CN112070233A - Model joint training method and device, electronic equipment and storage medium - Google Patents
Model joint training method and device, electronic equipment and storage medium Download PDFInfo
- Publication number
- CN112070233A CN112070233A CN202010866349.9A CN202010866349A CN112070233A CN 112070233 A CN112070233 A CN 112070233A CN 202010866349 A CN202010866349 A CN 202010866349A CN 112070233 A CN112070233 A CN 112070233A
- Authority
- CN
- China
- Prior art keywords
- model
- user
- similar
- target
- target task
- Prior art date
- Legal status (The legal status 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 status listed.)
- Granted
Links
- 238000012549 training Methods 0.000 title claims abstract description 106
- 238000000034 method Methods 0.000 title claims abstract description 59
- 238000012512 characterization method Methods 0.000 claims abstract description 76
- 238000004821 distillation Methods 0.000 claims abstract description 59
- 238000013140 knowledge distillation Methods 0.000 claims abstract description 16
- 230000015654 memory Effects 0.000 claims description 20
- 238000004422 calculation algorithm Methods 0.000 claims description 10
- 230000003993 interaction Effects 0.000 claims description 8
- 238000013136 deep learning model Methods 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 5
- 238000000605 extraction Methods 0.000 claims description 2
- 239000000126 substance Substances 0.000 claims description 2
- 238000013473 artificial intelligence Methods 0.000 abstract description 3
- 238000010586 diagram Methods 0.000 description 11
- 238000013135 deep learning Methods 0.000 description 8
- 230000007547 defect Effects 0.000 description 8
- 238000004891 communication Methods 0.000 description 4
- 238000004590 computer program Methods 0.000 description 4
- 230000006870 function Effects 0.000 description 4
- 238000010801 machine learning Methods 0.000 description 4
- 238000013507 mapping Methods 0.000 description 4
- 230000008569 process Effects 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 239000004973 liquid crystal related substance Substances 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000013526 transfer learning Methods 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000001413 cellular effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000010295 mobile communication Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000001953 sensory effect Effects 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Abstract
The application discloses a model joint training method and device, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence and cloud computing. The specific implementation scheme is as follows: performing knowledge distillation according to the user similar characteristics and the user similar labels under similar tasks to obtain a similar distillation model, and obtaining a similar characterization model according to the similar distillation model; and constructing a target task model according to the user target characteristics under the target task and the user target labels acquired from the label provider based on the similar characterization model, and predicting the target task for the user. The method and the device can improve the prediction accuracy of the target task model.
Description
Technical Field
The application relates to the technical field of artificial intelligence and cloud computing, in particular to the technical field of deep learning, and specifically relates to a model joint training method and device, electronic equipment and a storage medium.
Background
Machine learning is the core of artificial intelligence, and based on multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and the like, the existing knowledge structure is reorganized by a computer to continuously improve the performance of the computer.
Federated machine learning (Federated machine learning) is used to help organizations model data usage and machine learning while meeting the requirements of user privacy protection, data security, and government regulations.
Disclosure of Invention
The disclosure provides a method, a device, an electronic device and a storage medium for model joint training.
According to an aspect of the present disclosure, there is provided a model joint training method, including:
performing knowledge distillation according to the user similar characteristics and the user similar labels under similar tasks to obtain a similar distillation model, and obtaining a similar characterization model according to the similar distillation model;
and constructing a target task model according to the user target characteristics under the target task and the user target labels acquired from the label provider based on the similar characterization model, and predicting the target task for the user.
According to another aspect of the present disclosure, there is provided a model joint training method, including:
determining a user target label under a target task;
sending a model training request carrying the user target label to a model trainer for instructing the model trainer to execute the following steps: performing knowledge distillation according to the user similar characteristics and the user similar labels under similar tasks to obtain a similar distillation model, and obtaining a similar characterization model according to the similar distillation model; and constructing a target task model according to the user target characteristics and the user target labels under the target task based on the similar characterization model.
According to still another aspect of the present disclosure, there is provided a model joint training apparatus including:
the characterization model module is used for carrying out knowledge distillation according to the user similar characteristics and the user similar labels under similar tasks to obtain a similar distillation model, and obtaining a similar characterization model according to the similar distillation model;
and the target task model module is used for constructing a target task model according to the user target characteristics under the target task and the user target labels acquired from the label provider based on the similar characterization model and is used for predicting the target task of the user.
According to still another aspect of the present disclosure, there is provided a model joint training apparatus including:
the target label determining module is used for determining a user target label under a target task;
a training request sending module, configured to send a model training request carrying the user target label to a model trainer, and instruct the model trainer to execute the following: performing knowledge distillation according to the user similar characteristics and the user similar labels under similar tasks to obtain a similar distillation model, and obtaining a similar characterization model according to the similar distillation model; and constructing a target task model according to the user target characteristics and the user target labels under the target task based on the similar characterization model.
