Model training data construction method and device, large model, and probability prediction method

Abstract

The application relates to the technical field of artificial intelligence, and provides a model training data construction method and device, a large model and a probability prediction method, which comprise the following steps: inputting user behavior data in a first latest time period into a prediction model to obtain prediction probability of a second future time period output by the prediction model; wherein the prediction model is obtained by training user data in a complete historical period; converting the prediction probability into a prediction label; fusing an actual label of the user behavior data and the prediction label into a label sequence, and constructing training data based on the label sequence.The application directly incorporates the latest recent user behavior data into training samples, fills in future labels that have not yet occurred by using model prediction, constructs training data capable of improving the adaptability of a model to the latest behavior characteristics, breaks the limitation that traditional methods must wait until an observation period ends to use data, and improves the timeliness of a prediction model.

Description

Technical Field

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

[0002] In time-series forecasting scenarios, it is often necessary to predict user behavior within a specific period. To obtain reliable prediction results, a training set containing user behavior data and corresponding labels needs to be constructed to train the prediction model. However, obtaining training data with complete actual labels usually requires waiting until the entire observation period ends. This collection method inevitably leads to a lag in training data acquisition, making it difficult to meet actual business needs.

[0003] To address the issue of delayed training data acquisition, existing technologies typically utilize historical data from completed observation periods for model training. Specifically, this method extracts long-term historical data with fully generated behavioral labels as training samples to train a time-series prediction model. This model is then used to directly predict the user's future state, thus avoiding the process of waiting for the current user's observation period to end.

[0004] While the aforementioned existing technologies avoid waiting for data, user behavior evolves with the market, and models trained solely on complete historical data cannot adapt to the latest changes in a timely manner, leading to a decline in prediction accuracy. Conversely, while recent user data reflects the latest characteristics, it cannot be used directly due to the lack of subsequent actual labels. Therefore, how to construct training data that can improve the model's adaptability to the latest behavioral features has become a pressing problem in this field. Summary of the Invention

[0005] This invention provides a method and apparatus for constructing model training data, a large model, and a probability prediction method to solve the technical problem of how to construct training data that can improve the model's ability to adapt to the latest behavioral features.

[0006] This invention provides a method for constructing model training data, comprising: The user behavior data of the most recent first time period is input into the prediction model to obtain the predicted probability of the future second time period output by the prediction model; wherein, the prediction model is trained based on user data within a complete historical period; The predicted probability is converted into a predicted label; The actual labels and predicted labels of the user behavior data are fused into a label sequence, and training data is constructed based on the label sequence.

[0007] According to a method for constructing model training data provided by the present invention, the predicted probability is converted into a predicted label, comprising: The predicted probability is corrected to obtain the corrected probability; The correction probability is converted into the predicted label.

[0008] According to a model training data construction method provided by the present invention, the predicted probability is corrected to obtain the corrected probability, including: The predicted probabilities are input into each calibration model to obtain the calibration results output by each calibration model. The weighted average of each calibration result is determined to obtain the correction probability.

[0009] According to a method for constructing model training data provided by the present invention, the calibration model includes a bias function calibration model, a shape transfer calibration model, a regression calibration model, or an order-preserving regression calibration model.

[0010] According to a model training data construction method provided by the present invention, the predicted probability is corrected to obtain the corrected probability, including: Determine the average retention rate corresponding to the second time period within the complete historical cycle; The predicted probability is weighted and adjusted based on the average retention rate to obtain the corrected probability.

[0011] According to a method for constructing model training data provided by the present invention, the predicted probability is converted into a predicted label, comprising: Monte Carlo sampling is performed on the predicted probabilities to obtain the sampling results; Based on the sampling results, the continuous predicted probabilities are converted into discrete predicted labels.

[0012] According to a method for constructing model training data provided by the present invention, the actual labels and predicted labels of the user behavior data are fused into a label sequence, including: The actual labels and the predicted labels are concatenated in chronological order to obtain the label sequence.

[0013] The present invention also provides a model training data construction apparatus, comprising: The prediction module is used to input the user behavior data of the most recent first time period into the prediction model to obtain the predicted probability of the future second time period output by the prediction model; wherein, the prediction model is trained based on user data within a complete historical period; The conversion module is used to convert the predicted probability into a predicted label; The construction module is used to fuse the actual labels and the predicted labels of the user behavior data into a label sequence, and to construct training data based on the label sequence.

