by the processor, implements the steps of the above-mentioned prediction model training method or text data prediction method.

[0163] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the computing device embodiments are basically similar to the prediction model training method or text data prediction method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the prediction model training method or text data prediction method embodiments.

[0164] An embodiment of this specification also provides a computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the above-described prediction model training method or text data prediction method.

[0165] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the computer-readable storage medium embodiments are described simply because they are substantially similar to the prediction model training method or text data prediction method embodiments; relevant parts can be referred to the descriptions of the prediction model training method or text data prediction method embodiments.

[0166] An embodiment of this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described prediction model training method or text data prediction method.

[0167] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product belongs to the same concept as the technical solution of the above-described prediction model training method or text data prediction method. For details not described in detail in the technical solution of the computer program product, please refer to the description of the technical solution of the above-described prediction model training method or text data prediction method.

[0168] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0169] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0170] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0171] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0172] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A prediction model training method, applied to a training platform, the training platform comprising training nodes corresponding to each object, wherein the training nodes are isolated from each other; the method comprising: First sample data is obtained from the training node corresponding to the target object, wherein the first sample data includes a specified field, and the specified field represents the task characteristics of the target task; Based on the historical data accumulated by the training platform, second sample data is generated; The training node is invoked to perform feature matching on the second sample data based on the task features to obtain the third sample data, and a prediction model corresponding to the target task is trained based on the first sample data and the third sample data.

2. The method according to claim 1, wherein obtaining the first sample data from the training node corresponding to the target object includes: In response to a data acquisition request submitted by a target object, the field information and time information contained in the data acquisition request are parsed, wherein the field information represents the feature dimensions required by the target task, and the time information represents the historical time window required by the target task; Based on the field information and the time information, the first sample data is extracted from the training node.

3. The method according to claim 1, wherein generating the second sample data based on the historical data accumulated by the training platform includes: Obtain the fourth sample data accumulated by the training platform, wherein the fourth sample data is historical data of other objects besides the target object; Using the fourth sample data as a reference sample and the statistical information of the first sample data as constraints, a second sample data that conforms to the statistical information is generated.

4. The method according to claim 1, further comprising, before obtaining the first sample data from the training node corresponding to the target object: In response to a modeling request for a target object, virtual computing resources are allocated to the target object; Based on the virtual computing resources, a computing environment is constructed as the training node, wherein the computing environment allows the target object to access it through authorization; Before invoking the training node to perform feature matching on the second sample data based on the task features to obtain the third sample data, the method further includes: The second sample data is encapsulated with a preset migration procedure to form a data packet, wherein the migration procedure is used to perform feature matching on the second sample data; The data packet is sent to the training node.

5. The method according to claim 1, wherein the task is the feature distribution of sample data; the step of calling the training node to perform feature matching on the second sample data based on the task features to obtain the third sample data includes: The training node is invoked to perform a domain adaptation transformation on the second sample data based on the feature distribution of the first sample data to obtain the third sample data. The domain adaptation transformation is used to align the feature distribution of the third sample data with the feature distribution of the first sample data. The step of performing a domain adaptation transformation on the second sample data based on the feature distribution of the first sample data to obtain the third sample data includes: In the training node, the first sample data is parsed to obtain field definition information and field value range constraints, wherein the field definition information represents the field meta-information of each specified field, and the field value range constraints represent the effective value range of each specified field; Based on the field metadata, the second sample data is aligned with the field structure so that the second sample data contains a specified set of fields that match the first sample data; Based on the effective value range, the second sample data after field structure alignment is subjected to value range adaptation processing to obtain the third sample data.

6. The method according to claim 1, wherein training the prediction model corresponding to the target task based on the first sample data and the third sample data comprises: The training node is invoked to train a first prediction model based on the first sample data; The training node is invoked to train a second prediction model based on the combined sample data of the first sample data and the third sample data. Based on the performance evaluation rules corresponding to the target task, the prediction model corresponding to the target task is determined from the first prediction model and the second prediction model.

7. A text data prediction method, comprising: In response to a text data prediction request, retrieve the text to be predicted; The text to be predicted is input into the prediction model to obtain the predicted text data of the target task, wherein the prediction model is trained based on the prediction model training method according to any one of claims 1-6.

8. A computing device, comprising: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions. When the computer programs / instructions are executed by the processor, they implement the steps of the prediction model training method according to any one of claims 1-6 or the text data prediction method according to claim 7.

9. A computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the prediction model training method of any one of claims 1-6 or the text data prediction method of claim 7.

10. A computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the prediction model training method of any one of claims 1-6 or the text data prediction method of claim 7.

Images