According to a fifth aspect, there is provided an electronic device comprising:
at least one processor; and
a memory communicatively coupled to the at least one processor; wherein the content of the first and second substances,
the memory stores instructions executable by the at least one processor to enable the at least one processor to perform a model co-training method as described in any of the embodiments of the present application.
According to a sixth aspect, there is provided a non-transitory computer readable storage medium having stored thereon computer instructions for causing the computer to perform a model co-training method as described in any of the embodiments of the present application.
According to the technology of the application, the prediction accuracy of the target task model can be improved.
It should be understood that the statements in this section do not necessarily identify key or critical features of the embodiments of the present disclosure, nor do they limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description.
Drawings
The drawings are included to provide a better understanding of the present solution and are not intended to limit the present application. Wherein:
FIG. 1 is a schematic flow chart diagram illustrating a model joint training method according to an embodiment of the present disclosure;
FIG. 2 is a schematic flow chart diagram of another model joint training method provided in an embodiment of the present application;
FIG. 3 is a schematic flow chart diagram illustrating a further method for model joint training according to an embodiment of the present disclosure;
FIG. 4 is a schematic flow chart diagram illustrating a further method for model joint training according to an embodiment of the present disclosure;
FIG. 5 is a schematic flow chart diagram illustrating a further method for model joint training according to an embodiment of the present disclosure;
FIG. 6 is a schematic structural diagram of a model joint training device according to an embodiment of the present application;
FIG. 7 is a schematic structural diagram of another model joint training device provided in an embodiment of the present application;
FIG. 8 is a block diagram of an electronic device for implementing a model joint training method according to an embodiment of the present application.
Detailed Description
The following description of the exemplary embodiments of the present application, taken in conjunction with the accompanying drawings, includes various details of the embodiments of the application for the understanding of the same, which are to be considered exemplary only. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the present application. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.
Fig. 1 is a schematic flow chart of a model joint training method provided according to an embodiment of the present application. The embodiment can be applied to the situation that the label provider and the model trainer jointly train the target task model. The model joint training method disclosed in this embodiment may be executed by an electronic device, and specifically may be executed by a model joint training apparatus, which may be implemented by software and/or hardware and configured in an electronic device of a model training party. Referring to fig. 1, the model joint training method provided in this embodiment includes:
s110, performing knowledge distillation according to the user similar characteristics and the user similar labels under the similar tasks to obtain a similar distillation model, and obtaining a similar characterization model according to the similar distillation model.
The similar tasks and the target tasks can be similar in the scene, and the participants can be different, namely the prediction targets of the similar tasks and the target tasks are similar. The participants of the target task may be label providers and model trainers, while the participants of similar tasks may be model trainers.
The label provider is used for providing a user target label under a target task, and the model trainer is used for providing a user similar label under a similar task and user characteristics under the target task and the similar task. In addition, the tag extractor may also be used to provide part of the user features. Taking different wind control scenes as an example, if party A has a small number of wind control sub-label values of the wind control scene A, party B provides a wind control sub-prediction service of the wind control scene B, and the prediction service is determined based on a large number of wind control sub-label values of party B, party A can be used as a label provider, the wind control sub-label value of party A is used as a user target label, party B is used as a model trainer, the wind control sub-label value of party B is used as a user similar label, and a user feature associated with the user similar label is used as a user similar feature.
Specifically, the user similarity characteristics are used as input of a distillation model, and the user similarity labels are used as distillation targets to conduct distillation model training to obtain a similar distillation model. And the model before the middle layer of the similar distillation model can be used as a similar characterization model, namely the middle layer output of the similar distillation model is used as a similar characterization feature. For example, the output layers of a similar distillation model can be removed to obtain a similar characterization model. Moreover, the dimensionality of the similar characteristic features can be adjusted according to the sparsity of the characteristic features of the similar characteristics of the user, and if the characteristic features are sparse, the dimensionality of the similar characteristic features can be reduced; if dense, the dimensions of similar characterizing features may be increased. For example, the dimension of a similar characterizing feature may be set to 50. By distilling the similar task models under the similar tasks, the similar distillation models and the similar characterization models can keep the information of the similar labels of the users.
And S120, constructing a target task model based on the similar characterization model according to the user target characteristics under the target task and the user target label acquired from the label provider, and predicting the target task for the user.
In the combined modeling process, the training sample size greatly influences the modeling effect, the small training sample size can cause insufficient modeling, and the sample label data size generally provided by the label provider is limited, so that the modeling effect is limited. In the application, because the similar representation model retains the user similar label information in the similar scene, the target task model is built based on the similar representation model, namely the target task model is built by indirectly using the user similar label, so that the target task model not only covers the user target label of the label provider but also comprises the user similar label used in the similar task in the similar scene, the defect of insufficient number of the user target labels can be overcome, and the prediction accuracy of the target task model can be improved.
The user target characteristics under the target task can be determined by the model trainer, or can be determined by both the model trainer and the label provider, that is, both the model trainer and the label provider can provide part of the user target characteristics.