[0014] The present invention also provides a large model, which is obtained by training data constructed based on any model training data construction method.

[0015] The present invention also provides a probability prediction method, comprising: By inputting user behavior data regarding internet financial service products into the aforementioned large model, the user retention probability output by the large model is obtained.

[0016] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the model training data construction method as described above.

[0017] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the model training data construction method as described above.

[0018] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the model training data construction method as described above.

[0019] The model training data construction method and apparatus, large model, and probability prediction method provided by this invention input the most recent first time period user behavior data into a prediction model trained based on data within a complete historical period, obtain the predicted probability of the second future time period and convert it into a predicted label, and then fuse it with the actual label of the first time period to construct training data. Compared with the prior art, which can only rely on lagged historical data that has passed the complete observation period for training, this invention directly incorporates the latest recent user behavior data into the training samples and uses model prediction to fill in the future labels that have not yet occurred, constructing training data that can improve the model's adaptability to the latest behavioral features. This breaks the limitation of traditional methods that must wait for the observation period to end before data can be used, and improves the timeliness of the prediction model. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the model training data construction method provided by the present invention.

[0022] Figure 2This is a schematic diagram illustrating the principle of the model training data construction method provided by the present invention.

[0023] Figure 3 This is a schematic diagram of the structure of the model training data construction device provided by the present invention.

[0024] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0026] In various short-term subscription services or businesses with observation periods, a buffer period or observation period is usually set up to allow users to continue enjoying the service even if they do not renew or confirm on time. To improve operational efficiency, it is necessary to predict the user's continued usage status during the observation period. In this prediction scenario, model training usually relies on complete historical data. The typical approach is to use user data that has completed the observation period (such as users who have generated complete status or behavior for 1 to 30 days in the past month) for training. However, user behavior continues to evolve with market dynamics. On the one hand, to capture the latest changes in user behavior, it is necessary to use recent data to train the model; on the other hand, recent users have not yet completed their complete observation period, lacking actual labels for subsequent time periods, and cannot be directly used for training. For example, when predicting the status of users in the next 1 to 30 days, users who registered in the most recent week only have actual status data for days 1 to 7, and the status for days 8 to 30 is unknown. This missing data makes it impossible to directly train the model with new data. If only complete historical data is relied upon, the model cannot adapt to the evolution of user behavior patterns in a timely manner, and the prediction accuracy will continue to decline.

[0027] The following is combined with Figures 1 to 4 The present invention describes a method and apparatus for constructing model training data.

[0028] Figure 1 This is a flowchart illustrating the model training data construction method provided by the present invention, as shown below. Figure 1 As shown, the method includes, but is not limited to, steps S1, S2 and S3.

[0029] Step S1: Input the user behavior data of the most recent first time period into the prediction model to obtain the predicted probability of the future second time period output by the prediction model; wherein, the prediction model is trained based on user data within a complete historical period.

[0030] The most recent first time period is a relatively short historical time interval before the current time point, which may include the most recent week (e.g., day 1 to day 7), the most recent three days, the most recent five days, or the most recent month, etc.

[0031] User behavior data refers to objective data with recordable value generated by users during their use of products or services. This data may include user login activity data, click data, purchase data, or browsing time data.

[0032] The predictive model is an older model trained using long-term data, used to infer the future target state based on input features. A complete historical period is a long time span that has completely ended and generated complete real-world labels, such as the past year or the past six months. User data is a dataset containing user behavior data and complete real-world labels collected within each complete historical period.

[0033] The second future time period is a time interval that follows the first time period but has not yet been fully experienced and needs to be estimated. This can be a longer prediction period, such as day 8 to day 30, the next week, or the next month. The predicted probability is the likelihood of an event occurring on a future day, as output by the model.

[0034] For example, short-term labeled data from the past week can be extracted for a newly registered user, such as data from days 1-7 of the past week. This data represents real events and has high reliability. This data from days 1-7 of the past week is then input into a pre-trained old model. This old model was trained using historical data from 1-30 days over the past year. While its predictive ability for the latest changes is somewhat reduced, it still provides basic predictive capabilities. Subsequently, the retention probability for days 8-30 is estimated based on the recent short-term data, thus obtaining the model's output retention probability for the user from day 8 to day 30. For example, a probability of 0.6 on day 8, a probability of 0.4 on day 15, etc. This value represents probability, not a definitive 0 / 1 label.

