A physical examination user consultation conversion prediction method and system based on a time sequence network

Abstract

The application discloses a physical examination user consultation conversion prediction method and system based on a time sequence network, relates to the consultation data conversion prediction field, and forms a precise effect evaluation mechanism by training a prediction model based on historical behavior data, comparing a push result with actual user behavior in real time, and forming a precise effect evaluation mechanism. When the push accuracy is lower than a threshold value, a multidimensional optimization process is automatically triggered: first, the optimal configuration in the limited resources is found by adjusting the hyperparameters iteratively, and if it cannot be satisfied, the software performance index is further optimized, and finally the hardware configuration is upgraded when necessary; the hierarchical progressive optimization strategy ensures the timeliness of model updating and avoids resource waste. By including the training data volume, software and hardware indicators and other key parameters into the effect prediction model, the optimization direction can be predicted in advance, the adjustment efficiency is significantly improved, and finally the user behavior change can be continuously adapted, the push accuracy is ensured, and the resource utilization is maximized.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of consultation data conversion prediction, and particularly relates to a consultation conversion prediction method and system for physical examination users based on a time sequence network. BACKGROUND

[0002] A user consultation intention prediction method and system and a computer device are disclosed in Chinese patent application CN115796333A. The method comprises the following steps: obtaining an operation breakpoint encountered by a target user in a business interaction process, and then obtaining a score ranking list of the operation breakpoint through big data classification analysis; obtaining operation content of the target user in the business interaction process, and then obtaining a score ranking list of the operation content through big data classification analysis; obtaining page content, page attributes and page attribute corresponding page times of a page accessed by the target user in the business interaction process, and then obtaining a score ranking list of a business type attribute corresponding to the page through big data classification analysis; matching the score ranking lists of the operation breakpoint, the operation content and the business type attribute with user attribute labels, and obtaining an intention score ranking list through weighted calculation; and predicting a consultation intention of the target user based on the intention score ranking list.

[0003] The existing user consultation conversion prediction system has the problem of performance degradation of the prediction model caused by dynamic changes in user behavior, which cannot timely discover the decrease in accuracy. In traditional resource scheduling, there is the problem of over-configuration of hardware or performance bottleneck of software, which cannot realize accurate allocation of hardware and software resources. Meanwhile, in the model updating process, there is the problem of low adjustment efficiency caused by the lack of a prediction mechanism. SUMMARY

[0004] To solve the problems in the related art, the application provides a consultation conversion prediction method and system for physical examination users based on a time sequence network to overcome the above technical problems in the existing related art.

[0005] To solve the above technical problems, the application is implemented through the following technical solutions:

[0006] The application is a consultation conversion prediction method for physical examination users based on a time sequence network, which comprises the following steps:

[0007] S1, acquiring behavior sequence data of users at a plurality of historical time points and corresponding time point data;

[0008] S2, constructing a final user behavior sequence data prediction model based on the data collected in S1;

[0009] S3, collecting historical model training data volume, training hyperparameters, hardware and software performance indicators, training duration and model test accuracy data in the model updating process to construct a final behavior sequence prediction training effect mapping model;

[0010] S4, input the historical behavior sequence data of the user to be consulted into the mapping model in S2, map and select the behavior sequence with the highest probability for message pushing; then calculate the pushing accuracy, and if the accuracy is lower than the threshold, perform model daily update;

[0011] S5, collect the corresponding training data amount, hyperparameters and software and hardware performance index data of the day and input them into the mapping model in S3; if the mapping result does not meet the requirements, repeatedly adjust the hyperparameters and remap, if the training length and accuracy requirements are met within a limited number of times, complete the configuration, and the adjustment is completed; otherwise, perform S6;

[0012] S6, repeatedly adjust the software performance index in S5, if the training length and accuracy requirements are met within a limited number of times, complete the configuration; otherwise, further repeatedly adjust the hardware performance index until the requirements are met.

[0013] Preferably, the S1 comprises the following steps:

[0014] S11, set a plurality of user behavior sequence types to obtain a user behavior sequence type set;

[0015] S12, according to the user behavior sequence type set, obtain historical user behavior sequence data and corresponding time point data of a plurality of time points to obtain a historical user behavior sequence data set and a historical user behavior sequence time point data set;

[0016] By constructing a user behavior sequence collection system, fine-grained behavior trajectories of users in the platform interaction process can be comprehensively captured, and originally scattered user touch points can be converted into structured data assets with time sequence correlation, which can effectively support user behavior pattern mining, personalized recommendation optimization, product interaction experience improvement and precise evaluation of operation strategy, and provide a high-quality data basis for subsequent construction of user behavior prediction models and intelligent decision systems.

[0017] Preferably, the S2 comprises the following steps:

[0018] S21, according to the historical user behavior sequence data set and the historical user behavior sequence time point data set, construct a prediction model for user behavior sequence data at a future time to obtain a final user behavior sequence data prediction model;

[0019] By constructing a prediction model based on historical user behavior sequence data, discrete user behavior data can be converted into a forward-looking intelligent decision support tool. It can not only help the platform to identify the risk of loss of high-value users in advance, predict the response rate of marketing activities, and optimize content recommendation strategies, but also provide data-driven decision-making basis for product function iteration. The model can dynamically adapt to changes in behavior patterns and continuously optimize prediction accuracy through continuous learning. Ultimately, it forms a complete closed loop from data collection, behavior analysis to prediction application, significantly improving user retention rate, conversion efficiency and personalized service level.

[0020] Preferably, the end-user behavior sequence data prediction model in S21 adopts a Transformer-based time series prediction model.

[0021] The model effectively captures complex temporal dependence relationships in long sequences through self-attention mechanisms, and can handle ultra-long sequences of more than 1000 behavior points and identify non-linear behavior patterns compared to traditional methods.

[0022] Preferably, S3 includes the following steps:

[0023] S31, the end-user behavior sequence data prediction model is used in the actual user resource conversion prediction process; a number of groups of model training data volume data, model training hyperparameter data, hardware performance index data, software performance index data, training duration data, and model test accuracy rate data corresponding to the model during the daily update process are obtained, to obtain a historical model training data volume dataset, a historical model training hyperparameter dataset, a historical hardware performance index dataset, a historical software performance index dataset, a historical training duration dataset, and a historical model test accuracy rate dataset;

[0024] S32, according to the historical model training data volume dataset, the historical model training hyperparameter dataset, the historical hardware performance index dataset, the historical software performance index dataset, the historical training duration dataset, and the historical model test accuracy rate dataset, a mapping model between model training data volume, model training hyperparameter data, hardware performance index data, software performance index data, and training duration data and model test accuracy rate data is constructed, to obtain an end behavior sequence prediction training effect mapping model;

[0025] By integrating model training data volume, hyperparameter configuration, hardware performance index, software environment and other multi-dimensional historical data, a comprehensive training effect mapping model is constructed, which can accurately reveal the internal relationship between training duration and model accuracy under different resource configurations, thereby providing data support for resource allocation optimization. It also has forward-looking prediction capability, which can evaluate the training cost and expected accuracy of different hardware combinations before deployment, effectively avoiding resource waste.

