CN114969078A - Method for updating expert research interest of federated learning through real-time online prediction - Google Patents
Method for updating expert research interest of federated learning through real-time online prediction Download PDFInfo
- Publication number
- CN114969078A CN114969078A CN202210653137.1A CN202210653137A CN114969078A CN 114969078 A CN114969078 A CN 114969078A CN 202210653137 A CN202210653137 A CN 202210653137A CN 114969078 A CN114969078 A CN 114969078A
- Authority
- CN
- China
- Prior art keywords
- information
- expert
- model
- text
- global
- 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
- 238000000034 method Methods 0.000 title claims abstract description 36
- 238000011160 research Methods 0.000 title claims abstract description 34
- 239000013598 vector Substances 0.000 claims abstract description 32
- 238000012549 training Methods 0.000 claims abstract description 25
- 230000007246 mechanism Effects 0.000 claims abstract description 24
- 238000004891 communication Methods 0.000 claims abstract description 12
- 239000000284 extract Substances 0.000 claims abstract description 11
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 10
- 230000002776 aggregation Effects 0.000 claims abstract description 5
- 238000004220 aggregation Methods 0.000 claims abstract description 5
- 238000003062 neural network model Methods 0.000 claims abstract description 4
- 239000010410 layer Substances 0.000 claims description 30
- 210000002569 neuron Anatomy 0.000 claims description 17
- 238000013528 artificial neural network Methods 0.000 claims description 11
- 230000006870 function Effects 0.000 claims description 9
- 230000001965 increasing effect Effects 0.000 claims description 7
- 239000011159 matrix material Substances 0.000 claims description 7
- 230000004931 aggregating effect Effects 0.000 claims description 5
- 238000012935 Averaging Methods 0.000 claims description 4
- 230000008569 process Effects 0.000 claims description 4
- 230000014509 gene expression Effects 0.000 claims description 3
- 239000002356 single layer Substances 0.000 claims description 3
- 238000005457 optimization Methods 0.000 abstract description 8
- 238000005516 engineering process Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 230000002708 enhancing effect Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000000379 polymerizing effect Effects 0.000 description 1
- 238000012552 review Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/23—Updating
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/20—Ensemble learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Medical Informatics (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Databases & Information Systems (AREA)
- Machine Translation (AREA)
Abstract
The invention discloses a method for updating real-time online prediction of research interest of experts in federated learning, which comprises the following steps: each participant locally extracts expert text information for modeling to obtain semantic information and time information; learning the expert characteristics by using a text convolution neural network model, and introducing a self-attention mechanism to strengthen semantic information and learning time information of a text; by adopting a transverse federal learning mode, after each participant trains a model locally, training parameters are submitted to a global server, optimization is carried out by using a federal matching average algorithm, a hierarchical matching and communication mechanism is added, the server carries out global aggregation of the parameters, returns the global parameters to each local model and updates the local model; and clustering the coding vectors obtained by training the text convolutional neural network model to obtain the research interest of experts. The method solves the problem of data isolated island, protects data privacy and improves the accuracy of prediction of expert research interest.
Description
Technical Field
The invention belongs to the technical field of data processing, and particularly relates to a method for updating real-time online prediction of research interests of experts in federated learning.
Background
At present, the quantity of relevant data resources of scientific research experts is huge, the data resources are scattered in distribution, and deep processing is not carried out, so that it is difficult for users to quickly and accurately obtain academic information with actual value. The research interest of experts is particularly prominent. In searching for experts in different fields, it is not accurate enough according to the subject filled by the experts, the experts may have more prominent research in other fields, and the cooperation between different fields is more and more intimate, so in some demands, for example, project review department wants to find suitable reviewers, scientific research user seeks collaborators, professional questions need to consult domain expert help, etc., it will take a lot of time and effort and the result of the query is not necessarily valuable. The research fields of the experts are extracted from the academic experiences of the experts, the papers are published, and the cooperation projects are participated in, and the research interest of the updated experts is predicted according to the time information, so that the method is greatly helpful for accurately searching the needed experts.
The heterogeneity and privacy of expert data pose a great challenge to expert information integration, and mainly face the following problems: data security problems, huge text quantity, data island problems and the like. The existing federal learning techniques are mainly of two types, namely, the horizontal federal learning: under the condition that the user features of the two data sets are overlapped more and the user overlap less, transversely segmenting the data sets; longitudinal federal learning: and under the condition that the users of the two data sets overlap more and the user features overlap less, longitudinally segmenting the data sets.
Due to the fact that user data are not independently and simultaneously distributed and the user data amount is unbalanced, a traditional federal learning technology has large deviation when local data are updated, meanwhile, a model automatically deviates to equipment with good network conditions, and the training result is influenced by inconsistency of a system and the data.
Disclosure of Invention
In order to solve the problems, the invention provides a method for updating the expert research interest of federated learning in real time on-line prediction, which is characterized in that a text convolutional neural network model and an attention mechanism are utilized to learn the characteristics of experts, the expert research interest is obtained from expert information through the text convolutional neural network, a transverse federated learning technology is introduced, a hierarchical average optimization is utilized, a communication mechanism training model is added, the problem of data island is solved, the data privacy is protected, and meanwhile, the accuracy of the prediction of the expert research interest is improved.
In order to achieve the purpose, the invention adopts the technical scheme that: a method for updating the real-time online prediction of research interests of experts in federated learning comprises the following steps:
s10, aiming at expert information data of heterogeneous and non-public talent expert libraries in various regions, each participant respectively extracts expert text information locally for modeling to obtain semantic information and time information;
s20; learning the characteristics of the experts by using a text convolution neural network model, introducing a self-attention mechanism to strengthen semantic information of the text, and learning time information;
s30, adopting a transverse federal learning mode, submitting training parameters to a global server after each participant trains the model locally, optimizing by using a federal matching average algorithm, adding a hierarchical matching and communication mechanism, performing global aggregation of the parameters by the server, returning the global parameters to each local model, and updating the local model;
and S40, clustering the coding vectors obtained by training the text convolutional neural network model to obtain the research interest of experts.
Further, in step S10, facing expert information data of heterogeneous and non-public talent expert libraries in various regions, each participant extracts expert text information locally for modeling, and obtains semantic information and time information, including the steps of:
s11, aiming at expert information data of heterogeneous and non-public talent expert libraries in various regions, each participant respectively extracts expert text information including expert personal academic experience information, expert published paper information, expert participation project information and expert acquired prize information;
and S12, inputting the expert personal academic experience, the expert published paper information, the expert participation project information and the expert acquired reward information into a word vector model in a document form for training, and performing unified coding to obtain a word vector comprising a semantic information vector x and timestamp information t.
Further, the Word vector model adopts a Word2vec pre-training model.
Further, in step S20, learning the expert features by using the text convolutional neural network model includes the steps of:
s211, representing the text semantic information, dividing the text information into sentences, wherein each sentence is n words, each word is a k-dimensional vector, and the obtained word vector matrix is represented as:
s212, generating characteristic c for each word by using a filter i Expressed as:
c i =f(w·x i:i+h-1 +b);
wherein h is the height of the filter, b is a bias term, w is the weight, and f is a nonlinear function;
s213, generating sentence features from the obtained word features c as follows:
c=[c 1 ,c 2 ,...,c n-h+1 ];
s214, regularizing the penultimate layer of the text convolution neural network by using dropout, and finally outputting:
where w is the regularization weight, z is the sentence feature set,is the bitwise multiplication operator, r is the height of the bernoulli random variable filter, and b is a bias term.
Further, in step S20, the method for enhancing semantic information representation and learning time information by using a multi-head attention mechanism includes the steps of:
s221, for each semantic information vector x, a piece of timestamp information t corresponds to the semantic information vector x, and the timestamp information t is added into word vector semantics to obtain a new feature matrix Y (x, t);
s222, calculating a similarity coefficient between each sentence and an adjacent sentence, and normalizing to obtain an attention coefficient;
s223, fusing neighborhood information according to the obtained attention coefficient, and aggregating the neighborhood information with the original characteristics of the sentence to form new characteristics;
and S224, introducing a multi-head self-attention mechanism, simultaneously using a plurality of shared parameters W, combining the expressions of a plurality of newly generated sentences into one, and obtaining the information code with the time stamp.
Further, in step S30, the extracted hidden elements with similar features are matched and averaged by using a federal learning average algorithm, and a shared global model is built layer by layer.
Further, in each iteration, a corresponding global model is found according to a given weight matrix estimation, and then the global model is matched with local neurons on a client according to a Hungarian algorithm to obtain a new expanded global model, and the method comprises the following steps:
s321, finding the parameter arrangement before averaging the parameters of the neural network, and combining the weights of the neurons with similar characteristics;
s322, given an increasing function, optimizing the Nippon average algorithm to obtain an optimized average matching formula as follows:
wherein: l represents the number of nodes of the hidden layer, c (w) jl ,θ i ) Representing the similarity relationship between the ith neuron learned on the client j and the ith neuron in the global server model, f (i) representing an increasing function, σ representing a threshold,represents an arrangement of parameters;
s323, given the weight W provided by J clients j,1 ,W j,2 Calculating the weight of the federal neural network:
and S324, layering and matching, and finally aggregating to obtain model updating parameters.
Further, if the cost of the optimal matching is larger than a set threshold, no matching is performed, and a new global neuron is created in the corresponding local neuron.
Further, the hierarchical matching scheme includes the steps of:
s331, the global server only collects the weight of the first layer from the client, and executes the single-layer matching described above to obtain the weight of the first layer of the federal model;
s332, the global server broadcasts the obtained first-layer weight of the federated model to the client, and the client continues to train all continuous layers on a data set of the client and keeps the federated matching layer in a frozen state;
s333, continuously repeating the process until the last layer is reached; and for the last layer, carrying out weighted average on each client data point according to the class proportion of the client data point to obtain a final parameter.
Further, a communication mechanism is added into the federal mean learning algorithm, and the method comprises the following steps:
and the local client receives the matched global model when a new round is started, and reconstructs the local model with the same size as the original local model on the basis of the previous round of matching result, so that the size of the global model is kept in a smaller range.
The beneficial effects of the technical scheme are as follows:
according to the invention, aiming at expert information data of heterogeneous and non-public talent expert libraries in various regions, each participant extracts expert text information locally for modeling; learning the characteristics of the experts by using a text convolutional neural network model, introducing a self-attention mechanism to learn semantic information of an enhanced text, learning time information, and obtaining codes; and (3) adopting a transverse federal learning mode, carrying out model training locally by each participant, submitting training parameters to a global server, optimizing by using a federal matching average algorithm, and adding a hierarchical matching and communication mechanism. The server carries out global aggregation of the parameters, returns the global parameters to each local model and updates the local models; and clustering the coding vectors obtained by the network model by the training text convolution spirit to obtain the research interest of experts. According to the method, the semantic and structural information in the expert information is fully utilized, the expert research interest is obtained from the expert information through the text convolution neural network, the horizontal federal learning technology is introduced, the problem of data islanding is solved, the data privacy is protected, and meanwhile, the accuracy of prediction of the expert research interest is improved.
The invention fully mines the information of expert data from different sources in order to more effectively utilize various elements in the expert information; the elements and the time stamps in the expert information are vectorized, a feature vector containing text features and time features is constructed, different elements in the expert information and the relation between the elements can be effectively expressed, and therefore more knowledge is obtained, and the model effect is superior to that of a common information representation method.
The text convolutional neural network model is introduced to encode information, and has the greatest advantages of simple network structure and fewer parameters, so that the calculation efficiency is higher than that of other models. And adding a dropout regularization method to prevent the overfitting of the whole body from optimizing the updated model of the parameters by using a gradient method. By utilizing a multi-head self-attention mechanism, semantic information is strengthened, time information is learned, and the influence of the time information on research interest is considered. Meanwhile, a federal learning method is introduced, so that each party of data is trained on the premise of protecting privacy, the data volume is increased, and the accuracy of a prediction result is improved.
The invention focuses on the situation of data deviation, and each domain is regarded as a client, so that the local model is not influenced by aggregated data deviation, and the meaningful relation between functions and classes is learned. Algorithms can then be used to learn a good global model without bias. After considering the limitation that the parameters of the local model are evenly distributed according to the element average and the weight is in proportion to the size of the client data set in the prior federal learning algorithm (limited by the size of the client data set), the weights of the neurons with similar characteristics are combined based on the replacement invariance of the neurons, and a global sharing model is constructed. The training result of the text convolution neural network is optimized, and the communication efficiency is improved.
In order to be more suitable for a real online environment, a communication mechanism is added in a horizontal federal learning algorithm, the whole data cannot be trained due to the privacy of the data in the real environment, the communication mechanism is added in a global server and various local clients, wherein the local client receives a matched global model when a new round begins, and reconstructs a local model with the size equal to that of an original local model based on a matching result of the previous round. This process allows keeping the size of the global model small. Thereby reducing the performance requirements on the local client network. When a newly added client side is used for training, parameters of previous training can be directly transmitted to a newly added client side model, then new training transmission is carried out, and newly added data can be more conveniently integrated into overall data. Therefore, in a heterogeneous data scenario, our algorithm outperforms other federal learning methods.
Drawings
FIG. 1 is a schematic flow chart of an update method for real-time online prediction of research interests of experts in federated learning according to the present invention;
FIG. 2 is a block diagram of an expert research interest prediction model according to the present invention;
FIG. 3 is a schematic diagram of the principle framework for lateral federal learning of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described with reference to the accompanying drawings.
In this embodiment, referring to fig. 1 and fig. 2, the present invention provides a real-time online prediction updating technique for research interests of experts in bang learning, including the steps of:
s10, aiming at expert information data of heterogeneous and non-public talent expert libraries in various regions, each participant respectively extracts expert text information locally for modeling to obtain a semantic information vector X and a timestamp vector t;
s20; learning expert features by using a text convolution neural network model, introducing a self-attention mechanism to strengthen semantic information of a text, and learning time information;
s30, training models locally by each participant in a horizontal federal learning mode, submitting training parameters to a global server, optimizing by using a federal matching average algorithm, adding a hierarchical matching and communication mechanism, performing global aggregation on the parameters by the server, returning the global parameters to each local model, and updating the local models;
and S40, clustering the coding vectors obtained by the network model by the training text convolution spirit to obtain the research interest of experts.
As an optimization scheme 1 of the above example, in step S10, in the face of expert information data of heterogeneous and non-public talent expert libraries in various regions, each participant respectively extracts expert text information locally for modeling, and obtains semantic information and time information, including the steps of:
s11, aiming at expert information data of heterogeneous and non-public talent expert libraries in various regions, each participant respectively extracts expert text information including expert personal academic experience information, expert published paper information, expert participation project information and expert acquired award information locally;
and S12, inputting the expert personal academic experience, the expert published paper information, the expert participation project information and the award information obtained by the expert into a word vector model in a document form for training, and performing unified coding to obtain a semantic information vector x and timestamp information t.
As an optimization scheme 2 of the above example, in the step S20, learning the expert features by using the text convolutional neural network model includes the steps of:
s212, generating characteristic c for each word by using a filter i Expressed as:
c i =f(w·x i:i+h-1 +b);
where h is the height of the filter, b is a bias term, w is the weight, and f is a non-linear function.
And S213, generating sentence characteristics according to the obtained word characteristics c, wherein the sentence characteristics are as follows:
c=[c 1 ,c 2 ,...,c n-h+1 ];
s214, regularizing the penultimate layer of the text convolution neural network by using dropout, and finally outputting:
where w is the regularization weight, z is the sentence feature set,is the bitwise multiplication operator, r is the height of the bernoulli random variable filter, and b is a bias term.
The pre-training Word vector model of the optimization scheme adopts a Word2vec pre-training model.
As the optimization scheme 3 of the above example, in the step S20, the method for enhancing semantic information representation and learning time information by using a multi-head self-attention mechanism includes the steps of:
s221, for each semantic information vector x, a piece of timestamp information t corresponds to the semantic information vector x, and the timestamp information t is added into word vector semantics to obtain a new feature matrix Y (x, t);
s222, then calculating a similarity coefficient between each sentence and an adjacent sentence, and obtaining an attention coefficient after normalization;
s223, fusing neighborhood information according to the obtained attention coefficient, and aggregating the neighborhood information with the original characteristics of the sentence to form new characteristics;
and S224, introducing a multi-head self-attention mechanism, simultaneously using a plurality of sharing parameters W, and combining the expressions of a plurality of newly generated sentences into one to obtain a code.
As an optimization scheme 4 of the above example, as shown in fig. 3, in step S30, S321, before averaging the parameters of the neural network, find the arrangement of the parameters, and combine the weights of neurons with similar features;
s322, given an increasing function, optimizing the Nippon average algorithm to obtain an optimized average matching formula as follows:
wherein: l represents the number of nodes of the hidden layer, c (w) jl ,θ i ) Representing the similarity relation between the ith neuron learned on the client j and the ith neuron in the global server model; f (i) represents an increasing function, σ represents a threshold value,representing the arrangement of the parameters.
S323, given the weight W provided by J clients j,1 ,W j,2 Calculating the weight of the federal neural network:
and S324, layering and matching, and finally polymerizing to obtain the required parameters.
As the optimization scheme 4 of the above example, a hierarchical matching scheme is adopted, which includes the steps of:
s331, the global server collects only the first-layer weights from the client and performs the single-layer matching described above to obtain the first-layer weights of the federated model.
S332, the global server broadcasts the weights to the client, the client continues to train all continuous layers on the data set of the client, and the federated matching layer is kept in a frozen state.
S333, this process is repeated until the last layer. For the last layer, weighted average is carried out on each client data point according to the class proportion of the client data point to obtain a final parameter.
Finally, a communication mechanism is added into the federal mean learning algorithm, and the method comprises the following steps:
and the local client receives the matched global model when a new round is started, and reconstructs a local model with the same size as the original local model on the basis of the previous round of matching result. Thereby keeping the size of the global model within a small range.
The foregoing shows and describes the general principles and broad features of the present invention and advantages thereof. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are given by way of illustration of the principles of the present invention, but that various changes and modifications may be made without departing from the spirit and scope of the invention, and such changes and modifications are within the scope of the invention as claimed. The scope of the invention is defined by the appended claims and equivalents thereof.
Claims (10)
1. A method for updating the real-time online prediction of the research interest of experts in the United nations study is characterized by comprising the following steps:
s10, aiming at expert information data of heterogeneous and non-public talent expert libraries in various regions, each participant respectively extracts expert text information locally for modeling to obtain semantic information and time information;
s20; learning expert features by using a text convolution neural network model, introducing a self-attention mechanism to strengthen semantic information of a text, and learning time information;
s30, adopting a transverse federal learning mode, submitting training parameters to a global server after each participant trains the model locally, optimizing by using a federal matching average algorithm, adding a hierarchical matching and communication mechanism, performing global aggregation of the parameters by the server, returning the global parameters to each local model, and updating the local model;
and S40, clustering the coding vectors obtained by training the text convolutional neural network model to obtain the research interest of experts.
2. The method for updating the real-time online prediction of the research interests of the experts in federated learning according to claim 1, wherein in step S10, in the face of expert information data of heterogeneous and unpublished talent expert bases in various regions, each participant respectively extracts expert text information locally for modeling to obtain semantic information and time information, comprising the steps of:
s11, aiming at expert information data of heterogeneous and non-public talent expert libraries in various regions, each participant respectively extracts expert text information including expert personal academic experience information, expert published paper information, expert participation project information and expert acquired prize information;
and S12, inputting the expert personal academic experience, the expert published paper information, the expert participation project information and the expert acquired reward information into a word vector model in a document form for training, and performing unified coding to obtain a word vector comprising a semantic information vector x and timestamp information t.
3. The method of claim 2, wherein the Word vector model is a Word2vec pre-training model.
4. The method for updating the real-time online prediction of the research interest of the expert in federated learning according to claim 1, wherein in step S20, learning the expert characteristics using a text convolutional neural network model includes the steps of:
s211, representing the text semantic information, dividing the text information into sentences, wherein each sentence is n words, each word is a k-dimensional vector, and the obtained word vector matrix is represented as:
s212, generating characteristic c for each word by using a filter i Expressed as:
c i =f(w·x i:i+h-1 +b);
wherein h is the height of the filter, b is a bias term, w is the weight, and f is a nonlinear function;
s213, generating sentence features from the obtained word features c as follows:
c=[c 1 ,c 2 ,...,c n-h+1 ];
s214, regularizing the penultimate layer of the text convolution neural network by using dropout, and finally outputting:
5. The method for updating real-time online prediction of research interest of experts in federated learning according to claim 1 or 4, wherein in step S20, semantic information representation and learning time information are enhanced by a multi-head self-attention mechanism, comprising the steps of:
s221, for each semantic information vector x, a piece of timestamp information t corresponds to the semantic information vector x, and the timestamp information t is added into word vector semantics to obtain a new feature matrix Y (x, t);
s222, calculating a similarity coefficient between each sentence and an adjacent sentence, and normalizing to obtain an attention coefficient;
s223, fusing neighborhood information according to the obtained attention coefficient, and aggregating the neighborhood information with the original characteristics of the sentence to form new characteristics;
and S224, introducing a multi-head self-attention mechanism, simultaneously using a plurality of shared parameters W, combining the expressions of a plurality of newly generated sentences into one, and obtaining the information code with the time stamp.
6. The method for updating the real-time online prediction of the research interests of the experts in federated learning according to claim 1, wherein in step S30, the extracted hidden elements with similar features are matched and averaged by using a federated learning averaging algorithm, and a shared global model is constructed layer by layer.
7. The method for updating the real-time online prediction of the research interests of the experts in the federated learning according to claim 6, wherein in each iteration, a corresponding global model is found according to a given weight matrix estimation, and then the global model is matched with local neurons on a client according to the Hungarian algorithm to obtain a new expanded global model, comprising the steps of:
s321, finding the parameter arrangement before averaging the parameters of the neural network, and combining the weights of the neurons with similar characteristics;
s322, given an increasing function, optimizing the Nippon average algorithm to obtain an optimized average matching formula as follows:
wherein: l represents the number of nodes of the hidden layer, c (w) jl ,θ i ) Representing the similarity relationship between the ith neuron learned on client j and the ith neuron in the global server model, f (i) representing an increasing function, σ representing a threshold,represents an arrangement of parameters;
s323, given the weight W provided by J clients j,1 ,W j,2 Calculating the weight of the federal neural network:
and S324, layering and matching, and finally aggregating to obtain model updating parameters.
8. The method of claim 7, wherein if the cost of optimal matching is greater than a set threshold, no matching is performed, and a new global neuron is created in the corresponding local neuron.
9. The method for updating the expert research interest of federated learning on-line in real time prediction according to claim 7 or 8, wherein the hierarchical matching scheme comprises the steps of:
s331, the global server only collects the weight of the first layer from the client, and executes the single-layer matching described above to obtain the weight of the first layer of the federal model;
s332, the global server broadcasts the obtained first-layer weight of the federated model to the client, the client continues to train all continuous layers on a data set of the client, and the federated matching layer is kept in a frozen state;
s333, continuously repeating the process until the last layer is reached; and for the last layer, carrying out weighted average on each client data point according to the class proportion of the client data point to obtain a final parameter.
