WO2020024716A1 - 一种针对新场景的预测模型训练方法及装置 - Google Patents

一种针对新场景的预测模型训练方法及装置 Download PDF

Info

Publication number
WO2020024716A1
WO2020024716A1 PCT/CN2019/091658 CN2019091658W WO2020024716A1 WO 2020024716 A1 WO2020024716 A1 WO 2020024716A1 CN 2019091658 W CN2019091658 W CN 2019091658W WO 2020024716 A1 WO2020024716 A1 WO 2020024716A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
migrated
training sample
sample set
models
Prior art date
Application number
PCT/CN2019/091658
Other languages
English (en)
French (fr)
Inventor
张天翼
陈明星
郭龙
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 WO2020024716A1 publication Critical patent/WO2020024716A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION 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
    • G06Q10/00Administration; Management
    • G06Q10/06Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063Operations research, analysis or management
    • G06Q10/0637Strategic management or analysis, e.g. setting a goal or target of an organisation; Planning actions based on goals; Analysis or evaluation of effectiveness of goals
    • G06Q10/06375Prediction of business process outcome or impact based on a proposed change

Definitions

  • the embodiments of the present specification relate to the field of Internet application technologies, and in particular, to a method and device for training a prediction model for a new scenario.
  • the model can be trained through machine learning based on accumulated sample data to achieve the required decision-making functions. For example, in the financial risk control scenario, a large amount of transaction data can be used as sample data to train a risk control model through machine learning, so that the trained risk control model can be used to automatically make risk decisions for new transactions.
  • accumulating sample data and training models to achieve the deployment of machine learning models often takes a long time.
  • data accumulation and training for risk control models generally take more than six months.
  • one solution is to deploy and use a historical model in a new scenario.
  • the historical model is trained based on historical data in other scenarios, but due to the differences between the sample data of each scenario, the historical model is used in the new scenario. The effect is usually poor.
  • the embodiments of the present specification provide a method and device for training a prediction model for a new scenario.
  • the technical solution is as follows:
  • a method for training a prediction model for a new scene includes:
  • the models to be migrated are: models that are deployed and used in the old scene and can be migrated to the new scene;
  • the model to be migrated is updated to obtain a model suitable for a new scenario.
  • a prediction model training device for a new scene includes:
  • a model to be migrated acquisition module configured to obtain a set of models to be migrated, the models to be migrated are: models that are deployed and used in the old scene and can be migrated to the new scene;
  • a labeling model selection module configured to select at least one model from the set of models to be migrated for predictive labeling of unlabeled samples in a new scene
  • a sample set acquisition module configured to obtain an initial training sample set in a new scene, where the initial training sample set includes unlabeled samples
  • a sample labeling module for adding predicted labels to unlabeled samples in the initial training sample set using the selected model
  • a model update module is used to update a model to be migrated based on a supervised learning algorithm using an initial training sample set to which a predictive label has been added to obtain a model suitable for a new scenario.
  • the technical solutions provided in the embodiments of this specification can migrate the models used in the old scene to the new scene, and the sample accumulation time in the new scene is relatively short. Therefore, if the samples have no or only a few actual labels, The migration model performs label prediction, thereby further optimizing the models to be migrated, making these models more suitable for use in new scenarios, and providing a more efficient and accurate prediction model training solution for new scenarios.
  • FIG. 1 is a schematic flowchart of a prediction model training method for a new scenario according to an embodiment of the present specification
  • FIG. 2 is a schematic flowchart of a wind control model training method for a new scenario according to an embodiment of the present specification
  • FIG. 3 is a schematic structural diagram of a prediction model training device for a new scenario according to an embodiment of the present specification
  • FIG. 4 is a schematic structural diagram of a device for configuring a device according to an embodiment of the present specification.
  • An embodiment of the present specification provides a method for training a prediction model for a new scenario. As shown in FIG. 1, the method may include the following steps:
  • This specification does not limit the specific way to obtain the set of models to be migrated.
  • the feature vector input by each model in the old scene and the feature vector extractable by the training samples in the new scene may be compared to determine whether each model in the old scene can be migrated to the new scene.
  • a first feature set is first obtained, and the set includes: a predetermined number of feature vectors that can be extracted from a new scene training sample; and then for any model deployed in the old scene: a second feature set is obtained, and the set Including: several feature vectors input by the model; determining that the model is a model to be migrated if the model conforms to a preset migration rule; the preset migration rule includes: the first feature set and the second feature set The feature vectors included in the intersection meet preset migration conditions.
  • the above-mentioned preset migration conditions may specifically compare the first feature set and the second feature set from various perspectives in various forms.
  • the preset migration condition may be: the number of feature vectors included in the intersection is not less than a preset threshold, that is, whether the model can be migrated to by comparing the number of feature vectors in the intersection of the first feature set and the second feature set. New scene. If the number of feature vectors in the intersection is small, the model has a higher probability of performing poorly in the new scene, so it can be considered that the model cannot be migrated to the new scene; otherwise, it is considered that the model can be migrated to the new scene.
  • a preset threshold that is, whether the model can be migrated to by comparing the number of feature vectors in the intersection of the first feature set and the second feature set. New scene. If the number of feature vectors in the intersection is small, the model has a higher probability of performing poorly in the new scene, so it can be considered that the model cannot be migrated to the new scene; otherwise, it is considered that the model can be migrated to the new scene.
  • the preset migration conditions can be: A weighted score calculated according to a preset weight of each feature vector included in the intersection is not less than a preset threshold.
  • the feature vectors that are more important for model training can be preset with higher weights, and the more important the preset weights are higher. Therefore, if the important feature vector included in the intersection is high, the weighted score finally calculated is also high, and the model can be considered to be migrated to a new scene.
  • the preset migration conditions can also be in other forms, and each migration condition can be used alone or in combination. Those skilled in the art can flexibly set according to actual needs, and this specification does not specifically limit this.
  • the type of prediction model trained in the new scene can be determined and specified by the R & D staff based on experience or algorithms in advance. Then, in order to further measure the models in the old scene based on the comparison of feature vectors, can it be migrated to the new scene, When the set of models to be migrated is obtained, at least one type specified in advance for the new scene prediction model may be further obtained, and the preset migration rule may further include that the type of the model is included in at least one type specified in advance .
  • the preset migration rules may also include rules of other dimensions, which are not limited in the embodiments of the present specification.
  • the R & D personnel can also specify the model to be migrated from the old scene to the new scene, and when designating, the R & D personnel can also measure whether the various models can be migrated to the feature vector, model type and other dimensions based on experience or algorithm. New scenes, performance after migration, and more.
  • Training samples can often be labeled in multiple ways. For example, you can label manually, and manual labeling is usually more accurate, but the amount of training sample data used for model training is usually large, and manual labeling is less efficient; for example, in some scenarios, labels can be generated according to actual conditions, such as credit card wind
  • labels can be generated according to actual conditions, such as credit card wind
  • the control scenario when a bank verifies that a credit card has been stolen, the credit card and the corresponding transaction can be marked as black samples, but in this scenario, a black sample label may not be obtained in the short term.
  • At least one model is selected from a set of models to be migrated for predictive labeling of unlabeled samples in a new scene, thereby improving labeling efficiency and shortening labeling cycles.