[0026] Preferably, the final behavior sequence prediction training effect mapping model in S32 adopts a multilayer perception regression model.

[0027] The model structure can accurately capture the complex mapping relationship between hardware resource configuration, software environment, data size and model performance through deep nonlinear transformation, thereby providing a quantitative basis for resource allocation and training efficiency optimization.

[0028] Preferably, the S4 comprises the following steps:

[0029] S41, selecting a user to be consulted; obtaining user behavior sequence data of the user to be consulted corresponding to a plurality of time points in history according to the user behavior sequence type set, to obtain a user behavior sequence data set of the user to be consulted in history and a user to be consulted historical time point set;

[0030] S42, inputting the user behavior sequence data set of the user to be consulted in history and the user to be consulted historical time point set into a final user behavior sequence data prediction model for mapping, to obtain a current user behavior sequence probability data set of the user to be consulted; selecting a user behavior sequence corresponding to the maximum probability data in the current user behavior sequence probability data set of the user to be consulted for pushing; after the pushing of the day is completed, calculating the probability data of the pushed message being the same as the actual behavior of the user, to obtain current message pushing accuracy data;

[0031] S43, setting a current message pushing accuracy threshold; if the current message pushing accuracy data is greater than or equal to the current message pushing accuracy threshold, the final behavior sequence prediction training effect mapping model does not need to be updated on the day; otherwise, S5 is executed;

[0032] By introducing a real-time feedback mechanism, the coincidence degree of the actual behavior and the predicted behavior is automatically calculated after each pushing, to form a quantitative accuracy evaluation index, thereby providing data support for subsequent optimization. In addition, the model updating demand is dynamically judged by an adaptive threshold, which not only avoids resource waste caused by frequent updating, but also timely triggers the model optimization process when the pushing effect is not up to standard, thereby ensuring that the system always adapts to the change of user behavior mode.

[0033] Preferably, the S5 comprises the following steps:

[0034] S51, setting a user consultation conversion prediction interval corresponding to the day in S43, denoted as a current consultation conversion prediction interval; then collecting model training data volume data corresponding to the model daily update in S43, preset model training hyperparameter data, hardware performance index data and software performance index data used for training the model, to obtain daily model training data volume data, a daily model training hyperparameter data set, a daily hardware performance index data set and a daily software performance index data set.

[0035] S52, inputting the daily model training data volume data, the daily model training hyperparameter data set, the daily hardware performance index data set, and the daily software performance index data set into the final behavior sequence prediction training effect mapping model for mapping to obtain daily model training duration data and daily model test accuracy rate data;

[0036] S53, setting a daily model test accuracy rate threshold and a first maximum repeated adjustment number; if the daily model training duration data is greater than the current consultation conversion prediction interval or the daily model test accuracy rate data is less than the daily model test accuracy rate threshold, repeatedly adjusting the daily model training hyperparameter data set, otherwise, no adjustment is needed;

[0037] After each adjustment, a daily adjusted training hyperparameter data set is obtained, and the daily adjusted training hyperparameter data set, the daily model training data volume data, the daily hardware performance index data set, the daily software performance index data set, and the daily model training data volume data are input into the final behavior sequence prediction training effect mapping model for mapping to obtain first adjusted training duration data and first adjusted test accuracy rate data; if the repeated number is less than or equal to the first maximum repeated adjustment number, the first adjusted training duration data is less than or equal to the current consultation conversion prediction interval, and the first adjusted test accuracy rate data is greater than or equal to the daily model test accuracy rate threshold, the adjustment is completed, a daily final training hyperparameter data set is obtained, and the hyperparameters in the model daily updating process are set according to the daily final training hyperparameter data set; otherwise, S6 is executed;

[0038] The prediction model is used to map the training duration and accuracy in advance, effectively predicting the update feasibility; a double threshold judgment system is introduced, and the time efficiency and model quality are considered at the same time, and the hyperparameter optimization cycle is automatically triggered when any condition is not met; both the deployment risk of inefficient models is avoided and the business interruption caused by training timeout is prevented; through presetting the maximum adjustment number, controllable optimization is realized, the resource utilization rate is maximized on the premise of ensuring the model performance; secondly, the training hyperparameter data corresponding to the adjustment difficulty is the lowest, so the adjustment is performed preferentially, and if the adjustment does not meet the requirements, the adjustment is performed in order from small to large according to the adjustment difficulty, so that the time consumption and adjustment cost of the adjustment process can be saved to the greatest extent.

[0039] Preferably, S6 comprises the following steps:

[0040] S61, set the second maximum number of repeated adjustments; repeatedly adjust the daily software performance indicator dataset, obtain a daily adjusted software performance indicator dataset after each adjustment, and input the daily adjusted software performance indicator dataset, the daily adjusted training hyperparameter dataset, the daily hardware performance indicator dataset, and the daily model training data volume data into the final behavior sequence prediction training effect mapping model for mapping to obtain second adjusted training duration data and second adjusted test accuracy data;

[0041] If the number of repetitions is less than or equal to the second maximum number of repeated adjustments, the second adjusted training duration data is less than or equal to the current consultation conversion prediction interval, and the second adjusted test accuracy data is greater than or equal to the daily model test accuracy threshold, the adjustment is complete, and a daily final software performance indicator dataset is obtained. The software indicators and hyperparameters in the model daily updating process are set according to the daily final software performance indicator dataset and the daily adjusted training hyperparameter dataset; otherwise, S62 is performed;

[0042] S62, repeatedly adjust the daily hardware performance indicator dataset, obtain a daily adjusted hardware performance indicator dataset after each adjustment, and input the daily adjusted hardware performance indicator dataset, the daily adjusted training hyperparameter dataset, the daily adjusted software performance indicator dataset, and the daily model training data volume data into the final behavior sequence prediction training effect mapping model for mapping to obtain third adjusted training duration data and third adjusted test accuracy data; until the third adjusted training duration data is less than or equal to the current consultation conversion prediction interval and the third adjusted test accuracy data is greater than or equal to the daily model test accuracy threshold, a current final hardware performance indicator dataset is obtained;

[0043] The software indicators, hyperparameters, and hardware indicators in the model daily updating process are set according to the daily adjusted software performance indicator dataset, the daily adjusted training hyperparameter dataset, and the current final hardware performance indicator dataset;

[0044] A hierarchical and progressive adjustment strategy is adopted. First, the software performance indicators are finely tuned, and the optimal software configuration scheme is found through iterative adjustment. When software optimization cannot meet the requirements, the hardware resource adjustment stage is automatically entered, forming a complete resource adaptation closed loop. Thus, the potential of existing software and hardware resources can be fully utilized, and the training time can be significantly compressed under the premise of ensuring model prediction accuracy. The preset maximum number of adjustments is used to realize controllable optimization and avoid infinite loops.

[0045] A physical examination user consultation conversion prediction system based on a time sequence network comprises a historical user sequence time data acquisition module, a user behavior sequence data prediction model construction module, a behavior sequence prediction training effect mapping model construction module, a current day model update determination module, a current day model training hyperparameter update determination module, and a software and hardware performance index adjustment determination module.

[0046] The present application has the following advantages:

[0047] 1. In the present application, a closed-loop intelligent optimization system is constructed to achieve dynamic balance of user behavior prediction and resource scheduling. First, a prediction model is trained based on historical behavior data, and a precise effect evaluation mechanism is formed by comparing the push result with the actual user behavior in real time. When the push accuracy is lower than the threshold, a multi-dimensional optimization process is automatically triggered: first, the optimal configuration within the limited resources is found by iterative adjustment of hyperparameters, and if it cannot be met, the software performance index is further optimized, and finally the hardware configuration is upgraded if necessary. The hierarchical progressive optimization strategy not only ensures the timeliness of model updating, but also avoids resource waste. By including key parameters such as training data volume and software and hardware indicators in the effect prediction model, the optimization direction can be predicted in advance, significantly improving the adjustment efficiency; ultimately, it can continuously adapt to changes in user behavior, maximize resource utilization while ensuring push accuracy, and provide a sustainable solution for long-term stable operation of the intelligent recommendation system.

[0048] 2. In the present application, both time efficiency and model quality are considered, and when either condition is not met, the hyperparameter optimization cycle is automatically triggered; this avoids the risk of deploying inefficient models and prevents business interruptions caused by training timeouts; by presetting the maximum number of adjustments, controllable optimization is achieved, maximizing resource utilization while ensuring model performance, forming an intelligent decision-making closed loop that balances efficiency and effectiveness, significantly improving system stability and operational intelligence level; secondly, the training hyperparameter data has the lowest adjustment difficulty, so it is adjusted first, and if it does not meet the requirements after adjustment, it is adjusted in order of increasing difficulty, thereby saving time and cost in the adjustment process to the greatest extent.

[0049] 3. In the present application, the software performance index is first fine-tuned, and the optimal software configuration scheme is found through iterative adjustment; when software optimization cannot meet the requirements, the hardware resource adjustment stage is automatically entered, forming a complete resource adaptation closed loop; thereby making full use of the potential of existing software and hardware resources, significantly compressing the training time while ensuring the accuracy of model prediction.

[0050] Of course, any product implementing the present application does not necessarily need to achieve all the advantages described above at the same time. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed for the description of the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of these drawings.

[0052] Figure 1 A flowchart of a health examination user consultation conversion prediction method based on a time sequence network according to the present application is shown in the figure.

[0053] Figure 2 A flowchart of a final user behavior sequence data prediction model constructed according to the present application is shown in the figure.

[0054] Figure 3 A flowchart of a final behavior sequence prediction training effect mapping model constructed according to the present application is shown in the figure.

[0055] Figure 4 A flowchart of a daily model update determination according to the present application is shown in the figure.

[0056] Figure 5 A flowchart of a daily model training hyperparameter update determination according to the present application is shown in the figure.

[0057] Figure 6 A flowchart of a daily model training software and hardware performance index adjustment determination according to the present application is shown in the figure.

[0058] Figure 7 A module schematic diagram of a health examination user consultation conversion prediction system based on a time sequence network according to the present application is shown in the figure. DETAILED DESCRIPTION

[0059] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application. However, the present application can be practiced in a variety of ways beyond the specific details set forth herein without departing from the scope of the present application, and it is understood that the present application is intended to cover all modifications and variations of this application. Accordingly, the application is not limited to the specific embodiments described below.

[0060] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. The "in one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0061] Embodiment one

[0062] Please refer to Figure 1 The present embodiment is a health examination user consultation conversion prediction method based on a time sequence network, which comprises the following steps:

[0063] S1, acquire historical user behavior sequence data at multiple time points and corresponding time point data;

[0064] Referring to Figure 2 , the S1 includes the following steps:

[0065] S11, set several user behavior sequence types to obtain a user behavior sequence type set; the user behavior sequence type set includes report viewing duration, abnormal indicator clicking, coupon obtaining, and doctor card dwell time, etc.

[0066] Among them, the report viewing duration can capture the user's behavior of viewing the report through the front-end burying point technology, and the specific implementation includes: using Mutation Observer to listen to the DOM changes of the report page, recording the timestamps of the user entering and leaving the report page; the abnormal indicator clicking can be captured by deploying a click event listener in the abnormal indicator display area (such as the dashboard, alarm list); the coupon obtaining can be obtained by binding the obtaining event at the coupon obtaining button or entrance, and when triggered, it is reported to the back-end server; the doctor card dwell time can be monitored by the front-end Intersection Observer API whether the doctor card enters the user's visual area, and the start and end time of the dwell is recorded;

[0067] S12, according to the user behavior sequence type set, acquire historical user behavior sequence data at multiple time points and corresponding time point data, obtain historical user behavior sequence data set and historical user behavior sequence time point data set;

[0068] By constructing a user behavior sequence collection system, the fine-grained behavior trajectory of the user in the platform interaction process can be comprehensively captured, and the originally scattered user touch points can be converted into structured data assets with time sequence correlation; by using the Mutation Observer to monitor the page life cycle, the Intersection Observer to accurately monitor the visual area dwell, and the event listener to capture the active interaction behavior, etc. Multi-dimensional technical means, not only realize the quantitative analysis of key indicators such as report viewing depth, abnormal focus, marketing conversion node and content attraction, but also establish a complete behavior link portrait from user entry, browsing, interaction to exit; This fine-grained data collection method can effectively support user behavior pattern mining, personalized recommendation optimization, product interaction experience improvement and accurate evaluation of operation strategy, provide a high-quality data foundation for subsequent construction of user behavior prediction model and intelligent decision system, and finally realize product iteration and continuous optimization of user experience under the driving of data;

[0069] S2, based on the data collected in S1, construct a final user behavior sequence data prediction model;

[0070] The S2 comprises the following steps:

[0071] S21, constructing a prediction model for user behavior sequence data at a future time according to the historical user behavior sequence data set and the historical user behavior sequence time point data set, to obtain a final user behavior sequence data prediction model;

[0072] The final user behavior sequence data prediction model in S21 adopts a time series prediction model based on Transformer;

[0073] The S21 comprises the following steps:

[0074] S211, constructing an initial user behavior sequence data prediction model and setting a first training data ratio (such as 8:2 or 7:3, which can be adaptively adjusted according to actual training); dividing the historical user behavior sequence data set and the historical user behavior sequence time point data set according to the first training data ratio to obtain a first training data set and a first test data set;

[0075] S212, setting a first training error threshold (10%~15%, which can be adaptively adjusted according to actual training); inputting the first training data set into the initial user behavior sequence data prediction model for training; during the training process, if the training error is less than the first training error threshold, stopping the training to obtain a trained user behavior sequence data prediction model; otherwise, continuing the training until the training error is less than the first training error threshold;

[0076] S213, setting a first test accuracy threshold (90%~95%, which can be adaptively adjusted according to actual test); inputting the first test data set into the trained user behavior sequence data prediction model for testing; after the testing, obtaining first test accuracy data; if the first test accuracy data is greater than or equal to the first test accuracy threshold, taking the trained user behavior sequence data prediction model as the final user behavior sequence data prediction model; otherwise, returning to S212 to continue training the trained user behavior sequence data prediction model and repeating S213 until the first test accuracy data is greater than or equal to the first test accuracy threshold;

[0077] The structure of the initial user behavior sequence data prediction model can refer to Table 1 as follows:

[0078] Table 1: Structure of initial user behavior sequence data prediction model

[0079] Model Name Model Type Model Structure Initial User Behavior Sequence Data Prediction Model Transformer-based Time Series Prediction Model Encoder Structure: Input Embedding Layer: maps behavior sequence into 256-dimensional vectors, including position encoding and timestamp encoding; 8 layers of Transformer encoding blocks, each containing: multi-head self-attention mechanism (8 attention heads, attention dimension 64); feedforward neural network (2 fully connected layers, hidden layer dimension 1024, using GELU activation function) layer normalization and residual connection; Decoder Structure: 4 layers of Transformer decoding blocks, each containing: masked multi-head self-attention mechanism (to prevent information leakage); encoder-decoder attention layer: feedforward neural network (hidden layer dimension 512); Output Layer: fully connected layer (256 dimensions -> 128 dimensions -> number of behavior types); Softmax activation function outputs probability distribution for each behavior type;

[0080] The model effectively captures complex temporal dependencies in long sequences through self-attention mechanisms, can handle ultra-long sequences of more than 1000 behavior points compared to traditional methods, and can identify non-linear behavior patterns; the total number of model parameters is about 15 million, and the teacher forced strategy is used during training, the AdamW optimizer is used, and the cosine annealing mechanism is used for learning rate scheduling;

[0081] By building a prediction model based on historical user behavior sequence data, discrete user behavior data can be converted into a forward-looking intelligent decision support tool; by analyzing the behavior patterns of users at different time points and their evolution rules, the user interest migration path, behavior cycle characteristics and potential conversion motivation can be accurately captured, thereby realizing the probabilistic prediction of future user behavior; such prediction capability can not only help the platform to identify the risk of loss of high-value users in advance, predict the response rate of marketing activities, and optimize content recommendation strategies, but also provide data-driven decision-making basis for product function iteration; more importantly, the model can dynamically adapt to changes in behavior patterns, continuously optimize prediction accuracy through continuous learning, and ultimately form a complete closed loop from data collection, behavior analysis to prediction application, significantly improving user retention rate, conversion efficiency and personalized service level, and building a sustainable intelligent driving force for business growth;

[0082] S3, by collecting historical model training data volume, training hyperparameters, hardware and software performance indicators, training time and model test accuracy data during the model updating process to build a final behavior sequence prediction training effect mapping model;

[0083] Please refer to Figure 3 , the S3 includes the following steps:

[0084] S31, the final user behavior sequence data prediction model is used in the actual user resource conversion prediction process; the corresponding model training data volume data, model training hyperparameter data (including model training learning rate data, etc.), hardware performance indicator data (including CPU core number, memory size, GPU frequency, etc.), software performance indicator data (model framework version and CUDA compatibility, etc.), training time data and model test accuracy data during the model updating process of the historical actual user resource conversion prediction process are obtained, to obtain the historical model training data volume dataset, the historical model training hyperparameter dataset, the historical hardware performance indicator dataset, the historical software performance indicator dataset, the historical training time dataset and the historical model test accuracy dataset;

[0085] S32, constructing a mapping model between the model training data volume, the model training hyperparameter data, the hardware performance indicator data, the software performance indicator data, the training time length data, and the model test accuracy rate data according to the historical model training data volume data set, the historical model training hyperparameter data set, the historical hardware performance indicator data set, the historical software performance indicator data set, the historical training time length data set, and the historical model test accuracy rate data set, to obtain a final behavior sequence prediction training effect mapping model;

[0086] The final behavior sequence prediction training effect mapping model in S32 adopts a multilayer perception machine regression model.

[0087] The S32 includes the following steps:

[0088] S321, constructing an initial behavior sequence prediction training effect mapping model and setting a second training data proportion (such as 8:2 or 7:3, which can be adaptively adjusted according to the actual training situation); dividing the historical user behavior sequence data set and the historical user behavior sequence time point data set according to the second training data proportion to obtain a second training data set and a second test data set;

[0089] S322, setting a second training error threshold (10%~15%, which can be adaptively adjusted according to the actual training situation); inputting the second training data set into the initial behavior sequence prediction training effect mapping model for training; during the training process, if the training error is less than the second training error threshold, stopping the training to obtain a trained behavior sequence prediction training effect mapping model; otherwise, continuing the training until the training error is less than the second training error threshold;

[0090] S323, setting a second test accuracy rate threshold (90%~95%, which can be adaptively adjusted according to the actual test situation); inputting the second test data set into the trained behavior sequence prediction training effect mapping model for testing; after the testing is completed, obtaining second test accuracy rate data; if the second test accuracy rate data is greater than or equal to the second test accuracy rate threshold, taking the trained behavior sequence prediction training effect mapping model as the final behavior sequence prediction training effect mapping model; otherwise, returning to S322 to continue training the trained behavior sequence prediction training effect mapping model and repeating S323 until the second test accuracy rate data is greater than or equal to the second test accuracy rate threshold;

[0091] The structure of the initial behavior sequence prediction training effect mapping model can refer to Table Two shown in the following table:

[0092] Table Two: Structure of the initial behavior sequence prediction training effect mapping model

[0093] Model Name Model Type Model Structure Initial Behavior Sequence Prediction Training Effect Mapping Model Multi-layer Perceptron Regression Model Input Layer: Node Number: dynamically adjusted according to total number of features, including training data volume (1-dimensional), hyperparameters (such as learning rate, etc., n-dimensional), hardware indicators (CPU core number, memory size, GPU frequency, etc., m-dimensional), software indicators (framework version encoding, CUDA compatibility, etc., k-dimensional), total dimension 1+n+m+k; input features need to be standardized (Z-score normalization); Hidden Layer Design (3 fully connected layers): First Hidden Layer: 512 neurons, using ReLU activation function, adding Dropout regularization (ratio 0.3) to prevent overfitting; Second Hidden Layer: 256 neurons, using ReLU activation function, Dropout ratio 0.2; Third Hidden Layer: 128 neurons, using ReLU activation function; Output Layer: Node Number: 2 neurons (corresponding to training duration and test accuracy two prediction targets respectively); Activation Function: training duration output unit uses linear activation, test accuracy output unit uses Sigmoid activation (to constrain the predicted value to the [0, 1] interval) Parameter Configuration: Weight Initialization: uses He normal distribution initialization, suitable for ReLU activation function characteristics; Optimizer: Adam optimizer, initial learning rate set to 0.001, supports dynamic adjustment; Loss Function: weighted combination of mean square error and binary cross-entropy (weight ratio 6:4); Batch Size: 32 samples per batch; Training Epochs: 200 epochs, combined with early stopping method (patience=15) to prevent overfitting;

[0094] The model structure can accurately capture the complex mapping relationship between hardware resource configuration, software environment, data size and model performance through deep nonlinear transformation, providing quantitative basis for resource allocation and training efficiency optimization;

[0095] By constructing a multi-dimensional historical training data mapping model, the complex relationship between model training effect and resource consumption can be systematically quantitatively analyzed and predicted. Specifically, by integrating model training data volume, hyperparameter configuration, hardware performance indicators (such as CPU core number, memory size and GPU frequency), software environment (such as model framework version and CUDA compatibility) and other multi-dimensional historical data, a comprehensive training effect mapping model is constructed, which can accurately reveal the internal correlation between training time and model accuracy under different resource configurations, thereby providing data support for resource allocation optimization - for example, dynamically adjusting hyperparameters according to real-time hardware status to balance training efficiency and model performance, or quickly predicting the optimal training configuration in a new environment. In addition, this scheme also has the ability of forward-looking prediction, which can evaluate the training cost and expected accuracy of different hardware combinations before deployment, effectively avoiding resource waste; at the same time, by continuously monitoring the compatibility influence of software framework and hardware, it can timely identify the performance risks brought by environment upgrade, ensuring the stability and repeatability of training process; finally, this data-driven decision mechanism significantly improves the model iteration efficiency, reduces the experimental cost, and enhances the intelligent level of the entire machine learning operation process;

[0096] S4, input the historical behavior sequence data of the user to be consulted into the mapping model in S2, and select the behavior sequence with the highest probability for message pushing; then calculate the pushing accuracy, and if the accuracy is lower than the threshold, update the model daily;

[0097] Please refer to Figure 4 , the S4 includes the following steps:

[0098] S41, select a user to be consulted; according to the user behavior sequence type set, obtain the user behavior sequence data of the user to be consulted at multiple time points in history, obtain the user behavior sequence data set of the user to be consulted and the historical time point set of the user to be consulted;

[0099] S42, input the user behavior sequence data set of the user to be consulted and the historical time point set of the user to be consulted into the final user behavior sequence data prediction model for mapping, obtain the current user behavior sequence probability data set of the user to be consulted; select the user behavior sequence corresponding to the maximum probability data in the current user behavior sequence probability data set of the user to be consulted for pushing; after the day of pushing, calculate the probability data of the pushed message and the actual behavior of the user, and obtain the current message pushing accuracy data;

[0100] S43, set the current message push accuracy rate threshold (can be adaptively set according to actual push requirements); if the current message push accuracy rate data is greater than or equal to the current message push accuracy rate threshold, the final behavior sequence prediction training effect mapping model does not need to be updated; otherwise, S5 is executed;

[0101] By deeply mining user historical behavior sequence data, a user interest portrait can be accurately constructed, the most potential behavior path is identified based on a probability mapping model for targeted push, and the matching precision of push content and user real needs is significantly improved. Secondly, by introducing a real-time feedback mechanism, the coincidence degree of actual behavior and predicted behavior is automatically calculated after each push, forming a quantitative accuracy rate evaluation index to provide data support for subsequent optimization. In addition, the system designs an intelligent model update determination mechanism, which dynamically determines the model update demand through an adaptive threshold, avoiding resource waste caused by frequent updates, and triggering the model optimization process in time when the push effect is not up to standard, ensuring that the system always adapts to changes in user behavior patterns. Thus, the entire push system has the ability of self-learning and continuous evolution, greatly reducing the operation and maintenance cost while ensuring the push accuracy, and finally forming a complete intelligent operation closed loop from data collection, behavior prediction, accurate push to effect evaluation and model iteration, realizing the dual improvement of push efficiency and user experience.

[0102] S5, collect the training data volume, hyperparameters and hardware and software performance index data corresponding to the day and input them into the mapping model in S3; if the mapping result does not meet the requirements, the hyperparameters are adjusted repeatedly and mapped again, if the training time and accuracy requirements are met within a limited number of times, the configuration is completed, and the adjustment is completed; otherwise, S6 is executed.

[0103] Please refer to Figure 5 , the S5 includes the following steps:

[0104] S51, set the user consultation conversion prediction interval corresponding to the day in S43, denoted as the current consultation conversion prediction interval (i.e. the maximum idle time left for model daily update); then collect the model training data volume data, pre-set model training hyperparameter data, hardware performance index data and software performance index data used for model training corresponding to the model daily update on the day in S43, to obtain the model training data volume data on the day, the model training hyperparameter data set on the day, the hardware performance index data set on the day and the software performance index data set on the day;

[0105] S52, input the daily model training data volume data, daily model training hyperparameter dataset, daily hardware performance index dataset and daily software performance index dataset into the final behavior sequence prediction training effect mapping model for mapping to obtain daily model training time length data and daily model test accuracy rate data;

[0106] S53, set a daily model test accuracy rate threshold and a first maximum repeated adjustment number (which can be adaptively set according to actual model update requirements); if the daily model training time length data is greater than the current consultation conversion prediction interval or the daily model test accuracy rate data is less than the daily model test accuracy rate threshold, repeatedly adjust the daily model training hyperparameter dataset, otherwise, no adjustment is needed;

[0107] After each adjustment, obtain the daily adjusted training hyperparameter dataset and input the daily adjusted training hyperparameter dataset, daily model training data volume data, daily hardware performance index dataset, daily software performance index dataset and daily model training data volume data into the final behavior sequence prediction training effect mapping model for mapping to obtain first adjusted training time length data and first adjusted test accuracy rate data; if the number of repetitions is less than or equal to the first maximum repeated adjustment number, the first adjusted training time length data is less than or equal to the current consultation conversion prediction interval, and the first adjusted test accuracy rate data is greater than or equal to the daily model test accuracy rate threshold, the adjustment is completed, the daily final training hyperparameter dataset is obtained, and the hyperparameters in the model daily update process are set according to the daily final training hyperparameter dataset; otherwise, S6 is executed;

[0108] By constructing an intelligent model update decision mechanism, the precise balance of resource optimization and model performance is realized; first, based on the business demand, the time interval is set to ensure that the model update does not affect the normal business operation; by comprehensively collecting multi-dimensional data such as training data volume, hyperparameter configuration and software and hardware performance indicators, a complete training environment portrait is established; the prediction model is used to map the training time length and accuracy rate in advance to effectively predict the update feasibility; by introducing a double threshold judgment system, the time efficiency and model quality are considered at the same time, and the hyperparameter optimization cycle is automatically triggered when any condition is not met; both the deployment risk of inefficient models and the business interruption caused by training overtime are avoided; by presetting the maximum number of adjustments, controllable optimization is realized, the resource utilization rate is maximized on the premise of ensuring the model performance, an intelligent decision-making closed loop considering efficiency and effect is formed, and the system stability and operation intelligent level are significantly improved; secondly, the training hyperparameter data has the lowest adjustment difficulty, so it is adjusted first, and if the adjustment does not meet the requirements, the adjustment is performed in order from small to large according to the adjustment difficulty, so that the time consumption and adjustment cost of the adjustment process can be saved to the greatest extent;

[0109] S6, repeatedly adjusting the software performance index in S5, if the training length and accuracy requirements are met within a limited number of times, the configuration is completed; otherwise, further repeatedly adjusting the hardware performance index until the requirements are met;

[0110] Please refer to Figure 6 , the S6 comprises the following steps:

[0111] S61, set the second maximum number of repeated adjustments; repeatedly adjust the daily software performance index dataset, obtain a daily adjusted software performance index dataset after each adjustment, and input it into the final behavior sequence prediction training effect mapping model together with the daily adjusted training hyperparameter dataset, the daily hardware performance index dataset, and the daily model training data volume data to obtain second adjusted training length data and second adjusted test accuracy data;

[0112] If the number of repetitions is less than or equal to the second maximum number of repeated adjustments, and the second adjusted training length data is less than or equal to the current consultation conversion prediction interval, and the second adjusted test accuracy data is greater than or equal to the daily model test accuracy threshold, the adjustment is completed, and the daily final software performance index dataset is obtained. Set the software index and hyperparameter during the model day update process according to the daily final software performance index dataset and the daily adjusted training hyperparameter dataset; otherwise, perform S62;

[0113] S62, repeatedly adjust the daily hardware performance index dataset, obtain a daily adjusted hardware performance index dataset after each adjustment, and input it into the final behavior sequence prediction training effect mapping model together with the daily adjusted training hyperparameter dataset, the daily adjusted software performance index dataset, and the daily model training data volume data to obtain third adjusted training length data and third adjusted test accuracy data; until the third adjusted training length data is less than or equal to the current consultation conversion prediction interval and the third adjusted test accuracy data is greater than or equal to the daily model test accuracy threshold, the current final hardware performance index dataset is obtained;

[0114] Set the software index, hyperparameter, and hardware index during the model day update process according to the daily adjusted software performance index dataset, the daily adjusted training hyperparameter dataset, and the current final hardware performance index dataset;

[0115] For example, in the process of optimizing the user consultation conversion prediction model, as follows:

[0116] Initial data preparation:

[0117] User to consult: Randomly select user ID U20231101 as the target user for the day; Historical behavior data: User's past 30-day behavior sequence: View Report A (3 times) → Coupon Collection (2 times) → Abnormal Indicator Click (1 time) → View Report B (5 times) → Doctor Card Stay Time (0 times); Corresponding time points: 2023-10-01 to 2023-10-30;

[0118] First push and evaluation:

[0119] Behavior prediction: Model calculates the probability of each behavior sequence: Sequence 1: View Report A → Coupon Collection → Abnormal Indicator Click (probability 68%); Sequence 2: View Report B → Coupon Collection → View Report B (probability 22%); Other sequence probabilities are less than 10%; Push decision: Select the highest probability sequence 1 for push; Actual feedback: User's actual behavior is to view Report A → Coupon Collection → Abnormal Indicator Click; Accuracy calculation: Push accuracy = actual matching behavior number / total push behavior number = 3 / 3 = 100%;

[0120] Model update determination:

[0121] Threshold setting: The accuracy threshold for the day is 85%; Decision result: 100%>85%, so no model update is needed for the day;

[0122] Hypothetical model update process (when the accuracy is lower than the threshold):

[0123] Resource evaluation: Consultation conversion prediction interval: 2 hours, training data volume: 500,000 user behavior records, initial hyperparameter values: learning rate = 0.001, hidden layer node number = 128, hardware configuration: CPU core number = 8, memory = 32GB, software environment: Python3.8, TensorFlow2.4;

[0124] Hyperparameter optimization: Initial training duration: 1.5 hours (<2 hour threshold); Initial test accuracy: 82% (<85% threshold); Adjusted hyperparameters: learning rate = 0.0005, hidden layer node number = 256; Adjusted training duration: 1.8 hours (<2 hours); Adjusted test accuracy: 86% (>85%); Final hyperparameter configuration: learning rate = 0.0005, hidden layer node number = 256;

[0125] Software optimization: Initial training duration: 1.8 hours (<2 hours); Initial test accuracy: 86% (>85%); Final software configuration: Enable CUDA acceleration, optimize memory allocation strategy;

[0126] Hardware optimization: initial training duration: 1.8 hours (<2 hours); initial test accuracy: 86% (>85%); final hardware configuration: maintain existing configuration, no adjustment needed;

[0127] Final configuration:

[0128] Hyperparameters: learning rate = 0.0005, number of hidden layer nodes = 256; software configuration: enable CUDA acceleration, optimize memory allocation; hardware configuration: CPU core number = 8, memory = 32GB;

[0129] By constructing a systematic resource configuration optimization mechanism, the dynamic balance between model training efficiency and quality is achieved; a hierarchical progressive adjustment strategy is adopted, first the software performance indicators are fine-tuned, and the optimal software configuration scheme is found through iterative adjustment; when software optimization cannot meet the requirements, it automatically enters the hardware resource adjustment stage, forming a complete resource adaptation closed loop; thereby making full use of the potential of existing software and hardware resources, significantly compressing the training time while ensuring the model prediction accuracy; among them, through the pre-set maximum number of adjustments, controllable optimization is achieved, avoiding infinite loops; finally, through multi-dimensional parameter linkage configuration, the best practice scheme of software and hardware collaborative optimization is established, which not only improves the model update efficiency, but also ensures the service quality, while reducing resource waste, providing a sustainable resource configuration solution for long-term stable operation of the system.

[0130] Embodiment two

[0131] Please refer to Figure 7 The embodiment discloses a physical examination user consultation conversion prediction system based on a time sequence network, which can realize the method of the above-mentioned embodiment, comprising a historical user sequence time data acquisition module, a user behavior sequence data prediction model construction module, a behavior sequence prediction training effect mapping model construction module, a daily model update determination module, a daily model training hyperparameter update determination module, and a software and hardware performance index adjustment determination module.

[0132] The historical user sequence time data acquisition module acquires the behavior sequence data of the user at multiple time points and the corresponding time point data;

[0133] The user behavior sequence data prediction model construction module constructs a final user behavior sequence data prediction model based on the data collected by the historical user sequence time data acquisition module;

[0134] The behavior sequence prediction training effect mapping model construction module acquires historical model training data volume, training hyperparameters, hardware and software performance indicators, training duration, and model test accuracy data during model updating to construct a final behavior sequence prediction training effect mapping model;

[0135] The daily model update determination module inputs the historical behavior sequence data of the user to be consulted into the user behavior sequence data prediction model construction module, maps in the mapping model, selects the behavior sequence with the highest probability according to the mapping result, and pushes the message; the prediction behavior is compared with the actual user behavior to calculate the push accuracy, and if the accuracy is lower than the threshold, the model is updated daily;

[0136] The daily model training hyperparameter update determination module collects the corresponding training data amount, hyperparameters and software and hardware performance index data of the day and inputs them into the mapping model in the behavior sequence prediction training effect mapping model construction module; if the mapping result does not meet the requirements, the hyperparameters are repeatedly adjusted and the updated hyperparameters are input into the mapping model in the behavior sequence prediction training effect mapping model construction module, if the hyperparameters that meet the time interval or accuracy threshold at the same time are found within the limited number of repetitions, the adjustment is completed; otherwise, the software and hardware performance index adjustment determination module is executed;

[0137] The software and hardware performance index adjustment determination module repeatedly adjusts the software performance index in the daily model training hyperparameter update determination module, and if the training time and accuracy requirements are met within the limited number of times, the configuration is completed; otherwise, the hardware performance index is further repeatedly adjusted until the time and accuracy requirements are met at the same time.

[0138] It should be noted that for the foregoing method embodiments, in order to simply describe, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited by the action sequence described, because according to the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily required by the present application.

Claims

1. A method for predicting the conversion of health checkup user consultations based on time-series networks, characterized in that, Includes the following steps: S1. Obtain user behavior sequence data and corresponding time point data from multiple historical time points; S2. Construct a prediction model for the final user behavior sequence data based on the data collected in S1; Specifically, it includes: S21. Based on the historical user behavior sequence dataset and the historical user behavior sequence time point dataset, construct a prediction model for user behavior sequence data at future moments to obtain the final user behavior sequence data prediction model. S3. By collecting several sets of historical data on the daily update process of the model, including the amount of training data, training hyperparameters, hardware and software performance indicators, training time, and model test accuracy, a final behavior sequence prediction training effect mapping model is constructed. S4. Input the historical behavior sequence data of the user to be consulted into the mapping model in S2 and select the behavior sequence with the highest probability for message push; then calculate the push accuracy. If the accuracy is lower than the threshold, the model is updated daily. S5. Collect the training data volume, hyperparameters, and hardware / software performance metrics for the day and input them into the mapping model in S3. If the mapping result does not meet the requirements, repeatedly adjust the hyperparameters and remap. If the training duration and accuracy requirements are met within the limited number of attempts, the configuration and adjustment are complete. Otherwise, proceed to S6. S6. Repeatedly adjust the software performance indicators in S5. If the training time and accuracy requirements are met within a limited number of times, the configuration is complete. Find the optimal software configuration scheme through iterative adjustment. Otherwise, continue to repeatedly adjust the hardware performance indicators until the requirements are met.

2. The method for predicting user consultation conversion based on time-series networks according to claim 1, characterized in that, S1 includes the following steps: S11. Define several types of user behavior sequence to obtain a set of user behavior sequence types; S12. Based on the user behavior sequence type set, obtain user behavior sequence data and corresponding time point data at multiple historical time points to obtain historical user behavior sequence dataset and historical user behavior sequence time point dataset.

3. The method for predicting user consultation conversion based on time-series networks according to claim 1, characterized in that: The end-user behavior sequence data prediction model described in S21 adopts a Transformer-based time series prediction model.

4. The method for predicting user consultation conversion based on time-series networks according to claim 3, characterized in that, S3 includes the following steps: S31. The final user behavior sequence data prediction model is used in the actual user resource conversion prediction process; several sets of model training data, model training hyperparameter data, hardware performance index data, software performance index data, training duration data and model test accuracy data corresponding to the daily updates of the model in the historical actual user resource conversion prediction process are obtained to obtain historical model training data dataset, historical model training hyperparameter dataset, historical hardware performance index dataset, historical software performance index dataset, historical training duration dataset and historical model test accuracy dataset. S32. Based on the historical model training data volume dataset, historical model training hyperparameter dataset, historical hardware performance index dataset, historical software performance index dataset, historical training duration dataset, and historical model test accuracy dataset, construct a mapping model between the model training data volume, model training hyperparameter data, hardware performance index data, software performance index data, training duration data, and model test accuracy data to obtain the final behavior sequence prediction training effect mapping model.

5. The method for predicting user consultation conversion based on time-series networks according to claim 4, characterized in that: The final behavior sequence prediction training effect mapping model described in S32 adopts a multilayer perceptron regression model.

6. The method for predicting user consultation conversion based on time-series networks according to claim 5, characterized in that, S4 includes the following steps: S41. Select users to be consulted; obtain user behavior sequence data of users to be consulted corresponding to multiple time points in history according to the user behavior sequence type set, and obtain the historical behavior sequence dataset of users to be consulted and the historical time point set of users to be consulted. S42. Input the historical behavior sequence dataset of the user to be consulted and the historical time point set of the user to be consulted into the final user behavior sequence data prediction model for mapping, to obtain the current user behavior sequence probability dataset, and select the user behavior sequence corresponding to the highest probability data for push; after the push ends on the same day, calculate the probability data that the pushed message is the same as the user's actual behavior, and obtain the current message push accuracy data. S43. Set the current message push accuracy threshold; if the current message push accuracy data is greater than or equal to the current message push accuracy threshold, there is no need to update the final behavior sequence prediction training effect mapping model on that day; otherwise, execute S5.

7. The method for predicting user consultation conversion based on time-series networks according to claim 6, characterized in that, S5 includes the following steps: S51. Set the user consultation conversion prediction interval corresponding to the day mentioned in S43, and record it as the current consultation conversion prediction interval; then collect the model training data volume data, preset model training hyperparameter data, hardware performance index data and software performance index data corresponding to the daily model update mentioned in S43, to obtain the model training data volume data, model training hyperparameter dataset, hardware performance index dataset and software performance index dataset for the day. S52. Input the daily model training data volume, daily model training hyperparameter dataset, daily hardware performance index dataset, and daily software performance index dataset into the final behavior sequence prediction training effect mapping model for mapping, and obtain the daily model training duration data and the daily model test accuracy data. S53. Set the daily model test accuracy threshold and the first maximum number of repeated adjustments; if the daily model training duration data is greater than the current consultation conversion prediction interval or the daily model test accuracy data is less than the daily model test accuracy threshold, repeatedly adjust the daily model training hyperparameter dataset; otherwise, no adjustment is required. After each adjustment, the adjusted training hyperparameter dataset for that day is obtained and combined with the model training data volume, hardware performance index dataset, software performance index dataset, and model training data volume for that day. These are then input into the final behavior sequence prediction training effect mapping model for mapping, resulting in the first adjusted training duration data and the first adjusted test accuracy data. If the number of repetitions is less than or equal to the first maximum number of repetitions, the first adjusted training duration data is less than or equal to the current consultation conversion prediction interval, and the first adjusted test accuracy data is greater than or equal to the model test accuracy threshold for that day, the adjustment is complete, the final training hyperparameter dataset for that day is obtained, and the hyperparameters are set during the daily model update process; otherwise, proceed to step S6.

8. The method for predicting user consultation conversion based on time-series networks according to claim 7, characterized in that, S6 includes the following steps: S61. Set a second maximum number of repeated adjustments; repeatedly adjust the software performance index dataset for the day, and after each adjustment, obtain the adjusted software performance index dataset for the day and combine it with the adjusted training hyperparameter dataset for the day, the hardware performance index dataset for the day and the model training data volume for the day, and input it into the final behavior sequence prediction training effect mapping model for mapping to obtain the second adjusted training duration data and the second adjusted test accuracy data. If the number of repetitions is less than or equal to the second maximum number of repetitions, the second adjusted training duration is less than or equal to the current consultation conversion prediction interval, and the second adjusted test accuracy is greater than or equal to the daily model test accuracy threshold, the adjustment is complete, and the daily final software performance index dataset is obtained. Based on the daily final software performance index dataset and the daily adjusted training hyperparameter dataset, the software indicators and hyperparameters in the daily model update process are set; otherwise, proceed to step S62. S62. Repeatedly adjust the hardware performance index dataset for the day. After each adjustment, obtain the adjusted hardware performance index dataset for the day and combine it with the adjusted training hyperparameter dataset for the day, the adjusted software performance index dataset for the day, and the model training data volume for the day. Input the data into the final behavior sequence prediction training effect mapping model for mapping to obtain the third adjusted training duration data and the third adjusted test accuracy data. Continue until the third adjusted training duration data is less than or equal to the current consultation conversion prediction interval and the third adjusted test accuracy data is greater than or equal to the model test accuracy threshold for the day, and obtain the current final hardware performance index dataset. The software metrics, hyperparameters, and hardware metrics are set based on the daily adjusted software performance metric dataset, the daily adjusted training hyperparameter dataset, and the current final hardware performance metric dataset during the daily model update process.

9. A system for implementing the time-series network-based method for predicting the conversion of physical examination user consultations as described in any one of claims 1-8, characterized in that: It includes a historical user sequence time data acquisition module, a user behavior sequence data prediction model construction module, a behavior sequence prediction training effect mapping model construction module, a daily model update determination module, a daily model training hyperparameter update determination module, and a software and hardware performance index adjustment determination module. The historical user sequence time data acquisition module acquires user behavior sequence data and corresponding time point data from multiple historical time points. The user behavior sequence data prediction model building module constructs the final user behavior sequence data prediction model based on the data collected by the historical user sequence time data acquisition module. The behavior sequence prediction training effect mapping model construction module collects several sets of historical data on the model training data volume, training hyperparameters, hardware and software performance indicators, training time and model test accuracy during the daily model update process to construct the final behavior sequence prediction training effect mapping model. The daily model update determination module inputs the historical behavior sequence data of the user to be consulted into the mapping model of the user behavior sequence data prediction model construction module, and selects the behavior sequence with the highest probability for message push based on the mapping result; then compares the predicted behavior with the actual user behavior to calculate the push accuracy. If the accuracy is lower than the threshold, the model is updated daily. The daily model training hyperparameter update determination module collects the corresponding training data volume, hyperparameters, and hardware / software performance index data for the day and inputs them into the behavior sequence prediction training effect mapping model construction module for mapping. If the mapping result does not meet the requirements, the hyperparameters are repeatedly adjusted and the updated hyperparameters are input into the behavior sequence prediction training effect mapping model construction module for mapping. If a hyperparameter that simultaneously meets the time interval or accuracy threshold is found within a limited number of repetitions, the adjustment is complete; otherwise, the hardware / software performance index adjustment determination module is executed. The software and hardware performance indicator adjustment and judgment module repeatedly adjusts the software performance indicators in the model training hyperparameter update and judgment module for the day. If the training time and accuracy requirements are met within a limited number of times, the configuration is completed; otherwise, the hardware performance indicators are further repeatedly adjusted until both time and accuracy requirements are met.

Images