CN114611631A - Method, system, device and medium for fast training a model from a partial training set - Google Patents
Method, system, device and medium for fast training a model from a partial training set Download PDFInfo
- Publication number
- CN114611631A CN114611631A CN202210389521.5A CN202210389521A CN114611631A CN 114611631 A CN114611631 A CN 114611631A CN 202210389521 A CN202210389521 A CN 202210389521A CN 114611631 A CN114611631 A CN 114611631A
- Authority
- CN
- China
- Prior art keywords
- model
- forgetting
- data
- forgotten
- training
- 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.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/60—Protecting data
- G06F21/62—Protecting access to data via a platform, e.g. using keys or access control rules
- G06F21/6218—Protecting access to data via a platform, e.g. using keys or access control rules to a system of files or objects, e.g. local or distributed file system or database
- G06F21/6245—Protecting personal data, e.g. for financial or medical purposes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/04—Inference or reasoning models
Abstract
The invention discloses a method, a system, equipment and a medium for quickly training a model from a part of training sets, wherein the method comprises the following steps: selecting a data subset to be forgotten, deleting forgotten data from the data subset, taking the remaining data subset after deleting the forgotten data as a training sample, and training to obtain a forgotten model; evaluating the forgetting model by utilizing member reasoning attack, calculating output distribution of the forgetting model, judging whether deleted forgetting data is successfully forgotten or not according to the output distribution, and taking the forgetting model as a final forgetting model if the forgetting is successful; evaluating the forgetting effect of the final forgetting model by utilizing backdoor attack, and if the evaluation is qualified, taking the final forgetting model as a target model; the method only needs to make small disturbance to the weight of the target model and does not need to retrain, thereby greatly saving time, and the method does not change in the training process of the data sample or the target model, is universal to all training models and does not need to add extra constraint conditions.
Description
Technical Field
The invention relates to the technical field of deep learning, in particular to a method, a system, equipment and a medium for quickly training a model from a part of training set.
Background
In recent years, deep learning has become a hot topic, and automated driving, face recognition, medical evaluation, and the like bring various aspects of convenience to people. In order to provide better service, a number of deep learning models consuming more power and higher performance are developed. The service providers need to collect a large amount of data information to provide a basis for the deep learning models. People are also constantly providing personal information for service providers to obtain more convenience. But the system can calculate and obtain more data after collecting the original information while enjoying the convenience of personal information exchange, the original information is deeply aggregated into statistical information, and more data are derived for system construction. However, the purchase record may reveal the user's deposit and the medical history indicia may reveal the user's family genetic medical history. The private information of the user can flow out in a potential mode, appears in places unknown to the user in various modes, intuitively, if some data are forgotten, the simplest method is to delete the data from an input sample and retrain a machine learning model by using the rest input sample, but the training time is extremely consumed. The predecessors also proposed some statistical query-based work, and a method of using statistical query to obtain data set features rather than training directly on the data set is proposed, so that forgetting is performed in less time than retraining, but the statistical query method is only applicable to a non-adaptive machine learning model (late training does not depend on early training), and forgetting effect is difficult to achieve on a neural network.
Another Bourtoule method divides the data set into sections, each section is trained as a separate sub-model and stored, trains the total model by incremental learning, and in order to forget a sample, retrains starting with the first intermediate model containing the sample contribution. However, this method reduces the training time, but consumes a large amount of memory space; another data forgetting work based on federal learning saves updating parameters of each round in a normal training model aggregation stage, then reduces iteration times of client training when forgetting data is deleted and retraining is carried out, and combines parameters of a current client and the updating parameters saved in the previous training to construct a server model in model aggregation. Because the updated parameters need to be stored, the training process of the target model is modified, and the stored parameters carry the information to be forgotten, the forgetting cannot be guaranteed to be thorough theoretically.
Disclosure of Invention
The invention aims to provide a novel method, a novel system, novel equipment and a novel medium for quickly training a model from a partial training set, and aims to solve the technical problems caused by the fact that data forgetting is carried out by adopting the prior art in the background technology by adopting a novel technical concept.