  • Each model to be migrated is a model that can be migrated to a new scene, but specifically, because the feature vectors input by each model are different from the model type, etc., some of the models can be directly applied to the new scene, and some models require After the update, it can be applied to the new scene well. Therefore, from the set of models to be migrated, some models that perform better in the new scene can be selected for predictive labeling.
  • Selecting at least one model from the set of models to be migrated may be specifically implemented in multiple ways.
  • a third feature set may be obtained first.
  • the set includes: a plurality of feature vectors specified in advance for predicting sample labels in a new scene; and then each feature set corresponding to each model to be migrated is obtained.
  • a feature set includes: a number of feature vectors corresponding to the model; and at least one model is selected from the set of models to be migrated according to a preset selection rule.
  • the number of feature vectors in the intersection, the number of important feature vectors, and whether the model type is the same can be used to measure whether a model is selected. It is used for predictive labeling, which is not described in this specification.
  • the model to be migrated can be specified by the R & D personnel, and in this step, when selecting a model for predicting and labeling the unlabeled samples in the new scene from the set of models to be migrated, the R & D personnel can also The choice of experience or algorithm will not be repeated here.
  • the initial training sample set in the new scene can include unlabeled samples that need to be labeled with predictive labels, or labeled samples that can be labeled with actual labels (which can be white and / or black samples).
  • the selected model is used for Predict labels for unlabeled samples.
  • prediction tags can be added in multiple ways.
  • the corresponding relationship between different values and different prediction labels may be set in advance, for example, when the value is larger or smaller than a preset value, it corresponds to a black sample label, otherwise it corresponds to a white sample label.
  • a corresponding predictive label can be obtained directly according to the predicted value (that is, equal to the weighted sum) output by the model.
  • the weights corresponding to the output values of each model can be set in advance. For example, models with better performance have higher or lower weights.
  • the weights of each model are the same, that is, it is equivalent to not setting a weight for each model.
  • the predicted labels added by the selected model can also be manually checked and corrected to improve the accuracy of the predicted labels.
  • the field can be flexibly set according to the actual situation, which is not limited in this specification.
  • the model to be migrated is updated to obtain a model applicable to the new scenario.
  • the training sample set in the old scene can also be obtained, and the training sample set includes labeled samples with actual labels added; the initial sample set in the new scene and the old scene And use the combined training sample set to update the model to be migrated based on the supervised learning algorithm.
  • a large number of training samples have been accumulated in the old scene, and the training samples are labeled samples that have been added with actual labels. Therefore, it can be used to assist the migration in the new scene when the number of training samples accumulated in the new scene is small.
  • the model is updated.
  • the training samples in the old scene are not necessarily completely applicable to the model update of the new scene. Some of the samples may have higher similarity with the training samples in the new scene, while other samples have lower similarity, so it can be Different training samples in the training sample set after the initial sample set in the new scene and the training sample set in the old scene are combined with different weights.
  • each training sample in the initial sample set has the highest weight
  • the training sample set in the old scene has the weight of each training sample with a high degree of similarity to each training sample in the initial sample set
  • the optimized training sample set includes Add labeled samples with actual labels; combine the initial training sample set with the predicted labels added and the optimized training sample set with the actual labels added, and use the combined training sample set to update the migration model based on a supervised learning algorithm.
  • each model to be migrated can be directly applied to the new scenario, and it is updated according to this solution while being applied, and a model more suitable for the new scenario can also be obtained. After the time is updated, it is applied to the new scene, and the update can be continued after the application, which is not limited in this manual.
  • a large amount of accumulated transaction data can be used as sample data, and the risk control model can be trained through machine learning, so that based on the trained risk control model, risk decisions can be made on new transactions in a timely and accurate manner.
  • the amount of sample data is usually related to the transaction volume and cumulative time of the new scene.
  • the sample training set needs to include a certain amount of black sample data.
  • the bank ’s card theft business Reasons such as long processing time make it impossible to accumulate the amount of black sample data required for training in a short time.
  • the existing risk control model in the old scene can be migrated to the new scene.
  • the new scenario and the old scenario can be trading markets in different countries and regions.
  • the risk control models deployed in the old scenario can include: stolen card risk control model, stolen account risk control model, hidden case identification model, and so on.
  • the control model can be trained based on transaction data from multiple countries and regions.
  • multiple models that can be deployed and used in new and old scenes can be trained in the cloud in advance based on data collected from various old scenes.
  • the risk control model for stolen cards and the risk control model for stolen accounts are used to control the risk of stolen credit cards and stolen payment accounts, and can be supervised and trained.
  • the hidden case recognition model uses more targeted feature vectors to identify transactions that have not yet been determined as cases by the bank (ie, non-obvious cases) but have the characteristics of the case.
  • an abnormal transaction (Such as abnormal transaction volume, transaction time, transaction location, etc.) equipment, accounts, credit cards, network environment, etc., there is a higher risk of stolen cards and account theft
  • the hidden case identification model can identify the corresponding transaction as Black samples.
  • the hidden case recognition model can be trained through unsupervised learning, so that it can be applied to scenarios where there are no actual cases (labels).
  • the above model can be delivered to the new scenario locally in the form of a model file.
  • the deployed model can be used locally to score trading events and make risk decisions.
  • the models deployed in the cloud are obtained through training samples from multiple countries and regions.
  • the advantage is that the training samples are comprehensive and versatile.
  • the disadvantage is that compared with the new model trained with local data in new scenarios, It is not completely suitable for new scenarios, so you need to update these models after a certain number of training samples have been accumulated in the new scenarios.
  • the accumulation time in the new scene is relatively short. For example, in the 1 week after deployment, there are fewer accumulated training samples, and each sample has no labels, and the model cannot be updated. Therefore, in the first stage, the cloud-trained and unupdated models are used to make risk control decisions for transactions in the new scenario.
  • Training data can be updated for each model.
  • the hidden case recognition model can be used to add prediction samples to the initial training sample set.
  • the new scene is the market of Malaysia
  • the old scene includes markets such as Thailand, the United States, and Japan.
  • Thailand and Malaysia have closer consumption levels and habits.
  • the transaction data is more similar, while the US, Japan and Malaysia are less similar. Therefore, you can set the highest weight for training samples accumulated in Malaysia, set higher weights for training samples from Thailand, and set lower weights for training samples from the United States and Japan. Therefore, by means of dynamic weighting, the updated models can be made more suitable for the new scene when there is less data in the new scene.
  • the updated models in the second stage can still be used for trading decisions in new scenarios.
  • each model can be further updated.
  • the training samples used by the update may include only training samples with actual labels in the new scene, or training samples with predictive labels added by the hidden case recognition model in the new scene, and may also include training samples in the old scene. A lot of training samples, and so on.
  • risk control models are deployed and updated in new scenarios.
  • the data accumulated in new scenarios can also be uploaded to the cloud for updating existing models and training other new models. And deploy to other new scenarios, etc.
  • the models used in the old scene can be migrated to the new scene, and the sample accumulation time in the new scene is relatively short. Therefore, when the samples have no or only a few actual labels, the labels are migrated through the model to be migrated Prediction, thereby further optimizing the models to be migrated, making these models more suitable for use in new scenarios, and providing a more efficient and accurate prediction model training solution for new scenarios.