Wherein the target task model is a joint training result. After the target task model is built, the characteristics of the user to be predicted are input into the target task model, and the prediction result of the user to be predicted under the target task can be obtained.
According to the technical scheme of the embodiment of the application, the similar scene and the similar task prediction model are distilled, the similar representation model is extracted from the similar distillation model, and the combined modeling is carried out based on the similar representation model, namely the target task model is constructed by indirectly using the user similar labels, so that the defect of insufficient user target labels can be overcome, and the prediction accuracy of the target task model can be improved.
Fig. 2 is a schematic flowchart of a model joint training method according to an embodiment of the present application. The present embodiment is an alternative proposed on the basis of the above-described embodiments. Referring to fig. 2, the model joint training method provided in this embodiment includes:
s210, training the deep learning model according to the user similar characteristics and the user similar labels under the similar tasks to obtain a similar distillation model, and obtaining a similar characterization model according to the similar distillation model.
In the related technology of joint learning, under the condition of ensuring data security, the data of each participant can not go out of the local through security strategies such as homomorphic encryption and the like, and joint training is carried out. Since homomorphic encryption only supports addition and multiplication, the related art can only use a joint modeling model including only addition and multiplication, such as an XGBoost tree model, and the training efficiency of the model using the homomorphic encryption algorithm is low.
In the application, the user similarity feature and the user similarity label are provided by the model training party, so that the model training party does not need to adopt a homomorphic encryption algorithm to construct the similar distillation model, and is not limited by an operator and training efficiency supported by the homomorphic encryption algorithm. The similar distillation model is constructed on the basis of the deep learning network structure, compared with a tree model, the method has better generalization and feature characterization capability, and therefore the similar characterization model obtained according to the similar distillation model has better expression on similar scenes.
And S220, constructing a target task model based on the similar characterization model according to the user target characteristics under the target task and the user target label acquired from the label provider, and predicting the target task for the user.
In an alternative embodiment, S220 includes: connecting the similar representation model with a prediction output layer of a target task to obtain a joint prediction model of joint modeling; and training the combined prediction model according to the user target characteristics under the target task and the user target labels acquired from the label provider to obtain the target task model.
Specifically, a joint prediction model is built according to the similar identification model, user target characteristics are used as input of the joint prediction model, a user target label is used as output of the joint prediction model, the joint prediction model is trained, and a training result is used as a target task model. And in the transfer learning, the simulated representation model is used as a pre-training model, and then the supervised transfer learning is carried out based on the user target label, so that the over-fitting problem caused by the limited user target label in the supervised learning can be relieved.
In an alternative embodiment, S220 includes: taking the user target characteristics under the target task as the input of the similar characterization model to obtain the similar characterization characteristics of the user; fusing the user target characteristics and the similar characteristic characteristics under the target task to obtain user joint characteristics; and constructing a target task model according to the user joint characteristics and the user target label acquired from the label provider.
Specifically, the user joint features can be obtained by splicing the user target features and the similar characterization features, the user joint features are used as input, and the user target labels are used as output to perform model training. The model may be a tree model. Similar characterization features are spliced to user target features as new features, a more flexible application mode is provided, more models are selected in combined modeling, and additional characterization information is provided for assisting in modeling by the similar identification features.
According to the technical scheme of the embodiment of the application, the deep learning distillation model is adopted, so that the generalization and characteristic characterization capabilities are better; and by providing two joint modeling modes, the overfitting problem caused by limited label data in supervised learning can be relieved or the flexibility of model selection is provided.
Fig. 3 is a schematic flowchart of a model joint training method according to an embodiment of the present application. The present embodiment is an alternative proposed on the basis of the above-described embodiments. Referring to fig. 3, the model joint training method provided in this embodiment includes:
s310, sampling training samples of the similar tasks according to the user label value distribution of the similar tasks.
Specifically, training samples of similar tasks can be sampled in proportion according to the distribution of different user label intervals. And taking a wind control scene as an example, distributing according to different fraction sections of wind control fractions, and sampling according to a proportion. By extracting samples in proportion according to the user label value distribution, the sample balance degree of the subsequent similar distillation model can be improved, and therefore the accuracy of the similar distillation model is improved.
And S320, determining the user similar characteristics and the user similar labels under the similar tasks according to the extracted training samples.
Specifically, the features in the sampled training samples are used as the user similarity features, and the labels in the sampled training samples are used as the user similarity labels.
S330, determining the original user identification text according to the user identification ciphertext acquired from the label provider.
Specifically, according to the user identification ciphertext, mapping is carried out to obtain a user identification original text of the model training party. Taking the user identification ciphertext as the user mobile phone number ciphertext as an example, the model training party obtains the user mobile phone number original text through mapping.
S340, determining the user target characteristics according to the user identification original text.
Specifically, the user target characteristics of the model training party are obtained according to the user identification original text. The label provides a user identification ciphertext to the direction model training party instead of the user identification original text, so that the user identification is prevented from being leaked in the transmission process, and the data security can be further improved.