The invention provides a method for quickly training a model from a part of training set, which comprises the following steps:
selecting a data subset to be forgotten, deleting forgotten data from the data subset, taking the remaining data subset after deleting the forgotten data as a training sample, and training to obtain a forgotten model;
evaluating the forgetting model by utilizing member reasoning attack, calculating output distribution of the forgetting model, judging whether deleted forgetting data is successfully forgotten or not according to the output distribution, and taking the forgetting model as a final forgetting model if the forgetting is successful;
and evaluating the forgetting effect of the final forgetting model by utilizing backdoor attack, and if the evaluation is qualified, taking the final forgetting model as a target model.
The invention also provides a system for quickly training a model from a partial training set, which implements the method, and comprises the following modules which are connected and communicated with each other:
the forgetting model acquisition module is used for selecting a data subset to be forgotten, deleting forgotten data from the data subset, and training the data subset left after deleting the forgotten data to obtain a forgetting model;
the final forgetting model obtaining module is used for evaluating the forgetting model by utilizing member reasoning attack, calculating the output distribution of the forgetting model, judging whether the deleted forgetting data is successfully forgotten or not according to the output distribution, and if the forgetting data is successfully forgotten, taking the forgetting model as the final forgetting model;
and the target model acquisition module is used for evaluating the forgetting effect of the final forgetting model by utilizing backdoor attack, and if the evaluation is qualified, the final forgetting model is used as the target model.
The present invention also provides a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method for fast training a model from a partial training set according to the first aspect when executing the computer program.
The present invention also provides a computer readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the method for fast training a model from a partial training set as described in the first aspect above.
The method, the system, the equipment and the medium for quickly training the model from the partial training set have the advantages that:
1. because the method only slightly perturbs the weight of the target model and does not need retraining, the time is greatly saved;
2. the method of the invention does not change the training process of the data sample or the target model, so that the method is universal to all training models and does not need additional constraint conditions;
3. the method of the invention does not need any information except the original data set and the model, thereby being very efficient and convenient;
4. the method provided by the invention uses various indexes to evaluate the effect, and objectively proves the effectiveness of the method from multiple angles.
Drawings
For better clarity of the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and it is obvious for those skilled in the art that other drawings can be obtained according to these drawings without creative efforts.
FIG. 1 is a schematic flow chart of a method for fast training a model from a partial training set according to an embodiment of the present invention;
FIG. 2 is a schematic diagram illustrating a model training principle provided by an embodiment of the present invention;
FIG. 3 is a schematic diagram of a principle of a model inference attack method according to an embodiment of the present invention;
fig. 4 is a schematic diagram illustrating a back door attack principle provided by an embodiment of the present invention;
fig. 5 is a schematic diagram illustrating the effect of backdoor attack according to an embodiment of the present invention;
FIG. 6 is a graph of accuracy after forgetting different data sets provided by an embodiment of the present invention;
fig. 7a and 7b are comparative diagrams of the forgetting effect provided by the embodiment of the invention.
Detailed Description
The following further describes embodiments of the present invention with reference to the drawings. It should be noted that the description of the embodiments is provided to help understanding of the present invention, and is not intended to limit the present invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
Referring to fig. 1, the method for fast training a model from a partial training set provided in this embodiment includes the following steps:
s101, selecting a data subset to be forgotten, deleting forgotten data from the data subset, taking the remaining data subset after deleting the forgotten data as a training sample, and training to obtain a forgetting model.
Changing the representation of the forgotten data in the model to correct the weights of the model is a core idea. First, to forget, a direction of weight correction needs to be determined. In the case of neglecting the cost, retraining with the remaining data set from which the forgotten data is deleted is a simple and effective forgetting method. Based on this idea, the retrained model is not used as a reference standard for forgetting to guide the forgetting direction.
