WO2020177499A1 - Model prediction acceleration method and device - Google Patents

Model prediction acceleration method and device Download PDF

Info

Publication number
WO2020177499A1
WO2020177499A1 PCT/CN2020/073797 CN2020073797W WO2020177499A1 WO 2020177499 A1 WO2020177499 A1 WO 2020177499A1 CN 2020073797 W CN2020073797 W CN 2020073797W WO 2020177499 A1 WO2020177499 A1 WO 2020177499A1
Authority
WO
WIPO (PCT)
Prior art keywords
module
output data
prediction
quick
current
Prior art date
Application number
PCT/CN2020/073797
Other languages
French (fr)
Chinese (zh)
Inventor
杨永晟
Original Assignee
阿里巴巴集团控股有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 阿里巴巴集团控股有限公司 filed Critical 阿里巴巴集团控股有限公司
Publication of WO2020177499A1 publication Critical patent/WO2020177499A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
    • G06Q40/04—Trading; Exchange, e.g. stocks, commodities, derivatives or currency exchange

Definitions

  • the present disclosure relates to data processing, and in particular to methods and devices for forecasting foreign exchange transaction volume.
  • the initiation of the purchase of foreign exchange is carried out at the time agreed with the bank, so it is very important that the modeling and training of the predictive model system is completed before the time of purchase of foreign exchange.
  • the business requirements are given before the time of purchase Forecast value, so there is a strong requirement for the time stability of the entire forecasting process.
  • the timeliness and stability of model operation is affected by the timeliness of upstream data and the availability of resources in the operating environment, so how to ensure that the forecast value output data is completed before the time of business purchase is particularly important.
  • the present disclosure provides a method for accelerating prediction, and the method includes:
  • the input data is received by a predictive model system, wherein the predictive model system includes one or more predictive modules connected in series, one or more rapid modules are connected to the predictive model system, and the input of the rapid module is the one or The input of one prediction module of the plurality of prediction modules and the fast module provides fast output data;
  • the quick module is generated according to at least one prediction module in the prediction model system.
  • the predictive model system generates model output data
  • the method further includes:
  • the model output data is determined as final predictive output data.
  • the fast module is connected between two prediction modules of the prediction model system, and the at least one prediction module includes the same prediction module as the input of the fast module and one or more subsequent prediction modules Forecast module.
  • the quickness module is generated by adjusting parameters for the at least one prediction module and/or deleting one or more sub-modules of the at least one prediction module.
  • the parameters include training step size, training times and/or error accuracy.
  • the method further includes:
  • the method further includes:
  • the current quick output data, the current cycle number and the current module number will be overwritten in the memory with the previous quick output data, the previous cycle number and the current module number. Module number first;
  • the input data is historical transaction data of the foreign exchange exchange business
  • the prescribed time is the expiration time of the foreign exchange purchase settlement period
  • Another aspect of the present invention provides an apparatus for accelerating prediction, including:
  • a prediction model system that receives input data, wherein the prediction model system includes one or more prediction modules connected in series;
  • One or more fast modules connected to the prediction model system the input of the fast module is the input of one of the one or more prediction modules, and the fast module provides fast output data, the fast The module is generated according to at least one prediction module in the prediction model system;
  • a memory the memory storing at least one quick output data
  • the processor acquires the at least one quick output data from the one or more quick output data at a prescribed time, stores the at least one quick output data in the memory, and stores the latest one among the at least one predicted output data
  • the predicted output data of is determined as the final predicted output data.
  • the predictive model system generates model output data
  • the processor is further configured to:
  • the model output data is received from the prediction model system at the prescribed time, the model output data is determined to be the final prediction output data.
  • the fast module is connected between two prediction modules of the prediction model system, and the at least one prediction module includes the same prediction module as the input of the fast module and one or more subsequent prediction modules Forecast module.
  • the quickness module is generated by adjusting parameters for the at least one prediction module and/or deleting one or more sub-modules of the at least one prediction module.
  • the parameters include training step size, training times and/or error accuracy.
  • the processor is further configured to:
  • the processor is further configured to:
  • the current quick output data, the current cycle number and the current module number will be overwritten in the memory with the previous quick output data, the previous cycle number and the current module number. Module number first;
  • the input data is historical transaction data of the foreign exchange exchange business
  • the prescribed time is the expiration time of the foreign exchange purchase settlement period
  • a further aspect of the present invention provides a computer device including:
  • a memory arranged to store computer-executable instructions that, when executed, cause the processor to perform the following operations:
  • the input data is received by a predictive model system, wherein the predictive model system includes one or more predictive modules connected in series, one or more rapid modules are connected to the predictive model system, and the input of the rapid module is the one or The input of one prediction module of the plurality of prediction modules and the fast module provides fast output data;
  • the quick module is generated according to at least one prediction module in the prediction model system.
  • the present disclosure accelerates model prediction by incorporating a fast module in the prediction model system, and the fast module can shorten the running time for obtaining the predicted value.
  • the fast module can reduce model features, reduce the number of model training times, reduce the number of training samples, etc., thereby speeding up the prediction speed and ensuring that there is prediction output data within a specified time. Further, by selecting the latest (or the most accurate) forecast output data, it is guaranteed that the most accurate forecast output data currently is obtained.
  • Figure 1 is a system diagram for model prediction.
  • Fig. 2 is a structural diagram of a system for accelerating model prediction according to an embodiment of the present disclosure.
  • Fig. 3 shows an example of constructing a fast module according to an embodiment of the present disclosure.
  • Fig. 4 shows a structural diagram of a system for accelerating model prediction according to various aspects of the present disclosure.
  • Figure 5 illustrates a flowchart of a method for accelerating model prediction according to aspects of the present disclosure.
  • Figure 1 shows a system diagram for model prediction.
  • the system for model prediction may include a computing platform 100 and a timer 102.
  • the timer 102 can also be incorporated into the computing platform 100.
  • the computing platform 100 is, for example, an Alipay server, and may include a predictive model system 101.
  • the predictive model system 101 may receive sample data.
  • the prediction model system 101 may include one or more cascaded prediction modules. The sample data is processed by each prediction module in turn to obtain the prediction result.
  • historical transaction data before the current forecast period can be input into the pre-trained forecast model system 101 for forecasting.
  • the i th cycle can generate a prediction result of the prediction V i
  • i + 1-th cycle can generate a prediction result of the prediction V i + 1.
  • new sample data will be input to the prediction model system for the next period of prediction.
  • the forecast period may be one hour.
  • the timer 102 clocks the operation of the computing platform 100. Specifically, the operation of the computing platform 100 may be periodically controlled.
  • the timer 102 sends a signal to the computing platform 100 at the end of the current prediction period to instruct the prediction model system 101 to perform prediction for the next period.
  • model prediction needs to be completed before the end of each period, otherwise the prediction of the next period will interrupt the prediction of the current period, and the output data of the current period will be missing.
  • the forecasting model system needs to output data forecasts before the end of the current foreign exchange purchase cycle.
  • the timeliness of model operation is affected by the timeliness of upstream data and the resource availability of the operating environment.
  • the forecasting model system fails to complete the forecasting operation of the current cycle, thus losing the current forecast output data.
  • the purpose of the present disclosure is to solve the technical problem that the predicted value of the data cannot be output on time when the prediction model system runs overtime.
  • the prediction model system can be constructed as a series connection of several prediction modules.
  • the foreign exchange transaction volume prediction model system may include a data cleaning module, a quantitative modeling module, an abnormality detection module, and a business decision adjustment module.
  • these modules need to be completed within a given time (for example, one hour) and finally generate predicted values.
  • a given time for example, one hour
  • the operating link is affected by the timeliness of upstream data and the availability of operating environment resources, it is prone to fail to complete the operation of all modules within a given time.
  • the embodiments of the present disclosure are explained in detail below.
  • Fig. 2 is a diagram of a system for accelerating model prediction according to an embodiment of the present disclosure.
  • the system for accelerating model prediction includes the conventional prediction part on the left and the quick part (the part in the box) on the right.
  • the period is one hour. It should be understood that other cycles are also within the concept of this disclosure.
  • the conventional forecasting part can be split into one or more forecasting modules connected in series according to the needs of the actual forecasting process, for example, including a data cleaning module, a quantitative modeling module, an anomaly detection module, and a business decision module.
  • the data cleaning module can complete the cleaning of the latest data (for example, the data of the last hour). For example, sample data is extracted from multiple business systems and contains historical data. Some of the data is wrong data, and some data conflicts with each other. These wrong or conflicting data are called “dirty data.” . The "dirty data” needs to be “washed out” according to certain rules, that is, data that does not meet the requirements is filtered.
  • the quantitative modeling module is the core code module, which can complete functions such as feature construction, model training, and output data of predicted values.
  • the anomaly detection module can perform anomaly detection on the predicted value of the previous module.
  • the business decision-making module can adjust the predicted value accordingly based on business experience.
  • the data input is processed by the modules in the regular part to generate the final predicted value.
  • the data cleaning module, quantitative modeling module, anomaly detection module and business decision module are connected in series. Each module needs to wait for the output data of the upstream module for processing. At the same time, the processing speed of each module will be affected by the availability of operating environment resources Therefore, the final predicted value of the data may not be output within the specified time.
  • the present disclosure takes into account the timeliness of prediction, and adds a quick part to the system, and the quick part includes one or more quick modules.
  • the quick part includes the quick module 1, the quick module 2 and the quick module 3, which respectively quickly generate prediction values V 1 , V 2 and V 3 with lower accuracy.
  • the fast module sacrifices a certain prediction accuracy to speed up the prediction.
  • the fast module 1 is connected to the output data of the data cleaning module, and can be generated according to the subsequent modules of the data cleaning module in the regular part, such as a simplified version of the quantitative modeling module, anomaly detection module, and business decision module.
  • the fast module 1 can complete the training related to the quantitative modeling module, anomaly detection module and the business decision module in a short time, and output the data prediction value V 1 .
  • the prediction value V 1 has a lower accuracy than the final prediction value V 4 , but the output data time is earlier.
  • the prediction path including the data cleaning module and the fast module 1 can quickly generate the prediction value V 1 with lower accuracy.
  • the fast module 2 is connected to the output data of the quantitative modeling module, and can be generated according to the subsequent modules of the quantitative modeling module in the regular part, such as a simplified version of the anomaly detection module and the business decision module.
  • the fast module 2 can complete the training related to the anomaly detection module and the business decision module in a short time, and output the data prediction value V 2 .
  • the prediction value V 2 has lower accuracy than the final prediction value V 4 , but the output data is faster; it has higher accuracy than the prediction value V 1 , but the output data is slow.
  • the fast module 3 is connected to the output data of the anomaly detection module, and can be generated according to the subsequent modules of the anomaly detection module in the regular part, for example, a simplified version of the business decision module.
  • the fast module 3 can complete the training related to the business decision module in a short time, and output the data prediction value V 3 .
  • the prediction value V 3 has a lower accuracy than the final prediction value V 4 , but the output data is faster; it has a higher accuracy than the prediction value V 2 but the output data is slow.
  • the fast module can complete the training and output data of the predicted value in a short time.
  • the purpose of the fast module is to reduce the running time to obtain the predicted value.
  • the simplification of the conventional prediction module by the fast module can include: reducing the model Features, reduce the number of model training, reduce the number of training samples, etc.
  • the fast module essentially sacrifices a certain prediction accuracy to speed up the prediction.
  • the quick part includes three modules. However, those skilled in the art can understand that the quick part may also include less than 3 modules, such as 1 or 2 modules. In a further embodiment, according to the number N of modules included in the conventional partial prediction process, the quick part may include 1 to N modules.
  • the system can output the final predicted value V 4 of the data. If the conventional prediction model system has not output the final data prediction value V 4 , the system can output the data prediction value of the quickest module output data that is currently the most accurate, for example, V 3 .
  • the final predicted value V 4 of the previous cycle can be used as the output data.
  • the system for accelerating model prediction of the present disclosure can ensure that there is a predicted value output data at a specified time, and the predicted value of the current most accurate version of the output data.
  • Fig. 3 shows an example of constructing a fast module according to an embodiment of the present disclosure.
  • the conventional part of the system for accelerating model prediction includes a quantitative modeling module, an anomaly detection module, and a business decision module.
  • the quantitative modeling module can include the AUTO_ML (automated learning framework) sub-module, the classic time series prediction sub-module, the ARIMA (Autoregressive Integrated Moving Average Model, autoregressive integrated moving average model) sub-module, Holt-Winters (three times Exponential smoothing) sub-module, same-ring comparison sub-module.
  • AUTO_ML automated learning framework
  • ARIMA Automatic Integrated Moving Average Model
  • autoregressive integrated moving average model autoregressive integrated moving average model
  • Holt-Winters three times Exponential smoothing sub-module
  • same-ring comparison sub-module same-ring comparison sub-module.
  • the anomaly detection module includes an anomaly detection sub-module based on real-time data sources and an anomaly detection sub-module based on offline data sources.
  • the business decision module includes an adjustment sub-module based on information input from an external operator and an adjustment sub-module based on information input from an internal operator.
  • the quick module 1 may be a simplified version of the quantitative modeling module, anomaly detection module, and business decision module.
  • the fast module 1 can obtain different operating speeds by adjusting parameters for the quantitative modeling module (such as adjusting the training step size, training times and error accuracy), thereby speeding up the processing time of the module; the fast module 1 can remove the based on the abnormal detection module
  • the anomaly detection sub-module of real-time data sources is only included in the processing of the anomaly detection sub-module based on offline data sources; the fast module 1 can remove the adjustment sub-module based on the information input of the external operator for the business decision module, and only include the sub-module based on internal operation
