Business data prediction method, system and device and storage medium
By combining large language models and prediction models, and utilizing dual-channel networks and Bayesian inference weighted fusion technology, the challenges of dynamic changes and multi-factor relationships in freight business forecasting are solved, achieving high-precision business data forecasting and adapting to complex and ever-changing freight scenarios.
Patent Information
- Application Number
- CN202511090144.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-05
- Publication Date
- 2025-11-14
AI Technical Summary
Existing technologies are ill-suited for adapting to dynamic changes in freight forecasting, have limited ability to characterize complex relationships among multiple factors, and are sensitive to new data patterns and outliers, exhibiting poor stability.
By combining large language models and prediction models, business transaction data is obtained through offline and online preprocessing, features are extracted and updated in real time, global and local features are extracted using a dual-channel network, and weighted fusion is performed by combining Bayesian inference and attention networks to adjust feature weights to improve prediction accuracy.
It improves the accuracy of business data prediction, enables timely response to minute-level business changes and supports long-term planning, and adapts to complex and ever-changing freight scenarios.
Smart Images
Figure CN120952855A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of business data processing, and in particular to business data prediction methods, systems, devices, and storage media. Background Technology
[0002] Currently, existing technologies in the market have many shortcomings in freight business forecasting and urgently need improvement. Traditional statistical models rely on historical data to construct simple linear or nonlinear models, which have limited ability to characterize the complex relationships between multiple factors in freight scenarios and are difficult to adapt to dynamic changes. At the same time, these models are sensitive to new data patterns and outliers, and have poor stability. Therefore, how to improve the predictive ability for business data and improve the accuracy of business data processing is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0003] The purpose of this application is to provide a business data prediction method, system, computer-readable storage medium, and electronic device that can improve the prediction accuracy of business data.
[0004] To address the aforementioned technical problems, this application provides a business data prediction method, the specific technical solution of which is as follows:
[0005] Acquire business transaction data; the business transaction data includes offline data and online real-time data;
[0006] Business-related features are extracted from the online real-time data, and the business-related features are updated in real time by combining them with the offline data to obtain business time-series data;
[0007] During the offline prediction process, the business time series data and its semantic information are collected. The business time series data and the semantic data are respectively input into the large language model and the prediction model to obtain the semantic features output by the large language model and the local time series features output by the prediction model.
[0008] During the online prediction process, the business time series data and its semantic information are input into a dual-channel network, and the global impact features are output.
[0009] The feature weights of the semantic features, the local temporal features, and the global impact features are adjusted based on the historical prediction accuracy to obtain the business prediction results.
[0010] Optionally, after obtaining the business transaction data, the following may also be included:
[0011] The business transaction data is subjected to offline preprocessing and online preprocessing; the offline preprocessing includes data cleaning and feature normalization; the online preprocessing includes data standardization and noise removal.
[0012] Optionally, after obtaining the semantic features output by the large language model and the local temporal features output by the prediction model, the method further includes:
[0013] The semantic features and the local temporal features are weighted using Bayesian inference to obtain the updated weights;
[0014] The semantic features and the local temporal features are weighted and fused using the updated weights.
[0015] Optionally, Bayesian inference is applied to the semantic features and the local temporal features to obtain updated weights, including:
[0016] Obtain the first output mean and first standard deviation of the large language model, and the second output mean and second standard deviation of the prediction model;
[0017] The semantic features and the local temporal features are standardized based on the first output mean, the first standard deviation, the second output mean, and the second standard deviation.
[0018] The prior weights of the large language model and the prediction model are set according to the historical performance of the models, respectively.
[0019] An attention network is constructed by taking the standardized outputs of the large language model and the prediction model as inputs.
[0020] A multilayer perceptron is set as the attention network; the attention network is used to output the attention weights of the large language model and the prediction model respectively.
[0021] The attention weights are used as likelihoods, and the posterior weights are calculated using Bayes' theorem.
[0022] The updated weights are obtained by weighting and fusing the outputs of the large language model and the prediction model based on the posterior weights.
[0023] Optionally, the prior weights of the large language model and the prediction model can be set according to their historical performance, including:
[0024] Calculate the first prediction error index of the large language model and the second prediction error index of the prediction model;
[0025] The first prior weights of the large language model are determined based on the first prediction error index.
[0026] The second prior weight of the prediction model is determined based on the second prediction error index.