And S350, performing knowledge distillation according to the user similar characteristics and the user similar labels under the similar tasks to obtain a similar distillation model, and obtaining a similar characterization model according to the similar distillation model.
And S360, based on the similar characterization model, constructing a target task model according to the user target characteristics under the target task and the user target label acquired from the label provider, and predicting the target task for the user.
In an alternative embodiment, a model trainer and the label provider perform data interaction by adopting a homomorphic encryption algorithm. Specifically, a homomorphic encryption algorithm is adopted for user identification data interaction or partial user target feature interaction. After the similar representation model is constructed, a homomorphic encryption algorithm is introduced in the construction process of the target task model, so that the safety of interactive data can be further improved.
According to the technical scheme, the target task model is established indirectly by using the user similar label, the prediction accuracy of the target task model can be improved, and the prediction accuracy of the target task model and the data safety of each participant are further improved by determining the user similar characteristics and the user target characteristics.
Fig. 4 is a schematic flowchart of a model joint training method provided in an embodiment of the present application. The embodiment can be applied to the situation that the label provider and the model trainer jointly train the target task model. The model joint training method disclosed in this embodiment may be executed by an electronic device, and specifically, may be executed by a model joint training apparatus, where the apparatus may be implemented by software and/or hardware and configured in an electronic device of a tag provider. Referring to fig. 4, the model joint training method provided in this embodiment includes:
and S410, determining a user target label under the target task.
S420, sending a model training request carrying the user target label to a model training party, and instructing the model training party to execute the following steps: performing knowledge distillation according to the user similar characteristics and the user similar labels under similar tasks to obtain a similar distillation model, and obtaining a similar characterization model according to the similar distillation model; and constructing a target task model according to the user target characteristics and the user target labels under the target task based on the similar characterization model.
The similar tasks and the target tasks can be similar in the scene, and the participants can be different, namely the prediction targets of the similar tasks and the target tasks are similar. The user similarity label and the user target label refer to the user label under the similar task and the target task respectively. It should be noted that samples used in the process of constructing the similar distillation model and the target task model may use the same user, or belong to different users.
The similar representation model reserves the user similar label information in the similar scene, and the target task model is built based on the similar representation model, namely the target task model is built by indirectly using the user similar label, so that the defect of insufficient user target labels can be overcome, and the prediction accuracy of the target task model can be improved.
In an alternative embodiment, the similar distillation model is a deep learning model. The similar distillation model is constructed on the basis of the deep learning network structure, compared with a tree model, the method has better generalization and feature characterization capability, and therefore the similar characterization model obtained according to the similar distillation model has better expression on similar scenes.
According to the technical scheme, the target task model is established by indirectly using the user similar tags, so that the defect of insufficient user target tags can be overcome, and the prediction accuracy of the target task model can be improved. In addition, the deep learning distillation model is adopted, so that the generalization and characteristic characterization capabilities are better, and the prediction accuracy of the target task model is further improved.
Fig. 5 is a schematic flowchart of a model joint training method according to an embodiment of the present application. The embodiment is a specific implementation scheme provided on the basis of the above embodiment. Referring to fig. 5, the model joint training method provided in this embodiment includes:
s510, the label provider determines a user target label under the target task.
S520, the label providing direction model trainer sends the user target label.
And S530, carrying out knowledge distillation by the model training party according to the user similar characteristics and the user similar labels under the similar tasks to obtain a similar distillation model.
And S540, the model training party obtains a similar characterization model according to the similar distillation model.
And S550, constructing a target task model according to the user target characteristics and the user target labels under the target task by the model training party based on the similar representation model.
In an alternative embodiment, S550 may include: mapping according to a user id ciphertext acquired from a tag provider to obtain a user id original text of a model trainer, and acquiring a user target characteristic of the model trainer according to the user id original text; constructing a joint prediction model based on the similarity characterization model; and training the joint prediction model by adopting the user target characteristics and the user target labels to obtain a target task model.
In an alternative embodiment, S550 may include: mapping according to a user id ciphertext acquired from a tag provider to obtain a user id original text of a model trainer, and acquiring a user target characteristic of the model trainer according to the user id original text; taking the user target characteristics as the input of a similar characterization model to obtain similar characterization characteristics of the user; splicing the target characteristics of the user and the similar characteristic characteristics of the user to obtain user joint characteristics; and constructing a target task model according to the user joint characteristics and the user target labels.
According to the technical scheme, any two-party data safety fusion and modeling calculation are supported, and the full-flow service from feature analysis processing, model training and effect evaluation to model application deployment is provided. The representation information of the similar tasks under the similar scene is obtained through representation learning, and the representation information of the similar tasks is applied to the combined modeling scene with insufficient label information, so that the defect that the model is over-fit or insufficiently trained due to insufficient labels is overcome to a certain extent; moreover, the obtained characterization information is dense, and the feature expression capacity is stronger; the characterization model can adopt a deep learning model, and has stronger generalization and feature expression capability compared with the traditional tree-type combined modeling model; the obtained characterization features can participate in combined modeling in a plurality of ways, so that the target task model not only covers the target labels of users of the label provider, but also comprises the similar labels of the users used in the similar tasks under the similar scenes; the expansibility is strong, and the method can be expanded to different service scenes.