  • the fast module 2 can be a simplified version of the anomaly detection module and the business decision module.
  • the fast module 2 can remove the anomaly detection sub-module based on real-time data sources for the anomaly detection module, and only incorporate the processing of the anomaly detection sub-module based on offline data sources; for the business decision module, remove the adjustment sub-module based on external operator information input. Module, and only include the processing of the adjustment sub-module based on the information input of the internal operator.
  • the quick module 3 can be a simplified version of the business decision module.
  • the fast module 3 can remove the adjustment sub-module based on the information input of the external operator for the business decision module, and only incorporate the processing of the adjustment sub-module based on the information input of the internal operator.
  • the prediction model system may include various prediction modules, and each quick module may also adjust the parameters of the anomaly detection module and/or the business decision module, and remove other sub-modules.
  • the quick module can also remove the subsequent modules.
  • the fast module 1 can be generated based on the quantitative modeling module and the business decision module, without including the anomaly detection module.
  • the conventional part of the prediction system may have different prediction modules, so the system has correspondingly different quick modules, where each quick module can be generated according to one or more corresponding conventional prediction modules, for example, the quick module It can be a simplified version of the corresponding one or more conventional prediction modules.
  • FIG. 4 shows a structural diagram of a system 400 for accelerating model prediction according to various aspects of the present disclosure.
  • the system 400 may include a predictive model system part 401, a fast module part 402, a memory 403, and a processor 404.
  • the prediction model system part 401 may include a prediction module 1, a prediction module 2, ... a prediction module N-1, and a prediction module N connected in series. That is, the input sample is input to the prediction module 1, the prediction result of the prediction module 1 is input to the prediction module 2, the prediction result of the prediction module 2 is input to the prediction module 3, ... the prediction result of the prediction module N-1 is input to the prediction module N , The output data of the prediction module N is the final output data of the prediction model system.
  • the data is sequentially processed by the prediction modules 1-N to produce the final prediction result V N (that is, the output data of the last prediction module N).
  • the timeliness of the forecasting model system is often affected by the timeliness of upstream data and the resource availability of the operating environment.
  • the forecasting model system fails to complete the forecasting operation of the current cycle, thus missing the current forecast output data.
  • a fast module part 402 is added to the system.
  • the fast module part 402 may include one or more of the fast module 0, the fast module 1, the fast module 2, ... the fast module N-1.
  • the input sample can be input to the fast module 0, and the fast module 0 can be generated according to the prediction module 1—the prediction module N, for example, it may be a simplified version of the prediction module 1—the prediction module N. For example, adjusting the parameters of one or more modules in the prediction module 1-prediction module N and/or deleting one or more sub-modules of the one or more modules.
  • the quick module 1 receives the output data of the prediction module 1, and can generate it according to the subsequent prediction module 2—the prediction module N, for example, it may be a simplified version of the prediction module 2—the prediction module N. For example, adjust the parameters of one or more modules in the prediction module 2—prediction module N and/or delete one or more sub-modules of the one or more modules.
  • the fast module 2 receives the output data of the prediction module 2, and can be generated according to the subsequent prediction module 3—prediction module N, for example, it may be a simplified version of prediction module 3—prediction module N; and so on. For example, adjust the parameters of one or more modules in the prediction module 3—prediction module N and/or delete one or more sub-modules of the one or more modules.
  • the output data of fast module 0 is V 0
  • the output data of fast module 1 is V 1
  • the output data of fast module 2 is V 2
  • ... the output data of fast module N-1 is V N-1
  • the prediction model system The output data of the last module N is V N.
  • V 0 , V 1 , V 2 , ... V N-1 , V N can be input to the memory 403 for storage.
  • the processor 404 may number V 0 , V 1 , V 2 , ... V N-1 , V N , and store each predicted output data together with its number in the memory.
  • the number can include the module number and the cycle number.
  • processor 404 may identify predicted output data from a module V i which rapid or a prediction from the last module. For example, when the fast module i or the prediction module N outputs the data prediction result, the prediction result and the module identifier may be sent to the memory 403 and the processor 404 together. The processor 404 uses the module identification to identify which module the prediction result comes from.
  • the processor 404 can then determine the module number of Vi according to the identification of the fast module or the last prediction module.
  • the module number of V i corresponds to the position in the predictive model system of the node of the predictive model system connected to the fast module generating V i .
  • the module number of the fast module connected to the predictive model system node downstream can be greater than the module number of the fast module connected to the upstream predictive model system node (closer to the input end) .
  • the next module of the prediction model system node connected to the fast module 0 is the prediction module 1, then the number of the prediction output data from the fast module 0 can be 0; the fast module 1 is connected to the prediction module 1 and the prediction module 2. If the number of the predicted output data from the fast module 1 can be 1; if the fast module 2 is connected between the prediction module 2 and the prediction module 3, the number of the predicted output data from the fast module 2 is 2, and so on ; numbering from the predicted output data prompt module N-1 is may be N-1; if V i from the last prediction module, the number i is V is greater than all numbers predicted output data from the fast block, e.g., can be fast The number of modules plus one (for example, N).
  • the prediction module predicts a model system nodes as long as the fast block number V i V i correspond to the generating module to be connected in The order in the prediction model system is sufficient.
  • processor 404 may identify the predicted output data in which the period V i is generated. For example, when the quick module i or the prediction module N outputs the data prediction result, the prediction result is sent to the memory 403 and the processor 404 together with the period number of the period when the prediction result is generated. The processor 404 uses the cycle number to identify the cycle in which the prediction result is generated.
  • the quick module i or the prediction module N may have a cycle number counter, and each time a new cycle is entered, the cycle number value in the counter can be incremented.
  • the processor 404 determines the latest version of the output data V 0 , V 1 , V 2 , ... V N-1 , or V N stored in the memory 403 as the model prediction
  • the result is output data. For example, if the fast module with the current cycle output data V 0 , V 1 , V 2 , V 3 in the memory, the output data V 3 is used as the model prediction result.
  • the prediction of the model is periodic. If the last module of any quick module/prediction model system has no output data for the current period within a specified time, the final prediction value of the previous period can be used as the output data.
  • the foreign exchange transactions may be one hour prediction period, if no current hour fast data module output V i and no output data model system predictive V N, may be the final one hour predicted value as output data.
  • Each entry in the memory includes the output data of the last module of the fast module/predictive model system and its module number and/or cycle number.
  • Table 1 shows an example of a memory storing predicted values.
  • Cycle number Module number Predictive value 0 0 V 0 0 1 V 1 ... ... ... M-1 N V N M 0 V 0 M 1 V 1 M 2 V 2 ... ... ... M N-1 V N-1 M N V N M+1 0 V 0 M+1 1 V 1 ... ... ...
  • search for one or more entries with the largest period number in the memory that is, one or more predicted output data for the latest period
  • search for one or more entries with the largest period number that is, one or more predicted output data for the latest period
  • the entry with the largest module number that is, the latest predicted output data of the latest period
  • the fast module part may include one or more of the fast modules 0-N-1 , As long as its number reflects the sequence of the corresponding modules. That is, those skilled in the art can select one or more of these quick modules to speed up model prediction according to actual needs. For example, the fast module 0 can be omitted, and thus the samples are only input to the prediction module 1 for subsequent processing.
  • the memory may only store the latest predicted output data.
  • processor 404 may identify the current prediction data output V i of the current period and the current module ID number, the previous period preceding the current cycle of the predicted output data and the number stored in memory 403 is compared numbers, if the current cycle number Greater than the previous period number, that is, the current predicted output data is the output data of the predicted period newer than the stored predicted output data, then the current predicted output data, the current period number and the current module number are overwritten in the memory. Forecast output data and previous period number and previous module number.
  • the current cycle number is equal to the previous cycle number, that is, the current predicted output data and the stored predicted output data are generated in the same cycle
  • the current module number is compared with the previous stored in the memory 403
  • the previous module numbers of the predicted output data are compared; if the current module number is greater than the previous module number, that is, the current predicted output data is a newer version than the previous predicted output data, the current predicted output data is stored in the memory 403
  • the current cycle number and the current module number cover the previous forecast output data, the previous cycle number and the previous module number. In this way, storage space can be effectively saved.
  • V i module number, cycle number
  • Figure 5 illustrates a flowchart of a method for accelerating model prediction according to aspects of the present disclosure.
  • the prediction model system may include multiple prediction modules connected in series, the input samples and/or the output data of the prediction module can be input to the next prediction module and the corresponding fast module, and the fast module is based on the prediction model Generated by at least one prediction module in the system, for example, a simplified version of one or more prediction modules behind (downstream) the connected prediction model system node.
  • the simplified version of the prediction module can be obtained by adjusting the parameters of the prediction module and/or deleting the sub-modules in the prediction module.
  • the adjusted parameters may include training step size, training times, and/or error accuracy.
  • the method includes in step 502, receiving input data by a predictive model system.
  • the input data can be samples to be processed.
  • the data can be input to the prediction module 1, and if the fast module 0 exists, it can also be input to the fast module 0.
  • each prediction module can be input to the next prediction module and the corresponding quick module (if any), and the quick module can be a simplified version of the next prediction module and one or more subsequent prediction modules.
  • step 504 at a prescribed time (for example, at the expiration of the current period), the predicted output data of multiple quick modules (which can be called quick output data) and the predicted output data of the last prediction module (which can be called model Output data) at least one of them.
  • the quick output data and model output data can be numbered (module number and cycle number), where the module number of the quick output data corresponds to the prediction model system node connected to the quick module that generates the quick output data in the prediction model
  • the position in the system, the module number of the output data of the fast module connected to the downstream node of the predictive model system is greater than the module number of the output data of the fast module connected to the upstream node of the predictive model system; and the module number of the model output data is greater than each fast
  • the module number of the predicted output data, and the period number is the number of the period in which the predicted output data is generated.
  • the aforementioned predicted output data can be stored in the memory together with the corresponding module number and cycle number.
  • At a predetermined time it can be determined that at least one of the predicted output data of the multiple fast modules and the predicted output data of the last prediction module is stored in the memory.
  • step 506 the latest prediction output data of the at least one of the prediction output data of the multiple quick modules and the prediction output data of the last prediction module is determined as the model output data.
  • the cycle number, the predicted output data number, and the corresponding output data of the fast module can be stored in the memory together, and the predicted output data with the largest cycle number and the largest predicted output data number can be found in the memory at a specified time as the model output data.
  • the present disclosure uses a fast module to shorten the running time for obtaining the predicted value.
  • the fast module can speed up the prediction by reducing model features, reducing the number of model training times, and reducing the number of training samples to ensure that there is predictive output data within a specified time. Further, by selecting the latest (or the most accurate) forecast output data, it is guaranteed that the most accurate forecast output data currently is obtained.
  • a general-purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine.
  • the processor may also be implemented as a combination of computing devices (for example, a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors in cooperation with a DSP core, or any other such configuration).
  • each function can be stored as one or more instructions or codes on a computer-readable medium or transmitted therethrough.
  • Other examples and implementations fall within the scope of this disclosure and the appended claims.
  • the functions described above can be implemented using software, hardware, firmware, hardwired, or any combination thereof executed by a processor.
  • the features that implement the function may also be physically located in various locations, including being distributed so that various parts of the function are implemented at different physical locations.
  • Computer-readable media includes both non-transitory computer storage media and communication media, including any media that facilitates the transfer of a computer program from one place to another.
  • the non-transitory storage medium may be any available medium that can be accessed by a general-purpose or special-purpose computer.
  • non-transitory computer readable media may include RAM, ROM, electrically erasable programmable read-only memory (EEPROM), compact disk (CD) ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices , Or any other non-transitory medium that can be used to carry or store desired program code means in the form of instructions or data structures and can be accessed by a general-purpose or special-purpose computer, or a general-purpose or special-purpose processor.
  • any connection is also properly called a computer-readable medium.
  • the software is transmitted from a web site, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave Yes
  • the coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave are included in the definition of the medium.
  • Disks and discs as used herein include CDs, laser discs, optical discs, digital versatile discs (DVD), floppy discs and Blu-ray discs, in which discs often reproduce data magnetically and discs reproduce data optically with lasers . Combinations of the above media are also included in the scope of computer-readable media.

Landscapes

  • Business, Economics & Management (AREA)
  • Engineering & Computer Science (AREA)
  • Economics (AREA)
  • Strategic Management (AREA)
  • General Business, Economics & Management (AREA)
  • Human Resources & Organizations (AREA)
  • Marketing (AREA)
  • Finance (AREA)
  • Development Economics (AREA)
  • Physics & Mathematics (AREA)
  • Accounting & Taxation (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Game Theory and Decision Science (AREA)
  • Technology Law (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

A method for accelerating prediction. The method comprises: a prediction model system (401) receiving input data (502), the prediction model system (401) comprising one or more prediction modules connected in series, one or more swift modules being connected to the prediction model system (401), an input of the swift module being an input of one of the one or more prediction modules and the swift module providing swift output data; acquiring at least one piece of swift output data from the one or more swift modules at a specified time; and determining the latest swift output data among the at least one piece of swift output data as final prediction output data, the swift module being generated according to at least one prediction module in the prediction model system (401).