[0035] Step S1 uses an old model trained on long-term historical data to estimate recent short-term data that lacks real labels due to the observation period not yet being over. This provides a data foundation for generating a complete training label sequence in the future and effectively overcomes the defect that recent data cannot be used directly due to missing labels.

[0036] Step S2: Convert the predicted probability into a predicted label.

[0037] Predicted labels represent the final category to which the target belongs, and can include retention labels (e.g., 1 for retention, 0 for non-retention), churn labels, activity status labels, etc. They can be truncated by setting a fixed probability threshold (e.g., greater than 0.5 is recorded as 1, otherwise as 0), or by using a probability sampling algorithm to map or discretize the continuous range of predicted probabilities into specific category labels.

[0038] For example, the retention probabilities of the old model output from day 8 to day 30, such as 0.6 or 0.4, can be smoothly transformed into definite retention labels of 0 or 1 through specific mapping logic or algorithm. For example, the probability of 0.6 can be transformed into label 1.

[0039] Step S2 can transform the continuous probability values ​​representing possibilities into discrete real state forms that meet the training requirements of supervised learning models, so that the predicted data are consistent with the actual observed labels in terms of data structure, and can be successfully used for subsequent model training.

[0040] Step S3: The actual labels and predicted labels of the user behavior data are fused into a label sequence, and training data is constructed based on the label sequence.

[0041] Actual tags are the target states that actually occur and are recorded within the first time period. These can be actual retention tags (0 or 1) from day 1 to day 7, actual churn tags, actual purchase tags, etc. Tag sequences are collections of multiple tags arranged according to a certain dimension. These can be complete user retention tag time series, user churn tag time series, activity status sequences, etc.

[0042] For example, the actual retention tags of users from day 1 to day 7 are merged with the generated predicted tags from day 8 to day 30 to form a complete user retention tag sequence spanning 30 days, and a new training sample set is constructed based on this sequence.

[0043] Step S3 combines real and reliable short-term recent tags with predicted and completed long-term future tags to construct a sample set containing the latest user behavior characteristics and a complete tag system, effectively solving the problem of unavailable data for newly registered users.

[0044] As described above, this invention inputs the user behavior data of the most recent first time period into a prediction model trained on data from a complete historical period, obtains the predicted probability of the second time period, converts it into a predicted label, and then merges it with the actual label of the first time period to construct training data. Compared with existing technologies that can only rely on lagging historical data that has completed the full observation period for training, this invention directly incorporates the latest recent user behavior data into the training samples and uses the model prediction to fill in the future labels that have not yet occurred, constructing training data that can improve the model's ability to adapt to the latest behavioral features. This breaks the limitation of traditional methods that must wait for the observation period to end before data can be used, and improves the timeliness of the prediction model.

[0045] In one embodiment, step S2 may further include: The predicted probability is corrected to obtain the corrected probability. The correction probability is converted into a predicted label.

[0046] The corrected probability is a more accurate probability output after error compensation or distribution adjustment. It can utilize historical prior knowledge, statistical averages, or additional calibration algorithms to perform numerical addition, subtraction, multiplication, scaling operations on the original predicted probability, or use a calibration function for nonlinear mapping.

[0047] For example, since the probability predicted by the old model for days 8 to 30 usually has an overall bias due to the data distribution shift, it is necessary to correct the short-term 8-30 day prediction probability. Combine this with a certain long-term data trend to perform mathematical correction on the initial prediction probability, so as to obtain a more accurate corrected probability. Then, the corrected probability is converted into a prediction label of 0 or 1.

[0048] This invention corrects the predicted probabilities before converting them into predicted labels. Since the predicted probabilities directly output by older models often have systematic biases, correcting the probabilities improves their reliability and accuracy, resulting in training labels that better reflect the real data distribution and helping to reduce label noise during model training.

[0049] In one embodiment, correcting the predicted probability to obtain the corrected probability may further include: The predicted probabilities are input into each calibration model to obtain the calibration results output by each calibration model. The weighted average of the calibration results is determined to obtain the correction probability.

[0050] A calibration model is a mathematical aid specifically designed to correct the discrepancy between the model's output probability and the true posterior probability. It can include the sub-models within an integrated calibration model. The weighted average is the sum of the values ​​obtained by multiplying each calibration result by its corresponding weighting coefficient.