Referring to the schematic diagram of the model training principle shown in fig. 2, the specific steps are as follows:
first, a reference model is trained. And randomly selecting a part of subsets from the rest data as training samples of the reference model in consideration of time cost, wherein the initial model is the same as the target model. And calculating the output distribution of the forgotten data on the reference model as a reference. And (4) keeping the posterior distribution of the reference model unchanged, and adjusting the forgetting model in iteration to enable the distribution of the forgetting model to be close to the distribution iteration of the reference model. At the end of the iteration, the two distributions are nearly the same, iterating to the last forgotten forgetting model, as shown in fig. 2. The iterative loss function here uses the KL divergence to calculate the distance between the two distributions.
S102, evaluating the forgetting model by utilizing member reasoning attack, calculating output distribution of the reference model, judging whether the deleted forgetting data is successfully forgotten or not according to the output distribution, and if the forgetting is successful, taking the forgetting model as a final forgetting model.
The precision is an important index of target model training, and in order to apply the model to the practice without influencing the practical effect of the target model, the precision before and after the model is forgotten is essential to be evaluated. Therefore, after the forgetting model is obtained through training, the forgetting model also needs to be evaluated, and in the invention, the forgetting model is evaluated by using member reasoning attack.
Referring to FIG. 3, a schematic diagram of a member inference attack method is shown;
the member reasoning attack can judge whether a certain sample exists in a training set of a certain model according to distribution, the member reasoning is realized based on a shadow model, firstly, data similar to the distribution of a training set of a forgetting model is prepared to be integrated into a shadow data set, and the shadow data set is used as the training set to train the shadow model. At this time, the shadow model, the shadow training set, and the shadow untraining set are continuously prepared. Based on this, a two-class attack model is trained, and after an input is given, the attack model can judge whether the attack model is a training set member of a forgetting model.
In the experiment of the invention, data to be forgotten is input into a forgetting model, output distribution is calculated, and an attack model judges whether the data is member data of the forgetting model or not according to the distribution. If the attack model infers that the forgotten data is non-member data, the forgetting is successful.
In the method, member reasoning attack is adopted to evaluate the residual forgetting information in the model after forgetting, the member reasoning attack model is trained based on the original model before forgetting, and the training set is the training set of the original model minus the forgetting data. The method can accurately distinguish the learned or unlearned data, and judge whether the forgotten data is member data or not for the target model according to the posterior distribution difference of the input sample and the posterior distribution difference of the remaining data.
The accuracy of different forgotten data sets is shown in fig. 4, for the mnist data set, the precision of the first column before forgetting is 98.97%, the second column obtains the precision of 98.96% after the 1/100 samples are deleted, the precision is not obviously reduced because the sample amount is not greatly reduced, and the precision of the third column after forgetting by the method of the present invention is 98.12%, compared with the retrained model, the precision is not obviously reduced, which indicates that the method of the present invention does not affect the performance of the target model. For the performance of the cifar10 dataset on resnet, the accuracy of the first column before forgetting is 90.87%, and the accuracy of the second column after retraining is 4% lower than that before forgetting because of the reduced number of samples, whereas in the method of the present invention, the accuracy of the model after forgetting is 89.64%, which is slightly higher than that of the model after retraining. For the performance of the cifar10 data set on the VGG16, the pre-forgetting and post-forgetting trends are the same as those of the resnet18, but the overall accuracy is lower than that of the resnet by 3%, because the model of the VGG16 trains the cifar10 data set to have the architectural limitation compared with the resnet18 model, and the data set is not influenced by the forgetting method.
And S103, evaluating the forgetting effect of the final forgetting model by utilizing backdoor attack, and if the evaluation is qualified, taking the final forgetting model as a target model.
See FIG. 5 for a schematic diagram of a model backdoor attack method;
the backdoor attack is a common attack method and is beneficial to evaluating the forgetting effect. And implanting the back door into partial data, wherein the data implanted into the back door is uniformly labeled with a specific label. The model generated through backdoor data training "remembers" backdoor information and is triggered when the same backdoor is met, and the same prediction result, namely a fixed label, is generated.