  • an embodiment of the present specification further provides a prediction model training device for a new scene.
  • the device may include:
  • a model to be migrated acquisition module 110 is configured to obtain a set of models to be migrated, where the models to be migrated are: models that are deployed and used in the old scene and can be migrated to the new scene;
  • a labeling model selection module 120 configured to select at least one model from the set of models to be migrated for predictive labeling of unlabeled samples in a new scene;
  • a sample set acquisition module 130 is configured to obtain an initial training sample set in a new scene, where the initial training sample set includes unlabeled samples;
  • a sample labeling module 140 configured to use the selected model to add predictive labels to the unlabeled samples in the initial training sample set
  • a model update module 150 is configured to use an initial training sample set to which a predictive label has been added and update a model to be migrated based on a supervised learning algorithm to obtain a model applicable to a new scenario.
  • model acquisition module 110 to be migrated may include:
  • the feature to-be-migrated obtaining unit 111 is configured to obtain a first feature set, the set includes: a number of feature vectors that can be extracted from a predetermined new scene training sample; for any model used in the old scene deployment: obtaining a second feature set
  • the set includes: several feature vectors input by the model;
  • the model to be migrated selecting unit 112 is configured to determine the model as a model to be migrated if the model meets a preset migration rule; the preset migration rule includes: an intersection of a first feature set and a second feature set The included feature vectors meet preset migration conditions.
  • the preset migration conditions may include:
  • the number of feature vectors included in the intersection is not less than a preset threshold; and / or a weighted score calculated according to a preset weight of each feature vector included in the intersection is not less than a preset threshold.
  • the model to be migrated acquisition module 110 may further include: a type to be migrated acquisition unit 113 configured to obtain at least one type specified in advance for a new scenario prediction model;
  • the preset migration rule may further include that at least one type specified in advance includes a type of the model.
  • annotation model selection module 120 may include:
  • the labeled feature acquisition unit 121 is configured to obtain a third feature set, which includes: a number of feature vectors specified in advance for predicting sample labels in a new scene; and a feature set corresponding to each model to be migrated, including any feature
  • the set includes: several feature vectors input by the corresponding model;
  • the label model selection unit 122 is configured to select at least one model from the set of models to be migrated according to a preset selection rule.
  • the sample labeling module 140 may include:
  • the predicted value determining unit 141 is configured to: for any selected model: input the unlabeled samples in the initial training sample set into the model, and obtain the output predicted value;
  • a predictive label determination unit 142 is configured to: for any input unlabeled sample: determine the weight of the predicted value output by each model; calculate the weighted sum of each predicted value, and determine the weighted sum corresponding predicted label; The sample adds the predictive label.
  • the sample set obtaining module 130 may be further configured to obtain an optimized training sample set in a new scenario, where the optimized training sample set includes labeled samples to which actual labels have been added. ;
  • the model update module 150 may be specifically configured to: combine the initial training sample set with the predicted labels added and the optimized training sample set with the actual labels added, and use the combined training sample set based on a supervised learning algorithm to treat migration The model is updated.
  • the sample set obtaining module 130 may be further configured to obtain a training sample set in an old scene, where the training sample set includes labeled samples to which actual labels have been added;
  • the model updating module 150 may be specifically configured to: merge the initial sample set in the new scene with the training sample set in the old scene, and use the combined training sample set to update the model to be migrated based on a supervised learning algorithm.
  • An embodiment of the present specification further provides a computer device including at least a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the foregoing program for a new scenario when the processor executes the program.
  • Prediction model training method includes at least:
  • the models to be migrated are: models that are deployed and used in the old scene and can be migrated to the new scene;
  • the model to be migrated is updated to obtain a model suitable for new scenarios.
  • FIG. 4 shows a more specific schematic diagram of the hardware structure of a computing device provided by an embodiment of the present specification.
  • the device may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050.
  • the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040 implement a communication connection within the device through a bus 1050.
  • the processor 1010 may be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an Application Specific Integrated Circuit (ASIC), or one or more integrated circuits, etc., for performing related operations. Program to implement the technical solutions provided by the embodiments of this specification.
  • a general-purpose CPU Central Processing Unit
  • ASIC Application Specific Integrated Circuit
  • the memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage devices, dynamic storage devices, and the like.
  • the memory 1020 may store an operating system and other application programs.
  • related program codes are stored in the memory 1020 and are called and executed by the processor 1010.
  • the input / output interface 1030 is used to connect an input / output module to implement information input and output.
  • the input / output / module can be configured in the device as a component (not shown in the figure), or it can be externally connected to the device to provide corresponding functions.
  • the input device may include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc.
  • the output device may include a display, a speaker, a vibrator, and an indicator light.
  • the communication interface 1040 is used to connect a communication module (not shown in the figure) to implement communication interaction between the device and other devices.
  • the communication module can implement communication through a wired method (such as USB, network cable, etc.), and can also implement communication through a wireless method (such as mobile network, WIFI, Bluetooth, etc.).
  • the bus 1050 includes a path for transmitting information between various components of the device (for example, the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040).
  • the device may also include necessary for achieving normal operation Other components.
  • the foregoing device may also include only components necessary to implement the solutions of the embodiments of the present specification, and does not necessarily include all the components shown in the drawings.
  • An embodiment of the present specification also provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the foregoing prediction model training method for a new scenario is implemented.
  • the method includes at least:
  • the models to be migrated are: models that are deployed and used in the old scene and can be migrated to the new scene;
  • the model to be migrated is updated to obtain a model suitable for new scenarios.
  • Computer-readable media includes both permanent and non-persistent, removable and non-removable media.
  • Information can be stored by any method or technology.
  • Information may be computer-readable instructions, data structures, modules of a program, or other data.
  • Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), and read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, read-only disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, Magnetic tape cartridges, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media may be used to store information that can be accessed by computing devices.
  • computer-readable media does not include temporary computer-readable media, such as modulated data signals and carrier waves.
  • the embodiments of the present specification can be implemented by means of software plus a necessary universal hardware platform. Based on such an understanding, the technical solutions of the embodiments of the present specification may be embodied in the form of software products that are essentially or contribute to the existing technology.
  • the computer software product may be stored in a storage medium, such as ROM / RAM, Magnetic disks, optical disks, and the like include a number of instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform the methods described in various embodiments or portions of the embodiments of this specification.
  • the system, device, module, or unit described in the foregoing embodiments may be specifically implemented by a computer chip or entity, or a product with a certain function.
  • a typical implementation device is a computer, and the specific form of the computer may be a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email sending and receiving device, and a game control Desk, tablet computer, wearable device, or a combination of any of these devices.