Fig. 6 is a schematic structural diagram of a model training apparatus according to an embodiment of the present disclosure, which may be configured in an electronic device of a model training party. Referring to fig. 6, a model joint training apparatus 600 provided in an embodiment of the present application may include:
the characterization model module 601 is used for performing knowledge distillation according to the user similarity characteristics and the user similarity labels under the similar tasks to obtain a similar distillation model, and obtaining a similar characterization model according to the similar distillation model;
and a target task model module 602, configured to construct a target task model according to the user target feature under the target task and the user target tag obtained from the tag provider, based on the similar characterization model, and configured to perform target task prediction on the user.
Optionally, the characterization model module 601 is specifically configured to:
and training the deep learning model according to the user similar characteristics and the user similar labels under the similar tasks to obtain a similar distillation model.
Optionally, the apparatus 600 further includes a similar data module, where the similar data module includes:
the sample extraction unit is used for sampling the training samples of the similar tasks according to the user label value distribution of the similar tasks;
and the similar data unit is used for determining the user similar characteristics and the user similar labels according to the extracted training samples.
Optionally, the apparatus 600 further includes a target feature module, where the target feature module includes:
the identification original text unit is used for determining a user identification original text according to a user identification ciphertext acquired from the label provider;
and the target characteristic unit is used for determining the target characteristics of the user according to the original text of the user identification.
Optionally, the target task model module 602 includes:
the combined model unit is used for connecting the similar representation model with a prediction output layer of a target task to obtain a combined prediction model of combined modeling;
and the target task model unit is used for training the combined prediction model according to the user target characteristics under the target task and the user target labels acquired from the label provider to obtain the target task model.
Optionally, the target task model module 602 includes:
the characterization feature unit is used for taking the user target features under the target task as the input of the similar characterization model to obtain the similar characterization features of the user;
the combined feature unit is used for fusing the user target features and the similar characterization features under the target task to obtain user combined features;
and the target task model unit is used for constructing a target task model according to the user joint characteristics and the user target labels acquired from the label provider.
Optionally, a homomorphic encryption algorithm is used for data interaction between the model trainer and the label provider.
According to the technical scheme, the target task model is established by indirectly using the user similar labels through the model training party, the defect that the user target labels provided by the label providing party are insufficient can be overcome, and the prediction accuracy of the target task model can be improved. By adopting the deep learning distillation model, the method has better generalization and characteristic characterization capabilities, and further improves the prediction accuracy of the target task model. In addition, the prediction accuracy can be further improved by providing two joint learning modes based on the similar characterization model.
Fig. 7 is a schematic structural diagram of a model training apparatus according to an embodiment of the present disclosure, which may be configured in an electronic device of a tag provider. Referring to fig. 7, a model joint training apparatus 700 provided in an embodiment of the present application may include:
a target tag determination module 701, configured to determine a user target tag in a target task;
a training request sending module 702, configured to send a model training request carrying the user target label to a model trainer, and instruct the model trainer to execute the following: performing knowledge distillation according to the user similar characteristics and the user similar labels under similar tasks to obtain a similar distillation model, and obtaining a similar characterization model according to the similar distillation model; and constructing a target task model according to the user target characteristics and the user target labels under the target task based on the similar characterization model.
According to the technical scheme, the target task model is established by indirectly using the user similar labels through the model training party, the defect that the user target labels provided by the label providing party are insufficient can be overcome, and the prediction accuracy of the target task model can be improved. By adopting the deep learning distillation model, the method has better generalization and characteristic characterization capabilities, and further improves the prediction accuracy of the target task model. In addition, the prediction accuracy can be further improved by providing two joint learning modes based on the similar characterization model.
According to an embodiment of the present application, an electronic device and a readable storage medium are also provided.
Fig. 8 is a block diagram of an electronic device for a method of model joint training according to an embodiment of the present application. Electronic devices are intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the present application that are described and/or claimed herein.
As shown in fig. 8, the electronic apparatus includes: one or more processors 801, memory 802, and interfaces for connecting the various components, including a high speed interface and a low speed interface. The various components are interconnected using different buses and may be mounted on a common motherboard or in other manners as desired. The processor may process instructions for execution within the electronic device, including instructions stored in or on the memory to display graphical information of a GUI on an external input/output apparatus (such as a display device coupled to the interface). In other embodiments, multiple processors and/or multiple buses may be used, along with multiple memories and multiple memories, as desired. Also, multiple electronic devices may be connected, with each device providing portions of the necessary operations (e.g., as a server array, a group of blade servers, or a multi-processor system). Fig. 8 illustrates an example of a processor 801.