Description

模型预测加速方法和装置Model prediction acceleration method and device 技术领域Technical field
本公开涉及数据处理，尤其涉及外汇交易量预测方法和装置。The present disclosure relates to data processing, and in particular to methods and devices for forecasting foreign exchange transaction volume.
背景技术Background technique
在国际汇兑业务中，需要通过提前购买下一个购汇结算周期的各外汇交易量，减少潜在的汇率敞口波动风险，进行损益控制，为了进行损益控制，需要对每个购汇结算周期的外汇交易量进行预测。In the international exchange business, it is necessary to purchase each foreign exchange transaction volume in the next foreign exchange purchase settlement cycle in advance to reduce the risk of potential exchange rate exposure fluctuations, and to control the profit and loss. Forecast of transaction volume.
在实际业务中，发起购汇操作是在与银行约定好的时间点进行的，所以预测模型系统建模与训练工作在购汇时间点之前完成非常重要，业务要求在购汇时间点之前给出预测值，所以对整个预测流程的时效稳定性有较强的要求。实际运行中，模型运行的时效稳定性受到上游数据时效性、运行环境的资源可用度影响，所以如何保障在业务购汇时间点之前完成预测值输出数据是尤为重要的。In actual business, the initiation of the purchase of foreign exchange is carried out at the time agreed with the bank, so it is very important that the modeling and training of the predictive model system is completed before the time of purchase of foreign exchange. The business requirements are given before the time of purchase Forecast value, so there is a strong requirement for the time stability of the entire forecasting process. In actual operation, the timeliness and stability of model operation is affected by the timeliness of upstream data and the availability of resources in the operating environment, so how to ensure that the forecast value output data is completed before the time of business purchase is particularly important.
发明内容Summary of the invention
为解决上述技术问题，本公开提供了一种用于加快预测的方法，所述方法包括：In order to solve the above technical problems, the present disclosure provides a method for accelerating prediction, and the method includes:
由预测模型系统接收输入数据，其中所述预测模型系统包括串接的一个或多个预测模块，一个或多个迅捷模块连接到所述预测模型系统，所述迅捷模块的输入是所述一个或多个预测模块中的一个预测模块的输入并且所述迅捷模块提供迅捷输出数据；The input data is received by a predictive model system, wherein the predictive model system includes one or more predictive modules connected in series, one or more rapid modules are connected to the predictive model system, and the input of the rapid module is the one or The input of one prediction module of the plurality of prediction modules and the fast module provides fast output data;
在规定时间从所述一个或多个迅捷模块获取至少一个迅捷输出数据；以及Obtain at least one quick output data from the one or more quick modules at a specified time; and
将所述至少一个迅捷输出数据中最新的迅捷输出数据确定为最终预测输出数据，Determining the latest quick output data in the at least one quick output data as the final predicted output data,
其中所述迅捷模块是根据所述预测模型系统中的至少一个预测模块生成的。The quick module is generated according to at least one prediction module in the prediction model system.
可任选地，所述预测模型系统产生模型输出数据，并且所述方法进一步包括：Optionally, the predictive model system generates model output data, and the method further includes:
如果在所述规定时间所述预测模型系统产生了模型输出数据，则将所述模型输出数据确定为最终预测输出数据。If the predictive model system generates model output data at the prescribed time, the model output data is determined as final predictive output data.
可任选地，所述迅捷模块连接在所述预测模型系统的两个预测模块之间，并且所述至少一个预测模块包括与所述迅捷模块的输入相同的预测模块及其后续一个或多个 预测模块。Optionally, the fast module is connected between two prediction modules of the prediction model system, and the at least one prediction module includes the same prediction module as the input of the fast module and one or more subsequent prediction modules Forecast module.
可任选地，所述迅捷模块是通过对所述至少一个预测模块调整参数和/或删除所述至少一个预测模块的一个或多个子模块来生成的。Optionally, the quickness module is generated by adjusting parameters for the at least one prediction module and/or deleting one or more sub-modules of the at least one prediction module.
可任选地，所述参数包括训练步长、训练次数和/或误差精度。Optionally, the parameters include training step size, training times and/or error accuracy.
可任选地，该方法进一步包括：Optionally, the method further includes:
确定所述至少一个迅捷输出数据中的每一者的模块编号和周期编号，其中连接到预测模型系统下游的迅捷模块的迅捷输出数据的模块编号大于连接到预测模型系统上游的迅捷模块的迅捷输出数据的模块编号，所述周期编号是生成预测输出数据的周期的编号；Determine the module number and period number of each of the at least one quick output data, wherein the module number of the quick output data of the quick output data connected to the downstream of the predictive model system is greater than the quick output of the quick module connected to the upstream of the predictive model system The module number of the data, where the period number is the number of the period in which the predicted output data is generated;
将所述至少一个迅捷输出数据中的每一者与其周期编号、模块编号一起存储在存储器中；Storing each of the at least one quick output data in the memory together with its cycle number and module number;
在所述规定时间在所述存储器中查找周期编号最大的一个或多个迅捷输出数据；以及Searching for one or more quick output data with the largest cycle number in the memory at the specified time; and
选择所述一个或多个预测输出数据中模块编号最大的迅捷输出数据作为所述最终预测输出数据。Select the quick output data with the largest module number among the one or more predicted output data as the final predicted output data.
可任选地，该方法进一步包括：Optionally, the method further includes:
确定所获取的当前迅捷输出数据的当前周期编号和当前模块编号；Determine the current cycle number and current module number of the acquired current quick output data;
将所述当前周期编号与存储器中所存储的在先迅捷输出数据的在先周期编号进行比较；Comparing the current cycle number with the previous cycle number of the previous quick output data stored in the memory;
如果所述当前周期编号大于所述在先周期编号，则在所述存储器中将所述当前迅捷输出数据以及当前周期编号和当前模块编号覆盖所述在先迅捷输出数据以及在先周期编号和在先模块编号；If the current cycle number is greater than the previous cycle number, the current fast output data, current cycle number and current module number will be overwritten in the memory with the previous fast output data, previous cycle number and previous cycle number. First module number;
如果所述当前周期编号等于所述在先周期编号，则将所述当前模块编号与所述在先模块编号进行比较；If the current cycle number is equal to the previous cycle number, comparing the current module number with the previous module number;
如果所述当前模块编号大于所述在先模块编号，则在所述存储器中将所述当前迅捷输出数据以及当前周期编号和当前模块编号覆盖所述在先迅捷输出数据以及在先周期编号和在先模块编号；以及If the current module number is greater than the previous module number, the current quick output data, the current cycle number and the current module number will be overwritten in the memory with the previous quick output data, the previous cycle number and the current module number. Module number first; and
在所述规定时间获取所述存储器中所存储的迅捷输出数据。Acquire the quick output data stored in the memory at the prescribed time.
可任选地，所述输入数据是外汇汇兑业务的历史交易数据，并且所述规定时间是购汇结算周期期满时间。Optionally, the input data is historical transaction data of the foreign exchange exchange business, and the prescribed time is the expiration time of the foreign exchange purchase settlement period.
本发明的另一方面提供了一种用于加快预测的装置，包括：Another aspect of the present invention provides an apparatus for accelerating prediction, including:
预测模型系统，接收输入数据，其中所述预测模型系统包括串接的一个或多个预测模块；A prediction model system that receives input data, wherein the prediction model system includes one or more prediction modules connected in series;
连接至所述预测模型系统的一个或多个迅捷模块，所述迅捷模块的输入是所述一个或多个预测模块中的一个预测模块的输入并且所述迅捷模块提供迅捷输出数据，所述迅捷模块是根据所述预测模型系统中的至少一个预测模块生成的；One or more fast modules connected to the prediction model system, the input of the fast module is the input of one of the one or more prediction modules, and the fast module provides fast output data, the fast The module is generated according to at least one prediction module in the prediction model system;
存储器，所述存储器存储至少一个迅捷输出数据；A memory, the memory storing at least one quick output data;
处理器，在规定时间从所述一个或多个迅捷模块获取所述至少一个迅捷输出数据，将所述至少一个迅捷输出数据存储在所述存储器中，以及将所述至少一个预测输出数据中最新的预测输出数据确定为最终预测输出数据。The processor acquires the at least one quick output data from the one or more quick output data at a prescribed time, stores the at least one quick output data in the memory, and stores the latest one among the at least one predicted output data The predicted output data of is determined as the final predicted output data.
可任选地，所述预测模型系统产生模型输出数据，并且所述处理器被进一步配置成：Optionally, the predictive model system generates model output data, and the processor is further configured to:
如果在所述规定时间从所述预测模型系统接收到模型输出数据，则将所述模型输出数据确定为最终预测输出数据。If the model output data is received from the prediction model system at the prescribed time, the model output data is determined to be the final prediction output data.
可任选地，所述迅捷模块连接在所述预测模型系统的两个预测模块之间，并且所述至少一个预测模块包括与所述迅捷模块的输入相同的预测模块及其后续一个或多个预测模块。Optionally, the fast module is connected between two prediction modules of the prediction model system, and the at least one prediction module includes the same prediction module as the input of the fast module and one or more subsequent prediction modules Forecast module.
可任选地，所述迅捷模块是通过对所述至少一个预测模块调整参数和/或删除所述至少一个预测模块的一个或多个子模块来生成的。Optionally, the quickness module is generated by adjusting parameters for the at least one prediction module and/or deleting one or more sub-modules of the at least one prediction module.
可任选地，所述参数包括训练步长、训练次数和/或误差精度。Optionally, the parameters include training step size, training times and/or error accuracy.
可任选地，所述处理器被进一步配置成：Optionally, the processor is further configured to:
确定所述至少一个迅捷输出数据中的每一者的模块编号和周期编号，其中连接到预测模型系统下游的迅捷模块的迅捷输出数据的模块编号大于连接到预测模型系统上游的迅捷模块的迅捷输出数据的模块编号，所述周期编号是生成预测输出数据的周期的编号；Determine the module number and period number of each of the at least one quick output data, wherein the module number of the quick output data of the quick output data connected to the downstream of the predictive model system is greater than the quick output of the quick module connected to the upstream of the predictive model system The module number of the data, where the period number is the number of the period in which the predicted output data is generated;
将所述至少一个迅捷输出数据中的每一者与其周期编号、模块编号一起存储在存储器中；Storing each of the at least one quick output data in the memory together with its cycle number and module number;
在所述规定时间在所述存储器中查找周期编号最大的一个或多个迅捷输出数据；以及Searching for one or more quick output data with the largest cycle number in the memory at the specified time; and
选择所述一个或多个预测输出数据中模块编号最大的迅捷输出数据作为所述最终预测输出数据。Select the quick output data with the largest module number among the one or more predicted output data as the final predicted output data.
可任选地，所述处理器被进一步配置成：Optionally, the processor is further configured to:
确定所获取的当前迅捷输出数据的当前周期编号和当前模块编号；Determine the current cycle number and current module number of the acquired current quick output data;
将所述当前周期编号与存储器中所存储的在先迅捷输出数据的在先周期编号进行比较；Comparing the current cycle number with the previous cycle number of the previous quick output data stored in the memory;
如果所述当前周期编号大于所述在先周期编号，则在所述存储器中将所述当前迅捷输出数据以及当前周期编号和当前模块编号覆盖所述在先迅捷输出数据以及在先周期编号和在先模块编号；If the current cycle number is greater than the previous cycle number, the current fast output data, current cycle number and current module number will be overwritten in the memory with the previous fast output data, previous cycle number and previous cycle number. First module number;
如果所述当前周期编号等于所述在先周期编号，则将所述当前模块编号与所述在先模块编号进行比较；If the current cycle number is equal to the previous cycle number, comparing the current module number with the previous module number;
如果所述当前模块编号大于所述在先模块编号，则在所述存储器中将所述当前迅捷输出数据以及当前周期编号和当前模块编号覆盖所述在先迅捷输出数据以及在先周期编号和在先模块编号；以及If the current module number is greater than the previous module number, the current quick output data, the current cycle number and the current module number will be overwritten in the memory with the previous quick output data, the previous cycle number and the current module number. Module number first; and
在所述规定时间获取所述存储器中所存储的迅捷输出数据。Acquire the quick output data stored in the memory at the prescribed time.
可任选地，所述输入数据是外汇汇兑业务的历史交易数据，并且所述规定时间是购汇结算周期期满时间。Optionally, the input data is historical transaction data of the foreign exchange exchange business, and the prescribed time is the expiration time of the foreign exchange purchase settlement period.
本发明的进一步方面提供了一种计算机设备，包括：A further aspect of the present invention provides a computer device including:
处理器；以及Processor; and
被安排成存储计算机可执行指令的存储器，所述可执行指令在被执行时使所述处理器执行以下操作：A memory arranged to store computer-executable instructions that, when executed, cause the processor to perform the following operations:
由预测模型系统接收输入数据，其中所述预测模型系统包括串接的一个或多个预测模块，一个或多个迅捷模块连接到所述预测模型系统，所述迅捷模块的输入是所述一 个或多个预测模块中的一个预测模块的输入并且所述迅捷模块提供迅捷输出数据；The input data is received by a predictive model system, wherein the predictive model system includes one or more predictive modules connected in series, one or more rapid modules are connected to the predictive model system, and the input of the rapid module is the one or The input of one prediction module of the plurality of prediction modules and the fast module provides fast output data;
在规定时间从所述一个或多个迅捷模块获取至少一个迅捷输出数据；以及Obtain at least one quick output data from the one or more quick modules at a specified time; and
将所述至少一个迅捷输出数据中最新的迅捷输出数据确定为最终预测输出数据，Determining the latest quick output data in the at least one quick output data as the final predicted output data,
其中，所述迅捷模块是根据所述预测模型系统中的至少一个预测模块生成的。Wherein, the quick module is generated according to at least one prediction module in the prediction model system.
本公开通过在预测模型系统中纳入迅捷模块来加速模型预测，迅捷模块可以缩短获得预测值的运行时长。例如，迅捷模块可以减少模型特征、减少模型训练次数、减少训练样本数等，从而加快预测速度，保证在规定时间内有预测输出数据。进一步，通过选择最新(或即最精确)的预测输出数据，保证得到当前最精确的预测输出数据。The present disclosure accelerates model prediction by incorporating a fast module in the prediction model system, and the fast module can shorten the running time for obtaining the predicted value. For example, the fast module can reduce model features, reduce the number of model training times, reduce the number of training samples, etc., thereby speeding up the prediction speed and ensuring that there is prediction output data within a specified time. Further, by selecting the latest (or the most accurate) forecast output data, it is guaranteed that the most accurate forecast output data currently is obtained.
附图说明Description of the drawings
图1是用于模型预测的系统示图。Figure 1 is a system diagram for model prediction.
图2是根据本公开的一个实施例的用于加速模型预测的系统结构图。Fig. 2 is a structural diagram of a system for accelerating model prediction according to an embodiment of the present disclosure.
图3示出了根据本公开的一个实施例的构建迅捷模块的示例。Fig. 3 shows an example of constructing a fast module according to an embodiment of the present disclosure.
图4示出了根据本公开的各方面的用于加速模型预测的系统的结构图。Fig. 4 shows a structural diagram of a system for accelerating model prediction according to various aspects of the present disclosure.
图5解说了根据本公开的各方面的用于加速模型预测的方法的流程图。Figure 5 illustrates a flowchart of a method for accelerating model prediction according to aspects of the present disclosure.