Landscapes

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

Abstract

公开了一种针对新场景的预测模型训练方法及装置。一种针对新场景的预测模型训练方法,该方法包括:获得待迁移模型的集合;从所述待迁移模型的集合中选择至少一个模型,以用于对新场景中的无标签样本进行预测标注;获得新场景中的初始训练样本集;利用所选择的模型,为初始训练样本集中的无标签样本添加预测标签;利用已添加预测标签的初始训练样本集,基于有监督学习算法,对待迁移模型进行更新,得到可适用于新场景的模型。

Description

一种针对新场景的预测模型训练方法及装置 技术领域
本说明书实施例涉及互联网应用技术领域,尤其涉及一种针对新场景的预测模型训练方法及装置。
背景技术
大数据时代可以基于积累的样本数据,通过机器学习训练模型,从而实现需要的决策功能。例如,在金融风险控制场景,可以将大量的交易数据作为样本数据,通过机器学习训练风控模型,从而可以使用所训练的风控模型,自动对新的交易进行风险决策等。
但是,在某些场景中,积累样本数据并训练模型,从而实现机器学习模型的部署,往往需要较长的时间,如风控模型的数据积累与训练一般需要半年以上。对此,一种解决方案是,在新场景中部署使用历史模型,该历史模型是基于其他场景中的历史数据所训练,但由于各场景样本数据间存在差异,历史模型应用于在新场景中的效果通常较差。
基于现有技术,需要更高效、更准确的针对新场景的预测模型训练方案。
发明内容
针对上述技术问题,本说明书实施例提供一种针对新场景的预测模型训练方法及装置,技术方案如下:
一种针对新场景的预测模型训练方法,该方法包括:
获得待迁移模型的集合,所述待迁移模型为:在旧场景部署使用、且可迁移至新场景的模型;
从所述待迁移模型的集合中选择至少一个模型,以用于对新场景中的无标签样本进行预测标注;
获得新场景中的初始训练样本集,所述初始训练样本集中包括无标签样本;
利用所选择的模型,为初始训练样本集中的无标签样本添加预测标签;
利用已添加预测标签的初始训练样本集,基于有监督学习算法,对待迁移模型进行更新, 得到可适用于新场景的模型。
一种针对新场景的预测模型训练装置,该装置包括:
待迁移模型获取模块,用于获得待迁移模型的集合,所述待迁移模型为:在旧场景部署使用、且可迁移至新场景的模型;
标注模型选取模块,用于从所述待迁移模型的集合中选择至少一个模型,以用于对新场景中的无标签样本进行预测标注;
样本集获取模块,用于获得新场景中的初始训练样本集,所述初始训练样本集中包括无标签样本;
样本标注模块,用于利用所选择的模型,为初始训练样本集中的无标签样本添加预测标签;
模型更新模块,用于利用已添加预测标签的初始训练样本集,基于有监督学习算法,对待迁移模型进行更新,得到可适用于新场景的模型。
本说明书实施例所提供的技术方案,可以将旧场景中部署使用的模型迁移至新场景中,并且在新场景中样本积累时间较短,因而样本没有或只有少数实际标签的情况下,通过待迁移模型进行标签预测,从而进一步优化待迁移模型,使这些模型更适于在新场景中使用,为新场景提供一种更高效且更准确的预测模型训练方案。
应当理解的是,以上的一般描述和后文的细节描述仅是示例性和解释性的,并不能限制本说明书实施例。
此外,本说明书实施例中的任一实施例并不需要达到上述的全部效果。
附图说明
为了更清楚地说明本说明书实施例或现有技术中的技术方案,下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本说明书实施例中记载的一些实施例,对于本领域普通技术人员来讲,还可以根据这些附图获得其他的附图。
图1是本说明书实施例的针对新场景的预测模型训练方法的流程示意图;
图2是本说明书实施例的针对新场景的风控模型训练方法的流程示意图;
图3是本说明书实施例的针对新场景的预测模型训练装置的结构示意图;
图4是用于配置本说明书实施例装置的一种设备的结构示意图。
具体实施方式
为了使本领域技术人员更好地理解本说明书实施例中的技术方案,下面将结合本说明书实施例中的附图,对本说明书实施例中的技术方案进行详细地描述,显然,所描述的实施例仅仅是本说明书的一部分实施例,而不是全部的实施例。基于本说明书中的实施例,本领域普通技术人员所获得的所有其他实施例,都应当属于保护的范围。
本说明书实施例提供一种针对新场景的预测模型训练方法,参见图1所示,该方法可以包括以下步骤:
S101,获得待迁移模型的集合;
由于新场景与旧场景的差异,在旧场景部署使用的各个模型,其中部分可能不适用于新场景,而部分可能适用于新场景,可以迁移至新场景,待迁移模型即为在旧场景部署使用、且可迁移至新场景的模型。
本说明书不限定获得待迁移模型的集合的具体方式。
在本说明书实施例中,可以比较旧场景中各个模型输入的特征向量、与新场景中训练样本可提取的特征向量,从而确定旧场景中各个模型是否可以迁移至新场景。具体地,首先获得第一特征集合,该集合中包括:预先确定的新场景训练样本可提取的若干特征向量;然后针对在旧场景部署使用的任一模型:获得第二特征集合,该集合中包括:该模型所输入的若干特征向量;在该模型符合预设迁移规则的情况下,将该模型确定为待迁移模型;所述预设迁移规则包括:第一特征集合与第二特征集合的交集中包括的特征向量满足预设迁移条件。
上述预设迁移条件具体可以通过多种形式,从多种角度比较第一特征集合与第二特征集合。
例如,预设迁移条件可以为:交集中包括的特征向量的数量不小于预设阈值,即通过比较第一特征集合与第二特征集合的交集中特征向量的数量,确定该模型是否可以迁移至新场景。如果交集中特征向量数量较少,则该模型在新场景中表现较差的概率较高,因而可以认为该模型不可以被迁移至新场景;反之,则认为该模型可以被迁移至新场景。
又例如,新场景中的某些特征向量对于模型训练较为重要,则可以在衡量旧场景中模型是否适合迁移到新场景时,重点考虑是否包括这些特征向量,因此,预设迁移条件可以为:根据交集中包括的各特征向量的预设权重计算的加权分值不小于预设阈值。对于模型训练较 为重要的特征向量可以预设较高的权重,且越重要预设权重越高。从而,如果交集中包括的重要特征向量较高,则最终计算的加权分值也较高,并且可以认为该模型可以迁移至新场景。
预设迁移条件也可以为其他形式,并且各迁移条件可以单独使用,也可以搭配使用,本领域技术人员可以根据实际需求灵活地设定,本说明书对此不做具体限定。
此外,预设迁移规则中也可以包括其他具体规则。新场景中所训练的预测模型的类型,可以由研发人员预先根据经验或算法确定并指定,那么,为了在比较特征向量的基础上,进一步地衡量旧场景中各模型,可否迁移至新场景,在获得待迁移模型的集合时,还可以进一步地获得预先为新场景预测模型指定的至少一个类型,并且,预设迁移规则中还可以进一步包括:预先指定的至少一个类型中包括该模型的类型。
从特征向量及模型类型两个维度,确定旧场景中的某个模型,是否可以迁移至新场景,从而使迁移至新场景的模型可以通过进一步的训练,更好地应用于新场景。当然,预设迁移规则中也可以包括其他维度的规则,本说明书实施例对此不做限定。
当然,也可以由研发人员指定旧场景迁移至新场景的待迁移模型,并且研发人员在指定时,也可以根据经验或算法,通过从特征向量、模型类型等维度,衡量各个模型是否可以迁移至新场景、以及迁移后的表现、等等。
S102,从所述待迁移模型的集合中选择至少一个模型,以用于对新场景中的无标签样本进行预测标注;
S103,获得新场景中的初始训练样本集,所述初始训练样本集中包括无标签样本;
S104,利用所选择的模型,为初始训练样本集中的无标签样本添加预测标签;
为了便于描述,将S102至S104结合进行说明。
基于有监督学习训练样本时,需要训练样本为有标签样本。训练样本通常可以通过多种方式标注标签。例如,可以人工进行标注,且人工标注通常较为准确,但是用于模型训练的训练样本数据量通常很大,人工标注效率较低;又如,部分场景中可以根据实际情况产生标签,如信用卡风控场景中,当银行核实某信用卡被盗后,该信用卡及对应交易均可以被标注为黑样本,但这种场景中短期内可能无法获得黑样本标签。
本说明书实施例中,从待迁移模型的集合中选择至少一个模型,用于对新场景中的无标签样本进行预测标注,从而提高标注效率、缩短标注周期。
各个待迁移模型均为可迁移至新场景的模型,但具体地,由于各模型输入的特征向量与 模型类型等存在区别,其中部分模型可以直接较好地应用于新场景,而部分模型则需要更新后才可以较好地应用于新场景,因此,可以从待迁移模型的集合中,选择在新场景中表现较好的部分模型,用于预测标注。
从所述待迁移模型的集合中选择至少一个模型具体可以通过多种方式实现。
在本说明书实施例中,可以首先获得第三特征集合,该集合中包括:预先指定的用于预测新场景中样本标签的若干特征向量;然后获得各待迁移模型对应的各特征集合,其中任一特征集合中包括:对应模型所输入的若干特征向量;根据预设选择规则,从所述待迁移模型的集合中,选择至少一个模型。