The memory 802 is a non-transitory computer readable storage medium as provided herein. Wherein the memory stores instructions executable by at least one processor to cause the at least one processor to perform the method of model joint training provided herein. The non-transitory computer readable storage medium of the present application stores computer instructions for causing a computer to perform the method of model joint training provided herein.
The memory 802, as a non-transitory computer readable storage medium, may be used to store non-transitory software programs, non-transitory computer executable programs, and modules, such as program instructions/modules corresponding to the method for model joint training in the embodiments of the present application (e.g., the characterization model module 601 and the target task model module 602 shown in fig. 6, and the target tag determination module 701 and the training request sending module 702 shown in fig. 7). The processor 801 executes various functional applications of the server and model co-training, i.e., a method for implementing model co-training in the above-described method embodiments, by running non-transitory software programs, instructions, and modules stored in the memory 802.
The memory 802 may include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application program required for at least one function; the storage data area may store data created from use of the model co-trained electronic device, and the like. Further, the memory 802 may include high-speed random access memory, and may also include non-transitory memory, such as at least one magnetic disk storage device, flash memory device, or other non-transitory solid state storage device. In some embodiments, memory 802 optionally includes memory located remotely from processor 801, and such remote memory may be connected to the electronics of the model co-training via a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.
The electronic device of the method for model joint training may further include: an input device 803 and an output device 804. The processor 801, the memory 802, the input device 803, and the output device 804 may be connected by a bus or other means, and are exemplified by a bus in fig. 8.
The input device 803 may receive input numeric or character information and generate key signal inputs related to user settings and function controls of the electronic device with which the model is co-trained, such as a touch screen, keypad, mouse, track pad, touch pad, pointer stick, one or more mouse buttons, track ball, joystick, or other input device. The output devices 804 may include a display device, auxiliary lighting devices (e.g., LEDs), and haptic feedback devices (e.g., vibrating motors), among others. The display device may include, but is not limited to, a Liquid Crystal Display (LCD), a Light Emitting Diode (LED) display, and a plasma display. In some implementations, the display device can be a touch screen.
Various implementations of the systems and techniques described here can be realized in digital electronic circuitry, integrated circuitry, application specific ASICs (application specific integrated circuits), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include: implemented in one or more computer programs that are executable and/or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, receiving data and instructions from, and transmitting data and instructions to, a storage system, at least one input device, and at least one output device.
These computer programs (also known as programs, software applications, or code) include machine instructions for a programmable processor, and may be implemented using high-level procedural and/or object-oriented programming languages, and/or assembly/machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and/or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and/or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal used to provide machine instructions and/or data to a programmable processor.
To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to a user; and a keyboard and a pointing device (e.g., a mouse or a trackball) by which a user can provide input to the computer. Other kinds of devices may also be used to provide for interaction with a user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form, including acoustic, speech, or tactile input.
The systems and techniques described here can be implemented in a computing system that includes a back-end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front-end component (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: local Area Networks (LANs), Wide Area Networks (WANs), blockchain networks, and the internet.
The computer system may include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.
According to the technical scheme, the target task model is established by indirectly using the user similar labels through the model training party, the defect that the user target labels provided by the label providing party are insufficient can be overcome, and the prediction accuracy of the target task model can be improved. By adopting the deep learning distillation model, the method has better generalization and characteristic characterization capabilities, and further improves the prediction accuracy of the target task model. In addition, the prediction accuracy can be further improved by providing two joint learning modes based on the similar characterization model.
The above-described embodiments should not be construed as limiting the scope of the present application. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may be made in accordance with design requirements and other factors. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims (18)
1. A model co-training method, comprising:
performing knowledge distillation according to the user similar characteristics and the user similar labels under similar tasks to obtain a similar distillation model, and obtaining a similar characterization model according to the similar distillation model;
and constructing a target task model according to the user target characteristics under the target task and the user target labels acquired from the label provider based on the similar characterization model, and predicting the target task for the user.
2. The method of claim 1, wherein the knowledge distillation from user-similar features and user-similar labels under similar tasks to obtain a similar distillation model comprises:
and training the deep learning model according to the user similar characteristics and the user similar labels under the similar tasks to obtain a similar distillation model.
3. The method of claim 1, further comprising:
sampling training samples of similar tasks according to the user label value distribution of the similar tasks;
and determining the user similar features and the user similar labels according to the extracted training samples.
4. The method of claim 1, further comprising:
determining a user identification original text according to a user identification ciphertext acquired from the tag provider;
and determining the user target characteristics according to the user identification original text.
5. The method of claim 1, wherein constructing a target task model from user target features under a target task and user target tags obtained from tag providers based on the similar characterization model comprises:
connecting the similar representation model with a prediction output layer of a target task to obtain a joint prediction model of joint modeling;
and training the combined prediction model according to the user target characteristics under the target task and the user target labels acquired from the label provider to obtain the target task model.