10. The method for updating the real-time online prediction of the research interests of the experts in federated learning according to claim 7, wherein a communication mechanism is added to the federated average learning algorithm, comprising the steps of:
and the local client receives the matched global model when a new round is started, and reconstructs the local model with the same size as the original local model on the basis of the previous round of matching result, so that the size of the global model is kept in a smaller range.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210653137.1A CN114969078A (en) | 2022-06-09 | 2022-06-09 | Method for updating expert research interest of federated learning through real-time online prediction |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210653137.1A CN114969078A (en) | 2022-06-09 | 2022-06-09 | Method for updating expert research interest of federated learning through real-time online prediction |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114969078A true CN114969078A (en) | 2022-08-30 |
Family
ID=82961790
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210653137.1A Pending CN114969078A (en) | 2022-06-09 | 2022-06-09 | Method for updating expert research interest of federated learning through real-time online prediction |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114969078A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117350373A (en) * | 2023-11-30 | 2024-01-05 | 艾迪恩(山东)科技有限公司 | Personalized federal aggregation algorithm based on local self-attention mechanism |
CN117494191A (en) * | 2023-10-17 | 2024-02-02 | 南昌大学 | Point-of-interest micro-service system and method for information physical security |
-
2022
- 2022-06-09 CN CN202210653137.1A patent/CN114969078A/en active Pending
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117494191A (en) * | 2023-10-17 | 2024-02-02 | 南昌大学 | Point-of-interest micro-service system and method for information physical security |
CN117350373A (en) * | 2023-11-30 | 2024-01-05 | 艾迪恩(山东)科技有限公司 | Personalized federal aggregation algorithm based on local self-attention mechanism |
CN117350373B (en) * | 2023-11-30 | 2024-03-01 | 艾迪恩(山东)科技有限公司 | Personalized federal aggregation algorithm based on local self-attention mechanism |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111414461B (en) | Intelligent question-answering method and system fusing knowledge base and user modeling | |
CN114969078A (en) | Method for updating expert research interest of federated learning through real-time online prediction | |
CN112967088A (en) | Marketing activity prediction model structure and prediction method based on knowledge distillation | |
Pérez et al. | Data augmentation through multivariate scenario forecasting in Data Centers using Generative Adversarial Networks | |
CN111274790A (en) | Chapter-level event embedding method and device based on syntactic dependency graph | |
CN113065649A (en) | Complex network topology graph representation learning method, prediction method and server | |
CN112257841A (en) | Data processing method, device and equipment in graph neural network and storage medium | |
CN112884045B (en) | Classification method of random edge deletion embedded model based on multiple visual angles | |
CN112988917A (en) | Entity alignment method based on multiple entity contexts | |
CN113344615A (en) | Marketing activity prediction method based on GBDT and DL fusion model | |
CN112417289A (en) | Information intelligent recommendation method based on deep clustering | |
Liu et al. | Research and citation analysis of data mining technology based on Bayes algorithm | |
CN116386899A (en) | Graph learning-based medicine disease association relation prediction method and related equipment | |
CN117131933A (en) | Multi-mode knowledge graph establishing method and application | |
Xu et al. | Improving knowledge tracing via a heterogeneous information network enhanced by student interactions | |
Chen et al. | Feature extraction method of 3D art creation based on deep learning | |
CN116541593B (en) | Course recommendation method based on hypergraph neural network | |
CN115588487B (en) | Medical image data set manufacturing method based on federal learning and antagonism network generation | |
Zhang | [Retracted] Analysis of College Students’ Network Moral Behavior by the History of Ideological and Political Education under Deep Learning | |
Zhang et al. | Probabilistic matrix factorization recommendation of self-attention mechanism convolutional neural networks with item auxiliary information | |
CN115063251A (en) | Social communication propagation dynamic network representation method based on relationship strength and feedback mechanism | |
Liang et al. | [Retracted] Painting Classification in Art Teaching under Machine Learning from the Perspective of Emotional Semantic Analysis | |
CN109299291A (en) | A kind of Ask-Answer Community label recommendation method based on convolutional neural networks | |
Shen et al. | Online teaching course recommendation based on autoencoder | |
CN114970684A (en) | Community detection method for extracting network core structure by combining VAE |
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 |