与S101中确定待迁移模型类似地,在选择用于预测标注的模型时,也可以从交集中特征向量的数量、重要特征向量的数量、及模型类型是否相同等维度,衡量是否选择某个模型用于预测标注,本说明书在此不再赘述。
此外,仅通过数量或加权分值是否大于预设阈值、模型类型是否相同等硬性条件,可能会出现待迁移模型的集合中,不存在符合预设选择规则的模型的情况,因此,还可以预设各种优先级排序条件,并根据排序结果,选择1个或多个模型用于预测标注。
如S101所述,待迁移模型可以由研发人员指定,而本步骤中从待迁移模型的集合中,选择用于对新场景中的无标签样本进行预测标注的模型时,也可以由研发人员根据经验或算法进行选择,在此不再赘述。
新场景中的初始训练样本集中,可以包括需要被标注预测标签的无标签样本,也可以包括已标注实际标签的有标签样本(可以为白样本和/或黑样本),所选择的模型用于对无标签样本进行预测标注。
具体可以通过多种方式添加预测标签。
在本说明书实施例中,可以预先设定不同数值与不同预测标签的对应关系,如数值大于或小于某预设值时,对应黑样本标签,反之对应白样本标签等。对于所选择的任一模型:将初始训练样本集中的无标签样本输入该模型,得到输出的预测值;对于所输入的任一无标签样本:确定各模型输出的预测值的权重;计算各预测值的加权和,并确定该加权和对应的预测标签;为该无标签样本添加该预测标签。
例如,如果仅从待迁移模型的集合中,选择了1个模型进行预测标注,则可以直接根据该模型输出的预测值(即等于加权和),得到对应的预测标签。
又如,如果从待迁移模型的集合中,选择了多个模型进行预测标注,则可以预先设定各模型输出值对应的权重,如表现更好的模型对应权重更高或更低,当然,也可以预先设定各模型权重相同,即相当于未对各模型设定权重。
此外,利用所选择的模型添加的预测标签,还可以通过人工检查及修正,以提高预测标签的准确度。
对于上述情况,本领域可以根据实际情况灵活地设定,本说明书不做限定。
S105,利用已添加预测标签的初始训练样本集,基于有监督学习算法,对待迁移模型进行更新,得到可适用于新场景的模型。
对待迁移模型进行更新时,可以仅将初始训练样本集中,已添加预测标签的训练样本输入待迁移模型。
如果新场景中所积累的训练样本数量较少,还可以获得旧场景中的训练样本集,该训练样本集中包括已添加实际标签的有标签样本;将新场景中的初始样本集与旧场景中的训练样本集合并,利用合并后的训练样本集,基于有监督学习算法,对待迁移模型进行更新。
旧场景中已经积累了大量的训练样本,且训练样本为已添加实际标签的有标签样本,因此可以在新场景中所积累的训练样本数量较少的情况下,用于辅助新场景中对待迁移模型进行更新。
当然,旧场景中的训练样本,并不一定完全适用于新场景的模型更新,其中可能有部分样本与新场景中的训练样本相似度较高,而其他样本的相似度较低,因此可以为新场景中的初始样本集与旧场景中的训练样本集合并后的训练样本集中的不同训练样本,预设不同的权重。
例如,初始样本集中的各训练样本权重最高,旧场景中的训练样本集中,与初始样本集中的各训练样本相似度较高的各训练样本权重次之,而相似度较低的各训练样本权重最低。
此外,随着时间的推移,新场景中也将积累到已添加实际标签的有标签样本,从而构成优化训练样本集,可以获得新场景中的优化训练样本集,所述优化训练样本集中包括已添加实际标签的有标签样本;将已添加预测标签的初始训练样本集与已添加实际标签的优化训练样本集合并,利用合并后的训练样本集,基于有监督学习算法,对待迁移模型进行更新。
可以理解的是,根据新场景对预测模型的需求,各个待迁移模型可以直接应用于新场景中,并在应用的同时按照本方案进行更新,得到更适用于新场景的模型,也可以经过一段时 间的更新后,再应用于新场景中,并且还可以在应用之后继续进行更新,本说明书不做限定。
下面结合一个更为具体的实例,对本说明书提供的针对新场景的预测模型训练方法进行说明。
金融风险控制领域,可以将积累的大量交易数据作为样本数据,通过机器学习训练风控模型,从而可以基于所训练的风控模型,及时、准确地对新的交易进行风险决策等。
但是,在新场景中建立风险控制模型时,积累训练模型所需的大量样本数据,往往需要较长的时间。例如,样本数据量通常与新场景的交易量及累计时间有关,此外,样本训练集中需要包括一定量的黑样本数据,而以信用卡盗卡等情况作为黑样本数据时,由于银行对盗卡业务处理时间较长等原因,使得短时间内无法积累到训练所需的黑样本数据量。
针对上述问题,可以将旧场景中已有的风控模型,迁移到新场景中。
新场景与旧场景可以为不同国家、地区的交易市场,旧场景中部署使用的风控模型可以包括:盗卡风控模型、盗账户风控模型、隐案识别模型、等等,并且这些风控模型可以是基于多个国家、地区的交易数据训练得到。
如图2所示,可以预先在云端,基于从各个旧场景中集中的数据,训练可在新、旧场景中部署使用的多个模型。
其中,盗卡风控模型与盗账户风控模型,分别针对盗信用卡与盗支付账户的情况进行风险控制,并且可以通过有监督学习训练。
隐案识别模型通过输入针对性更强的特征向量,用于识别银行尚未定为案件(即,非显案)、但具备案件特征的交易。
例如,如果同一台设备(如手机)、或同一网络环境等,同时使用多个信用卡或支付账户,则该设备或环境中存在批量盗卡、盗账户的风险较高;又例如,对于已被列为黑名单的设备、账户、信用卡、网络环境等,与之发生关联的设备、账户、信用卡、网络环境等,存在盗卡、盗账户的风险也较高;在例如,进行了异常交易(如交易量、交易时间、交易地点等异常)的设备、账户、信用卡、网络环境等,存在盗卡、盗账户的风险也较高;隐案识别模型可以基于上述特征,将对应的交易识别为黑样本。
并且,隐案识别模型可以通过无监督学习训练,从而可以应用于尚无实际案件(标签)的场景。
当新场景中需要部署盗卡风控模型、盗账户风控模型及隐案识别模型时,可以以模型文 件的形式,将上述模型下发到新场景本地。并且,本地可以直接使用所部署的模型,对交易事件进行打分、进行风险决策等。
云端下发部署的模型,是通过多个国家、地区的训练样本训练得到,其优点是训练样本全面,通用性较强,而其缺点是与通过新场景本地的数据训练的全新模型相比,并非完全适用于新场景,因此,还需要在新场景积累一定训练样本后,更新这些模型。
各模型部署到新场景本地并使用后,从训练样本积累的角度,可以分为多个阶段。
在第一阶段中,可以认为新场景中积累时间较短,如在部署后的1个周内,积累的训练样本较少,且各样本均无标签,无法进行模型的更新。因此,在第一阶段中,使用云端训练并且未更新的各模型,对新场景中的交易进行风控决策。
在第二阶段中,如在部署后的1个周至1个月之间,可以认为新场景中积累了一定量的训练样本,构成初始训练样本集,如果结合云端下发的旧场景中的大量训练数据,可以对各模型进行更新。但是,由于金融机构处理盗卡、盗账户的周期较长,此时还未积累到带有实际标签的有标签样本,因此可以通过隐案识别模型,为初始训练样本集添加预测样本。
此外,可以对新、旧场景中的训练样本设置不同的权重,例如,新场景为马来西亚的市场,旧场景包括泰国、美国、日本等市场,其中,泰国与马来西亚的消费水平、习惯更接近,交易数据相似度更高,而美国、日本与马来西亚的交易数据相似度更低。因此,可以为马来西亚本地积累的训练样本设置最高的权重,为来自泰国的训练样本设置较高的权重,而为来自美国、日本的训练样本设置更低的权重。从而,通过动态加权的方式,可以在新场景中数据较少的情况下,使更新训练后的各模型更适于新场景。
在第二阶段中更新后的各模型,仍可以用于新场景的交易决策。
在第三阶段中,如在部署1个月之后,可以认为新场景中已经积累了足够量的训练样本,并且积累到了带有实际标签的有标签样本,则可以进一步地更新各模型。更新所使用的训练样本,可以仅包括新场景中、带有实际标签的训练样本,也可以包括新场景中、带有隐案识别模型添加的预测标签的训练样本,还可以包括旧场景中的大量训练样本,等等。
除了通过云端预先训练的模型及积累的数据,在新场景中部署与更新风控模型,新场景中所积累的数据,也可以上传至云端,以用于更新已有模型、训练其他新模型,并部署到其他新场景等。
可见,应用上述方案,可以将旧场景中部署使用的模型迁移至新场景中,并且在新场景中样本积累时间较短,因而样本没有或只有少数实际标签的情况下,通过待迁移模型进行标 签预测,从而进一步优化待迁移模型,使这些模型更适于在新场景中使用,为新场景提供一种更高效且更准确的预测模型训练方案。
相应于上述方法实施例,本说明书实施例还提供一种针对新场景的预测模型训练装置,参见图3所示,该装置可以包括:
待迁移模型获取模块110,用于获得待迁移模型的集合,所述待迁移模型为:在旧场景部署使用、且可迁移至新场景的模型;
标注模型选取模块120,用于从所述待迁移模型的集合中选择至少一个模型,以用于对新场景中的无标签样本进行预测标注;
样本集获取模块130,用于获得新场景中的初始训练样本集,所述初始训练样本集中包括无标签样本;
样本标注模块140,用于利用所选择的模型,为初始训练样本集中的无标签样本添加预测标签;
模型更新模块150,用于利用已添加预测标签的初始训练样本集,基于有监督学习算法,对待迁移模型进行更新,得到可适用于新场景的模型。