[0051] For example, the retention prediction probabilities for days 8 to 30 can be input into multiple parallel calibration models to obtain the calibration results output independently by each model. Then, the principle of the integrated calibration model is called to set the weight coefficients of each calibration result and determine the weighted average of these results. The result of this weighted sum is used as the final corrected probability after correcting the short-term 8-30 day prediction probabilities.

[0052] This invention employs integrated processing for weighted averaging, which can reduce the error risk caused by overfitting or limited applicability of a single calibration model, and effectively improve the stability and generalization ability of the correction probability.

[0053] In one embodiment, the calibration model may include a deviation function calibration model, a shape migration calibration model, a regression calibration model, or an order-preserving regression calibration model.

[0054] Bias function calibration models correct for discrepancies between predicted and true values ​​by fitting a distribution function of the deviation between the predicted and true values; examples include multinomial bias fitting models. Shape transfer calibration models align distributions using the shape characteristics of the data distribution; examples include histogram matching models. Regression calibration models remap probabilities using linear or nonlinear regression algorithms; examples include logistic regression calibration models. Ordination-preserving regression calibration models maintain the monotonicity of probabilities by fitting a step function; examples include regression models based on strict ordination constraints.

[0055] For example, to address the limitations of a single model, the integrated calibration model can specifically set up and employ four sub-models: bias function calibration, shape transfer calibration, regression calibration model, and ordinal-preserving regression calibration. Depending on actual business needs, the predicted probabilities can be corrected using one of these models or a flexible combination of multiple models.

[0056] This invention provides a variety of mathematical calibration models that can be flexibly combined and the most suitable calibration strategy can be selected according to different data deviation characteristics, thereby improving the model correction effect.

[0057] In one embodiment, correcting the predicted probability to obtain the corrected probability may further include: Determine the average retention rate corresponding to the second time period within the complete historical cycle; The predicted probability is adjusted by weighting the average retention rate to obtain the corrected probability.

[0058] Average retention rate is the average proportion of users who remain within a specific time period over a longer historical statistical period. Weighted adjustment can be achieved by multiplying the predicted probability output by the model by the average retention rate according to a set proportional coefficient and then summing the results.

[0059] For example, based on long-term data (such as the trend of user retention rate changes over the past year), the average retention rate on day 8 in history is determined to be 0.7 using aggregate statistical methods. If the current old model predicts a user's probability on day 8 to be 0.6, then based on the average retention rate of 0.7, the weight ratio formula is used to adjust 0.6 to obtain the adjusted corrected probability. For example, with a weight of 0.5 to 0.5, the corrected probability is 0.65.

[0060] This invention introduces historical statistical averages as anchoring standards, which allows the overall distribution of the corrected probability to converge toward the macroscopic true statistical trend while preserving individual prediction differences, thereby avoiding extreme deviations in the predicted probability.

[0061] In one embodiment, step S2 may further include: Monte Carlo sampling is performed on the predicted probabilities to obtain the sampling results; Based on the sampling results, the continuous prediction probabilities are transformed into discrete prediction labels.

[0062] Monte Carlo sampling is a method of repeated sampling using a random distribution. Monte Carlo sampling may include: calling a pseudo-random number generator to generate a uniformly distributed random number between 0 and 1, comparing the random number with the predicted probability to be processed, and outputting a predicted label 1 if the random number is less than the predicted probability, otherwise outputting a predicted label 0.

[0063] For example, Monte Carlo sampling is performed on the corrected continuous probability (e.g., 0.6) to generate random numbers. If the random number is less than 0.6, the sampling result is 1; otherwise, it is 0. That is, based on a probability of 0.6, the sampling result may be 0 or 1. However, since the probability is biased towards 0.6, the higher the sampling result, the more likely it is to be 1. This generates 0 / 1 labels for short-term data retention of 8-30 days, smoothly transforming continuous predicted probabilities into discrete 0 or 1 predicted labels.

[0064] This invention uses a random sampling mechanism instead of a single fixed threshold segmentation, so that the overall statistical distribution of the discrete labels after batch conversion can approximate the distribution of the original continuous prediction probability without loss, preserving the uncertainty characteristics of the data and effectively improving the generalization representativeness of the training data.

[0065] In one embodiment, step S3, fusing the actual labels and predicted labels of the user behavior data into a label sequence, may further include: By concatenating the actual labels and predicted labels in chronological order, a label sequence is obtained.

[0066] The chronological order is the rule for arranging events according to their occurrence time, and can include ascending or descending order. For example, when preparing to build a complete sequence for final training, the actual labels of short-term (0-10 days) user behavior data and the predicted labels of short-term (8-30 days) generated through the aforementioned stages are strictly concatenated according to the chronological order of time (i.e., day 1, day 2, ... up to day 30) to obtain a complete and coherent sequence of user retention labels.

[0067] This invention combines and merges data in chronological order, which preserves the temporal correlation characteristics of user behavior tags over time. This ensures that the constructed training data conforms to the input logic of the temporal model, which is beneficial for the model to learn the correct temporal dependencies.

[0068] Based on the preceding text, we can conclude that... Figure 2 The complete technical solution is shown below. This solution can be specifically divided into the following four descriptive stages: In stage 1, the invention trains a prediction model using long-term historical data, constructs a basic prediction model using historical periodic data containing complete labels, and predicts user behavior within a short observation period, thereby obtaining potential user status labels in advance. Simultaneously, to address potential data distribution bias issues in the basic model, model prediction correction is performed in stage 2, specifically as follows... Figure 2 As shown, an integrated calibration model is employed, including multiple operators such as bias function calibration, shape offset calibration, regression calibration model, and ordinal-preserving regression calibration. A weighted average is used to obtain the corrected probability that accurately anchors the current market distribution. Subsequently, in stage 3, Monte Carlo sampling is combined, using a random mapping mechanism instead of a single fixed threshold segmentation to retain the uncertainty characteristics of the data. Representative samples are then selected from the actual data that has not yet passed the observation period and added to the training set, improving the model's generalization ability and prediction accuracy. This approach effectively alleviates the problem of insufficient data timeliness and provides a feasible solution for real-time or near-real-time user status prediction.

[0069] This invention introduces recent user data to estimate the probability of missing time periods using an old model and samples it to generate labels. It transforms recent data without complete labels into usable training samples, breaking through the limitation of traditional methods that require waiting for the observation period to end before data can be used. This enables the model to capture the latest changes in user behavior and market dynamics in a timely manner, improving the model's adaptability to the current user group and the accuracy of prediction.

[0070] The present invention also provides a large model, which is obtained by training on training data constructed based on the above-described model training data construction method.

[0071] In Phase 4, after constructing an augmented dataset by concatenating and fusing actual data nodes (e.g., short-term 1-7 days of actual data) with predicted labels in chronological order, this dataset can be used to build a sample set containing the latest user features and retrain the model. The constructed augmented dataset can be divided into training and validation sets, fed into the model in batches, and the model weights updated by calculating the loss function and using backpropagation. Using the newly generated training data sample set containing recent features and long-term predicted labels, the model can be retrained or a new model can be trained. Because its training data incorporates recent changes, the newly trained model will have better timeliness.

[0072] The model of this invention can adapt to the current dynamic evolution of the market and capture the latest user behavior patterns, with high predictive timeliness.

[0073] The present invention also provides a probability prediction method, comprising: By inputting user behavior data related to internet financial service products into the aforementioned large model, the user retention probability output by the large model is obtained.

[0074] Internet financial service products can be short-term or long-term financial services provided, processed, and disputed online through internet channels. User retention probability is the likelihood that a financial service user will continue to hold their policy, not terminate the service, or continue to pay service fees within a specific period; it can be the probability of renewal the following month, the retention probability during the service waiting period, etc.

[0075] This invention can be practically deployed and applied to the prediction business scenarios of short-term Internet financial service products. In this case, the prediction probability calculated and output by the model is specifically the user retention probability of whether users stay on the platform in each period.

[0076] This invention is applicable to the specific vertical scenario of estimating the waiting period of financial service products, enabling the constructed training data to directly serve the task of predicting user behavior during the waiting period of short-term financial service products, thereby improving the operational efficiency of financial service products in this field.

[0077] The model training data construction apparatus provided by the present invention is described below. The model training data construction apparatus described below can be referred to in correspondence with the model training data construction method described above.

[0078] Based on any of the above embodiments Figure 3 This is a schematic diagram of the structure of the model training data construction device provided by the present invention, as shown below. Figure 3 As shown, the device includes: The prediction module 310 is used to input the user behavior data of the most recent first time period into the prediction model to obtain the prediction probability of the future second time period output by the prediction model; wherein, the prediction model is trained based on user data within a complete historical period; Conversion module 320 is used to convert predicted probabilities into predicted labels; Module 330 is used to fuse the actual and predicted labels of user behavior data into a label sequence, and to build training data based on the label sequence.

[0079] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, communications interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a model training data construction method. This method includes: inputting user behavior data from the most recent first time period into a prediction model to obtain the predicted probability of a future second time period output by the prediction model; wherein the prediction model is trained based on user data from a complete historical period; converting the predicted probability into predicted labels; fusing the actual labels of the user behavior data and the predicted labels into a label sequence, and constructing training data based on the label sequence.

[0080] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0081] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the model training data construction method provided by the above methods. The method includes: inputting user behavior data of the most recent first time period into a prediction model to obtain the prediction probability of a future second time period output by the prediction model; wherein the prediction model is trained based on user data within a complete historical period; converting the prediction probability into prediction labels; fusing the actual labels of the user behavior data and the prediction labels into a label sequence, and constructing training data based on the label sequence.

[0082] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for constructing model training data provided by the methods described above. This method includes: inputting user behavior data from a recent first time period into a prediction model to obtain a predicted probability for a future second time period output by the prediction model; wherein the prediction model is trained based on user data within a complete historical period; converting the predicted probability into predicted labels; fusing the actual labels of the user behavior data and the predicted labels into a label sequence, and constructing training data based on the label sequence.

[0083] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0084] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing model training data, characterized in that, include: The user behavior data of the most recent first time period is input into the prediction model to obtain the predicted probability of the future second time period output by the prediction model; wherein, the prediction model is trained based on user data within a complete historical period; The predicted probability is converted into a predicted label; The actual labels and predicted labels of the user behavior data are fused into a label sequence, and training data is constructed based on the label sequence.

2. The model training data construction method according to claim 1, characterized in that, Converting the predicted probability into a predicted label includes: The predicted probability is corrected to obtain the corrected probability; The correction probability is converted into the predicted label.

3. The method for constructing model training data according to claim 2, characterized in that, The predicted probability is corrected to obtain the corrected probability, including: The predicted probabilities are input into each calibration model to obtain the calibration results output by each calibration model. The weighted average of each calibration result is determined to obtain the correction probability.

4. The model training data construction method according to claim 3, characterized in that, The calibration model includes a deviation function calibration model, a shape migration calibration model, a regression calibration model, or an order-preserving regression calibration model.

5. The method for constructing model training data according to claim 2, characterized in that, The predicted probability is corrected to obtain the corrected probability, including: Determine the average retention rate corresponding to the second time period within the complete historical cycle; The predicted probability is weighted and adjusted based on the average retention rate to obtain the corrected probability.

6. The method for constructing model training data according to claim 1, characterized in that, Converting the predicted probability into a predicted label includes: Monte Carlo sampling is performed on the predicted probabilities to obtain the sampling results; Based on the sampling results, the continuous predicted probabilities are converted into discrete predicted labels.

7. The method for constructing model training data according to claim 1, characterized in that, The actual labels and predicted labels of the user behavior data are fused into a label sequence, including: The actual labels and the predicted labels are concatenated in chronological order to obtain the label sequence.

8. A model training data construction device, characterized in that, include: The prediction module is used to input the user behavior data of the most recent first time period into the prediction model to obtain the predicted probability of the future second time period output by the prediction model; wherein, the prediction model is trained based on user data within a complete historical period; The conversion module is used to convert the predicted probability into a predicted label; The construction module is used to fuse the actual labels and the predicted labels of the user behavior data into a label sequence, and to construct training data based on the label sequence.

9. A large model, characterized in that, The large model is trained based on the training data described in any one of claims 1-7.

10. A probability prediction method, characterized in that, include: By inputting user behavior data related to internet financial service products into the large model described in claim 9, the user retention probability output by the large model is obtained.

11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the model training data construction method as described in any one of claims 1 to 7.

12. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the model training data construction method as described in any one of claims 1 to 7.

13. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the model training data construction method as described in any one of claims 1 to 7.

Images