In the invention, the backdoor is firstly implanted into the forgotten data, the label is set as a specific label, and the specific label and normal data (namely residual data) are trained together to generate an original model. As shown in the left part of fig. 6, when normal test data is encountered, it can be predicted normally, and when test data having the same back door is encountered, a specific tag is set. The method is applied to a forgetting evaluation stage, the fact that the original model is trained by backdoor forgetting data and normal residual data is assumed, and the original model is forgotten by using the method of the invention to obtain the forgotten original model. Ideally, as shown in the right part of fig. 6, the forgotten model will not be affected by any more forgotten data, i.e. data with a back door has been forgotten.
At this time, the data carrying the back door is input into the calculation, and the obtained probability distribution is similar to that of normal data, and the same prediction result as that before the back door of the device is generated instead of a specific label marked on the back door. Inputting the test data carrying the back door, wherein the obtained precision is less than 20%, and the inference result of the model on 80% of back door data (namely data needing to be forgotten) is not matched with the set fixed label. The precision of the data to be forgotten is reduced from 98% to 20%, which shows that the target model really forgets the backdoor information which is remembered in the training stage by using the method of the invention, thereby proving that the forgetting is successful. As shown in the figure, in the forgetting process of the mnist data set, the forgotten data is not needed, the test precision thereof is kept at a high level, and the precision of the forgotten residual data is also reflected to be maintained at a high level. On the forgotten data, the test accuracy is reduced from the first 98% to a lower level, which represents the forgetting effect. The forgetting performance and the precision after forgetting are weighed, and 0.5 weight is respectively given to the weighed precision, and the higher the obtained result is, the best overall performance is represented, as shown by a line with a circle in the figure. cifar10.r dataset, as shown in fig. 7a and 7 b.
In addition to performance, efficiency is a key factor in data forgetting. The invention aims to consume less time than retraining on the basis of achieving the effect almost the same as retraining. The method of the invention can be improved by tens of times in time consumption with reference to the retrained model. As shown in table 1, for the mnist data set, 1/100 data to be forgotten account for the total data set, and the total number of the data to be forgotten is 600, and the forgetting is performed by the method of the present invention, which requires two iterations to reach the end point, and the required time is 3.81 s. When the reference model is trained, 42.81s is needed, and the total time length is 46.62 s. The total time spent in the method of retraining after deleting the forgetting information is 750.69 s. In contrast, the process of the present invention accelerates by a factor of 16.10. The other same principles are adopted. Therefore, the final forgetting model can be taken as the target model.
Table 1: different method forgetting efficiency comparison table
Because the method only slightly perturbs the weight of the target model and does not need to retrain, the time is greatly saved; the method of the invention does not change the training process of the data sample or the target model, so that the method is universal to all training models and does not need additional constraint conditions; the method of the invention does not need any information except the original data set and the model, thereby being very efficient and convenient; the method provided by the invention uses various indexes to evaluate the effect, and objectively proves the effectiveness of the method from multiple angles.
In one embodiment, the present invention also provides a system for fast training a model from a partial training set, the system comprising:
the forgetting model acquisition module is used for selecting a data subset to be forgotten, deleting forgotten data from the data subset, and training the data subset left after deleting the forgotten data to obtain a forgetting model;
the final forgetting model obtaining module is used for evaluating the forgetting model by utilizing member reasoning attack, calculating the output distribution of the forgetting model, judging whether the deleted forgetting data is successfully forgotten or not according to the output distribution, and if the forgetting data is successfully forgotten, taking the forgetting model as the final forgetting model;
and the target model acquisition module is used for evaluating the forgetting effect of the final forgetting model by utilizing backdoor attack, and if the evaluation is qualified, the final forgetting model is used as the target model.
In an embodiment, the present invention further provides a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, the processor implements the method for fast training a model from a partial training set according to any of the above embodiments.
In an embodiment, the present invention further provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the processor is caused to execute the method for fast training a model from a partial training set according to any one of the above embodiments.
The method provided by the invention only needs to make slight disturbance on the weight of the target model and does not need to retrain, thereby greatly saving time, and the method does not change in the training process of the data sample or the target model, is universal to all training models and does not need to add extra constraint conditions.