在本说明书提供的一种具体实施方式中,所述待迁移模型获取模块110,可以包括:
待迁移特征获取单元111,用于获得第一特征集合,该集合中包括:预先确定的新场景训练样本可提取的若干特征向量;针对在旧场景部署使用的任一模型:获得第二特征集合,该集合中包括:该模型所输入的若干特征向量;
待迁移模型选取单元112,用于在该模型符合预设迁移规则的情况下,将该模型确定为待迁移模型;所述预设迁移规则包括:第一特征集合与第二特征集合的交集中包括的特征向量满足预设迁移条件。
在本说明书提供的一种具体实施方式中,所述预设迁移条件可以包括:
交集中包括的特征向量的数量不小于预设阈值;和/或根据交集中包括的各特征向量的预设权重计算的加权分值不小于预设阈值。
在本说明书提供的一种具体实施方式中,所述待迁移模型获取模块110,还可以包括:待迁移类型获取单元113,用于获得预先为新场景预测模型指定的至少一个类型;
所述预设迁移规则,还可以包括:预先指定的至少一个类型中包括该模型的类型。
在本说明书提供的一种具体实施方式中,所述标注模型选取模块120,可以包括:
标注特征获取单元121,用于获得第三特征集合,该集合中包括:预先指定的用于预测新场景中样本标签的若干特征向量;获得各待迁移模型对应的各特征集合,其中任一特征集合中包括:对应模型所输入的若干特征向量;
标注模型选取单元122,用于根据预设选择规则,从所述待迁移模型的集合中,选择至少一个模型。
在本说明书提供的一种具体实施方式中,所述样本标注模块140,可以包括:
预测值确定单元141,用于对于所选择的任一模型:将初始训练样本集中的无标签样本输入该模型,得到输出的预测值;
预测标签确定单元142,用于对于所输入的任一无标签样本:确定各模型输出的预测值的权重;计算各预测值的加权和,并确定该加权和对应的预测标签;为该无标签样本添加该预测标签。
在本说明书提供的一种具体实施方式中,所述样本集获取模块130,还可以用于:获得新场景中的优化训练样本集,所述优化训练样本集中包括已添加实际标签的有标签样本;
所述模型更新模块150,具体可以用于:将已添加预测标签的初始训练样本集与已添加实际标签的优化训练样本集合并,利用合并后的训练样本集,基于有监督学习算法,对待迁移模型进行更新。
在本说明书提供的一种具体实施方式中,所述样本集获取模块130,还可以用于:获得旧场景中的训练样本集,该训练样本集中包括已添加实际标签的有标签样本;
所述模型更新模块150,具体可以用于:将新场景中的初始样本集与旧场景中的训练样本集合并,利用合并后的训练样本集,基于有监督学习算法,对待迁移模型进行更新。
上述装置中各个模块的功能和作用的实现过程具体详见上述方法中对应步骤的实现过程,在此不再赘述。
本说明书实施例还提供一种计算机设备,其至少包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序,其中,处理器执行所述程序时实现前述的针对新场景的预测模型训练方法。该方法至少包括:
获得待迁移模型的集合,所述待迁移模型为:在旧场景部署使用、且可迁移至新场景的模型;
从所述待迁移模型的集合中选择至少一个模型,以用于对新场景中的无标签样本进行预测标注;
获得新场景中的初始训练样本集,所述初始训练样本集中包括无标签样本;
利用所选择的模型,为初始训练样本集中的无标签样本添加预测标签;
利用已添加预测标签的初始训练样本集,基于有监督学习算法,对待迁移模型进行更新,得到可适用于新场景的模型。
图4示出了本说明书实施例所提供的一种更为具体的计算设备硬件结构示意图,该设备可以包括:处理器1010、存储器1020、输入/输出接口1030、通信接口1040和总线1050。其中处理器1010、存储器1020、输入/输出接口1030和通信接口1040通过总线1050实现彼此之间在设备内部的通信连接。
处理器1010可以采用通用的CPU(Central Processing Unit,中央处理器)、微处理器、应用专用集成电路(Application Specific Integrated Circuit,ASIC)、或者一个或多个集成电路等方式实现,用于执行相关程序,以实现本说明书实施例所提供的技术方案。
存储器1020可以采用ROM(Read Only Memory,只读存储器)、RAM(Random Access Memory,随机存取存储器)、静态存储设备,动态存储设备等形式实现。存储器1020可以存储操作系统和其他应用程序,在通过软件或者固件来实现本说明书实施例所提供的技术方案时,相关的程序代码保存在存储器1020中,并由处理器1010来调用执行。
输入/输出接口1030用于连接输入/输出模块,以实现信息输入及输出。输入输出/模块可以作为组件配置在设备中(图中未示出),也可以外接于设备以提供相应功能。其中输入设备可以包括键盘、鼠标、触摸屏、麦克风、各类传感器等,输出设备可以包括显示器、扬声器、振动器、指示灯等。
通信接口1040用于连接通信模块(图中未示出),以实现本设备与其他设备的通信交互。其中通信模块可以通过有线方式(例如USB、网线等)实现通信,也可以通过无线方式(例如移动网络、WIFI、蓝牙等)实现通信。
总线1050包括一通路,在设备的各个组件(例如处理器1010、存储器1020、输入/输出接口1030和通信接口1040)之间传输信息。
需要说明的是,尽管上述设备仅示出了处理器1010、存储器1020、输入/输出接口1030、通信接口1040以及总线1050,但是在具体实施过程中,该设备还可以包括实现正常运行所 必需的其他组件。此外,本领域的技术人员可以理解的是,上述设备中也可以仅包含实现本说明书实施例方案所必需的组件,而不必包含图中所示的全部组件。
本说明书实施例还提供一种计算机可读存储介质,其上存储有计算机程序,该程序被处理器执行时实现前述的针对新场景的预测模型训练方法。该方法至少包括:
获得待迁移模型的集合,所述待迁移模型为:在旧场景部署使用、且可迁移至新场景的模型;
从所述待迁移模型的集合中选择至少一个模型,以用于对新场景中的无标签样本进行预测标注;
获得新场景中的初始训练样本集,所述初始训练样本集中包括无标签样本;
利用所选择的模型,为初始训练样本集中的无标签样本添加预测标签;
利用已添加预测标签的初始训练样本集,基于有监督学习算法,对待迁移模型进行更新,得到可适用于新场景的模型。
计算机可读介质包括永久性和非永久性、可移动和非可移动媒体可以由任何方法或技术来实现信息存储。信息可以是计算机可读指令、数据结构、程序的模块或其他数据。计算机的存储介质的例子包括,但不限于相变内存(PRAM)、静态随机存取存储器(SRAM)、动态随机存取存储器(DRAM)、其他类型的随机存取存储器(RAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、快闪记忆体或其他内存技术、只读光盘只读存储器(CD-ROM)、数字多功能光盘(DVD)或其他光学存储、磁盒式磁带,磁带磁磁盘存储或其他磁性存储设备或任何其他非传输介质,可用于存储可以被计算设备访问的信息。按照本文中的界定,计算机可读介质不包括暂存电脑可读媒体(transitory media),如调制的数据信号和载波。
通过以上的实施方式的描述可知,本领域的技术人员可以清楚地了解到本说明书实施例可借助软件加必需的通用硬件平台的方式来实现。基于这样的理解,本说明书实施例的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来,该计算机软件产品可以存储在存储介质中,如ROM/RAM、磁碟、光盘等,包括若干指令用以使得一台计算机设备(可以是个人计算机,服务器,或者网络设备等)执行本说明书实施例各个实施例或者实施例的某些部分所述的方法。
上述实施例阐明的系统、装置、模块或单元,具体可以由计算机芯片或实体实现,或者由具有某种功能的产品来实现。一种典型的实现设备为计算机,计算机的具体形式可以是 个人计算机、膝上型计算机、蜂窝电话、相机电话、智能电话、个人数字助理、媒体播放器、导航设备、电子邮件收发设备、游戏控制台、平板计算机、可穿戴设备或者这些设备中的任意几种设备的组合。
本说明书中的各个实施例均采用递进的方式描述,各个实施例之间相同相似的部分互相参见即可,每个实施例重点说明的都是与其他实施例的不同之处。尤其,对于装置实施例而言,由于其基本相似于方法实施例,所以描述得比较简单,相关之处参见方法实施例的部分说明即可。以上所描述的装置实施例仅仅是示意性的,其中所述作为分离部件说明的模块可以是或者也可以不是物理上分开的,在实施本说明书实施例方案时可以把各模块的功能在同一个或多个软件和/或硬件中实现。也可以根据实际的需要选择其中的部分或者全部模块来实现本实施例方案的目的。本领域普通技术人员在不付出创造性劳动的情况下,即可以理解并实施。
以上所述仅是本说明书实施例的具体实施方式,应当指出,对于本技术领域的普通技术人员来说,在不脱离本说明书实施例原理的前提下,还可以做出若干改进和润饰,这些改进和润饰也应视为本说明书实施例的保护范围。

Claims (17)

  1. 一种针对新场景的预测模型训练方法,该方法包括:
    获得待迁移模型的集合,所述待迁移模型为:在旧场景部署使用、且可迁移至新场景的模型;
    从所述待迁移模型的集合中选择至少一个模型,以用于对新场景中的无标签样本进行预测标注;
    获得新场景中的初始训练样本集,所述初始训练样本集中包括无标签样本;
    利用所选择的模型,为初始训练样本集中的无标签样本添加预测标签;
    利用已添加预测标签的初始训练样本集,基于有监督学习算法,对待迁移模型进行更新,得到可适用于新场景的模型。
  2. 根据权利要求1所述的方法,所述获得待迁移模型的集合,包括:
    获得第一特征集合,该集合中包括:预先确定的新场景训练样本可提取的若干特征向量;
    针对在旧场景部署使用的任一模型:
    获得第二特征集合,该集合中包括:该模型所输入的若干特征向量;
    在该模型符合预设迁移规则的情况下,将该模型确定为待迁移模型;所述预设迁移规则包括:第一特征集合与第二特征集合的交集中包括的特征向量满足预设迁移条件。
  3. 根据权利要求2所述的方法,所述预设迁移条件包括:
    交集中包括的特征向量的数量不小于预设阈值;
    和/或
    根据交集中包括的各特征向量的预设权重计算的加权分值不小于预设阈值。
  4. 根据权利要求2所述的方法,
    所述获得待迁移模型的集合,还包括:获得预先为新场景预测模型指定的至少一个类型;
    所述预设迁移规则,还包括:预先指定的至少一个类型中包括该模型的类型。
  5. 根据权利要求1所述的方法,所述从所述待迁移模型的集合中选择至少一个模型,包括:
    获得第三特征集合,该集合中包括:预先指定的用于预测新场景中样本标签的若干特征向量;
    获得各待迁移模型对应的各特征集合,其中任一特征集合中包括:对应模型所输入的若干特征向量;
    根据预设选择规则,从所述待迁移模型的集合中,选择至少一个模型。
  6. 根据权利要求1所述的方法,所述利用所选择的模型,为初始训练样本集中的无标签 样本添加预测标签,包括:
    对于所选择的任一模型:将初始训练样本集中的无标签样本输入该模型,得到输出的预测值;
    对于所输入的任一无标签样本:确定各模型输出的预测值的权重;计算各预测值的加权和,并确定该加权和对应的预测标签;为该无标签样本添加该预测标签。
  7. 根据权利要求1所述的方法,所述利用已添加预测标签的初始训练样本集,基于有监督学习算法,对待迁移模型进行更新,包括:
    获得新场景中的优化训练样本集,所述优化训练样本集中包括已添加实际标签的有标签样本;
    将已添加预测标签的初始训练样本集与已添加实际标签的优化训练样本集合并,利用合并后的训练样本集,基于有监督学习算法,对待迁移模型进行更新。
  8. 根据权利要求1所述的方法,所述利用已添加预测标签的初始训练样本集,基于有监督学习算法,对待迁移模型进行更新,包括:
    获得旧场景中的训练样本集,该训练样本集中包括已添加实际标签的有标签样本;
    将新场景中的初始样本集与旧场景中的训练样本集合并,利用合并后的训练样本集,基于有监督学习算法,对待迁移模型进行更新。
  9. 一种针对新场景的预测模型训练装置,该装置包括:
    待迁移模型获取模块,用于获得待迁移模型的集合,所述待迁移模型为:在旧场景部署使用、且可迁移至新场景的模型;
    标注模型选取模块,用于从所述待迁移模型的集合中选择至少一个模型,以用于对新场景中的无标签样本进行预测标注;
    样本集获取模块,用于获得新场景中的初始训练样本集,所述初始训练样本集中包括无标签样本;
    样本标注模块,用于利用所选择的模型,为初始训练样本集中的无标签样本添加预测标签;
    模型更新模块,用于利用已添加预测标签的初始训练样本集,基于有监督学习算法,对待迁移模型进行更新,得到可适用于新场景的模型。
  10. 根据权利要求9所述的装置,所述待迁移模型获取模块,包括:
    待迁移特征获取单元,用于获得第一特征集合,该集合中包括:预先确定的新场景训练样本可提取的若干特征向量;针对在旧场景部署使用的任一模型:获得第二特征集合,该集合中包括:该模型所输入的若干特征向量;
    待迁移模型选取单元,用于在该模型符合预设迁移规则的情况下,将该模型确定为待迁移模型;所述预设迁移规则包括:第一特征集合与第二特征集合的交集中包括的特征向量满足预设迁移条件。
  11. 根据权利要求10所述的装置,所述预设迁移条件包括:
    交集中包括的特征向量的数量不小于预设阈值;
    和/或
    根据交集中包括的各特征向量的预设权重计算的加权分值不小于预设阈值。
  12. 根据权利要求10所述的装置,
    所述待迁移模型获取模块,还包括:待迁移类型获取单元,用于获得预先为新场景预测模型指定的至少一个类型;
    所述预设迁移规则,还包括:预先指定的至少一个类型中包括该模型的类型。
  13. 根据权利要求9所述的装置,所述标注模型选取模块,包括:
    标注特征获取单元,用于获得第三特征集合,该集合中包括:预先指定的用于预测新场景中样本标签的若干特征向量;获得各待迁移模型对应的各特征集合,其中任一特征集合中包括:对应模型所输入的若干特征向量;
    标注模型选取单元,用于根据预设选择规则,从所述待迁移模型的集合中,选择至少一个模型。
  14. 根据权利要求9所述的装置,所述样本标注模块,包括:
    预测值确定单元,用于对于所选择的任一模型:将初始训练样本集中的无标签样本输入该模型,得到输出的预测值;
    预测标签确定单元,用于对于所输入的任一无标签样本:确定各模型输出的预测值的权重;计算各预测值的加权和,并确定该加权和对应的预测标签;为该无标签样本添加该预测标签。
  15. 根据权利要求9所述的装置,
    所述样本集获取模块,还用于:获得新场景中的优化训练样本集,所述优化训练样本集中包括已添加实际标签的有标签样本;
    所述模型更新模块,具体用于:将已添加预测标签的初始训练样本集与已添加实际标签的优化训练样本集合并,利用合并后的训练样本集,基于有监督学习算法,对待迁移模型进行更新。
  16. 根据权利要求9所述的装置,
    所述样本集获取模块,还用于:获得旧场景中的训练样本集,该训练样本集中包括已添 加实际标签的有标签样本;
    所述模型更新模块,具体用于:将新场景中的初始样本集与旧场景中的训练样本集合并,利用合并后的训练样本集,基于有监督学习算法,对待迁移模型进行更新。
  17. 一种计算机设备,包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序,其中,所述处理器执行所述程序时实现如权利要求1至8任一项所述的方法。
PCT/CN2019/091658 2018-08-03 2019-06-18 一种针对新场景的预测模型训练方法及装置 WO2020024716A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810875574.1A CN109359793B (zh) 2018-08-03 2018-08-03 一种针对新场景的预测模型训练方法及装置
CN201810875574.1 2018-08-03

Publications (1)

Publication Number Publication Date
WO2020024716A1 true WO2020024716A1 (zh) 2020-02-06

Family

ID=65349816

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/091658 WO2020024716A1 (zh) 2018-08-03 2019-06-18 一种针对新场景的预测模型训练方法及装置

Country Status (3)

Country Link
CN (1) CN109359793B (zh)
TW (1) TWI818999B (zh)
WO (1) WO2020024716A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111859872A (zh) * 2020-07-07 2020-10-30 中国建设银行股份有限公司 一种文本标注方法和装置