6. The method of claim 1, wherein constructing a target task model from user target features under a target task and user target tags obtained from tag providers based on the similar characterization model comprises:
taking the user target characteristics under the target task as the input of the similar characterization model to obtain the similar characterization characteristics of the user;
fusing the user target characteristics and the similar characteristic characteristics under the target task to obtain user joint characteristics;
and constructing a target task model according to the user joint characteristics and the user target label acquired from the label provider.
7. The method of claim 1, wherein a homomorphic encryption algorithm is adopted for data interaction between a model trainer and the label provider.
8. A model co-training method, comprising:
determining a user target label under a target task;
sending a model training request carrying the user target label to a model trainer for instructing the model trainer to execute the following steps: performing knowledge distillation according to the user similar characteristics and the user similar labels under similar tasks to obtain a similar distillation model, and obtaining a similar characterization model according to the similar distillation model; and constructing a target task model according to the user target characteristics and the user target labels under the target task based on the similar characterization model.
9. A model co-training apparatus comprising:
the characterization model module is used for carrying out knowledge distillation according to the user similar characteristics and the user similar labels under similar tasks to obtain a similar distillation model, and obtaining a similar characterization model according to the similar distillation model;
and the target task model module is used for constructing a target task model according to the user target characteristics under the target task and the user target labels acquired from the label provider based on the similar characterization model and is used for predicting the target task of the user.
10. The apparatus of claim 9, wherein the characterization model module is specifically configured to:
and training the deep learning model according to the user similar characteristics and the user similar labels under the similar tasks to obtain a similar distillation model.
11. The apparatus of claim 9, further comprising a similar data module comprising:
the sample extraction unit is used for sampling the training samples of the similar tasks according to the user label value distribution of the similar tasks;
and the similar data unit is used for determining the user similar characteristics and the user similar labels according to the extracted training samples.
12. The apparatus of claim 9, further comprising a target feature module, the target feature module comprising:
the identification original text unit is used for determining a user identification original text according to a user identification ciphertext acquired from the label provider;
and the target characteristic unit is used for determining the target characteristics of the user according to the original text of the user identification.
13. The apparatus of claim 9, wherein the target task model module comprises:
the combined model unit is used for connecting the similar representation model with a prediction output layer of a target task to obtain a combined prediction model of combined modeling;
and the target task model unit is used for training the combined prediction model according to the user target characteristics under the target task and the user target labels acquired from the label provider to obtain the target task model.
14. The apparatus of claim 9, wherein the target task model module comprises:
the characterization feature unit is used for taking the user target features under the target task as the input of the similar characterization model to obtain the similar characterization features of the user;
the combined feature unit is used for fusing the user target features and the similar characterization features under the target task to obtain user combined features;
and the target task model unit is used for constructing a target task model according to the user joint characteristics and the user target labels acquired from the label provider.
15. The apparatus of claim 9, wherein a homomorphic encryption algorithm is used for data interaction between the model trainer and the label provider.
16. A model co-training apparatus comprising:
the target label determining module is used for determining a user target label under a target task;
a training request sending module, configured to send a model training request carrying the user target label to a model trainer, and instruct the model trainer to execute the following: performing knowledge distillation according to the user similar characteristics and the user similar labels under similar tasks to obtain a similar distillation model, and obtaining a similar characterization model according to the similar distillation model; and constructing a target task model according to the user target characteristics and the user target labels under the target task based on the similar characterization model.
17. An electronic device, comprising:
at least one processor; and
a memory communicatively coupled to the at least one processor; wherein the content of the first and second substances,
the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-8.