[0027] Optionally, the business time-series data and its semantic information are input into a dual-channel network, and the output global impact features include:
[0028] The business time series data and its semantic information are input into a dual-channel network, and the dual-channel network is used to extract global dependency features and local real-time features respectively.
[0029] The global dependency features and the local real-time features are weighted and fused to output the global impact features.
[0030] Optionally, after adjusting the feature weights of the semantic features, the local temporal features, and the global influence features based on historical prediction accuracy to obtain the business prediction result, the method further includes:
[0031] The business forecast results are used to predict business indicators, and business verification indicators are used to calculate the accuracy of the forecast results.
[0032] If the deviation of the prediction result is greater than a set threshold, an adjustment strategy is invoked to update at least one of the large language model, the prediction model, and the dual-channel network.
[0033] This application also provides a business data prediction system, including:
[0034] The data acquisition module is used to acquire business transaction data; the business transaction data includes offline data and online real-time data.
[0035] The feature extraction module is used to extract business-related features from the online real-time data, and update the business-related features in real time by combining them with the offline data to obtain business time-series data;
[0036] The offline prediction module is used to collect the business time series data and its semantic information during the offline prediction process, and input the business time series data and the semantic data into the large language model and the prediction model respectively to obtain the semantic features output by the large language model and the local time series features output by the prediction model.
[0037] The online prediction module is used to input the business time series data and its semantic information into a dual-channel network during the online prediction process, and output the global impact features.
[0038] The business prediction module is used to adjust the feature weights of the semantic features, the local temporal features, and the global impact features based on the historical prediction accuracy to obtain the business prediction results.
[0039] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.
[0040] This application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method described above when it invokes the computer program in the memory.
[0041] This application provides a business data prediction method, comprising: acquiring business transaction data; the business transaction data including offline data and online real-time data; extracting business-related features from the online real-time data, and updating the business-related features in real time in conjunction with the offline data to obtain business time-series data; during offline prediction, collecting the business time-series data and its semantic information, and inputting the business time-series data and the semantic data into a large language model and a prediction model respectively to obtain semantic features output by the large language model and local temporal features output by the prediction model; during online prediction, inputting the business time-series data and its semantic information into a dual-channel network to output global impact features; adjusting the feature weights of the semantic features, the local temporal features, and the global impact features according to historical prediction accuracy to obtain business prediction results.
[0042] This application combines a large language model with traditional forecasting models when performing business forecasting. This improves the accuracy of extracting data from unstructured business transaction data while retaining the advantages of processing structured business transaction data, thus enhancing the accuracy of business forecasting. Furthermore, by combining offline and online forecasting, it can meet the forecasting needs of minute-level business operations, responding promptly to business changes, and also enabling long-term offline forecasting in advance, supporting long-term business planning.
[0043] This application also provides a business data prediction system, a computer-readable storage medium, and an electronic device, which have the above-mentioned beneficial effects, and will not be elaborated here. Attached Figure Description
[0044] To more clearly illustrate the technical solutions in the embodiments of this application 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 only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0045] Figure 1 A flowchart illustrating a business data prediction method provided in this application embodiment;
[0046] Figure 2 This is a schematic diagram of the structure of a business data prediction system provided in an embodiment of this application;
[0047] Figure 3 This is a structural diagram of an electronic device provided in an embodiment of this application. Detailed Implementation
[0048] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0049] The object information involved in this application, including but not limited to object device information, object personal information, and data, including but not limited to data used for analysis, stored data, and displayed data, are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of related data shall comply with the laws, regulations and standards of relevant countries and regions.
[0050] See Figure 1 , Figure 1 A flowchart of a business data prediction method provided in this application embodiment, the method including:
[0051] S101: Obtain business transaction data; the business transaction data includes offline data and online real-time data;
[0052] S102: Extract business-related features from the online real-time data, and update the business-related features in real time by combining them with the offline data to obtain business time-series data;
[0053] S103: During the offline prediction process, the business time series data and its semantic information are collected, and the business time series data and the semantic data are respectively input into the large language model and the prediction model to obtain the semantic features output by the large language model and the local time series features output by the prediction model.
[0054] S104: During the online prediction process, the business time series data and its semantic information are input into a dual-channel network, and the global impact features are output.
[0055] S105: Adjust the feature weights of the semantic features, the local temporal features, and the global influence features based on the historical prediction accuracy to obtain the business prediction results.
[0056] There are no restrictions on the type of business transaction data required. Data can be obtained from, but is not limited to, the business MySQL database directly connected to the transaction system. Enterprise data can be collected in real time using Flink, including key business metrics such as order quantity, paired order amount, and transaction value (GMV) for each province / region / city. Furthermore, there are no restrictions on the period or frequency of data acquisition.
[0057] To improve the efficiency and accuracy of business data prediction, after acquiring business transaction data, in one feasible implementation, offline preprocessing and online preprocessing can be performed on the business transaction data; the offline preprocessing includes data cleaning and feature normalization, while the online preprocessing includes data standardization and noise removal.
[0058] Data cleaning includes handling missing values, duplicate data, and outliers, and may also include data consistency checks and data type conversions. If missing values exist in the data, and the missing percentage is low, these records can be deleted directly. For example, in e-commerce transaction data, if a transaction record lacks payment method information, and such records constitute a small percentage, they can be removed from the dataset. For numerical data, the mean, median, or mode can be used to fill in the missing values. For example, for the transaction amount field, if there are a few missing values, the average of all transaction amounts can be calculated, and this average can be used to fill in the missing parts.
[0059] For categorized data, the mode can be used for filling missing values. For example, in a product category field, if there are many missing values, you can find the product category that appears most frequently (the mode) and use it to fill the missing values.
[0060] It can also perform deduplication on data. For example, it can identify duplicate records based on a combination of keys such as the unique identifier of the transaction record (e.g., transaction ID), transaction time, transaction amount, and information of both parties, and delete duplicate records, keeping only one valid transaction record.
[0061] When dealing with outliers, statistical methods or clustering methods can be used to detect them.
[0062] Data consistency checks are used to ensure that the format and type of data fields are consistent. For example, for a transaction time field, ensure that all time formats are uniform (e.g., "YYYY-MM-DD HH:MM:SS" format). Data consistency across fields can also be checked. For example, for a transaction, the transaction amount should match the product of the quantity and unit price of the goods; if inconsistencies exist, further verification of the data's accuracy is required.
[0063] In online preprocessing, the focus is on ensuring the quality of real-time data, including data standardization and noise reduction. In online environments, because the data is generated in real time, it's impossible to obtain the mean and standard deviation of the entire dataset in advance. Online algorithms can be used to calculate the mean and standard deviation in real time and standardize new data. Alternatively, a sliding window approach can be used, considering only the most recent N data points to calculate the mean and standard deviation for standardizing new data. This is suitable for scenarios requiring rapid adaptation to changes in data distribution.
[0064] When eliminating noise, methods such as Kalman filtering or median filtering can be used, and no specific limitations are made here.
[0065] In one feasible implementation, after obtaining the semantic features output by the large language model and the local temporal features output by the prediction model, Bayesian inference can be applied to the semantic features and the local temporal features to obtain updated weights, thereby using the updated weights to perform weighted fusion of the semantic features and the local temporal features.
[0066] The specific process may include the following steps:
[0067] The first step is to obtain the first output mean and first standard deviation of the large language model, and the second output mean and second standard deviation of the prediction model.
[0068] The second step is to standardize the semantic features and the local temporal features based on the first output mean, the first standard deviation, the second output mean, and the second standard deviation.
[0069] The third step is to set the prior weights of the large language model and the prediction model respectively based on the historical performance of the models.
[0070] Fourth step: Using the standardized outputs of the large language model and the prediction model as input, construct an attention network;
[0071] Step 5: Set up a multilayer perceptron as an attention network; the attention network is used to output the attention weights of the large language model and the prediction model respectively;
[0072] Step 6: Using the attention weights as likelihood, calculate the posterior weights using Bayes' theorem;
[0073] Step 7: Perform weighted fusion of the model outputs of the large language model and the prediction model based on the posterior weights to obtain updated weights.
[0074] First, obtain the first output mean and first standard deviation of the large language model. Statistical analysis can be performed on the output of the large language model on historical data to calculate the average of its output values and the degree of deviation from the average. Simultaneously, obtain the second output mean and second standard deviation of the prediction model to characterize the central tendency and dispersion of the prediction model's output.
[0075] These means and standard deviations are used to standardize semantic features and local temporal features. The purpose of standardization is to ensure that the outputs of different models are on the same comparable scale, eliminating the influence of differences in units and numerical ranges, and making the subsequent fusion process more reasonable and effective.
[0076] Then, prior weights are assigned to the large language model and the prediction model based on their historical performance. Prior weights are a measure of confidence in the past performance of these two models; models with better historical performance are assigned relatively higher prior weights. Prior weights can also be set by those skilled in the art. Alternatively, they can be calculated; this application provides a feasible calculation method:
[0077] Calculate a first prediction error index of the large language model and a second prediction error index of the prediction model, thereby determining a first prior weight of the large language model based on the first prediction error index and a second prior weight of the prediction model based on the second prediction error index.
[0078] An attention network is constructed using the standardized outputs of a large language model and a prediction model as input, with a multilayer perceptron serving as the specific form of the attention network. The attention network automatically learns the importance of each model's output in the current sample from the input data, outputting the attention weights for the large language model and the prediction model. These attention weights reflect the relative importance of each model in the current specific situation, thus dynamically adjusting the contribution of each model.
[0079] The obtained attention weights are used as likelihoods, and the posterior weights are calculated using Bayes' theorem. Bayes' theorem effectively combines prior weights and the current likelihood (attention weights) to calculate the posterior weights for each model given the current data. The posterior weights comprehensively consider both the model's historical performance and its performance on the current samples, providing weights for dynamically updating the model weights.
[0080] Finally, the outputs of the large language model and the prediction model are weighted and fused based on the calculated posterior weights to obtain updated weights. The fused result better reflects the advantages of each model and the characteristics of the current data, thereby improving the accuracy and reliability of the prediction and making the final prediction result more in line with actual business needs.
[0081] In step S104, the business time-series data and its semantic information can be input into a dual-channel network, and the dual-channel network can be used to extract global dependency features and local real-time features respectively. Then, a weighted fusion is performed on the global dependency features and the local real-time features to output the global impact features.
[0082] Globally dependent features mainly refer to periodic features, such as features affected by holidays or weekends, while local real-time features refer to minute-level order fluctuations, such as order surges during peak hours.
[0083] In practical applications, business time-series data and corresponding semantic information can be collected and organized to ensure data integrity and accuracy. The organized business time-series data and its semantic information are then input into the dual-channel network.
[0084] In the global dependency feature extraction process, one channel of the dual-channel network is dedicated to processing business time-series data. It utilizes internal mechanisms (such as the Transformer's self-attention mechanism) to capture global dependencies in the data and extract global dependency features. This process can identify long-term trends, periodic patterns, and correlations between different time points in the time-series data, providing global information.
[0085] During local real-time feature extraction, the other channel in the dual-channel network focuses on processing semantic information, extracting local real-time features through suitable methods (such as LSTM memory units and gating structures). This channel can capture local details such as fluctuations and changes in semantic information within a short time range, reflecting the semantic characteristics at the current moment or several adjacent moments.
[0086] The extracted global dependency features and local real-time features are then weighted and fused. This weighted fusion can be understood as comprehensively considering the importance or contribution of the two types of features, assigning them corresponding weights, and then combining them.
[0087] The weights can be determined based on prior knowledge, the model's historical performance, or through automatic learning during the training process. The goal is to enable the fused features to better reflect the overall picture of the business time series data, including both global macro trends and local detailed changes, and highlighting information that is more important to the prediction target.
[0088] The features obtained after weighted fusion are the global impact features, which are used as the output. This feature integrates key information on global dependencies and local real-time data, and can be used for subsequent analysis, prediction, and other tasks to help understand the inherent patterns and trends of business time series data more accurately.
[0089] This application's embodiments combine a large language model with traditional prediction models when performing business forecasting. This improves the extraction accuracy of unstructured business transaction data while retaining the advantages of processing structured business transaction data, thus enhancing the accuracy of business forecasting. Furthermore, the combination of offline and online forecasting meets the forecasting needs for minute-level business data, enabling timely responses to business changes. It also allows for long-term offline forecasting in advance, providing support for long-term business planning.
[0090] This application is applicable to scenarios such as logistics, e-commerce, express delivery, fresh food cold chain, urban public transportation planning, and ride-hailing operation scheduling.
[0091] For example, in e-commerce warehousing management scenarios, the technology of this invention can be used to predict product inventory demand within the e-commerce industry. By analyzing historical sales data, promotional activity arrangements, seasonal factors, and consumer search behavior text information, the sales volume of different products in the future can be predicted, thereby rationally allocating warehouse space, optimizing inventory replenishment strategies, and reducing the risk of inventory backlog or stockouts.
[0092] In the context of last-mile delivery scheduling, this technology is applied to the express delivery industry. By comprehensively considering factors such as changes in the number of packages at express delivery outlets, traffic conditions in the delivery area, weather conditions, and peak shopping periods during holidays, it predicts the express delivery demand for different time periods and regions. Based on this, it rationally allocates courier resources, plans optimal delivery routes, and improves the timeliness and service quality of express delivery.
[0093] In the context of fresh food cold chain logistics, this approach forecasts the demand for fresh food cold chain logistics in different regions and at different times by considering factors such as product shelf life, market demand fluctuations, temperature changes during transportation, and traffic congestion. This helps businesses rationally allocate cold chain transport vehicles and warehousing facilities to ensure the quality and freshness of their products.
[0094] In the context of urban public transportation planning, this study analyzes urban transportation card swipe data, historical passenger flow data for bus routes, large-scale event schedules, and textual information about travel on social media to predict public transportation passenger flow at different times and on different road sections. This provides decision support for urban traffic management departments to optimize bus routes and allocate bus vehicles, thereby alleviating urban traffic congestion.
[0095] In ride-hailing operation and dispatch scenarios, ride-hailing platforms can utilize this technology to comprehensively consider factors such as real-time traffic conditions, user ride-hailing demand distribution, weather conditions, and special events to predict ride-hailing demand in different areas and at different times. Through the intelligent dispatch system, vehicle resources can be rationally allocated, improving the order acceptance rate and operational efficiency of ride-hailing services.
[0096] Furthermore, based on the above embodiments, the business prediction results can be used to predict business indicators, and business verification indicators can be used to calculate the accuracy of the prediction results. If the deviation of the prediction result is greater than a set threshold, an adjustment strategy is invoked to update at least one of the large language model, the prediction model, and the dual-channel network.
[0097] The system predicts the impact of business forecasts on order volume, response rate, and price sensitivity. It evaluates the forecast accuracy using metrics such as MSE, MAE, and RMSE, verifies the effectiveness of real-time and offline forecasts, and sets the forecast error rate to ≤5%.
[0098] The difference between predicted and actual values is measured using metrics such as MAE (mean absolute error), RMSE (root mean square error), and MAPE (mean absolute percentage error). The smaller the metric, the higher the prediction accuracy.
[0099] For example, an expected judgment can be set (80% of indicators meet the deviation M=5%). The MAPE of all current indicators (such as order volume, matching order volume, response rate, GMV, etc.) is calculated. If the MAPE of more than 80% of the business indicator prediction data is ≤5%, it is considered reasonable and the prediction accuracy meets the standard. At this time, the business prediction result can be adopted (after the business confirms it again, it will be implemented in actual application).
[0100] It can also perform secondary business determination, that is, determine whether the prediction effect meets business expectations. If it does, the business adopts the data and carries out business strategy operation and application; if it does not, it is not adopted and the next round of prediction optimization continues.
[0101] An exemplary execution process of this application may be as follows:
[0102] Step 1: Data preprocessing (including offline and real-time data):
[0103] Internal and external data collection and storage:
[0104] Data from internal and external systems within the group is synchronized to the business MaxCompute database via ETL tools (ETL stands for Extract, Transform, and Load, a software tool used for data integration) to store real-time business data, including orders, transaction amounts, transaction order volume, GMV, etc., with a data collection frequency of once per minute.
[0105] Data processing and updating (offline):
[0106] Clean and process the data, handle missing and outlier values, standardize the data format, extract features relevant to business forecasting, such as user information, product characteristics, transaction volume, price, and seasonal cycle, normalize the features, and update on T+1.
[0107] Data processing and updating (real-time):
[0108] The system cleans and standardizes data in real time, extracts business features such as order information, supply and demand indicators, holiday information, weather information, and historical congestion data, and stores them in a database that is updated every minute.
[0109] Step 2: Multi-model hierarchical architecture prediction (based on a large language model and self-developed algorithm):
[0110] Offline prediction: Large language model + temporal prediction (dynamic weight fusion mechanism):
[0111] Feature-level fusion: Combines the semantic understanding capabilities of large language models (Transformer) with the time-series feature extraction capabilities of traditional prediction models (LSTM / ARIMA), and dynamically allocates weights through a multi-head attention mechanism.
[0112] Decision-level fusion: Bayesian inference is applied to the prediction results of the two models to achieve adaptive adjustment of model weights through attention mechanism and Bayesian inference.
[0113] Real-time prediction of large language models (deep learning) + multi-model ensemble + real-time correction module:
[0114] Spatiotemporal feature extraction: A Transformer-LSTM dual-channel network is used to extract features. The Transformer module captures global dependencies, and the LSTM module extracts local temporal features.
[0115] Multi-model ensemble: Introducing a dynamic weighted fusion mechanism to combine the outputs of traditional models such as ARIMA and Prophet with deep learning results.
[0116] Real-time correction: Sliding window residual learning is used to compensate for prediction errors online.
[0117] Step 3: Prediction Performance Evaluation (Algorithm Performance Evaluation and Training Optimization):
[0118] Algorithm prediction performance evaluation:
[0119] The prediction accuracy is evaluated using metrics such as MAE, RMSE, and MAPE. The prediction error rates of real-time and offline predictions are verified, and a prediction error rate of < 5% is considered reasonable.
[0120] Business assessment (80% of indicators met):
[0121] The predicted data is used to predict the effects on order volume, response rate, price sensitivity, etc. Through the analysis of indicators such as MSE, MAE, and GM, if more than 80% of the business indicators have MAPE < 5%, it is considered reasonable and the next step is executed.
[0122] Step 4: Business Adoption of Forecast Data (After further confirmation by the business, specific implementation and practical application):
[0123] Second confirmation of business:
[0124] If the current business forecasts for all metrics (such as order volume, matched order volume, response rate, etc.) are met, they will be adopted; otherwise, the next round of forecasting will continue.
[0125] See Figure 2 , Figure 2 This application provides a schematic diagram of a business data prediction system structure, which also includes:
[0126] The data acquisition module is used to acquire business transaction data; the business transaction data includes offline data and online real-time data.
[0127] The feature extraction module is used to extract business-related features from the online real-time data, and update the business-related features in real time by combining them with the offline data to obtain business time-series data;
[0128] The offline prediction module is used to collect the business time series data and its semantic information during the offline prediction process, and input the business time series data and the semantic data into the large language model and the prediction model respectively to obtain the semantic features output by the large language model and the local time series features output by the prediction model.
[0129] The online prediction module is used to input the business time series data and its semantic information into a dual-channel network during the online prediction process, and output the global impact features.
[0130] The business prediction module is used to adjust the feature weights of the semantic features, the local temporal features, and the global impact features based on the historical prediction accuracy to obtain the business prediction results.
[0131] This application also provides an embodiment corresponding to a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method described in the above method embodiments.
[0132] It is understood that if the methods in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or 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 executes all or part of the steps of the methods described in the various embodiments of this application. 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.
[0133] The computer-readable storage medium provided in this embodiment includes the method mentioned above, and has the same effect.
[0134] This application also provides an electronic device, see [link to document]. Figure 3 The present application provides a structural diagram of an electronic device, such as... Figure 3 As shown, it may include a processor 1410 and a memory 1420.
[0135] The processor 1410 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 1410 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 1410 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 1410 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 1410 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.
[0136] The memory 1420 may include one or more computer-readable storage media, which may be non-transitory. The memory 1420 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In this embodiment, the memory 1420 is used to store at least the following computer program 1421, which, after being loaded and executed by the processor 1410, is capable of implementing the relevant steps in the methods executed by the electronic device side as disclosed in any of the foregoing embodiments. In addition, the resources stored in the memory 1420 may also include an operating system 1422 and data 1423, etc., and the storage method may be temporary storage or permanent storage. The operating system 1422 may include Windows, Linux, Android, etc.
[0137] In some embodiments, the electronic device may further include a display screen 1430, an input / output interface 1440, a communication interface 1450, a sensor 1460, a power supply 1470, and a communication bus 1480.
[0138] certainly, Figure 3 The structure of the electronic device shown does not constitute a limitation on the electronic device in the embodiments of this application. In practical applications, the electronic device may include more than [other components]. Figure 3 More or fewer components as shown, or combinations of certain components.