The embodiments of the present invention have been described in detail with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. It will be apparent to those skilled in the art that various changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, and the scope of protection is still within the scope of the invention.
Claims (10)
1. A method for fast training a model from a partial training set, comprising the steps of:
selecting a data subset to be forgotten, deleting forgotten data from the data subset, taking the remaining data subset after deleting the forgotten data as a training sample, and training to obtain a forgotten model;
evaluating the forgetting model by utilizing member reasoning attack, calculating output distribution of the forgetting model, judging whether deleted forgetting data is successfully forgotten or not according to the output distribution, and taking the forgetting model as a final forgetting model if the forgetting is successful;
and evaluating the forgetting effect of the final forgetting model by utilizing backdoor attack, and if the evaluation is qualified, taking the final forgetting model as a target model.
2. A method for fast training a model from a partial training set as claimed in claim 1, wherein said using the subset of data remaining after deleting the forgotten data as a training sample comprises:
and training a reference model, and randomly selecting a part of subsets from the data left after deleting the forgotten data as training samples of the reference model.
3. A method for rapid training of a model from a partial training set as claimed in claim 2, wherein said training results in a forgetting model comprising:
and calculating the output distribution of the forgotten data on the reference model as a reference, keeping the posterior distribution of the reference model unchanged, adjusting the forgotten model in iteration to enable the distribution of the forgotten model to be close to the distribution of the reference model in iteration, and taking the model iterated to the last as the forgotten model.
4. The method for fast training a model from a partial training set according to claim 1, wherein the evaluating the forgetting model by using membership inference attack, calculating an output distribution of the forgetting model, judging whether the deleted forgetting data is successfully forgotten according to the output distribution, and if the forgetting is successful, taking the forgetting model as a final forgetting model comprises:
firstly, preparing a shadow data set integrated with data distributed similarly to a forgetting model training set, training a shadow model by taking the shadow data set as the training set, then continuously preparing the shadow model, the shadow training set and a shadow non-training set, training a binary attack model on the basis of the shadow data set, inputting the data to be forgotten into the forgetting model, calculating output distribution, judging whether the data to be forgotten is member data of the forgetting model or not by the attack model according to the output distribution, and if the attacking model deduces that the data to be forgotten is non-member data, indicating that the forgetting is successful, and taking the forgetting model as a final forgetting model.
5. The method for fast training a model from a partial training set as claimed in claim 1, wherein before evaluating the forgetting effect of the final forgetting model using a back door attack, comprising:
and implanting a back door into the forgotten data, setting the label as a specific label, and training the specific label and the data subset remained after the forgotten data is deleted together to generate an original model, wherein the original model is a model before forgetting.
6. The method for fast training a model from a partial training set of claim 5, after said generating an original model, comprising:
and applying the generated original model to a forgetting evaluation stage, and forgetting the original model to obtain the forgotten original model by assuming that the original model is trained by the backdoor forgetting data and the residual data subset.
7. The method for fast training a model from a partial training set of claim 6,
inputting data carrying a back door into the forgotten original model for calculation, calculating the probability distribution of the obtained probability distribution and the probability distribution of normal data, calculating the test precision according to the probability distribution, if the precision is reduced to exceed a threshold value, indicating that the forgetting is successful, and taking the forgotten original model as a target model.
8. A system for rapid training of models from a partial training set for carrying out the method according to any one of claims 1 to 7, characterized in that it comprises the following modules connected and communicating with each other:
the forgetting model acquisition module is used for selecting a data subset to be forgotten, deleting forgotten data from the data subset, and training the data subset left after deleting the forgotten data to obtain a forgetting model;
the final forgetting model obtaining module is used for evaluating the forgetting model by utilizing member reasoning attack, calculating the output distribution of the forgetting model, judging whether the deleted forgetting data is successfully forgotten or not according to the output distribution, and if the forgetting data is successfully forgotten, taking the forgetting model as the final forgetting model;
and the target model acquisition module is used for evaluating the forgetting effect of the final forgetting model by utilizing backdoor attack, and if the evaluation is qualified, the final forgetting model is used as the target model.
9. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements a method of fast training a model from a partial training set as claimed in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program which, when executed by a processor, causes the processor to carry out the method of fast training a model from a partial training set as claimed in any one of claims 1 to 7.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210389521.5A CN114611631A (en) | 2022-04-14 | 2022-04-14 | Method, system, device and medium for fast training a model from a partial training set |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210389521.5A CN114611631A (en) | 2022-04-14 | 2022-04-14 | Method, system, device and medium for fast training a model from a partial training set |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114611631A true CN114611631A (en) | 2022-06-10 |
Family
ID=81869153
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210389521.5A Pending CN114611631A (en) | 2022-04-14 | 2022-04-14 | Method, system, device and medium for fast training a model from a partial training set |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114611631A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116522007A (en) * | 2023-07-05 | 2023-08-01 | 中国科学技术大学 | Recommendation system model-oriented data forgetting learning method, device and medium |
CN117349899A (en) * | 2023-12-06 | 2024-01-05 | 湖北省楚天云有限公司 | Sensitive data processing method, system and storage medium based on forgetting model |
-
2022
- 2022-04-14 CN CN202210389521.5A patent/CN114611631A/en active Pending
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116522007A (en) * | 2023-07-05 | 2023-08-01 | 中国科学技术大学 | Recommendation system model-oriented data forgetting learning method, device and medium |
CN116522007B (en) * | 2023-07-05 | 2023-10-20 | 中国科学技术大学 | Recommendation system model-oriented data forgetting learning method, device and medium |
CN117349899A (en) * | 2023-12-06 | 2024-01-05 | 湖北省楚天云有限公司 | Sensitive data processing method, system and storage medium based on forgetting model |
CN117349899B (en) * | 2023-12-06 | 2024-04-05 | 湖北省楚天云有限公司 | Sensitive data processing method, system and storage medium based on forgetting model |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Guu et al. | Traversing knowledge graphs in vector space | |
Huo et al. | Knowledge modeling via contextualized representations for LSTM-based personalized exercise recommendation | |
US8370280B1 (en) | Combining predictive models in predictive analytical modeling | |
Uncu et al. | A novel feature selection approach: combining feature wrappers and filters | |
Turkman et al. | Computational Bayesian statistics: an introduction | |
Sun et al. | Learning multiple-question decision trees for cold-start recommendation | |
CN114611631A (en) | Method, system, device and medium for fast training a model from a partial training set | |
Bilgic et al. | Effective label acquisition for collective classification | |
CN111914094B (en) | Knowledge graph representation learning method based on ternary interaction | |
Braylan et al. | Modeling and aggregation of complex annotations via annotation distances | |
Rad et al. | GP-RVM: Genetic programing-based symbolic regression using relevance vector machine | |
Bai et al. | RLAT: Multi-hop temporal knowledge graph reasoning based on Reinforcement Learning and Attention Mechanism | |
CN115018111A (en) | Wind power prediction method and system integrating deep learning and self-adaptive modeling mechanisms | |
Bai et al. | Multi-hop temporal knowledge graph reasoning with temporal path rules guidance | |
CN116910223A (en) | Intelligent question-answering data processing system based on pre-training model | |
Yovits et al. | The average decision maker and its properties utilizing the generalized information system model | |
Li et al. | On Leveraging Large Language Models for Enhancing Entity Resolution | |
Kang et al. | Self-paced annotations of crowd workers | |
Okanlawon et al. | Feature selection for learning to predict outcomes of compute cluster jobs with application to decision support | |
Byrne et al. | Man bites dog: Looking for interesting inconsistencies in structured news reports | |
CN116070804B (en) | Power system load prediction method and device based on knowledge graph and data driving | |
Davy | A review of active learning and co-training in text classification | |
US11755570B2 (en) | Memory-based neural network for question answering | |
Ramanan et al. | Active feature acquisition via human interaction in relational domains | |
US11928128B2 (en) | Construction of a meta-database from autonomously scanned disparate and heterogeneous sources |
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 |