具体实施方式detailed description
为让本申请的上述目的、特征和优点能更明显易懂，以下结合附图对本申请的具体实施方式作详细说明。In order to make the above objectives, features and advantages of this application more comprehensible, specific implementations of this application will be described in detail below with reference to the accompanying drawings.
在下面的描述中阐述了很多具体细节以便于充分理解本发明，但是本申请还可以采用其它不同于在此描述的其它方式来实施，因此本发明不受下面公开的具体实施例的限制。In the following description, many specific details are set forth in order to fully understand the present invention, but this application can also be implemented in other ways different from those described herein, so the present invention is not limited by the specific embodiments disclosed below.
图1示出了用于模型预测的系统示图。Figure 1 shows a system diagram for model prediction.
如图1所示，用于模型预测的系统可包括计算平台100和定时器102。定时器102也可被纳入计算平台100中。As shown in FIG. 1, the system for model prediction may include a computing platform 100 and a timer 102. The timer 102 can also be incorporated into the computing platform 100.
计算平台100例如是支付宝服务器，可包括预测模型系统101。预测模型系统101可接收样本数据。一般而言，预测模型系统101可包括一个或多个串接的预测模块。样 本数据依次通过各个预测模块进行处理以得到预测结果。The computing platform 100 is, for example, an Alipay server, and may include a predictive model system 101. The predictive model system 101 may receive sample data. Generally speaking, the prediction model system 101 may include one or more cascaded prediction modules. The sample data is processed by each prediction module in turn to obtain the prediction result.
例如，在国际汇兑业务中，可以将当前预测周期之前的历史交易数据输入预先训练好的预测模型系统101中进行预测。例如，第i个预测周期可产生预测结果V i，第i+1个预测周期可产生预测结果V i+1。在每个预测周期结束时，新的样本数据将输入预测模型系统进行下一周期的预测。作为一个示例，该预测周期可以是一个小时。 For example, in an international exchange business, historical transaction data before the current forecast period can be input into the pre-trained forecast model system 101 for forecasting. For example, the i th cycle can generate a prediction result of the prediction V i, i + 1-th cycle can generate a prediction result of the prediction V i + 1. At the end of each prediction period, new sample data will be input to the prediction model system for the next period of prediction. As an example, the forecast period may be one hour.
定时器102对计算平台100的操作进行时钟控制。具体而言，可对计算平台100的操作进行周期控制。定时器102在当前预测周期结束时向计算平台100发送信号指令预测模型系统101进行下一周期的预测。The timer 102 clocks the operation of the computing platform 100. Specifically, the operation of the computing platform 100 may be periodically controlled. The timer 102 sends a signal to the computing platform 100 at the end of the current prediction period to instruct the prediction model system 101 to perform prediction for the next period.
因此，模型预测需要在每个周期结束之前完成，否则下一周期的预测将中断当前周期的预测，由此将缺失当前周期输出数据。Therefore, the model prediction needs to be completed before the end of each period, otherwise the prediction of the next period will interrupt the prediction of the current period, and the output data of the current period will be missing.
例如，在国际汇兑业务中，预测模型系统需要在当前购汇周期结束前输出数据预测值。For example, in the international currency exchange business, the forecasting model system needs to output data forecasts before the end of the current foreign exchange purchase cycle.
但模型运行的时效受到上游数据时效性、运行环境的资源可用度的影响，有时在当前周期结束前，预测模型系统未能完成当前周期的预测操作，由此丢失了当前预测输出数据。However, the timeliness of model operation is affected by the timeliness of upstream data and the resource availability of the operating environment. Sometimes, before the end of the current cycle, the forecasting model system fails to complete the forecasting operation of the current cycle, thus losing the current forecast output data.
本公开的目的在于解决在预测模型系统运行超时的情况下无法准时输出数据预测值的技术问题。The purpose of the present disclosure is to solve the technical problem that the predicted value of the data cannot be output on time when the prediction model system runs overtime.
在预测链路中，预测模型系统可被构建成几个预测模块的串接。例如，外汇交易量预测模型系统可包括数据清洗模块、量化建模模块、异常检测模块、以及业务决策调整模块。在外汇业务中，这几个模块需要在给定时间(例如，一个小时)内完成运行并最终生成预测值。但是运行链路由于实际中受到上游数据时效性以及运行环境资源可用度影响，容易出现在给定时间内无法完成所有模块的运行的情形。以下详细解说本公开的实施例。In the prediction link, the prediction model system can be constructed as a series connection of several prediction modules. For example, the foreign exchange transaction volume prediction model system may include a data cleaning module, a quantitative modeling module, an abnormality detection module, and a business decision adjustment module. In the foreign exchange business, these modules need to be completed within a given time (for example, one hour) and finally generate predicted values. However, due to the fact that the operating link is affected by the timeliness of upstream data and the availability of operating environment resources, it is prone to fail to complete the operation of all modules within a given time. The embodiments of the present disclosure are explained in detail below.
图2是根据本公开的一个实施例的用于加速模型预测的系统示图。Fig. 2 is a diagram of a system for accelerating model prediction according to an embodiment of the present disclosure.
如图2所示，用于加速模型预测的系统包括左边的常规预测部分和右边的迅捷部分(框内的部分)。在该示例中，周期为一个小时。应理解，其它周期也在本公开的构想中。As shown in Figure 2, the system for accelerating model prediction includes the conventional prediction part on the left and the quick part (the part in the box) on the right. In this example, the period is one hour. It should be understood that other cycles are also within the concept of this disclosure.
常规预测部分根据实际预测流程的需要可拆分成串接的一个或多个预测模块，例 如包括数据清洗模块、量化建模模块、异常检测模块以及业务决策模块。The conventional forecasting part can be split into one or more forecasting modules connected in series according to the needs of the actual forecasting process, for example, including a data cleaning module, a quantitative modeling module, an anomaly detection module, and a business decision module.
数据清洗模块可完成最新的数据(例如，最近一个小时的数据)的清洗处理。例如，样本数据通过从多个业务系统中抽取而来而且包含历史数据，其中一些数据是错误数据、有的数据相互之间有冲突，这些错误的或有冲突的数据被称为“脏数据”。需要按照一定规则将“脏数据”“洗掉”，即，过滤不符合要求的数据。The data cleaning module can complete the cleaning of the latest data (for example, the data of the last hour). For example, sample data is extracted from multiple business systems and contains historical data. Some of the data is wrong data, and some data conflicts with each other. These wrong or conflicting data are called "dirty data." . The "dirty data" needs to be "washed out" according to certain rules, that is, data that does not meet the requirements is filtered.
量化建模模块是核心的代码模块，可完成特征构建、模型训练、预测值输出数据等功能。The quantitative modeling module is the core code module, which can complete functions such as feature construction, model training, and output data of predicted values.
异常检测模块可对前一模块的预测值进行异常检测。The anomaly detection module can perform anomaly detection on the predicted value of the previous module.
业务决策模块可根据业务经验对预测值进行相应调整。The business decision-making module can adjust the predicted value accordingly based on business experience.
数据输入经过常规部分的各模块处理之后生成最终预测值。The data input is processed by the modules in the regular part to generate the final predicted value.
数据清洗模块、量化建模模块、异常检测模块以及业务决策模块是串接的，每个模块需要等待上游模块的输出数据来进行处理，同时每个模块的处理速度会受到运行环境资源可用度的影响，因此在规定时间内不一定能够输出数据最终预测值。The data cleaning module, quantitative modeling module, anomaly detection module and business decision module are connected in series. Each module needs to wait for the output data of the upstream module for processing. At the same time, the processing speed of each module will be affected by the availability of operating environment resources Therefore, the final predicted value of the data may not be output within the specified time.
本公开考虑到预测的时效性，在系统中添加了迅捷部分，迅捷部分包括一个或多个迅捷模块。The present disclosure takes into account the timeliness of prediction, and adds a quick part to the system, and the quick part includes one or more quick modules.
在图2所示的示例中，迅捷部分包括迅捷模块1、迅捷模块2和迅捷模块3，分别快速地生成精度较低的预测值V 1、V 2和V 3。换言之，迅捷模块牺牲一定的预测精度以加快预测速度。 In the example shown in FIG. 2, the quick part includes the quick module 1, the quick module 2 and the quick module 3, which respectively quickly generate prediction values V 1 , V 2 and V 3 with lower accuracy. In other words, the fast module sacrifices a certain prediction accuracy to speed up the prediction.
如图所示，迅捷模块1连接至数据清洗模块的输出数据，并且可以根据常规部分中数据清洗模块的后续模块来生成，例如是量化建模模块、异常检测模块以及业务决策模块的简化版本。迅捷模块1可在短时间内完成与量化建模模块、异常检测模块以及业务决策模块相关的训练，并且输出数据预测值V 1。预测值V 1与最终预测值V 4相比精度较低，但输出数据时间早。 As shown in the figure, the fast module 1 is connected to the output data of the data cleaning module, and can be generated according to the subsequent modules of the data cleaning module in the regular part, such as a simplified version of the quantitative modeling module, anomaly detection module, and business decision module. The fast module 1 can complete the training related to the quantitative modeling module, anomaly detection module and the business decision module in a short time, and output the data prediction value V 1 . The prediction value V 1 has a lower accuracy than the final prediction value V 4 , but the output data time is earlier.
由此包括数据清洗模块和迅捷模块1的预测路径可快速地生成精度较低的预测值V 1。 Therefore, the prediction path including the data cleaning module and the fast module 1 can quickly generate the prediction value V 1 with lower accuracy.
迅捷模块2连接至量化建模模块的输出数据，并且可以根据常规部分中量化建模模块的后续模块来生成，例如是异常检测模块和业务决策模块的简化版本。迅捷模块2可在短时间内完成与异常检测模块以及业务决策模块相关的训练，并且输出数据预测值 V 2。预测值V 2与最终预测值V 4相比精度较低，但输出数据快；比预测值V 1精度高，但输出数据慢。 The fast module 2 is connected to the output data of the quantitative modeling module, and can be generated according to the subsequent modules of the quantitative modeling module in the regular part, such as a simplified version of the anomaly detection module and the business decision module. The fast module 2 can complete the training related to the anomaly detection module and the business decision module in a short time, and output the data prediction value V 2 . The prediction value V 2 has lower accuracy than the final prediction value V 4 , but the output data is faster; it has higher accuracy than the prediction value V 1 , but the output data is slow.
迅捷模块3连接至异常检测模块的输出数据，并且可以根据常规部分中异常检测模块的后续模块来生成，例如是业务决策模块的简化版本。迅捷模块3可在短时间内完成与业务决策模块相关的训练，并且输出数据预测值V 3。预测值V 3与最终预测值V 4相比精度较低，但输出数据快；比预测值V 2精度高，但输出数据慢。 The fast module 3 is connected to the output data of the anomaly detection module, and can be generated according to the subsequent modules of the anomaly detection module in the regular part, for example, a simplified version of the business decision module. The fast module 3 can complete the training related to the business decision module in a short time, and output the data prediction value V 3 . The prediction value V 3 has a lower accuracy than the final prediction value V 4 , but the output data is faster; it has a higher accuracy than the prediction value V 2 but the output data is slow.
迅捷模块与对应的常规预测模块相比，能在短时间内完成训练以及预测值输出数据，迅捷模块的目的在于减少运行时长来获得预测值，迅捷模块对常规预测模块的简化可包括：减少模型特征、减少模型训练次数、减少训练样本数等等。迅捷模块本质上是牺牲一定的预测精度以加快预测速度。Compared with the corresponding conventional prediction module, the fast module can complete the training and output data of the predicted value in a short time. The purpose of the fast module is to reduce the running time to obtain the predicted value. The simplification of the conventional prediction module by the fast module can include: reducing the model Features, reduce the number of model training, reduce the number of training samples, etc. The fast module essentially sacrifices a certain prediction accuracy to speed up the prediction.
在该实施例中，迅捷部分包括三个模块。但是，本领域技术人员可以理解，迅捷部分也可包括少于3个的模块，例如1个或2个。在进一步的实施例中，根据常规部分预测流程所包括的模块数N，迅捷部分可以包括1～N个模块。In this embodiment, the quick part includes three modules. However, those skilled in the art can understand that the quick part may also include less than 3 modules, such as 1 or 2 modules. In a further embodiment, according to the number N of modules included in the conventional partial prediction process, the quick part may include 1 to N modules.
在图2的实施例中，在规定时间，如果常规预测模型系统已输出数据了最终预测值V 4，则系统可以输出数据最终预测值V 4。如果常规预测模型系统尚未输出数据最终预测值V 4，则系统可以输出数据当前最精确的迅捷模块输出数据预测值，例如V 3。 In the embodiment of FIG. 2, at a prescribed time, if the conventional prediction model system has output the final predicted value V 4 of the data, the system can output the final predicted value V 4 of the data. If the conventional prediction model system has not output the final data prediction value V 4 , the system can output the data prediction value of the quickest module output data that is currently the most accurate, for example, V 3 .
进一步，如果在规定时间(例如，当前周期结束时)，当前周期没有任何迅捷模块输出数据，则可将上一周期的最终预测值V 4作为输出数据。 Further, if at a specified time (for example, at the end of the current cycle), there is no quick module output data in the current cycle, the final predicted value V 4 of the previous cycle can be used as the output data.
本公开的用于加速模型预测的系统能够保证在规定时间有预测值输出数据，并且输出数据当前最精确版本的预测值。The system for accelerating model prediction of the present disclosure can ensure that there is a predicted value output data at a specified time, and the predicted value of the current most accurate version of the output data.
图3示出了根据本公开的一个实施例的构建迅捷模块的示例。其中，用于加速模型预测的系统的常规部分包括量化建模模块、异常检测模块以及业务决策模块。Fig. 3 shows an example of constructing a fast module according to an embodiment of the present disclosure. Among them, the conventional part of the system for accelerating model prediction includes a quantitative modeling module, an anomaly detection module, and a business decision module.
如图3所示，量化建模模块可包括AUTO_ML(自动化学习框架)子模块、经典时序预测子模块、ARIMA(Autoregressive Integrated Moving Average Model，自回归积分滑动平均模型)子模块、Holt-Winters(三次指数平滑)子模块、同环比子模块。As shown in Figure 3, the quantitative modeling module can include the AUTO_ML (automated learning framework) sub-module, the classic time series prediction sub-module, the ARIMA (Autoregressive Integrated Moving Average Model, autoregressive integrated moving average model) sub-module, Holt-Winters (three times Exponential smoothing) sub-module, same-ring comparison sub-module.
异常检测模块包括基于实时数据源的异常检测子模块和基于离线数据源的异常检测子模块。The anomaly detection module includes an anomaly detection sub-module based on real-time data sources and an anomaly detection sub-module based on offline data sources.
业务决策模块包括基于外部运营方信息输入的调整子模块和基于内部运营方的信 息输入的调整子模块。The business decision module includes an adjustment sub-module based on information input from an external operator and an adjustment sub-module based on information input from an internal operator.
如上所述，迅捷模块1可以是量化建模模块、异常检测模块以及业务决策模块的简化版本。As mentioned above, the quick module 1 may be a simplified version of the quantitative modeling module, anomaly detection module, and business decision module.
例如，迅捷模块1可通过针对量化建模模块调整参数(例如调整训练步长、训练次数和误差精度)来得到不同的运行速度，从而加快模块处理时间；迅捷模块1可针对异常检测模块去掉基于实时数据源的异常检测子模块，而只纳入基于离线数据源的异常检测子模块的处理；迅捷模块1可针对业务决策模块去掉基于外部运营方信息输入的调整子模块，而只纳入基于内部运营方的信息输入的调整子模块的处理。For example, the fast module 1 can obtain different operating speeds by adjusting parameters for the quantitative modeling module (such as adjusting the training step size, training times and error accuracy), thereby speeding up the processing time of the module; the fast module 1 can remove the based on the abnormal detection module The anomaly detection sub-module of real-time data sources is only included in the processing of the anomaly detection sub-module based on offline data sources; the fast module 1 can remove the adjustment sub-module based on the information input of the external operator for the business decision module, and only include the sub-module based on internal operation The processing of the adjustment sub-module of the party’s information input.
迅捷模块2可以是异常检测模块和业务决策模块的简化版本。例如，迅捷模块2可针对异常检测模块去掉基于实时数据源的异常检测子模块，而只纳入基于离线数据源的异常检测子模块的处理；针对业务决策模块去掉基于外部运营方信息输入的调整子模块，而只纳入基于内部运营方的信息输入的调整子模块的处理。The fast module 2 can be a simplified version of the anomaly detection module and the business decision module. For example, the fast module 2 can remove the anomaly detection sub-module based on real-time data sources for the anomaly detection module, and only incorporate the processing of the anomaly detection sub-module based on offline data sources; for the business decision module, remove the adjustment sub-module based on external operator information input. Module, and only include the processing of the adjustment sub-module based on the information input of the internal operator.
迅捷模块3可以是业务决策模块的简化版本。例如，迅捷模块3可针对业务决策模块去掉基于外部运营方信息输入的调整子模块，而只纳入基于内部运营方的信息输入的调整子模块的处理。The quick module 3 can be a simplified version of the business decision module. For example, the fast module 3 can remove the adjustment sub-module based on the information input of the external operator for the business decision module, and only incorporate the processing of the adjustment sub-module based on the information input of the internal operator.
请注意，以上列出了构建迅捷模块的一个具体示例，但本领域技术人员知晓，其它实现也是可能的。例如，预测模型系统可包括各种预测模块，并且各迅捷模块也可以调整异常检测模块和/或业务决策模块的参数、去掉其它子模块。Please note that a specific example of constructing a quick module is listed above, but those skilled in the art know that other implementations are also possible. For example, the prediction model system may include various prediction modules, and each quick module may also adjust the parameters of the anomaly detection module and/or the business decision module, and remove other sub-modules.
在一些方面，迅捷模块也可以去掉后续模块。例如，迅捷模块1可以根据量化建模模块和业务决策模块生成，而无需纳入异常检测模块。并且在其它实现中，预测系统的常规部分可具有不同的预测模块，由此系统具有相应不同的迅捷模块，其中各迅捷模块可以根据对应的一个或多个常规预测模块来生成，例如，迅捷模块可以是对应的一个或多个常规预测模块的简化版本。In some aspects, the quick module can also remove the subsequent modules. For example, the fast module 1 can be generated based on the quantitative modeling module and the business decision module, without including the anomaly detection module. And in other implementations, the conventional part of the prediction system may have different prediction modules, so the system has correspondingly different quick modules, where each quick module can be generated according to one or more corresponding conventional prediction modules, for example, the quick module It can be a simplified version of the corresponding one or more conventional prediction modules.
图4示出了根据本公开的各方面的用于加速模型预测的系统400的结构图。FIG. 4 shows a structural diagram of a system 400 for accelerating model prediction according to various aspects of the present disclosure.
如图4所示，系统400可包括预测模型系统部分401、迅捷模块部分402、存储器403、以及处理器404。As shown in FIG. 4, the system 400 may include a predictive model system part 401, a fast module part 402, a memory 403, and a processor 404.
预测模型系统部分401可包括串接的预测模块1、预测模块2、……预测模块N-1、以及预测模块N。即，输入样本被输入到预测模块1，预测模块1的预测结果被输入预测模块2，预测模块2的预测结果被输入预测模块3，……预测模块N-1的预测结果被 输入预测模块N，预测模块N的输出数据为预测模型系统的最终输出数据。The prediction model system part 401 may include a prediction module 1, a prediction module 2, ... a prediction module N-1, and a prediction module N connected in series. That is, the input sample is input to the prediction module 1, the prediction result of the prediction module 1 is input to the prediction module 2, the prediction result of the prediction module 2 is input to the prediction module 3, ... the prediction result of the prediction module N-1 is input to the prediction module N , The output data of the prediction module N is the final output data of the prediction model system.
数据依次经过预测模块1-N的处理产生最终预测结果V N(即，最后一个预测模块N的输出数据)。 The data is sequentially processed by the prediction modules 1-N to produce the final prediction result V N (that is, the output data of the last prediction module N).
预测模型系统运行的时效往往受到上游数据时效性、运行环境的资源可用度的影响，有时在当前周期结束前，预测模型系统未能完成当前周期的预测操作，由此缺失了当前预测输出数据。为了确保在当前周期能够获得输出数据，在系统中加入迅捷模块部分402。The timeliness of the forecasting model system is often affected by the timeliness of upstream data and the resource availability of the operating environment. Sometimes, before the end of the current cycle, the forecasting model system fails to complete the forecasting operation of the current cycle, thus missing the current forecast output data. In order to ensure that the output data can be obtained in the current cycle, a fast module part 402 is added to the system.
迅捷模块部分402可包括迅捷模块0、迅捷模块1、迅捷模块2、……迅捷模块N-1中的一者或多者。The fast module part 402 may include one or more of the fast module 0, the fast module 1, the fast module 2, ... the fast module N-1.
输入样本可被输入到迅捷模块0，迅捷模块0可以根据预测模块1—预测模块N来生成，例如可以是预测模块1—预测模块N的简化版本。例如，调整预测模块1—预测模块N中的一个或多个模块的参数和/或删除该一个或多个模块的一个或多个子模块。The input sample can be input to the fast module 0, and the fast module 0 can be generated according to the prediction module 1—the prediction module N, for example, it may be a simplified version of the prediction module 1—the prediction module N. For example, adjusting the parameters of one or more modules in the prediction module 1-prediction module N and/or deleting one or more sub-modules of the one or more modules.
迅捷模块1接收预测模块1的输出数据，并且可以根据后续的预测模块2—预测模块N来生成，例如可以是预测模块2—预测模块N的简化版本。例如，调整预测模块2—预测模块N中的一个或多个模块的参数和/或删除该一个或多个模块的一个或多个子模块。The quick module 1 receives the output data of the prediction module 1, and can generate it according to the subsequent prediction module 2—the prediction module N, for example, it may be a simplified version of the prediction module 2—the prediction module N. For example, adjust the parameters of one or more modules in the prediction module 2—prediction module N and/or delete one or more sub-modules of the one or more modules.
迅捷模块2接收预测模块2的输出数据，并且可以根据是后续的预测模块3—预测模块N来生成，例如可以是预测模块3—预测模块N的简化版本；依此类推。例如，调整预测模块3—预测模块N中的一个或多个模块的参数和/或删除该一个或多个模块的一个或多个子模块。The fast module 2 receives the output data of the prediction module 2, and can be generated according to the subsequent prediction module 3—prediction module N, for example, it may be a simplified version of prediction module 3—prediction module N; and so on. For example, adjust the parameters of one or more modules in the prediction module 3—prediction module N and/or delete one or more sub-modules of the one or more modules.
迅捷模块0的输出数据为V 0，迅捷模块1的输出数据为V 1，迅捷模块2的输出数据为V 2，……迅捷模块N-1的输出数据为V N-1，预测模型系统的最后一个模块N的输出数据为V N。V 0、V 1、V 2、……V N-1、V N可被输入到存储器403中进行存储。 The output data of fast module 0 is V 0 , the output data of fast module 1 is V 1 , the output data of fast module 2 is V 2 , ... the output data of fast module N-1 is V N-1 , the prediction model system The output data of the last module N is V N. V 0 , V 1 , V 2 , ... V N-1 , V N can be input to the memory 403 for storage.
处理器404可对V 0、V 1、V 2、……V N-1、V N进行编号，并将每个预测输出数据与其编号一起存储在存储器中。 The processor 404 may number V 0 , V 1 , V 2 , ... V N-1 , V N , and store each predicted output data together with its number in the memory.
该编号可包括模块编号和周期编号。The number can include the module number and the cycle number.
例如，处理器404可识别预测输出数据V i来自哪个迅捷模块或来自最后一个预测模块。例如，迅捷模块i或预测模块N在输出数据预测结果时，可将预测结果与模块标 识一起发送给存储器403和处理器404。处理器404通过模块标识来识别出该预测结果来自哪个模块。 For example, processor 404 may identify predicted output data from a module V i which rapid or a prediction from the last module. For example, when the fast module i or the prediction module N outputs the data prediction result, the prediction result and the module identifier may be sent to the memory 403 and the processor 404 together. The processor 404 uses the module identification to identify which module the prediction result comes from.
处理器404随后可按照该迅捷模块或最后一个预测模块的标识确定V i的模块编号。V i的模块编号对应于生成V i的迅捷模块所连接的预测模型系统节点在所述预测模型系统中的位置。如图4所示，连接到下游(离输入端较远)的预测模型系统节点的迅捷模块的模块编号可大于连接到上游(离输入端较近)的预测模型系统节点的迅捷模块的模块编号。 The processor 404 can then determine the module number of Vi according to the identification of the fast module or the last prediction module. The module number of V i corresponds to the position in the predictive model system of the node of the predictive model system connected to the fast module generating V i . As shown in Figure 4, the module number of the fast module connected to the predictive model system node downstream (farther from the input end) can be greater than the module number of the fast module connected to the upstream predictive model system node (closer to the input end) .
作为一个示例，迅捷模块0所连接的预测模型系统节点的后一模块为预测模块1，则来自迅捷模块0的预测输出数据的编号可为0；迅捷模块1连接在预测模块1与预测模块2之间，则来自迅捷模块1的预测输出数据的编号可为1；迅捷模块2连接在预测模块2与预测模块3之间，则来自迅捷模块2的预测输出数据的编号为2，依此类推；来自迅捷模块N-1的预测输出数据的编号可为N-1；如果V i来自最后一个预测模块，则V i的编号大于所有来自迅捷模块的预测输出数据的编号，例如，可以为迅捷模块的数目加一(例如，N)。 As an example, the next module of the prediction model system node connected to the fast module 0 is the prediction module 1, then the number of the prediction output data from the fast module 0 can be 0; the fast module 1 is connected to the prediction module 1 and the prediction module 2. If the number of the predicted output data from the fast module 1 can be 1; if the fast module 2 is connected between the prediction module 2 and the prediction module 3, the number of the predicted output data from the fast module 2 is 2, and so on ; numbering from the predicted output data prompt module N-1 is may be N-1; if V i from the last prediction module, the number i is V is greater than all numbers predicted output data from the fast block, e.g., can be fast The number of modules plus one (for example, N).
请注意，以上取值仅仅是示例性的，其它取值也在本发明的构想中，只要V i的模块编号对应于生成V i的迅捷模块所连接的预测模型系统节点的后一预测模块在所述预测模型系统中的顺序即可。 Note that the above values are merely exemplary, and are also contemplated in the present invention, other values, the prediction module predicts a model system nodes as long as the fast block number V i V i correspond to the generating module to be connected in The order in the prediction model system is sufficient.
进一步，处理器404可识别预测输出数据V i是在哪个周期产生的。例如，迅捷模块i或预测模块N在输出数据预测结果时，将预测结果与产生该预测结果时的周期的周期编号一起发送给存储器403和处理器404。处理器404通过周期编号来识别出该预测结果是在哪个周期产生的。 Further, processor 404 may identify the predicted output data in which the period V i is generated. For example, when the quick module i or the prediction module N outputs the data prediction result, the prediction result is sent to the memory 403 and the processor 404 together with the period number of the period when the prediction result is generated. The processor 404 uses the cycle number to identify the cycle in which the prediction result is generated.
例如，迅捷模块i或预测模块N中可具有周期编号计数器，每进入一个新的周期，计数器中的周期编号值可递增。For example, the quick module i or the prediction module N may have a cycle number counter, and each time a new cycle is entered, the cycle number value in the counter can be incremented.
在规定时间(例如，预测周期期满时)，处理器404确定存储器403中所存储的最新版本的输出数据V 0、V 1、V 2、……V N-1、或V N作为模型预测结果进行输出数据。例如，如果存储器中具有当前周期的迅捷模块输出数据V 0、V 1、V 2、V 3，则输出数据V 3作为模型预测结果。 At a prescribed time (for example, when the prediction period expires), the processor 404 determines the latest version of the output data V 0 , V 1 , V 2 , ... V N-1 , or V N stored in the memory 403 as the model prediction The result is output data. For example, if the fast module with the current cycle output data V 0 , V 1 , V 2 , V 3 in the memory, the output data V 3 is used as the model prediction result.
进一步，模型的预测是周期性的，如果在规定时间内，任一迅捷模块/预测模型系统的最后一个模块关于当前周期均没有输出数据，则可将前一周期的最终预测值作为输 出数据。Further, the prediction of the model is periodic. If the last module of any quick module/prediction model system has no output data for the current period within a specified time, the final prediction value of the previous period can be used as the output data.
例如，外汇交易中的预测周期可以为一个小时，如果当前小时没有任何迅捷模块输出数据V i且没有预测模型系统输出数据V N，则可将上一小时的最终预测值作为输出数据。 For example, the foreign exchange transactions may be one hour prediction period, if no current hour fast data module output V i and no output data model system predictive V N, may be the final one hour predicted value as output data.
存储器中的每个条目包括迅捷模块/预测模型系统的最后一个模块的输出数据及其模块编号和/或周期编号。Each entry in the memory includes the output data of the last module of the fast module/predictive model system and its module number and/or cycle number.
表1示出了存储预测值的存储器的一个示例。Table 1 shows an example of a memory storing predicted values.
周期编号Cycle number 模块编号Module number 预测值Predictive value
00 00 V 0 V 0
00 11 V 1 V 1
……... ……... ……...
M-1M-1 NN V N V N
MM 00 V 0 V 0
MM 11 V 1 V 1
MM 22 V 2 V 2
……... ……... ……...
MM N-1N-1 V N-1 V N-1
MM NN V N V N
M+1M+1 00 V 0 V 0
M+1M+1 11 V 1 V 1
……... ……... ……...
表1Table 1
在规定需要输出数据的时间，可首先查找存储器中周期编号最大的一个或多个条目(即，最新周期的一个或多个预测输出数据)，随后在周期编号最大的一个或多个条目中查找模块编号最大的条目(即，最新周期的最新预测输出数据)，将该条目对应的预测值作为输出数据。At the time when data is required to be output, search for one or more entries with the largest period number in the memory (that is, one or more predicted output data for the latest period), and then search for one or more entries with the largest period number The entry with the largest module number (that is, the latest predicted output data of the latest period) uses the predicted value corresponding to the entry as the output data.
请注意，虽然表1列出了迅捷模块部分包括全部迅捷模块0-N-1的情形，但本领域技术人员知晓，迅捷模块部分可以包括迅捷模块0-N-1中的一者或多者，只要其编号体现对应模块的先后顺序即可。即，本领域技术人员能够根据实际需要来选择这些迅 捷模块中的一者或多者来加快模型预测。例如，迅捷模块0可被省略，由此样本仅被输入到预测模块1以进行后续处理。Please note that although Table 1 lists the cases where the fast module part includes all fast modules 0-N-1, those skilled in the art know that the fast module part may include one or more of the fast modules 0-N-1 , As long as its number reflects the sequence of the corresponding modules. That is, those skilled in the art can select one or more of these quick modules to speed up model prediction according to actual needs. For example, the fast module 0 can be omitted, and thus the samples are only input to the prediction module 1 for subsequent processing.