18. A non-transitory computer readable storage medium having stored thereon computer instructions for causing the computer to perform the method of any one of claims 1-8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010866349.9A CN112070233B (en) | 2020-08-25 | 2020-08-25 | Model joint training method, device, electronic equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010866349.9A CN112070233B (en) | 2020-08-25 | 2020-08-25 | Model joint training method, device, electronic equipment and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112070233A true CN112070233A (en) | 2020-12-11 |
| CN112070233B CN112070233B (en) | 2024-03-22 |
Family
ID=73659797
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010866349.9A Active CN112070233B (en) | 2020-08-25 | 2020-08-25 | Model joint training method, device, electronic equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112070233B (en) |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108334934A (en) * | 2017-06-07 | 2018-07-27 | 北京深鉴智能科技有限公司 | Convolutional neural networks compression method based on beta pruning and distillation |
| CN108898168A (en) * | 2018-06-19 | 2018-11-27 | 清华大学 | The compression method and system of convolutional neural networks model for target detection |
| CN109637546A (en) * | 2018-12-29 | 2019-04-16 | 苏州思必驰信息科技有限公司 | Knowledge distillating method and device |
| US20190130312A1 (en) * | 2017-10-27 | 2019-05-02 | Salesforce.Com, Inc. | Hierarchical and interpretable skill acquisition in multi-task reinforcement learning |
| CN110223281A (en) * | 2019-06-06 | 2019-09-10 | 东北大学 | A kind of Lung neoplasm image classification method when in data set containing uncertain data |
| CN110246487A (en) * | 2019-06-13 | 2019-09-17 | 苏州思必驰信息科技有限公司 | Optimization method and system for single pass speech recognition modeling |
| CN110428052A (en) * | 2019-08-01 | 2019-11-08 | 江苏满运软件科技有限公司 | Construction method, device, medium and the electronic equipment of deep neural network model |
| CN111198940A (en) * | 2019-12-27 | 2020-05-26 | 北京百度网讯科技有限公司 | FAQ method, question-answer search system, electronic device, and storage medium |
| CN111460150A (en) * | 2020-03-27 | 2020-07-28 | 北京松果电子有限公司 | Training method, classification method and device of classification model and storage medium |
-
2020
- 2020-08-25 CN CN202010866349.9A patent/CN112070233B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108334934A (en) * | 2017-06-07 | 2018-07-27 | 北京深鉴智能科技有限公司 | Convolutional neural networks compression method based on beta pruning and distillation |
| US20190130312A1 (en) * | 2017-10-27 | 2019-05-02 | Salesforce.Com, Inc. | Hierarchical and interpretable skill acquisition in multi-task reinforcement learning |
| CN108898168A (en) * | 2018-06-19 | 2018-11-27 | 清华大学 | The compression method and system of convolutional neural networks model for target detection |
| CN109637546A (en) * | 2018-12-29 | 2019-04-16 | 苏州思必驰信息科技有限公司 | Knowledge distillating method and device |
| CN110223281A (en) * | 2019-06-06 | 2019-09-10 | 东北大学 | A kind of Lung neoplasm image classification method when in data set containing uncertain data |
| CN110246487A (en) * | 2019-06-13 | 2019-09-17 | 苏州思必驰信息科技有限公司 | Optimization method and system for single pass speech recognition modeling |
| CN110428052A (en) * | 2019-08-01 | 2019-11-08 | 江苏满运软件科技有限公司 | Construction method, device, medium and the electronic equipment of deep neural network model |
| CN111198940A (en) * | 2019-12-27 | 2020-05-26 | 北京百度网讯科技有限公司 | FAQ method, question-answer search system, electronic device, and storage medium |
| CN111460150A (en) * | 2020-03-27 | 2020-07-28 | 北京松果电子有限公司 | Training method, classification method and device of classification model and storage medium |
Non-Patent Citations (2)
| Title |
|---|
| FIDA MOHAMMAD THOKER 等: "Cross-Modal Knowledge Distillation for Action Recognition", 2019 IEEE INTERNATIONAL CONFERENCE ON IMAGE PROCESSING (ICIP), 26 August 2019 (2019-08-26) * |
| 张随雨;杨成;: "一种多标签统一域嵌入的推荐模型", 哈尔滨工业大学学报, no. 05, 10 May 2020 (2020-05-10) * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112070233B (en) | 2024-03-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111325020B (en) | Event argument extraction method and device and electronic equipment | |
| CN111507104B (en) | Method and device for establishing label labeling model, electronic equipment and readable storage medium | |
| CN112036509A (en) | Method and apparatus for training image recognition models | |
| JP2021120874A (en) | User interaction information processing model generation method and device, user interaction information processing method and device, equipment, medium, and program | |
| CN112507090B (en) | Method, apparatus, device and storage medium for outputting information | |
| CN112102448B (en) | Virtual object image display method, device, electronic equipment and storage medium | |
| CN111079945B (en) | End-to-end model training method and device | |
| CN112235613B (en) | Video processing method and device, electronic equipment and storage medium | |
| CN111832613B (en) | Model training method and device, electronic equipment and storage medium | |
| CN110543558B (en) | Question matching method, device, equipment and medium | |
| CN111324715B (en) | Method and device for generating question-answering robot | |
| CN111241838B (en) | Semantic relation processing method, device and equipment for text entity | |
| US20220027575A1 (en) | Method of predicting emotional style of dialogue, electronic device, and storage medium | |
| CN111680517A (en) | Method, apparatus, device and storage medium for training a model | |
| CN112380104A (en) | User attribute identification method and device, electronic equipment and storage medium | |
| CN111582477B (en) | Training method and device for neural network model | |
| CN112507702A (en) | Text information extraction method and device, electronic equipment and storage medium | |
| CN112560499A (en) | Pre-training method and device of semantic representation model, electronic equipment and storage medium | |
| CN111666771A (en) | Semantic label extraction device, electronic equipment and readable storage medium of document | |
| CN112529180A (en) | Method and apparatus for model distillation | |
| CN112052390A (en) | Resource screening method and device, electronic equipment and storage medium | |
| CN112541145A (en) | Page display method, device, equipment and storage medium | |
| CN112070233B (en) | Model joint training method, device, electronic equipment and storage medium | |
| CN115688802A (en) | Text risk detection method and device | |
| CN111510376B (en) | Image processing method and device and electronic equipment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |