CN116226540A - End-to-end federation personalized recommendation method and system based on user interest domain - Google Patents
End-to-end federation personalized recommendation method and system based on user interest domain Download PDFInfo
- Publication number
- CN116226540A CN116226540A CN202310512758.2A CN202310512758A CN116226540A CN 116226540 A CN116226540 A CN 116226540A CN 202310512758 A CN202310512758 A CN 202310512758A CN 116226540 A CN116226540 A CN 116226540A
- Authority
- CN
- China
- Prior art keywords
- client
- user
- model
- training
- layer
- 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
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
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
- G06N3/088—Non-supervised learning, e.g. competitive learning
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Databases & Information Systems (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses an end-to-end federal personalized recommendation method and system based on a user interest domain, and belongs to the technical field of personalized recommendation. Based on the cluster-like personalized thought, the recommendation model with the personalized prediction network group and the client classification network based on the graph neural network are arranged, the classification network is embedded into the personalized recommendation model, and the recommendation model and the client classification network are used as a global model in a federal recommendation system for federal training. Through the design of the end-to-end personalized recommendation model, the clustering process of the client can be automatically completed in the forward calculation of the model, so that the training efficiency is improved; the client classification network can also be trained by recommending targets so that the clustering process becomes learnable. The invention solves the problems of low training efficiency and non-learning in the conventional cluster-like personalized recommendation method, and improves the performance of the personalized federal recommendation model.
Description
Technical Field
The invention belongs to the technical field of personalized recommendation, and particularly relates to an end-to-end federal personalized recommendation method and system based on a user interest domain.
Background
The recommendation system is an effective means for relieving information overload, and is also an engine for increasing users, and is visible everywhere on each large Internet information platform. In recent years, with the increase of the privacy awareness of users and the release of various data security regulations, it is recommended to face privacy protection problems due to the need to use personal data of users. The combination of federal learning and recommendation is currently the main idea to solve this problem. However, the general federal recommendation method only focuses on privacy protection, ignores the problem of non-independent co-distribution (non-IID) of client (user) data under federal settings, and the problem can affect training of a model, resulting in impaired recommendation performance of a final model.
At present, the non-IID problem is solved in a plurality of ways, one of which is to adopt a personalized idea, namely, conversely, different models are customized for different clients by utilizing the nature of the non-IID. Besides the individuation method of the granularity of the client, the individuation scheme mainly clusters the clients to form a plurality of class clusters, and endows different models to different class clusters, wherein the clients in the same class cluster share one model. The two existing personalization methods have the following three defects in federal recommendation:
(1) Personalization of client granularity easily results in model overfitting local data, compromises generalization performance of the model, and once interest of a user is transformed, the model is difficult to recommend with high quality against new interests.
(2) The training model under the cluster-like granularity has more abundant data and wider coverage interest, so that the fitting problem is relieved; however, the method requires that a certain number of clients are clustered in each round of federal training, which affects training efficiency.
(3) Besides the training efficiency problem, the implementation process of the cluster granularity individuation method basically adopts an unsupervised clustering method and a non-learning method.
Disclosure of Invention
In order to solve the problem of non-IID of client data in the traditional federation recommendation application, the invention also adopts a personalized solution idea, and provides an end-to-end federation personalized recommendation method and system based on a cluster personalized idea, thereby improving the problems of low training efficiency and non-learning in the existing cluster personalized recommendation method and improving the performance of a federation recommendation model.
The invention provides the following technical scheme:
one of the purposes of the invention is to provide an end-to-end federal personalized recommendation method which mainly comprises two large contents of an end-to-end personalized recommendation model and a federal training process.
The end-to-end personalized recommendation model performs personalized design on the prediction network group part on the basis of the depth recommendation model, namely, z prediction networks are set, and z client class clusters are respectively corresponding to the z prediction networks. The invention does not personalize the whole recommendation model, but only sets a personalized prediction network group, because (1) the object ID embedding layer of the recommendation model has the characteristic sparsity problem, and users in one class cluster cannot cover all objects, so personalizing the layer can lead to some hidden vectors in one class cluster not to be trained, which brings the object cold start problem and is not beneficial to the generalization of the model; (2) Among the parameters of the federal recommendation model, the item ID embedding layer accounts for a large part, and sharing the layer by all clients is beneficial to saving the storage cost of the parameters.
After the personalized prediction network group is set, the invention also sets a client classification network based on a graph neural network, mainly regards a client (user) and an object interacted with the client as a star graph, namely a client-object interaction graph, takes the interaction graph, a user ID embedded representation and an object ID embedded representation in the interaction graph as the input of the graph neural network, wherein the user ID embedded representation and the object ID embedded representation are the characteristic representation of each node in the graph, and outputs probability distribution vectors of the class cluster of the client corresponding to the interaction graph after passing through the graph neural network.
An end-to-end federation personalized recommendation method based on a user interest domain mainly comprises the following steps:
step 1: the method comprises the steps that a service end builds a global model based on the preset client class cluster number and initializes model parameters, wherein the global model consists of a client classification network and a recommendation model, and the recommendation model comprises a prediction network group, a user ID embedding layer and an article ID embedding layer; distributing the user ID embedded layer parameters in the global model to all clients participating in federal training;
step 2, constructing a local training data set by all clients of federal training;
step 3: the server randomly selects a part of clients to participate in the federal training of the current round, and distributes other parameters except the user ID embedding layer in the global model to the selected clients;
step 4: the selected client trains the received global model by using the local training data set, updates local user ID embedded layer parameters and sends the rest parameters back to the server;
step 5: the server side aggregates global model training parameters except the user ID embedded layer sent back by the client side, and takes an aggregate result as a latest global model;
step 6: repeating the steps 3 to 5 until the global model converges; and each client terminal carries out personalized recommendation of the target object according to the locally stored user ID embedded layer parameters and the latest global model parameters which are transmitted by the server terminal and are except the user ID embedded layer.
Further, the prediction network group is composed of a plurality of parallel multi-layer perceptron (MLP).
Further, the client classification network comprises a graph neural network, a multi-layer perceptron and a softmax layer; the input of the graphic neural network is a user-object interaction graphic, a user ID embedded representation and a history interaction object ID embedded representation; the user ID embedded representation and the historical interaction article ID embedded representation are respectively generated by a user ID embedded layer and an article ID embedded layer in the recommendation model; and the output of the graph neural network sequentially passes through a multi-layer perceptron and a softmax layer to obtain cluster probability distribution vectors of the client.
Further, the number of the client class clusters, the number of the parallel multi-layer perceptrons in the predicted network group and the dimension of the class cluster probability distribution vector output by the client classification network are the same.
Further, the cluster probability distribution vector output by the client classification network is used as the weight of the output result of the prediction network group, the weight is in a soft label form, and the weighted average value of the output result of the prediction network group is used as the final output of the recommendation model.
Further, in the recommendation model, the input of the user ID embedding layer is user ID, the input of the object ID embedding layer is history interaction object ID and target object ID, and the user ID embedding layer and the object ID embedding layer generate user ID embedding representation, history interaction object ID embedding representation and target object ID embedding representation; and splicing the user ID embedded representation, the average value of the historical interactive object ID embedded representation and the target object ID embedded representation to be used as the input of the prediction network group in the recommendation model.
Another object of the present invention is to provide an end-to-end federal personalized recommendation system based on a user interest domain for implementing the above method; the end-to-end federal personalized recommendation system comprises: the system comprises a server side and a plurality of clients, wherein the server side is used for constructing a global model based on the preset client class cluster number, initializing model parameters and aggregating the global model parameters except the user ID embedding layer after the training of each client side; the global model consists of a client classification network and a recommendation model, wherein the recommendation model comprises a prediction network group, a user ID embedding layer and an article ID embedding layer; the client is used for training the global model by using the local training data set and returning the global model training parameters except the user ID embedding layer to the server;
each client comprises:
the data construction module is used for constructing a local training data set of the client, marking the article actually interacted with the user as a positive sample based on the historical interaction data of the user and the article, setting a negative sampling rate, randomly sampling a plurality of negative samples in a global article set, and ensuring that no intersection exists between the positive sample and the negative sample;
the model training module is used for receiving global model parameters sent by the server side and then loading the latest global model, training the global model based on the training data set obtained by the data construction module, updating local user ID embedded layer parameters and sending the rest parameters back to the server side;
the data transmission module is used for data transmission between the server and the client and comprises model parameters sent by the client to the server after the client receives the model parameters sent by the server and model parameters sent by the client to the server after the local training is finished.
Further, the client classification network in the global model is used for generating a class cluster probability distribution vector of the client, the class cluster probability distribution vector is used as a weight of the output result of the prediction network group, and a weighted average value of the output result of the prediction network group is used as a final output of the recommendation model.
The invention has the technical effects that: according to the invention, through personalized recommendation of the predicted network group part of the model and embedding the cluster-like division process into the model, the clustering of the client is automatically completed in the forward calculation of the model, and the training efficiency is improved; the set client classification network can train through the recommended targets, so that the clustering process becomes learnable; in addition, the classification network is set as the graph neural network, so that the classification effect of the class clusters is enhanced, and the recommendation performance of the model is improved.
Drawings
FIG. 1 is a schematic diagram of a framework of an end-to-end federal personalized recommendation method according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a federal training process in an end-to-end federal personalized recommendation method according to an embodiment of the present invention;
FIG. 3 is a hierarchical structure diagram of an end-to-end federal personalized recommendation system according to an embodiment of the present invention;
fig. 4 shows training results on a public recommended dataset according to an embodiment of the present invention, wherein (a) the evaluation index is AUC, and (b) the evaluation index is hitrate@10.
Detailed Description
The invention is further described below with reference to the accompanying drawings.
The invention discloses an end-to-end training idea, which embeds a client classification network into a personalized recommendation model, and specifically comprises the following steps: designing a client classification network and a recommendation model with prediction network groups, taking each probability value in probability distribution vectors output by the client classification network as the weight output by each prediction network in the prediction network groups in the recommendation model, namely carrying out weighted average on the output of the prediction network groups through the probability distribution vectors, and taking the weighted average result as the final output of the personalized recommendation model. In this way, the client can calculate the probability distribution of the class cluster to which the client belongs according to the characteristics (interaction diagram) of the client when the model is calculated in the forward direction, and select the most suitable prediction network combination according to the probability distribution vector, and the whole process is completed in the forward direction calculation of the model, so that the end-to-end individuation is realized, and the training efficiency of the individuation of the class cluster is improved; when the model gradient is reversely updated, the client classification network can also update through the recommended target, so that the purpose of learning is realized. The output of the client classification network is set in the form of a soft label, and no hardening process is performed (e.g., [0.1, 0.1, 0.8] can be hardened to be [0, 1 ]), because different clusters are not completely irrelevant, and certain similarity exists, so that weighted average by the soft label is more reasonable.
As shown in fig. 1, the end-to-end personalized recommendation model provided by the invention mainly comprises two core parts, wherein one of the core parts is that a personalized prediction network group is arranged in the recommendation model, and the number z of the prediction networks is a super parameter, which means that clients are divided into z class clusters; the other core is to design a client classification network based on a graph neural network, and for each client, the input is a client (user) -object interaction graph and the output is a cluster probability distribution vector to which the corresponding client belongs. The invention embeds the client classification network into the personalized recommendation model, which is specifically as follows: the method not only realizes end-to-end class cluster division, but also enables the client classification network to learn through a recommendation target.
FIG. 2 is a federal training process in accordance with one embodiment of the present invention, wherein the global model is the model shown in FIG. 1, including a recommendation model and a client classification network; the parameters of the recommendation model can be divided into three parts, namely a predicted network group parameter, a user ID embedded layer parameter and an article ID embedded layer parameter, and the user ID embedded layer parameter cannot be acquired by a server side in consideration of privacy protection of a client side, so that the part of parameters need to be updated locally, namely after the server side is initialized, the user ID embedded layer parameter is distributed to a corresponding client side, the part of parameters are not local any more, and the rest of parameters of the global model are still uploaded to the server side for aggregation update after being trained locally.
The whole federal training process is as follows:
step 1: the server sets training parameters such as the sampling rate of the client, the local training round of the client, the batch size of the local training of the client and the like; setting the number z of class clusters, initializing model parameters, wherein the global model consists of a recommendation model and a client classification network, the recommendation model comprises a user ID embedding layer, an article ID embedding layer and a prediction network group, and the client classification network comprises a graph neural network and a multi-layer perceptron. The user ID embedded layer parameters in the global model are distributed to all clients participating in federal training.
Step 2: the client builds a local training dataset.
Step 3: the server randomly selects a part of clients to participate in the federal training of the round based on the client sampling rate, and distributes other parameters except the user ID embedding layer in the global model to the selected clients.
Step 4: the selected client trains the received global model locally according to training parameters set by the server to obtain parameter updating gradients, updates own user ID embedded layer parameters locally, and sends the trained parameters except the user ID embedded layer back to the server.
In the step, as shown in fig. 1, each client participating in the present training takes a user ID as a user ID embedding layer in a recommendation model, takes a history interaction article ID and a target article ID as an article ID embedding layer in the recommendation model, and respectively obtains a user ID embedding representation, a history interaction article ID embedding representation and a target article ID embedding representation; the user ID embedded representation, the average value of the historical interactive object ID embedded representation and the target object ID embedded representation are spliced to be used as the input of a prediction network group in the recommendation model; meanwhile, a user-object interaction diagram, a user ID embedded representation and a history interaction object ID embedded representation are used as inputs of a graphic neural network in a client classification network, an output result of the graphic neural network is used as inputs of a multi-layer perceptron, and then a cluster probability distribution vector corresponding to the client is output through softmax; the dimension of the class cluster probability distribution vector is consistent with the number of the prediction networks in the prediction network group, the class cluster probability distribution vector of the classification network is used as weight, the output of the prediction network group is subjected to weighted average, the output after weighted average is used as the final output of the recommendation model, and the gradient is updated according to the final output calculation parameter.
Step 5: the server adopts FedAvg algorithm to aggregate the global model training parameters except the user ID embedded layer sent back by the client, and takes the aggregate result as the latest global model.
Step 6: the server repeats the steps 3 to 5 until the whole training process is converged; and each client terminal carries out personalized recommendation of the target object according to the locally stored user ID embedded layer parameters and the latest global model parameters which are transmitted by the server terminal and are except the user ID embedded layer.
In one implementation of the invention, the global model is based on a PyTorch implementation. The local training parameters set by the server in the step 1 are 10% of the sampling rate of the client, the local training turns are 5, and the local training batch size is 1024. The number z of the predicted networks in the predicted network group of the end-to-end personalized recommendation model is 16, each predicted network is a four-layer perceptron with the size of [48, 200, 80, 1], reLu is sampled as a nonlinear activation function, only the ID characteristics (ID embedded representation extracted by an ID embedded layer) of a user and an article are utilized, the dimension of the ID embedded layer of the user and the article is set to be 16, and the input of the predicted network group in the recommendation model is a 48-dimensional vector formed by splicing three types of user ID embedded representation, average value of historical interaction article ID embedded representation and target article ID embedded representation. The client classification network part is provided with two layers of graph neural networks in total, wherein the two layers of graph neural networks are in a GCN neighborhood information aggregation form, and the two layers of graph neural networks are input into a user-object interaction graph, a user ID embedded representation and a history interaction object ID embedded representation by adopting a vector splicing mode as an aggregation function of different-order neighborhood, the output of each layer of the graph neural network part is a 16-dimensional vector, and the output of the graph neural network part is subjected to three-layer perceptrons and softmax layers of [16, 80, 16] to finally output a cluster probability distribution vector of the client.
The window of the user history interactive objects is set to be 50, and the situation that less than 50 interactive objects are replaced by the No. 0 object.
In one implementation of the present invention, step 2 is primarily a local dataset used to build a training recommendation model. The local interaction data of all clients are ordered according to time sequence, wherein the first 80% is used as a training set, and the second 20% is used as a testing set. The negative sampling rate of the training set is set to 1:4, i.e. 1 positive sample corresponds to 4 negative samples, and the negative sampling rate of the test set is set to 1:99, i.e. 1 positive sample corresponds to 99 negative samples.
In one embodiment of the present invention, the optimizer of the client local training model in step 4 uses Adam, the learning rate is set to 0.01, adam optimizer parameters are initialized before each client starts training in each round of federal training, and the parameters do not participate in federal aggregation.
In one embodiment of the present invention, the service end in step 5 adopts FedAvg algorithm aggregation model, that is, fedAvg algorithm, and the parameter aggregation process can be expressed as
Wherein t represents the t-th round of federal training, K represents the number of clients participating in the present round of federal training (which may be numbered 1,2, …, K), n k Representing the data set size of the local client, N is the overall data set size formed by the clients participating in the current round of federal training,
representing global model parameters to be aggregated, W, obtained by local training of kth client of t-th round t+1 Represents the initial aggregation parameters for the t+1st round of federal training.
In one implementation of the present invention, the movie rating dataset MovieLens-1M is employed as the experimental dataset and the interaction data is partitioned by user ID to simulate the federal recommended scenario. AUC and HitRate@10 were used as evaluation indices.
Because the user is recorded as the client, the MovieLens-1M data set contains 6040 users, and therefore 6040 intelligent terminals are required to perform experiments according to actual conditions, and the experiment resources are limited. Fig. 4 (a) and (b) each show experimental results of an embodiment of the present invention, the abscissa shows federal communication round, and the ordinate of (a) and (b) in fig. 4 is AUC and hitrate@10, respectively, wherein the dotted line shows a conventional recommendation model, only including an embedded layer and a multi-layer perceptron, and the solid line shows a global model shown in fig. 1 proposed by the present invention. The experimental result shows that the effect of the invention is obviously better than that of the traditional recommendation model, and the effectiveness of the invention is demonstrated.
In this embodiment, an end-to-end federal personalized recommendation system based on the user interest domain is also provided, as shown in fig. 1, and the system is used to implement the foregoing embodiments. The terms "module," "unit," and the like, as used below, may be a combination of software and/or hardware that performs a predetermined function. Although the system described in the following embodiments is preferably implemented in software, implementation in hardware, or a combination of software and hardware, is also possible.
The federal learning aggregation system based on no data distillation provided in this embodiment includes:
fig. 3 is a hierarchical structure diagram of the end-to-end federal personalized recommendation system proposed by the present invention, and the whole is composed of a server and a plurality of clients.
The server side represents a certain article recommendation platform in practical application, and needs to train to obtain a global model through a federal learning mode under the condition of privacy protection, and a specific flow is shown in fig. 2 and is already described in detail in the method section.
The client represents an intelligent terminal (such as a mobile phone) in practical application, and a registered user of the article recommendation platform browses articles through the intelligent terminal, so that one terminal can be considered to correspond to one user, namely one client corresponds to one user.
Each client comprises three modules:
and the data construction module is used for constructing a local training data set and a test data set of the client. The training set and the testing set are divided through the interaction time of the user and the article, and the training set and the testing set are specifically as follows: the local interaction data of the client are ordered in time sequence, the former part of data is used for constructing a training set, and the latter part of data is used for constructing a testing set.
In this embodiment, the training data set and the test data set are constructed according to implicit feedback, specifically: based on the interaction data of the user and the objects, the objects actually interacted with the user are marked as positive samples, the negative sampling rate is set, a plurality of negative samples are randomly sampled in the global object set, and no intersection of the positive samples and the negative samples is ensured.
The model training module is mainly used for locally training the global model, a model framework is arranged in the model training module, the latest global model is loaded after receiving the global model parameters sent by the server, the global model is trained based on the training data set obtained by the data construction module, and after the gradient is obtained, the user ID embedded layer parameters in the global model are locally updated.
The data transmission module is used for data transmission between the server and the client and comprises model parameters sent by the client to the server after the client receives the model parameters sent by the server and model parameters sent by the client to the server after the local training is finished.
For the system embodiment, since the system embodiment basically corresponds to the method embodiment, the relevant parts only need to be referred to in the description of the method embodiment, and the implementation methods of the remaining modules are not repeated herein. The system embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purposes of the present invention. Those of ordinary skill in the art will understand and implement the present invention without undue burden.
Embodiments of the system of the present invention may be applied to any device having data processing capabilities, such as a computer or the like. The system embodiment may be implemented by software, or may be implemented by hardware or a combination of hardware and software. Taking software implementation as an example, the device in a logic sense is formed by reading corresponding computer program instructions in a nonvolatile memory into a memory by a processor of any device with data processing capability.
The foregoing list is only illustrative of specific embodiments of the invention. Obviously, the invention is not limited to the above embodiments, but many variations are possible. All modifications directly derived or suggested to one skilled in the art from the present disclosure should be considered as being within the scope of the present invention.
Claims (10)
1. The end-to-end federal personalized recommendation method based on the user interest domain is characterized by comprising the following steps of:
step 1: the method comprises the steps that a service end builds a global model based on the preset client class cluster number and initializes model parameters, wherein the global model consists of a client classification network and a recommendation model, and the recommendation model comprises a prediction network group, a user ID embedding layer and an article ID embedding layer; distributing the user ID embedded layer parameters in the global model to all clients participating in federal training;
step 2, constructing a local training data set by all clients of federal training;
step 3: the server randomly selects a part of clients to participate in the federal training of the current round, and distributes other parameters except the user ID embedding layer in the global model to the selected clients;
step 4: the selected client trains the received global model by using the local training data set, updates local user ID embedded layer parameters and sends the rest parameters back to the server;
step 5: the server side aggregates global model training parameters except the user ID embedded layer sent back by the client side, and takes an aggregate result as a latest global model;
step 6: repeating the steps 3 to 5 until the global model converges; and each client terminal carries out personalized recommendation of the target object according to the locally stored user ID embedded layer parameters and the latest global model parameters which are transmitted by the server terminal and are except the user ID embedded layer.
2. The method for personalized end-to-end federation recommendation based on user interest domains according to claim 1, wherein in step 1, the server needs to preset a client sampling rate and training parameters, wherein the training parameters include a client local training round and a client local training batch size.
3. The end-to-end federal personalized recommendation method based on user interest domains according to claim 1, wherein the predictive network group is composed of a plurality of parallel multi-layer perceptrons.
4. The end-to-end federal personalized recommendation method based on user interest domains of claim 3, wherein said client classification network comprises a graph neural network, a multi-layer perceptron and a softmax layer; the input of the graphic neural network is a user-object interaction graphic, a user ID embedded representation and a history interaction object ID embedded representation; the user ID embedded representation and the historical interaction article ID embedded representation are respectively generated by a user ID embedded layer and an article ID embedded layer in the recommendation model; and the output of the graph neural network sequentially passes through a multi-layer perceptron and a softmax layer to obtain cluster probability distribution vectors of the client.
5. The end-to-end federation personalized recommendation method based on user interest domains according to claim 4, wherein the number of client class clusters, the number of parallel multi-layer perceptrons in the predicted network group and the class cluster probability distribution vector dimension output by the client classification network are the same.
6. The personalized end-to-end federation recommendation method based on user interest domain according to claim 5, wherein the cluster probability distribution vector output by the client classification network is used as the weight of the output result of the prediction network group, and the weighted average value of the output result of the prediction network group is used as the final output of the recommendation model.
7. The end-to-end federation personalized recommendation method based on user interest domain according to claim 1, wherein the input of the user ID embedding layer in the recommendation model is user ID, the input of the object ID embedding layer is history interactive object ID and target object ID, and the user ID embedding layer and the object ID embedding layer generate user ID embedding representation, history interactive object ID embedding representation and target object ID embedding representation; and splicing the user ID embedded representation, the average value of the historical interactive object ID embedded representation and the target object ID embedded representation to be used as the input of the prediction network group in the recommendation model.
8. The end-to-end federation personalized recommendation method based on user interest domain according to claim 1, wherein step 5 uses federal averaging algorithm to aggregate global model training parameters except for user ID embedded layer sent back by each client.
9. An end-to-end federal personalized recommendation system based on a user interest domain, comprising: the system comprises a server side and a plurality of clients, wherein the server side is used for constructing a global model based on the preset client class cluster number, initializing model parameters and aggregating the global model parameters except the user ID embedding layer after the training of each client side; the global model consists of a client classification network and a recommendation model, wherein the recommendation model comprises a prediction network group, a user ID embedding layer and an article ID embedding layer; the client is used for training the global model by using the local training data set and returning the global model training parameters except the user ID embedding layer to the server;
each client comprises:
the data construction module is used for constructing a local training data set of the client, marking the article actually interacted with the user as a positive sample based on the historical interaction data of the user and the article, setting a negative sampling rate, randomly sampling a plurality of negative samples in a global article set, and ensuring that no intersection exists between the positive sample and the negative sample;
the model training module is used for receiving global model parameters sent by the server side and then loading the latest global model, training the global model based on the training data set obtained by the data construction module, updating local user ID embedded layer parameters and sending the rest parameters back to the server side;
the data transmission module is used for data transmission between the server and the client and comprises model parameters sent by the client to the server after the client receives the model parameters sent by the server and model parameters sent by the client to the server after the local training is finished.
10. The end-to-end federation personalized recommendation system based on user interest domains of claim 9, wherein the client classification network in the global model is configured to generate a class cluster probability distribution vector of the client, the class cluster probability distribution vector is used as a weight of the output result of the prediction network group, and a weighted average of the output results of the prediction network group is used as a final output of the recommendation model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310512758.2A CN116226540B (en) | 2023-05-09 | 2023-05-09 | End-to-end federation personalized recommendation method and system based on user interest domain |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310512758.2A CN116226540B (en) | 2023-05-09 | 2023-05-09 | End-to-end federation personalized recommendation method and system based on user interest domain |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116226540A true CN116226540A (en) | 2023-06-06 |
| CN116226540B CN116226540B (en) | 2023-09-26 |
Family
ID=86587730
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310512758.2A Active CN116226540B (en) | 2023-05-09 | 2023-05-09 | End-to-end federation personalized recommendation method and system based on user interest domain |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116226540B (en) |
Citations (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110874484A (en) * | 2019-10-16 | 2020-03-10 | 众安信息技术服务有限公司 | Data processing method and system based on neural network and federal learning |
| CN112085171A (en) * | 2020-09-15 | 2020-12-15 | 中山大学 | Recommendation method based on clustering multi-entity graph neural network |
| CN112328908A (en) * | 2020-11-11 | 2021-02-05 | 北京工业大学 | Personalized recommendation method based on collaborative filtering |
| US20210073639A1 (en) * | 2018-12-04 | 2021-03-11 | Google Llc | Federated Learning with Adaptive Optimization |
| CN112836130A (en) * | 2021-02-20 | 2021-05-25 | 四川省人工智能研究院(宜宾) | Context-aware recommendation system and method based on federated learning |
| CN113420232A (en) * | 2021-06-02 | 2021-09-21 | 杭州电子科技大学 | Privacy protection-oriented graph neural network federal recommendation method |
| CN113609398A (en) * | 2021-08-17 | 2021-11-05 | 石家庄铁道大学 | Social recommendation method based on heterogeneous graph neural network |
| US20220076169A1 (en) * | 2020-09-08 | 2022-03-10 | International Business Machines Corporation | Federated machine learning using locality sensitive hashing |
| CN114398538A (en) * | 2021-12-08 | 2022-04-26 | 西安电子科技大学 | Cross-domain recommendation method and system for privacy protection, storage medium and computer equipment |
| CN114417172A (en) * | 2022-01-26 | 2022-04-29 | 西安交通大学 | Deep interest evolution recommendation method, device, equipment and storage medium |
| CN114510652A (en) * | 2022-04-20 | 2022-05-17 | 宁波大学 | Social collaborative filtering recommendation method based on federal learning |
| CN114595396A (en) * | 2022-05-07 | 2022-06-07 | 浙江大学 | Sequence recommendation method and system based on federal learning |
| CN114912627A (en) * | 2022-05-10 | 2022-08-16 | 平安科技（深圳）有限公司 | Recommendation model training method, system, computer device and storage medium |
| CN114925270A (en) * | 2022-05-09 | 2022-08-19 | 华南师范大学 | Session recommendation method and model |
| CN114925849A (en) * | 2022-05-09 | 2022-08-19 | 华东师范大学 | Federal learning optimization method on graph data |
| CN115049076A (en) * | 2022-07-13 | 2022-09-13 | 中国科学技术大学 | Iterative clustering type federal learning method based on prototype network |
| CN115391576A (en) * | 2022-07-15 | 2022-11-25 | 浙江工业大学 | Cluster-based federal graph learning method and system |
| CN115409203A (en) * | 2022-07-25 | 2022-11-29 | 中国科学院信息工程研究所 | Federal recommendation method and system based on model independent meta learning |
| CN115409204A (en) * | 2022-07-26 | 2022-11-29 | 中国科学院信息工程研究所 | Federal recommendation method based on fast Fourier transform and learnable filter |
| CN115544873A (en) * | 2022-09-27 | 2022-12-30 | 南京邮电大学 | Training efficiency and personalized effect quantitative evaluation method for personalized federal learning |
| CN115587633A (en) * | 2022-11-07 | 2023-01-10 | 重庆邮电大学 | Personalized federal learning method based on parameter layering |
| US20230023520A1 (en) * | 2020-10-27 | 2023-01-26 | Tencent Technology (Shenzhen) Company Limited | Training Method, Apparatus, and Device for Federated Neural Network Model, Computer Program Product, and Computer-Readable Storage Medium |
| CN115688913A (en) * | 2022-12-29 | 2023-02-03 | 中南大学 | Cloud-side collaborative personalized federal learning method, system, equipment and medium |
| CN115760271A (en) * | 2022-10-28 | 2023-03-07 | 江苏理工学院 | Electromechanical commodity personalized recommendation method and system based on graph neural network |
| CN115796310A (en) * | 2022-11-08 | 2023-03-14 | 北京百度网讯科技有限公司 | Information recommendation method, information recommendation device, information recommendation model training device, information recommendation equipment and storage medium |
| CN115964568A (en) * | 2023-02-10 | 2023-04-14 | 南京邮电大学 | Personalized recommendation method based on edge cache |
-
2023
- 2023-05-09 CN CN202310512758.2A patent/CN116226540B/en active Active
Patent Citations (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20210073639A1 (en) * | 2018-12-04 | 2021-03-11 | Google Llc | Federated Learning with Adaptive Optimization |
| CN110874484A (en) * | 2019-10-16 | 2020-03-10 | 众安信息技术服务有限公司 | Data processing method and system based on neural network and federal learning |
| US20220076169A1 (en) * | 2020-09-08 | 2022-03-10 | International Business Machines Corporation | Federated machine learning using locality sensitive hashing |
| CN112085171A (en) * | 2020-09-15 | 2020-12-15 | 中山大学 | Recommendation method based on clustering multi-entity graph neural network |
| US20230023520A1 (en) * | 2020-10-27 | 2023-01-26 | Tencent Technology (Shenzhen) Company Limited | Training Method, Apparatus, and Device for Federated Neural Network Model, Computer Program Product, and Computer-Readable Storage Medium |
| CN112328908A (en) * | 2020-11-11 | 2021-02-05 | 北京工业大学 | Personalized recommendation method based on collaborative filtering |
| CN112836130A (en) * | 2021-02-20 | 2021-05-25 | 四川省人工智能研究院(宜宾) | Context-aware recommendation system and method based on federated learning |
| CN113420232A (en) * | 2021-06-02 | 2021-09-21 | 杭州电子科技大学 | Privacy protection-oriented graph neural network federal recommendation method |
| CN113609398A (en) * | 2021-08-17 | 2021-11-05 | 石家庄铁道大学 | Social recommendation method based on heterogeneous graph neural network |
| CN114398538A (en) * | 2021-12-08 | 2022-04-26 | 西安电子科技大学 | Cross-domain recommendation method and system for privacy protection, storage medium and computer equipment |
| CN114417172A (en) * | 2022-01-26 | 2022-04-29 | 西安交通大学 | Deep interest evolution recommendation method, device, equipment and storage medium |
| CN114510652A (en) * | 2022-04-20 | 2022-05-17 | 宁波大学 | Social collaborative filtering recommendation method based on federal learning |
| CN114595396A (en) * | 2022-05-07 | 2022-06-07 | 浙江大学 | Sequence recommendation method and system based on federal learning |
| CN114925270A (en) * | 2022-05-09 | 2022-08-19 | 华南师范大学 | Session recommendation method and model |
| CN114925849A (en) * | 2022-05-09 | 2022-08-19 | 华东师范大学 | Federal learning optimization method on graph data |
| CN114912627A (en) * | 2022-05-10 | 2022-08-16 | 平安科技（深圳）有限公司 | Recommendation model training method, system, computer device and storage medium |
| CN115049076A (en) * | 2022-07-13 | 2022-09-13 | 中国科学技术大学 | Iterative clustering type federal learning method based on prototype network |
| CN115391576A (en) * | 2022-07-15 | 2022-11-25 | 浙江工业大学 | Cluster-based federal graph learning method and system |
| CN115409203A (en) * | 2022-07-25 | 2022-11-29 | 中国科学院信息工程研究所 | Federal recommendation method and system based on model independent meta learning |
| CN115409204A (en) * | 2022-07-26 | 2022-11-29 | 中国科学院信息工程研究所 | Federal recommendation method based on fast Fourier transform and learnable filter |
| CN115544873A (en) * | 2022-09-27 | 2022-12-30 | 南京邮电大学 | Training efficiency and personalized effect quantitative evaluation method for personalized federal learning |
| CN115760271A (en) * | 2022-10-28 | 2023-03-07 | 江苏理工学院 | Electromechanical commodity personalized recommendation method and system based on graph neural network |
| CN115587633A (en) * | 2022-11-07 | 2023-01-10 | 重庆邮电大学 | Personalized federal learning method based on parameter layering |
| CN115796310A (en) * | 2022-11-08 | 2023-03-14 | 北京百度网讯科技有限公司 | Information recommendation method, information recommendation device, information recommendation model training device, information recommendation equipment and storage medium |
| CN115688913A (en) * | 2022-12-29 | 2023-02-03 | 中南大学 | Cloud-side collaborative personalized federal learning method, system, equipment and medium |
| CN115964568A (en) * | 2023-02-10 | 2023-04-14 | 南京邮电大学 | Personalized recommendation method based on edge cache |
Non-Patent Citations (4)
| Title |
|---|
| VICTOR CACERES CHIAN; MARCEL HILDEBRANDT; THOMAS RUNKLER;: "Learning Through Structure: Towards Deep Neuromorphic Knowledge Graph Embeddings", 《2021 INTERNATIONAL CONFERENCE ON NEUROMORPHIC COMPUTING (ICNC)》 * |
| 任义;迟翠容;单菁;王佳英;: "基于元学习的推荐算法选择优化框架实证", 计算机工程与设计, no. 06 * |
| 王亚珅: "面向数据共享交换的联邦学习技术发展综述", 无人系统技术, no. 06 * |
| 王蓉;马春光;武朋;: "基于联邦学习和卷积神经网络的入侵检测方法", 信息网络安全, no. 04 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116226540B (en) | 2023-09-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111368210B (en) | Information recommendation method and device based on artificial intelligence and electronic equipment | |
| US11669744B2 (en) | Regularized neural network architecture search | |
| US11604985B2 (en) | Population based training of neural networks | |
| CN114265979B (en) | Method for determining fusion parameters, information recommendation method and model training method | |
| CN102170449A (en) | Web service QoS prediction method based on collaborative filtering | |
| US20220318412A1 (en) | Privacy-aware pruning in machine learning | |
| CN114764471A (en) | Recommendation method, recommendation device and storage medium | |
| CN115905687A (en) | Cold start-oriented recommendation system and method based on meta-learning graph neural network | |
| Geishauser et al. | Dynamic dialogue policy for continual reinforcement learning | |
| Cheng et al. | Dynamic games for social model training service market via federated learning approach | |
| CN116452263A (en) | Information recommendation method, device, equipment, storage medium and program product | |
| Hawkins et al. | Generalizing meanings from partners to populations: Hierarchical inference supports convention formation on networks | |
| CN116090504A (en) | Training method and device for graphic neural network model, classifying method and computing equipment | |
| CN116226540B (en) | End-to-end federation personalized recommendation method and system based on user interest domain | |
| US20230214676A1 (en) | Prediction model training method, information prediction method and corresponding device | |
| CN116340996A (en) | Model fine tuning method for privacy protection and risk control method | |
| CN116932873A (en) | Video account recommending method, device, equipment, storage medium and program product | |
| CN115130536A (en) | Training method of feature extraction model, data processing method, device and equipment | |
| CN113810212B (en) | Root cause positioning method and device for 5G slice user complaints | |
| Capanema et al. | FedPredict: Combining Global and Local Parameters in the Prediction Step of Federated Learning | |
| CN113626564B (en) | Concept label generation method and device, electronic equipment and storage medium | |
| CN116227580A (en) | Autonomous learning method for mobile terminal intelligence | |
| CN114741584A (en) | Method, device and equipment for obtaining user feature representation and target object recommendation | |
| CN117786196A (en) | Object recommendation method, device, electronic equipment and storage medium | |
| CN117009556A (en) | Content recommendation method and related device based on evaluation model |
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 |