在另一方面，存储器也可以仅存储最新的预测输出数据。例如，处理器404可识别当前预测输出数据V i的当前周期编号和当前模块编号，将当前周期编号与存储器403中所存储的在先预测输出数据的在先周期编号进行比较，如果当前周期编号大于所述在先周期编号，即，当前预测输出数据是比所存储的预测输出数据更新的预测周期的输出数据，则在存储器中将当前预测输出数据以及当前周期编号和当前模块编号覆盖在先预测输出数据以及在先周期编号和在先模块编号。 On the other hand, the memory may only store the latest predicted output data. For example, processor 404 may identify the current prediction data output V i of the current period and the current module ID number, the previous period preceding the current cycle of the predicted output data and the number stored in memory 403 is compared numbers, if the current cycle number Greater than the previous period number, that is, the current predicted output data is the output data of the predicted period newer than the stored predicted output data, then the current predicted output data, the current period number and the current module number are overwritten in the memory. Forecast output data and previous period number and previous module number.
如果所述当前周期编号等于所述在先周期编号，即，当前预测输出数据与所存储的预测输出数据是在相同的周期中生成的，则将当前模块编号与存储器403中所存储的在先预测输出数据的在先模块编号进行比较；如果当前模块编号大于所述在先模块编号，即，当前预测输出数据是比在先预测输出数据更新的版本，则在存储器403中将当前预测输出数据以及当前周期编号和当前模块编号覆盖在先预测输出数据以及在先周期编号和在先模块编号。使用该方式，可以有效地节省存储空间。If the current cycle number is equal to the previous cycle number, that is, the current predicted output data and the stored predicted output data are generated in the same cycle, then the current module number is compared with the previous stored in the memory 403 The previous module numbers of the predicted output data are compared; if the current module number is greater than the previous module number, that is, the current predicted output data is a newer version than the previous predicted output data, the current predicted output data is stored in the memory 403 And the current cycle number and the current module number cover the previous forecast output data, the previous cycle number and the previous module number. In this way, storage space can be effectively saved.
以上描述了存储预测输出数据的具体示例，但本领域技术人员将领会，其它存储方式也是可能的，例如，向量形式：V i(模块编号，周期编号)。 A specific example described above the predicted output data storage, those skilled in the art will appreciate that other storage are possible, e.g., in vector form: V i (module number, cycle number).
图5解说了根据本公开的各方面的用于加速模型预测的方法的流程图。Figure 5 illustrates a flowchart of a method for accelerating model prediction according to aspects of the present disclosure.
该方法可在如图4所示的系统中实现。在该系统中，预测模型系统可包括串接的多个预测模块，输入样本和/或预测模块的输出数据可被输入到下一预测模块和对应的迅捷模块，并且其中迅捷模块是根据预测模型系统中的至少一个预测模块生成的，例如，所连接的预测模型系统节点后面(下游)的一个或多个预测模块的简化版本。This method can be implemented in the system shown in Figure 4. In this system, the prediction model system may include multiple prediction modules connected in series, the input samples and/or the output data of the prediction module can be input to the next prediction module and the corresponding fast module, and the fast module is based on the prediction model Generated by at least one prediction module in the system, for example, a simplified version of one or more prediction modules behind (downstream) the connected prediction model system node.
预测模块的简化版本可以通过对预测模块调整参数和/或删除预测模块中的子模块来得到。调整的参数可包括训练步长、训练次数和/或误差精度。The simplified version of the prediction module can be obtained by adjusting the parameters of the prediction module and/or deleting the sub-modules in the prediction module. The adjusted parameters may include training step size, training times, and/or error accuracy.
该方法包括在步骤502，由预测模型系统接收输入数据。The method includes in step 502, receiving input data by a predictive model system.
该输入数据可以是待处理样本。The input data can be samples to be processed.
如图4所示，数据可被输入到预测模块1，如果存在迅捷模块0，还可输入到迅捷模块0。As shown in Figure 4, the data can be input to the prediction module 1, and if the fast module 0 exists, it can also be input to the fast module 0.
每个预测模块的处理结果可被输入到下一个预测模块和对应的迅捷模块(若有)， 迅捷模块可以是该下一个预测模块及其后续一个或多个预测模块的简化版本。The processing result of each prediction module can be input to the next prediction module and the corresponding quick module (if any), and the quick module can be a simplified version of the next prediction module and one or more subsequent prediction modules.
在步骤504，在规定时间(例如，在当前周期期满时)获取多个迅捷模块的预测输出数据(可被称为迅捷输出数据)和最后一个预测模块的预测输出数据(可被称为模型输出数据)中的至少一者。In step 504, at a prescribed time (for example, at the expiration of the current period), the predicted output data of multiple quick modules (which can be called quick output data) and the predicted output data of the last prediction module (which can be called model Output data) at least one of them.
例如，可以对迅捷输出数据和模型输出数据进行编号(模块编号和周期编号)，其中迅捷输出数据的模块编号对应于生成该迅捷输出数据的迅捷模块所连接的预测模型系统节点在所述预测模型系统中的位置，连接到预测模型系统下游节点的迅捷模块的输出数据的模块编号大于连接到预测模型系统上游节点的迅捷模块的输出数据的模块编号；并且模型输出数据的模块编号大于每个迅捷预测输出数据的模块编号，周期编号是生成预测输出数据的周期的编号。前述预测输出数据可与对应模块编号和周期编号一起存储在存储器中。For example, the quick output data and model output data can be numbered (module number and cycle number), where the module number of the quick output data corresponds to the prediction model system node connected to the quick module that generates the quick output data in the prediction model The position in the system, the module number of the output data of the fast module connected to the downstream node of the predictive model system is greater than the module number of the output data of the fast module connected to the upstream node of the predictive model system; and the module number of the model output data is greater than each fast The module number of the predicted output data, and the period number is the number of the period in which the predicted output data is generated. The aforementioned predicted output data can be stored in the memory together with the corresponding module number and cycle number.
在规定时间可确定存储器中存储有该多个迅捷模块的预测输出数据和最后一个预测模块的预测输出数据中的至少一者。At a predetermined time, it can be determined that at least one of the predicted output data of the multiple fast modules and the predicted output data of the last prediction module is stored in the memory.
在步骤506，将该多个迅捷模块的预测输出数据和该最后一个预测模块的预测输出数据中的该至少一者中最新的预测输出数据确定为模型输出数据。In step 506, the latest prediction output data of the at least one of the prediction output data of the multiple quick modules and the prediction output data of the last prediction module is determined as the model output data.
例如，可以将周期编号、预测输出数据编号和迅捷模块的对应输出数据一起存储在存储器中，在规定时间在存储器中查找周期编号最大且预测输出数据编号最大的预测输出数据作为模型输出数据。For example, the cycle number, the predicted output data number, and the corresponding output data of the fast module can be stored in the memory together, and the predicted output data with the largest cycle number and the largest predicted output data number can be found in the memory at a specified time as the model output data.
本公开通过迅捷模块来缩短获得预测值的运行时长。迅捷模块可以通过减少模型特征、减少模型训练次数、减少训练样本数来加快预测速度，保证在规定时间内有预测输出数据。进一步，通过选择最新(或即最精确)的预测输出数据，保证得到当前最精确的预测输出数据。The present disclosure uses a fast module to shorten the running time for obtaining the predicted value. The fast module can speed up the prediction by reducing model features, reducing the number of model training times, and reducing the number of training samples to ensure that there is predictive output data within a specified time. Further, by selecting the latest (or the most accurate) forecast output data, it is guaranteed that the most accurate forecast output data currently is obtained.
本文结合附图阐述的说明描述了示例配置而不代表可被实现或者落在权利要求的范围内的所有示例。本文所使用的术语“示例性”意指“用作示例、实例或解说”，而并不意指“优于”或“胜过其他示例”。本详细描述包括具体细节以提供对所描述的技术的理解。然而，可以在没有这些具体细节的情况下实践这些技术。在一些实例中，众所周知的结构和设备以框图形式示出以避免模糊所描述的示例的概念。The description set forth herein in conjunction with the accompanying drawings describes example configurations and does not represent all examples that can be implemented or fall within the scope of the claims. The term "exemplary" as used herein means "serving as an example, instance, or illustration", and does not mean "better" or "outperform other examples." This detailed description includes specific details to provide an understanding of the described technology. However, these techniques can be practiced without these specific details. In some instances, well-known structures and devices are shown in block diagram form to avoid obscuring the concepts of the described examples.
在附图中，类似组件或特征可具有相同的附图标记。此外，相同类型的各个组件可通过在附图标记后跟随短划线以及在类似组件之间进行区分的第二标记来加以区分。 如果在说明书中仅使用第一附图标记，则该描述可应用于具有相同的第一附图标记的类似组件中的任何一个组件而不论第二附图标记如何。In the drawings, similar components or features may have the same reference signs. In addition, individual components of the same type can be distinguished by a dash followed by a reference number and a second label that distinguishes between similar components. If only the first reference number is used in the specification, the description can be applied to any one of the similar components having the same first reference number regardless of the second reference number.
结合本文中的公开描述的各种解说性框以及模块可以用设计成执行本文中描述的功能的通用处理器、DSP、ASIC、FPGA或其他可编程逻辑器件、分立的门或晶体管逻辑、分立的硬件组件、或其任何组合来实现或执行。通用处理器可以是微处理器，但在替换方案中，处理器可以是任何常规的处理器、控制器、微控制器、或状态机。处理器还可被实现为计算设备的组合(例如，DSP与微处理器的组合、多个微处理器、与DSP核心协同的一个或多个微处理器，或者任何其他此类配置)。The various illustrative blocks and modules described in conjunction with the disclosure herein can be used as general-purpose processors, DSPs, ASICs, FPGAs or other programmable logic devices, discrete gates or transistor logic, discrete gates or transistor logic designed to perform the functions described herein. Hardware components, or any combination thereof. A general-purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices (for example, a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors in cooperation with a DSP core, or any other such configuration).
本文中所描述的功能可以在硬件、由处理器执行的软件、固件、或其任何组合中实现。如果在由处理器执行的软件中实现，则各功能可以作为一条或多条指令或代码存储在计算机可读介质上或藉其进行传送。其他示例和实现落在本公开及所附权利要求的范围内。例如，由于软件的本质，以上描述的功能可使用由处理器执行的软件、硬件、固件、硬连线或其任何组合来实现。实现功能的特征也可物理地位于各种位置，包括被分布以使得功能的各部分在不同的物理位置处实现。另外，如本文(包括权利要求中)所使用的，在项目列举(例如，以附有诸如“中的至少一个”或“中的一个或多个”之类的措辞的项目列举)中使用的“或”指示包含性列举，以使得例如A、B或C中的至少一个的列举意指A或B或C或AB或AC或BC或ABC(即，A和B和C)。同样，如本文所使用的，短语“基于”不应被解读为引述封闭条件集。例如，被描述为“基于条件A”的示例性步骤可基于条件A和条件B两者而不脱离本公开的范围。换言之，如本文所使用的，短语“基于”应当以与短语“至少部分地基于”相同的方式来解读。The functions described herein can be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, each function can be stored as one or more instructions or codes on a computer-readable medium or transmitted therethrough. Other examples and implementations fall within the scope of this disclosure and the appended claims. For example, due to the nature of software, the functions described above can be implemented using software, hardware, firmware, hardwired, or any combination thereof executed by a processor. The features that implement the function may also be physically located in various locations, including being distributed so that various parts of the function are implemented at different physical locations. In addition, as used herein (including in the claims), used in item listings (e.g., listings with items appended with terms such as "at least one of" or "one or more of") "Or" indicates an inclusive enumeration, such that, for example, enumeration of at least one of A, B, or C means A or B or C or AB or AC or BC or ABC (ie, A and B and C). Likewise, as used herein, the phrase "based on" should not be read as quoting a closed set of conditions. For example, an exemplary step described as “based on condition A” may be based on both condition A and condition B without departing from the scope of the present disclosure. In other words, as used herein, the phrase "based on" should be read in the same way as the phrase "based at least in part."
计算机可读介质包括非瞬态计算机存储介质和通信介质两者，其包括促成计算机程序从一地向另一地转移的任何介质。非瞬态存储介质可以是能被通用或专用计算机访问的任何可用介质。作为示例而非限定，非瞬态计算机可读介质可包括RAM、ROM、电可擦除可编程只读存储器(EEPROM)、压缩盘(CD)ROM或其他光盘存储、磁盘存储或其他磁存储设备、或能被用来携带或存储指令或数据结构形式的期望程序代码手段且能被通用或专用计算机、或者通用或专用处理器访问的任何其他非瞬态介质。任何连接也被正当地称为计算机可读介质。例如，如果软件是使用同轴电缆、光纤电缆、双绞线、数字订户线(DSL)、或诸如红外、无线电、以及微波之类的无线技术从web网站、服务器、或其它远程源传送而来的，则该同轴电缆、光纤电缆、双绞线、数字订户线(DSL)、或诸如红外、无线电、以及微波之类的无线技术就被包括在介质的定义之 中。如本文所使用的盘(disk)和碟(disc)包括CD、激光碟、光碟、数字通用碟(DVD)、软盘和蓝光碟，其中盘常常磁性地再现数据而碟用激光来光学地再现数据。以上介质的组合也被包括在计算机可读介质的范围内。Computer-readable media includes both non-transitory computer storage media and communication media, including any media that facilitates the transfer of a computer program from one place to another. The non-transitory storage medium may be any available medium that can be accessed by a general-purpose or special-purpose computer. By way of example and not limitation, non-transitory computer readable media may include RAM, ROM, electrically erasable programmable read-only memory (EEPROM), compact disk (CD) ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices , Or any other non-transitory medium that can be used to carry or store desired program code means in the form of instructions or data structures and can be accessed by a general-purpose or special-purpose computer, or a general-purpose or special-purpose processor. Any connection is also properly called a computer-readable medium. For example, if the software is transmitted from a web site, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave Yes, the coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave are included in the definition of the medium. Disks and discs as used herein include CDs, laser discs, optical discs, digital versatile discs (DVD), floppy discs and Blu-ray discs, in which discs often reproduce data magnetically and discs reproduce data optically with lasers . Combinations of the above media are also included in the scope of computer-readable media.
提供本文的描述是为了使得本领域技术人员能够制作或使用本公开。对本公开的各种修改对于本领域技术人员将是显而易见的，并且本文中定义的普适原理可被应用于其他变形而不会脱离本公开的范围。由此，本公开并非被限定于本文所描述的示例和设计，而是应被授予与本文所公开的原理和新颖特征相一致的最广范围。The description herein is provided to enable those skilled in the art to make or use the present disclosure. Various modifications to the present disclosure will be obvious to those skilled in the art, and the general principles defined herein can be applied to other modifications without departing from the scope of the present disclosure. Therefore, the present disclosure is not limited to the examples and designs described herein, but should be granted the widest scope consistent with the principles and novel features disclosed herein.