[0139] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. As the system provided in the embodiments corresponds to the method provided in the embodiments, the description is relatively simple; relevant parts can be found in the method section.
[0140] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.
[0141] It should also be noted that, in this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
Claims
1. A business data prediction method, characterized in that, include: Acquire business transaction data; the business transaction data includes offline data and online real-time data; Business-related features are extracted from the online real-time data, and the business-related features are updated in real time by combining them with the offline data to obtain business time-series data; During the offline prediction process, the business time series data and its semantic information are collected. The business time series data and the semantic data are respectively input into the large language model and the prediction model to obtain the semantic features output by the large language model and the local time series features output by the prediction model. During the online prediction process, the business time series data and its semantic information are input into a dual-channel network, and the global impact features are output. The feature weights of the semantic features, the local temporal features, and the global impact features are adjusted based on the historical prediction accuracy to obtain the business prediction results.
2. The method according to claim 1, characterized in that, After obtaining the business transaction data, the following is also included: The business transaction data is subjected to offline preprocessing and online preprocessing; the offline preprocessing includes data cleaning and feature normalization; the online preprocessing includes data standardization and noise removal.
3. The method according to claim 1, characterized in that, After obtaining the semantic features output by the large language model and the local temporal features output by the prediction model, the method further includes: The semantic features and the local temporal features are weighted using Bayesian inference to obtain the updated weights; The semantic features and the local temporal features are weighted and fused using the updated weights.
4. The method according to claim 3, characterized in that, The updated weights are obtained by performing Bayesian inference weighting on the semantic features and the local temporal features, resulting in the following: Obtain the first output mean and first standard deviation of the large language model, and the second output mean and second standard deviation of the prediction model; The semantic features and the local temporal features are standardized based on the first output mean, the first standard deviation, the second output mean, and the second standard deviation. The prior weights of the large language model and the prediction model are set according to the historical performance of the models, respectively. An attention network is constructed by taking the standardized outputs of the large language model and the prediction model as inputs. A multilayer perceptron is set as the attention network; the attention network is used to output the attention weights of the large language model and the prediction model respectively. The attention weights are used as likelihoods, and the posterior weights are calculated using Bayes' theorem. The updated weights are obtained by weighting and fusing the outputs of the large language model and the prediction model based on the posterior weights.
5. The method according to claim 4, characterized in that, The prior weights for the large language model and the prediction model are set based on their historical performance, including: Calculate the first prediction error index of the large language model and the second prediction error index of the prediction model; The first prior weights of the large language model are determined based on the first prediction error index. The second prior weight of the prediction model is determined based on the second prediction error index.
6. The method according to claim 1, characterized in that, The business time-series data and its semantic information are input into a dual-channel network, and the output global impact features include: The business time series data and its semantic information are input into a dual-channel network, and the dual-channel network is used to extract global dependency features and local real-time features respectively. The global dependency features and the local real-time features are weighted and fused to output the global impact features.
7. The method according to claim 1, characterized in that, After adjusting the feature weights of the semantic features, the local temporal features, and the global influence features based on historical prediction accuracy to obtain the business prediction result, the method further includes: The business forecast results are used to predict business indicators, and business verification indicators are used to calculate the accuracy of the forecast results. If the deviation of the prediction result is greater than a set threshold, an adjustment strategy is invoked to update at least one of the large language model, the prediction model, and the dual-channel network.
8. A business data prediction system, characterized in that, include: The data acquisition module is used to acquire business transaction data; the business transaction data includes offline data and online real-time data. The feature extraction module is used to extract business-related features from the online real-time data, and update the business-related features in real time by combining them with the offline data to obtain business time-series data; The offline prediction module is used to collect the business time series data and its semantic information during the offline prediction process, and input the business time series data and the semantic data into the large language model and the prediction model respectively to obtain the semantic features output by the large language model and the local time series features output by the prediction model. The online prediction module is used to input the business time series data and its semantic information into a dual-channel network during the online prediction process, and output the global impact features. The business prediction module is used to adjust the feature weights of the semantic features, the local temporal features, and the global impact features based on the historical prediction accuracy to obtain the business prediction results.
9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the method as claimed in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, implements the steps of the method as described in any one of claims 1 to 7.