Families Citing this family (33)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109359793B (zh) * 2018-08-03 2020-11-17 创新先进技术有限公司 一种针对新场景的预测模型训练方法及装置
CN110033276A (zh) * 2019-03-08 2019-07-19 阿里巴巴集团控股有限公司 一种针对转账的安全策略生成方法、装置及设备
CN111797289A (zh) * 2019-04-09 2020-10-20 Oppo广东移动通信有限公司 模型处理方法、装置、存储介质和电子设备
CN110083834B (zh) * 2019-04-24 2023-05-09 北京百度网讯科技有限公司 语义匹配模型训练方法、装置、电子设备及存储介质
CN110232403B (zh) * 2019-05-15 2024-02-27 腾讯科技(深圳)有限公司 一种标签预测方法、装置、电子设备及介质
CN110163296B (zh) * 2019-05-29 2020-12-18 北京达佳互联信息技术有限公司 图像识别的方法、装置、设备及存储介质
CN110263824B (zh) * 2019-05-29 2023-09-05 创新先进技术有限公司 模型的训练方法、装置、计算设备及计算机可读存储介质
CN110390425A (zh) * 2019-06-20 2019-10-29 阿里巴巴集团控股有限公司 预测方法以及装置
CN110458393B (zh) * 2019-07-05 2023-07-18 创新先进技术有限公司 一种风险识别方案的确定方法、装置及电子设备
CN110689135B (zh) * 2019-09-05 2022-10-11 第四范式(北京)技术有限公司 一种反洗钱模型的训练方法、装置及电子设备
CN110838020B (zh) * 2019-09-16 2023-06-23 平安科技(深圳)有限公司 基于向量迁移的推荐方法、装置、计算机设备及存储介质
CN110765876A (zh) * 2019-09-20 2020-02-07 深圳码隆科技有限公司 一种商品识别模型的训练方法及装置
CN112581250B (zh) * 2019-09-30 2023-12-29 深圳无域科技技术有限公司 模型生成方法、装置、计算机设备和存储介质
CN110705717B (zh) * 2019-09-30 2022-05-17 支付宝(杭州)信息技术有限公司 计算机执行的机器学习模型的训练方法、装置及设备
CN110928889A (zh) * 2019-10-23 2020-03-27 深圳市华讯方舟太赫兹科技有限公司 训练模型更新方法、设备以及计算机存储介质
CN110910864B (zh) * 2019-10-24 2023-02-03 深圳追一科技有限公司 训练样本的选取方法、装置、计算机设备和存储介质
CN111062563A (zh) * 2019-11-08 2020-04-24 支付宝(杭州)信息技术有限公司 风险预测模型的训练方法、风险预测方法及相关装置
CN112861892B (zh) * 2019-11-27 2023-09-01 杭州海康威视数字技术股份有限公司 图片中目标的属性的确定方法和装置
CN111309715B (zh) * 2020-01-15 2023-04-18 腾讯科技(深圳)有限公司 通话场景的识别方法和装置
CN111352965B (zh) * 2020-02-18 2023-09-08 腾讯科技(深圳)有限公司 序列挖掘模型的训练方法、序列数据的处理方法及设备
CN111428783B (zh) * 2020-03-23 2022-06-21 支付宝(杭州)信息技术有限公司 对推荐模型的训练样本进行样本域转换的方法及装置
CN111488972B (zh) * 2020-04-09 2023-08-08 北京百度网讯科技有限公司 数据迁移方法、装置、电子设备和存储介质
CN111598338B (zh) * 2020-05-18 2021-08-31 贝壳找房(北京)科技有限公司 用于更新预测模型的方法、装置、介质和电子设备
CN113780314A (zh) * 2020-05-20 2021-12-10 阿里巴巴集团控股有限公司 一种分类模型训练方法、装置和系统
CN112036509A (zh) * 2020-09-30 2020-12-04 北京百度网讯科技有限公司 用于训练图像识别模型的方法和装置
CN112270545A (zh) * 2020-10-27 2021-01-26 上海淇馥信息技术有限公司 基于迁移样本筛选的金融风险预测方法、装置和电子设备
CN112258068B (zh) * 2020-10-30 2022-03-29 武汉理工光科股份有限公司 一种安保活动方案生成及动态更新方法、装置
CN112417485B (zh) * 2020-11-30 2024-02-02 支付宝(杭州)信息技术有限公司 一种基于可信执行环境的模型训练方法、系统及装置
CN112417767B (zh) * 2020-12-09 2024-02-27 东软睿驰汽车技术(沈阳)有限公司 一种衰减趋势确定模型构建方法、衰减趋势确定方法
CN112711938B (zh) * 2021-03-26 2021-07-06 北京沃丰时代数据科技有限公司 一种阅读理解模型构建方法、装置、电子设备及存储介质
CN115396831A (zh) * 2021-05-08 2022-11-25 中国移动通信集团浙江有限公司 交互模型生成方法、装置、设备及存储介质
CN113325855B (zh) * 2021-08-02 2021-11-30 北京三快在线科技有限公司 基于迁移场景用于预测障碍物轨迹的模型训练方法
CN113780578B (zh) * 2021-09-08 2023-12-12 北京百度网讯科技有限公司 模型训练方法、装置、电子设备及可读存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106599922A (zh) * 2016-12-16 2017-04-26 中国科学院计算技术研究所 用于大规模数据标定的迁移学习方法及系统
CN107895177A (zh) * 2017-11-17 2018-04-10 南京邮电大学 一种保持图像分类稀疏结构的迁移分类学习方法
CN108304936A (zh) * 2017-07-12 2018-07-20 腾讯科技(深圳)有限公司 机器学习模型训练方法和装置、表情图像分类方法和装置
CN109359793A (zh) * 2018-08-03 2019-02-19 阿里巴巴集团控股有限公司 一种针对新场景的预测模型训练方法及装置

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2012243180A (ja) * 2011-05-23 2012-12-10 Sony Corp 学習装置および方法、並びにプログラム
TWI553573B (zh) * 2014-05-15 2016-10-11 財團法人工業技術研究院 面向口碑分析與檢視系統、裝置及方法
US20160253597A1 (en) * 2015-02-27 2016-09-01 Xerox Corporation Content-aware domain adaptation for cross-domain classification
US10504029B2 (en) * 2015-06-30 2019-12-10 Microsoft Technology Licensing, Llc Personalized predictive models
US10776693B2 (en) * 2017-01-31 2020-09-15 Xerox Corporation Method and system for learning transferable feature representations from a source domain for a target domain
CN106991439A (zh) * 2017-03-28 2017-07-28 南京天数信息科技有限公司 基于深度学习与迁移学习的图像识别方法
CN107679859B (zh) * 2017-07-18 2020-08-25 中国银联股份有限公司 一种基于迁移深度学习的风险识别方法以及系统
CN108021931A (zh) * 2017-11-20 2018-05-11 阿里巴巴集团控股有限公司 一种数据样本标签处理方法及装置
CN107944874B (zh) * 2017-12-13 2021-07-20 创新先进技术有限公司 基于迁移学习的风控方法、装置及系统

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106599922A (zh) * 2016-12-16 2017-04-26 中国科学院计算技术研究所 用于大规模数据标定的迁移学习方法及系统
CN108304936A (zh) * 2017-07-12 2018-07-20 腾讯科技(深圳)有限公司 机器学习模型训练方法和装置、表情图像分类方法和装置
CN107895177A (zh) * 2017-11-17 2018-04-10 南京邮电大学 一种保持图像分类稀疏结构的迁移分类学习方法
CN109359793A (zh) * 2018-08-03 2019-02-19 阿里巴巴集团控股有限公司 一种针对新场景的预测模型训练方法及装置

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111859872A (zh) * 2020-07-07 2020-10-30 中国建设银行股份有限公司 一种文本标注方法和装置

Also Published As

Publication number Publication date
TW202008237A (zh) 2020-02-16
TWI818999B (zh) 2023-10-21
CN109359793B (zh) 2020-11-17
CN109359793A (zh) 2019-02-19

Similar Documents

Publication Publication Date Title
WO2020024716A1 (zh) 一种针对新场景的预测模型训练方法及装置
JP6940646B2 (ja) 情報推奨方法、情報推奨装置、機器および媒体
US11314988B2 (en) Image aesthetic processing method and electronic device
JP6912591B2 (ja) リスク管理制御方法及びデバイス
US20220044289A1 (en) Document term recognition and analytics
US20180342004A1 (en) Cumulative success-based recommendations for repeat users
US20180268306A1 (en) Using Different Data Sources for a Predictive Model
US20200134716A1 (en) Systems and methods for determining credit worthiness of a borrower
CN111553488B (zh) 一种针对用户行为的风险识别模型训练方法及系统
CN109214421B (zh) 一种模型训练方法、装置、及计算机设备
US11238368B2 (en) Machine learning and security classification of user accounts
US20210350393A1 (en) Behavioral data analytics platform
JP2017509947A (ja) ネットワークサブジェクトの社会的関係タイプの取得
JP6990163B2 (ja) 予測装置、予測方法、及び予測プログラム
US11687839B2 (en) System and method for generating and optimizing artificial intelligence models
CN111125529A (zh) 产品匹配方法、装置、计算机设备及存储介质
CN111611390B (zh) 一种数据处理方法及装置
CN114187112A (zh) 账户风险模型的训练方法和风险用户群体的确定方法
CN109635953A (zh) 一种特征衍生方法、装置及电子设备
CN110134860B (zh) 用户画像生成方法、装置和设备
JP2019185595A (ja) 情報処理装置、情報処理方法、情報処理プログラム、判定装置、判定方法及び判定プログラム
US11861698B2 (en) Vehicle selection platform
US20220207420A1 (en) Utilizing machine learning models to characterize a relationship between a user and an entity
CN108446738A (zh) 一种聚类方法、装置及电子设备
US20170068977A1 (en) Estimating prospect lifetime values

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: 19844530

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: 19844530

Country of ref document: EP

Kind code of ref document: A1