Claims (17)

  1. 一种用于加快预测的方法，所述方法包括：A method for accelerating prediction, the method comprising:
    由预测模型系统接收输入数据，其中所述预测模型系统包括串接的一个或多个预测模块，一个或多个迅捷模块连接到所述预测模型系统，所述迅捷模块的输入是所述一个或多个预测模块中的一个预测模块的输入并且所述迅捷模块提供迅捷输出数据；The input data is received by a predictive model system, wherein the predictive model system includes one or more predictive modules connected in series, one or more rapid modules are connected to the predictive model system, and the input of the rapid module is the one or The input of one prediction module of the plurality of prediction modules and the fast module provides fast output data;
    在规定时间从所述一个或多个迅捷模块获取至少一个迅捷输出数据；以及Obtain at least one quick output data from the one or more quick modules at a specified time; and
    将所述至少一个迅捷输出数据中最新的迅捷输出数据确定为最终预测输出数据，其中所述迅捷模块是根据所述预测模型系统中的至少一个预测模块生成的。The latest quick output data in the at least one quick output data is determined as the final predicted output data, wherein the quick module is generated according to at least one prediction module in the prediction model system.
  2. 如权利要求1所述的方法，其特征在于，所述预测模型系统产生模型输出数据，并且所述方法进一步包括：The method of claim 1, wherein the predictive model system generates model output data, and the method further comprises:
    如果在所述规定时间所述预测模型系统产生了模型输出数据，则将所述模型输出数据确定为最终预测输出数据。If the predictive model system generates model output data at the prescribed time, the model output data is determined as final predictive output data.
  3. 如权利要求1所述的方法，其特征在于，所述迅捷模块连接在所述预测模型系统的两个预测模块之间，并且所述至少一个预测模块包括与所述迅捷模块的输入相同的预测模块及其后续一个或多个预测模块。The method of claim 1, wherein the fast module is connected between two prediction modules of the prediction model system, and the at least one prediction module includes the same prediction as the input of the fast module Module and its subsequent one or more prediction modules.
  4. 如权利要求1或3所述的方法，其特征在于，所述迅捷模块是通过对所述至少一个预测模块调整参数和/或删除所述至少一个预测模块的一个或多个子模块来生成的。The method according to claim 1 or 3, wherein the quick module is generated by adjusting parameters of the at least one prediction module and/or deleting one or more sub-modules of the at least one prediction module.
  5. 如权利要求4所述的方法，其特征在于，所述参数包括训练步长、训练次数和/或误差精度。The method according to claim 4, wherein the parameters include training step size, training times, and/or error accuracy.
  6. 如权利要求1所述的方法，其特征在于，进一步包括，The method of claim 1, further comprising:
    确定所述至少一个迅捷输出数据中的每一者的模块编号和周期编号，其中连接到预测模型系统下游的迅捷模块的迅捷输出数据的模块编号大于连接到预测模型系统上游的迅捷模块的迅捷输出数据的模块编号，所述周期编号是生成预测输出数据的周期的编号；Determine the module number and period number of each of the at least one quick output data, wherein the module number of the quick output data of the quick output data connected to the downstream of the predictive model system is greater than the quick output of the quick module connected to the upstream of the predictive model system The module number of the data, where the period number is the number of the period in which the predicted output data is generated;
    将所述至少一个迅捷输出数据中的每一者与其周期编号、模块编号一起存储在存储器中；Storing each of the at least one quick output data in the memory together with its cycle number and module number;
    在所述规定时间在所述存储器中查找周期编号最大的一个或多个迅捷输出数据；以及Searching for one or more quick output data with the largest cycle number in the memory at the specified time; and
    选择所述一个或多个预测输出数据中模块编号最大的迅捷输出数据作为所述最终预测输出数据。Select the quick output data with the largest module number among the one or more predicted output data as the final predicted output data.
  7. 如权利要求1所述的方法，其特征在于，进一步包括，The method of claim 1, further comprising:
    确定所获取的当前迅捷输出数据的当前周期编号和当前模块编号；Determine the current cycle number and current module number of the acquired current quick output data;
    将所述当前周期编号与存储器中所存储的在先迅捷输出数据的在先周期编号进行比较；Comparing the current cycle number with the previous cycle number of the previous quick output data stored in the memory;
    如果所述当前周期编号大于所述在先周期编号，则在所述存储器中将所述当前迅捷输出数据以及当前周期编号和当前模块编号覆盖所述在先迅捷输出数据以及在先周期编号和在先模块编号；If the current cycle number is greater than the previous cycle number, the current fast output data, current cycle number and current module number will be overwritten in the memory with the previous fast output data, previous cycle number and previous cycle number. First module number;
    如果所述当前周期编号等于所述在先周期编号，则将所述当前模块编号与所述在先模块编号进行比较；If the current cycle number is equal to the previous cycle number, comparing the current module number with the previous module number;
    如果所述当前模块编号大于所述在先模块编号，则在所述存储器中将所述当前迅捷输出数据以及当前周期编号和当前模块编号覆盖所述在先迅捷输出数据以及在先周期编号和在先模块编号；以及If the current module number is greater than the previous module number, the current quick output data, the current cycle number and the current module number will be overwritten in the memory with the previous quick output data, the previous cycle number and the current module number. Module number first; and
    在所述规定时间获取所述存储器中所存储的迅捷输出数据。Acquire the quick output data stored in the memory at the prescribed time.
  8. 如权利要求1所述的方法，其特征在于，所述输入数据是外汇汇兑业务的历史交易数据，并且所述规定时间是购汇结算周期期满时间。The method according to claim 1, wherein the input data is historical transaction data of a foreign exchange exchange business, and the prescribed time is the expiration time of a foreign exchange settlement period.
  9. 一种用于加快预测的装置，包括：A device for accelerating prediction, including:
    预测模型系统，接收输入数据，其中所述预测模型系统包括串接的一个或多个预测模块；A prediction model system that receives input data, wherein the prediction model system includes one or more prediction modules connected in series;
    连接至所述预测模型系统的至少一个迅捷模块，所述迅捷模块的输入是所述一个或多个预测模块中的一个预测模块的输入并且所述迅捷模块提供迅捷输出数据，所述迅捷模块是根据所述预测模型系统中的至少一个预测模块生成的；At least one fast module connected to the prediction model system, the input of the fast module is the input of one of the one or more prediction modules and the fast module provides fast output data, the fast module is Generated according to at least one prediction module in the prediction model system;
    存储器，所述存储器存储至少一个迅捷输出数据；A memory, the memory storing at least one quick output data;
    处理器，在规定时间从所述一个或多个迅捷模块获取所述至少一个迅捷输出数据，将所述至少一个迅捷输出数据存储在所述存储器中，以及将所述至少一个预测输出数据中最新的预测输出数据确定为最终预测输出数据。The processor acquires the at least one quick output data from the one or more quick output data at a prescribed time, stores the at least one quick output data in the memory, and stores the latest one among the at least one predicted output data The predicted output data of is determined as the final predicted output data.
  10. 如权利要求9所述的装置，其特征在于，所述预测模型系统产生模型输出数据，并且所述处理器被进一步配置成：9. The apparatus of claim 9, wherein the predictive model system generates model output data, and the processor is further configured to:
    如果在所述规定时间从所述预测模型系统接收到模型输出数据，则将所述模型输出数据确定为最终预测输出数据。If the model output data is received from the prediction model system at the prescribed time, the model output data is determined to be the final prediction output data.
  11. 如权利要求9所述的装置，其特征在于，所述迅捷模块连接在所述预测模型系统的两个预测模块之间，并且所述至少一个预测模块包括与所述迅捷模块的输入相同的预测模块及其后续一个或多个预测模块。The device according to claim 9, wherein the fast module is connected between two prediction modules of the prediction model system, and the at least one prediction module includes the same prediction as the input of the fast module Module and its subsequent one or more prediction modules.
  12. 如权利要求9或11所述的装置，其特征在于，所述迅捷模块是通过对所述至少一个预测模块调整参数和/或删除所述至少一个预测模块的一个或多个子模块来生成的。The device according to claim 9 or 11, wherein the quickness module is generated by adjusting parameters of the at least one prediction module and/or deleting one or more sub-modules of the at least one prediction module.
  13. 如权利要求12所述的装置，其特征在于，所述参数包括训练步长、训练次数和/或误差精度。The device according to claim 12, wherein the parameters include training step length, training times, and/or error accuracy.
  14. 如权利要求9所述的装置，其特征在于，所述处理器被进一步配置成：The apparatus of claim 9, wherein the processor is further configured to:
    确定所述至少一个迅捷输出数据中的每一者的模块编号和周期编号，其中连接到预测模型系统下游的迅捷模块的迅捷输出数据的模块编号大于连接到预测模型系统上游的迅捷模块的迅捷输出数据的模块编号，所述周期编号是生成预测输出数据的周期的编号；Determine the module number and period number of each of the at least one quick output data, wherein the module number of the quick output data of the quick output data connected to the downstream of the predictive model system is greater than the quick output of the quick module connected to the upstream of the predictive model system The module number of the data, where the period number is the number of the period in which the predicted output data is generated;
    将所述至少一个迅捷输出数据中的每一者与其周期编号、模块编号一起存储在存储器中；Storing each of the at least one quick output data in the memory together with its cycle number and module number;
    在所述规定时间在所述存储器中查找周期编号最大的一个或多个迅捷输出数据；以及Searching for one or more quick output data with the largest cycle number in the memory at the specified time; and
    选择所述一个或多个预测输出数据中模块编号最大的迅捷输出数据作为所述最终预测输出数据。Select the quick output data with the largest module number among the one or more predicted output data as the final predicted output data.
  15. 如权利要求9所述的装置，其特征在于，所述处理器被进一步配置成：The apparatus of claim 9, wherein the processor is further configured to:
    确定所获取的当前迅捷输出数据的当前周期编号和当前模块编号；Determine the current cycle number and current module number of the acquired current quick output data;
    将所述当前周期编号与存储器中所存储的在先迅捷输出数据的在先周期编号进行比较；Comparing the current cycle number with the previous cycle number of the previous quick output data stored in the memory;
    如果所述当前周期编号大于所述在先周期编号，则在所述存储器中将所述当前迅捷输出数据以及当前周期编号和当前模块编号覆盖所述在先迅捷输出数据以及在先周期编号和在先模块编号；If the current cycle number is greater than the previous cycle number, the current fast output data, current cycle number and current module number will be overwritten in the memory with the previous fast output data, previous cycle number and previous cycle number. First module number;
    如果所述当前周期编号等于所述在先周期编号，则将所述当前模块编号与所述在先模块编号进行比较；If the current cycle number is equal to the previous cycle number, comparing the current module number with the previous module number;
    如果所述当前模块编号大于所述在先模块编号，则在所述存储器中将所述当前迅捷输出数据以及当前周期编号和当前模块编号覆盖所述在先迅捷输出数据以及在先周期编号和在先模块编号；以及If the current module number is greater than the previous module number, the current quick output data, the current cycle number and the current module number will be overwritten in the memory with the previous quick output data, the previous cycle number and the current module number. Module number first; and
    在所述规定时间获取所述存储器中所存储的迅捷输出数据。Acquire the quick output data stored in the memory at the prescribed time.
  16. 如权利要求9所述的装置，其特征在于，所述输入数据是外汇汇兑业务的历史交易数据，并且所述规定时间是购汇结算周期期满时间。9. The device according to claim 9, wherein the input data is historical transaction data of foreign exchange exchange business, and the prescribed time is the expiration time of a foreign exchange settlement period.
  17. 一种计算机设备，包括：A computer device including:
    处理器；以及Processor; and
    被安排成存储计算机可执行指令的存储器，所述可执行指令在被执行时使所述处理器执行以下操作：A memory arranged to store computer-executable instructions that, when executed, cause the processor to perform the following operations:
    由预测模型系统接收输入数据，其中所述预测模型系统包括串接的一个或多个预测模块，一个或多个迅捷模块连接到所述预测模型系统，所述迅捷模块的输入是所述一个或多个预测模块中的一个预测模块的输入并且所述迅捷模块提供迅捷输出数据；The input data is received by a predictive model system, wherein the predictive model system includes one or more predictive modules connected in series, one or more rapid modules are connected to the predictive model system, and the input of the rapid module is the one or The input of one prediction module of the plurality of prediction modules and the fast module provides fast output data;
    在规定时间从所述一个或多个迅捷模块获取至少一个迅捷输出数据；以及Obtain at least one quick output data from the one or more quick modules at a specified time; and
    将所述至少一个迅捷输出数据中最新的迅捷输出数据确定为最终预测输出数据，其中所述迅捷模块是根据所述预测模型系统中的至少一个预测模块生成的。The latest quick output data in the at least one quick output data is determined as the final predicted output data, wherein the quick module is generated according to at least one prediction module in the prediction model system.
PCT/CN2020/073797 2019-03-01 2020-01-22 Model prediction acceleration method and device WO2020177499A1 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910157451.9 2019-03-01
CN201910157451.9A CN110059847B (en) 2019-03-01 2019-03-01 Model prediction acceleration method and device

Publications (1)

Publication Number Publication Date
WO2020177499A1 true WO2020177499A1 (en) 2020-09-10

Family

ID=67316049

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/073797 WO2020177499A1 (en) 2019-03-01 2020-01-22 Model prediction acceleration method and device

Country Status (2)

Country Link
CN (1) CN110059847B (en)
WO (1) WO2020177499A1 (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110059847B (en) * 2019-03-01 2021-07-13 创新先进技术有限公司 Model prediction acceleration method and device

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2009110341A (en) * 2007-10-31 2009-05-21 Nippon Telegr & Teleph Corp <Ntt> Prediction apparatus using time information, prediction method, prediction program, and recording medium recording the program
CN108460490A (en) * 2018-03-16 2018-08-28 阿里巴巴集团控股有限公司 A kind of prediction technique, device and the equipment of business occurrence quantity
CN108717867A (en) * 2018-05-02 2018-10-30 中国科学技术大学苏州研究院 Disease forecasting method for establishing model and device based on Gradient Iteration tree
CN109039727A (en) * 2018-07-24 2018-12-18 中国银行股份有限公司 Message queue monitoring method and device based on deep learning
CN109117352A (en) * 2017-06-23 2019-01-01 华为技术有限公司 Server performance prediction technique and device
CN110059847A (en) * 2019-03-01 2019-07-26 阿里巴巴集团控股有限公司 Model prediction acceleration method and device

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
AU2009202874B2 (en) * 2009-07-16 2012-08-16 Commonwealth Scientific And Industrial Research Organisation System and Method for Prediction of Patient Admission Rates
CN102393628A (en) * 2011-09-14 2012-03-28 哈尔滨工业大学 Dynamic matrix control (DMC) engineering method based on model simplification and prediction error correction
CN108846525A (en) * 2018-08-02 2018-11-20 阿里巴巴集团控股有限公司 Dealing amount of foreign exchange prediction technique and device

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2009110341A (en) * 2007-10-31 2009-05-21 Nippon Telegr & Teleph Corp <Ntt> Prediction apparatus using time information, prediction method, prediction program, and recording medium recording the program
CN109117352A (en) * 2017-06-23 2019-01-01 华为技术有限公司 Server performance prediction technique and device
CN108460490A (en) * 2018-03-16 2018-08-28 阿里巴巴集团控股有限公司 A kind of prediction technique, device and the equipment of business occurrence quantity
CN108717867A (en) * 2018-05-02 2018-10-30 中国科学技术大学苏州研究院 Disease forecasting method for establishing model and device based on Gradient Iteration tree
CN109039727A (en) * 2018-07-24 2018-12-18 中国银行股份有限公司 Message queue monitoring method and device based on deep learning
CN110059847A (en) * 2019-03-01 2019-07-26 阿里巴巴集团控股有限公司 Model prediction acceleration method and device

Also Published As

Publication number Publication date
CN110059847B (en) 2021-07-13
CN110059847A (en) 2019-07-26

Similar Documents

Publication Publication Date Title
US11163760B2 (en) Providing a data query service to a user based on natural language request data
US20150032674A1 (en) Parallel decision or regression tree growing
US11657302B2 (en) Model selection in a forecasting pipeline to optimize tradeoff between forecast accuracy and computational cost
CN111680998B (en) Payment routing method, computing device, and computer-readable storage medium
WO2020177499A1 (en) Model prediction acceleration method and device
US11392577B2 (en) Real-time anomaly detection
US20140201124A1 (en) Property reactive modifications in a rete network
WO2021025915A1 (en) Classification rules engine and api generator
US10838801B2 (en) Automatically solving computer system error states via a comparative historical analysis of error states, logs, profile configurations, and operator responses
US11887017B2 (en) Automatically predicting transaction likelihood information and transaction-related temporal information using machine learning techniques
CN107203545B (en) Data processing method and device
CN115391160B (en) Abnormal change detection method, device, equipment and storage medium
CN116628346A (en) Training method and device for search word recommendation model
US11222282B2 (en) Sourcing a new machine-learning project by reusing artifacts from reference machine learning projects
US10740168B2 (en) System maintenance using unified cognitive root cause analysis for multiple domains
CN112650829A (en) Customer service processing method and device
US20210248512A1 (en) Intelligent machine learning recommendation platform
US11757741B2 (en) Demand prediction apparatus, demand prediction method and program for predicting a demand of a path on a network using selected trend patterns
CN109977097B (en) Network configuration management method and system based on database
US11775869B1 (en) Validation of account identifier
US20170270170A1 (en) Eliminating false predictors in data-mining
CN110740059B (en) Online early warning processing method and system
AU2021318113B2 (en) Combining rules-based knowledge engineering with machine learning prediction
KR102585107B1 (en) Method for managing related search keyword
US11609970B1 (en) Enhanced time series forecasting

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20765945

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20765945

Country of ref document: EP

Kind code of ref document: A1