CN115409203A - Federal recommendation method and system based on model independent meta learning - Google Patents

Federal recommendation method and system based on model independent meta learning Download PDF

Info

Publication number
CN115409203A
CN115409203A CN202210879263.9A CN202210879263A CN115409203A CN 115409203 A CN115409203 A CN 115409203A CN 202210879263 A CN202210879263 A CN 202210879263A CN 115409203 A CN115409203 A CN 115409203A
Authority
CN
China
Prior art keywords
client
model
recommendation model
recommendation
training
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210879263.9A
Other languages
Chinese (zh)
Inventor
吴广君
艾正阳
亓子森
王勇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Institute of Information Engineering of CAS
Original Assignee
Institute of Information Engineering of CAS
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Institute of Information Engineering of CAS filed Critical Institute of Information Engineering of CAS
Priority to CN202210879263.9A priority Critical patent/CN115409203A/en
Publication of CN115409203A publication Critical patent/CN115409203A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • 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
    • 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
    • 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
    • G06F16/958—Organisation or management of web site content, e.g. publishing, maintaining pages or automatic linking

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses a federal recommendation method and a system based on model independent meta learning. The method comprises the following steps: 1) The server selects a plurality of clients and sends the selected recommendation model to each client; 2) Each client divides local data into a support set and a query set; 3) Each client side trains and updates the received recommendation model based on the local support set; 4) Each client verifies the updated recommendation model on the query set and calculates the gradient of the model; 5) The server collects the gradients calculated by each client, updates the recommendation model based on the average gradient and sends the recommendation model to each client for the next round of training; 6) Repeating the steps 3) -5) until a condition is set, and obtaining a common recommendation model by each client; 7) Each client trains the public recommendation model by using local data to obtain respective personalized recommendation models; 8) And the client m inputs the interactive data in the target application scene into the personalized recommendation model to obtain a personalized recommendation result.

Description

Federal recommendation method and system based on model independent meta learning
Technical Field
The invention relates to the field of recommendation systems, in particular to a federal recommendation method and system based on model independent meta learning.
Background
The recommendation system helps users to effectively discover the most useful information or services in the event of information overload by learning the tastes and preferences of the users. Existing deep learning-based recommendation methods typically require the centralized storage of all user-item interaction data to learn deep neural networks and representations of users and items, which means that private data of users needs to be uploaded and aggregated. However, user-item interaction data is highly sensitive to privacy, and its transmission can cause privacy problems and data leakage. Under the pressure of strict Data Protection regulations such as General Data Protection Regulation (GDPR for short), user behavior Data is not permitted to be used without clear permission of a user. Therefore, these centrally trained recommendation models may no longer be applicable in the future.
Federal learning is a machine learning technique that learns intelligent models jointly based on scattered user data. Different from the existing machine learning method based on centralized storage of user data, in federal learning, the user data is locally stored on user equipment so as to protect privacy to the maximum extent. In view of this feature of federal learning, many research efforts combine it with recommendation systems to achieve privacy-preserving recommendations. Federal recommendations train the global recommendation model in a decentralized manner and distribute it to the user's devices for personalized recommendations. For example, in the joint collaborative filtering method FCF, each user device locally computes user and project embedded gradients from personal scores stored on the device. The gradient of item embedding is then uploaded and aggregated to update global item embedding on a central server while updating user embedding locally. And finally, calculating and acquiring a recommendation result of the user according to the dot product of the user embedding and the project embedding.
However, there are still two unresolved problems with the existing federal recommendation methods. First, typical federal learning develops only one common output for all users, and does not adapt the model to each user. This is an important missing feature, especially in view of the heterogeneity of the distribution of historical interaction data among users. The high heterogeneity of the historical interaction data of the user can be embodied in two aspects: (i) User interactions with items depend largely on the user's own interests and preferences, and the interests of different users vary widely, resulting in a wide variety of item types for different user interactions. (ii) The number of interactive items also shows a large difference between users. Therefore, one global recommendation model cannot meet the personalized recommendation requirements of different users. Secondly, the recommendation models applied by the existing federal recommendation method are mostly traditional collaborative filtering and matrix decomposition methods, and the exploration and application of more advanced self-attention mechanism-based models are lacked. In recent years, inspired by the Transformer model for machine translation, the application of a self-attention mechanism to a recommendation system has become a research trend. A recommendation model based on a self-attention mechanism may emphasize truly relevant and important interactions in a sequence while reducing irrelevant interactions. Therefore, compared with the traditional collaborative filtering and matrix decomposition model, the collaborative filtering and matrix decomposition model has higher flexibility and expression capability. The existing federal recommendation method does not explore the implementation and application of a self-attention mechanism in the federal framework, which causes that although the protection of user privacy can be realized, the recommendation precision has a larger gap compared with the most advanced recommendation model.
Disclosure of Invention
In order to overcome the defects of the conventional federal recommendation method, the invention provides a federal recommendation method and a system based on model independent meta learning. Personalization of recommended models is improved by introducing model independent meta learning (MAML) into the federated learning framework. Meanwhile, a recommendation model based on a self-attention mechanism is realized in a federal learning framework, and the recommendation performance is effectively improved.
Specifically, the present invention treats the local training process of each user as a training task, and each training task is composed of a support set and a query set which are mutually disjoint. The recommendation model is trained on the support set and then its loss on the query set is computed. This lost gradient is then uploaded to the central server, where the global model is updated accordingly. And then distributing the updated global model to the user equipment for a new round of training and updating. The goal of the present invention is to find an initial sharing model (i.e., a common initial point learned based on MAML) that new users can easily adapt to their local data set by performing one or several gradient descent steps on their own data. Thus, while the initial model is derived in a distributed manner among all users, the final model implemented by each user is highly personalized based on the user's own data.
The technical content of the invention comprises:
a federal recommendation method based on model independent meta learning comprises the following steps:
the central server randomly extracts part of the user clients for the recommended model training of the turn and sends the global model parameters of the turn to the clients; the recommendation model may employ an industry classical self-attention mechanism-based recommendation model: SASRec.
The client participating in training divides the local data into a support set and a query set according to a set proportion;
the client trains and updates the received recommendation model based on the support set;
verifying the updated recommendation model on the query set and calculating the loss and gradient of the recommendation model;
the server collects and averages the gradients calculated by all the clients participating in training, updates the recommendation model based on the average gradient and performs the next training round until all the clients obtain a common recommendation model;
each client trains the public recommendation model by using local data to obtain an individual recommendation model of each client;
the client m inputs the interactive data in the local target application scene into the personalized recommendation model corresponding to the client m to obtain a corresponding personalized recommendation result; and the client m is the client corresponding to the user m selected by the server.
Further, the serverSelecting a set of user clients
Figure BDA0003763589100000021
A subset of
Figure BDA0003763589100000022
Model parameters of the recommended model are then compared
Figure BDA0003763589100000031
Sent to the subset
Figure BDA0003763589100000032
Each client in (1); wherein, the local model parameters of the client m
Figure BDA0003763589100000033
Further, the method for the client m to divide the local data into the support set and the query set includes:
21 To convert local interaction sequence data of user m
Figure BDA0003763589100000034
Divided into k interactive sessions, i.e.
Figure BDA0003763589100000035
Wherein for the k-th session
Figure BDA0003763589100000036
Figure BDA0003763589100000037
Figure BDA0003763589100000038
For a conversation
Figure BDA0003763589100000039
The total number of time steps (i.e. the session length),
Figure BDA00037635891000000310
is as follows
Figure BDA00037635891000000311
Items interacted with at a time step;
22 The k interactive sessions are segmented into two parts which are not intersected with each other to obtain a support set
Figure BDA00037635891000000312
And query set
Figure BDA00037635891000000313
Wherein (x) i ,y i ) Represents the ith sample, x, in the support set i Representing a sequence of items, y, of an interactive session a except the last item i Representing the last item in the interactive session a; (x' i ,y′ i ) Query the ith sample in set, x' i Sequence of items, y ', representing the last item except in one interactive session b' i Representing the last item in the interactive session b,
Figure BDA00037635891000000314
to a support set
Figure BDA00037635891000000315
The total number of samples in (a) is,
Figure BDA00037635891000000316
as a set of queries
Figure BDA00037635891000000317
Total number of samples in (1).
Further, the method for training and updating the received recommendation model by using the model independent meta learning method comprises the following steps:
31 The recommendation model includes an embedding layer and an attention layer, wherein the attention layer includes an attention mechanism and two layers of feedforward neural networks; the model parameters of the recommendation model include a parameter θ of the embedding layer e And the parameter theta of the attention layer a (ii) a Wherein, theta e ＝{M I ,P}，θ a ＝{W Q ,W K ,W V ,W (1) ,W (2) ,b (1) ,b (2) }；M I Embedding a matrix for an item, P is a learnable position matrix, W Q 、W K 、W V Weight matrices, W, corresponding to the query, key, value in the self-attention mechanism (1) Weight matrix for the first layer feedforward neural network, b (1) Is a bias vector of a first layer feedforward neural network, W (2) Weight matrix for the second layer feedforward neural network, b (2) A bias vector for a second layer feedforward neural network;
32 For supporting sets
Figure BDA00037635891000000318
X in the ith sample i Firstly, converting the sequence into a sequence with a fixed length, then converting each item in the sequence into a one-hot coded vector and embedding the vector and the item into a matrix M I Multiply to obtain x i The corresponding input is embedded into the matrix I; combining the learnable position matrix P with the input embedding matrix I to obtain the output E = I + P of the embedding layer;
33 Input the matrix E into the attention mechanism of the attention layer to obtain an interest representation S of the user m;
34 Two-layer feedforward neural network with S input of interest expression connected in series and Relu as an activation function, and obtaining output FFN (S) = ReLU (SW) (1) +b (1) )W (2) +b (2) ；
35 Processing the FFN (S) layer normalization unit, the residual connection unit and the dropout unit obtained in the step 34) in sequence to obtain a user interest representation S';
36 Predict the preference score of user m for item i based on the obtained S
Figure BDA0003763589100000041
Wherein the content of the first and second substances,
Figure BDA0003763589100000042
is the embedded vector corresponding to item i in the sequence;
37 Rank the items based on the preference scores of user m for the items; selecting K items with the highest preference scores to obtain a recommended item list; calculating the cross entropy loss according to the recommended item list and the actual item list in the training data
Figure BDA0003763589100000043
38 Based on cross entropy loss
Figure BDA0003763589100000044
Calculating gradient and updating local recommendation model to obtain updated recommendation model
Figure BDA0003763589100000045
Further, the cross entropy loss is
Figure BDA0003763589100000046
Where, l is the loss function,
Figure BDA0003763589100000047
is the current local recommendation model of the client m, and (x, y) is an example of a training sample in the support set.
Further, the gradient method for calculating the recommendation model comprises the following steps:
41 Client m inputs training samples in the query set into the updated recommendation model
Figure BDA0003763589100000048
And calculating to obtain cross entropy loss
Figure BDA0003763589100000049
42 Based on cross entropy loss
Figure BDA00037635891000000410
Calculating gradients
Figure BDA00037635891000000411
A federal recommendation system based on model independent meta learning is characterized by comprising a plurality of clients and a server; wherein the content of the first and second substances,
the client is used for dividing local data for training the recommendation model into a support set and a query set, and then training and updating the received recommendation model by adopting a model independent meta-learning method based on the local support set; verifying the updated recommendation model on a query set and calculating the gradient of the recommendation model;
the server is used for sending the selected recommendation model to each client; collecting and averaging the gradients calculated by the clients, updating the recommendation model based on the average gradient, and sending the updated recommendation model to the clients for the next round of training; stopping training when a set condition is reached, and obtaining a common recommendation model by each client;
the client trains the public recommendation model by using local data to obtain an individualized recommendation model corresponding to the client; and inputting the interactive data in the local target application scene into the corresponding personalized recommendation model to obtain a personalized recommendation result.
A storage medium having a computer program stored therein, wherein the computer program is arranged to perform the above method when executed.
An electronic device comprising a memory and a processor, wherein the memory stores a program that performs the above described method.
Compared with the prior art, the invention has the beneficial effects that:
1. the invention provides a personalized federal learning recommendation variant based on model independent meta learning, which can find a model initial point shared among all users, wherein the model initial point is good after each user performs one or more gradient descent steps to update a loss function of the user. According to the method, the generation of the common initial point among different users is realized based on the MAML, and each subsequent user trains a recommendation model based on the common initial point. The role of the federal learning is to protect the privacy data of the user from leaking in the whole learning process by an external framework, namely the invention takes the federal learning as an external framework and MAML as an internal training method.
2. The recommendation model based on the self-attention mechanism is explored and realized in the federal learning framework, so that the recommendation performance of the recommendation model is greatly improved compared with that of a traditional recommendation model-based method.
3. The evaluation results on a plurality of reference data sets show that the method is superior to the existing federal recommendation model, the most advanced level is achieved, and the proposed components play important roles respectively.
Drawings
FIG. 1 is a flow chart of an implementation of the federated recommendation method based on model-independent meta-learning of the present invention.
Fig. 2 is a specific architecture diagram of the federate recommendation method based on model-independent meta learning according to the present invention.
FIG. 3 is a schematic diagram illustrating the effect of various components on the experimental results according to an embodiment of the present invention;
(a) The influence of different components on the experimental result under the ML-1m data set in one embodiment of the invention is shown;
(b) Is the influence of different components under the Beauty data set on the experimental results according to one embodiment of the invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is to be understood that the described embodiments are merely specific embodiments of the present invention, rather than all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments of the invention without making creative efforts, are described.
The object of the invention is meta-training (meta-train) an algorithm
Figure BDA0003763589100000051
The algorithm can quickly train out the algorithm with good performance after being deployed to a new user clientAnd (5) recommending the model. Algorithm
Figure BDA0003763589100000052
Is parameterized, wherein its parameters
Figure BDA0003763589100000053
A set of training tasks is used for updating during meta-training. The training process for each user client is considered a training task, e.g., user m trains the model using his local data is training task m.
As shown in fig. 1, the present invention can be divided into 5 steps in total:
s1: and the central server randomly extracts part of the user clients for the recommendation model training of the round and sends the global model parameters of the recommendation model of the round to the clients.
The method comprises the following specific construction steps:
s1.1: in each iteration, the central server selects a set of user clients
Figure BDA0003763589100000061
A subset of
Figure BDA0003763589100000062
To participate in the training.
S1.2: the central server transmits the model parameters of the current round
Figure BDA0003763589100000063
Send to the collection
Figure BDA0003763589100000064
Each client in the set. For client m, its local model parameters
Figure BDA0003763589100000065
S2: the client participating in training divides the local data set into a support set and a query set according to a set proportion.
The specific construction steps of the data set division are as follows:
s2.1: for each user m, his interaction sequence is represented as:
Figure BDA0003763589100000066
wherein
Figure BDA0003763589100000067
Representing the items that user m interacted with at time step j (i.e., the user's interaction items, e.g., movies in a movie recommendation scene, items in an item recommendation scene). The present invention manually partitions a user interaction sequence into a plurality of interaction sessions according to a particular time idle threshold. Thus, the interaction sequence of user m is divided into k interaction sessions, e.g.
Figure BDA0003763589100000068
These sessions are considered as the basic unit of model training. For conversation
Figure BDA0003763589100000069
It contains the interactive items inside the session:
Figure BDA00037635891000000610
wherein the content of the first and second substances,
Figure BDA00037635891000000611
representing user m at time step
Figure BDA00037635891000000612
The items that are interacted with,
Figure BDA00037635891000000613
representing the length of the session.
S2.2: these sessions are split into two parts that are disjoint to each other: support set
Figure BDA00037635891000000614
And query set
Figure BDA00037635891000000615
Wherein (x) i ,y i ) Represents the ith sample, x, in the support set i Representing a sequence of items, y, of the interactive session except the last item i Representing the last item in the interactive session. (x' i ,y′ i ) The same meaning is applied. That is, the recommendation model is trained and predicted with the last item of the input session as the target and the remaining sequence of items as inputs.
S3: the client trains and updates the received recommendation model based on the support set.
The specific steps of the training based on the support set are as follows:
s3.1: the recommendation model adopted by the invention comprises an embedding layer and an attention layer, wherein the attention layer comprises an attention mechanism and two layers of feedforward neural networks; the model parameters distributed from the central server to the user client in step S1.2 can be divided into the parameters θ of the embedding layer e And the parameter theta of the attention layer a . They in turn contain several different parameters: theta e ＝{M I P } and theta a ＝{W Q ,W K ,W V ,W (1) ,W (2) ,b (1) ,b (2) }. Wherein M is I Embedding a matrix for an item, P is a learnable position matrix, W Q ,W K ,W V Weight matrices, W, corresponding to the query, key, value in the self-attention mechanism (1) As a weight matrix of a first layer feedforward neural network, b (1) Is a bias vector of a first layer feedforward neural network, W (2) As a weight matrix of a second layer feedforward neural network, b (2) Is the bias vector of the second layer feedforward neural network.
S3.2: the input embedding is first generated according to the embedding layer parameters. For supporting set
Figure BDA00037635891000000616
X in the ith sample i Corresponding training session data
Figure BDA0003763589100000071
Firstly, rotate itAs a fixed-length sequence: (s) 1 ,s 2 ,…,s l ) Then each item is converted into a one-hot coded vector and embedded into a matrix M with the item I Multiplication, obtaining the input embedding matrix:
Figure BDA0003763589100000072
finally, combining the learnable position matrix P with the input embedding matrix to obtain the output of the embedding layer: e = I + P.
S3.3: based on the input embedding E in S3.2 and the parameters of the attention layer, the user interest representation is calculated using the self-attention mechanism:
S＝SA(E)＝Attention(EW Q ,EW K ,EW V )
wherein, the standard self-attention model is in the form of:
Figure BDA0003763589100000073
wherein Q, K, V represent query, key, value, respectively, with their respective dimensions being
Figure BDA0003763589100000074
The proposed model used in the present invention uses a structure in which two attention layers are connected in series.
S3.4: after each attention layer, two layers of feedforward neural networks, connected in series, and Relu as the activation function are employed, which may impart model nonlinearity and account for the interaction between different potential dimensions:
FFN(S)＝ReLU(SW (1) +b (1) )W (2) +b (2)
where S is the attention matrix, W, obtained in step S3.3 (1) A weight matrix which is a first layer feedforward neural network; w (2) Is the weight of the second layer feedforward neural network, and the dimensions are all
Figure BDA0003763589100000075
b (1) Biasing for layer one feed-forward neural networkVector, b (2) Is the offset vector of the second layer feedforward neural network, and the dimensions are all
Figure BDA0003763589100000076
S3.5: as the stack of self-attention and feed-forward layers and the network deepens, some problems become more severe, including overfitting, gradient fading, and slower training processes. The invention introduces layer normalization, residual connection and dropout technologies respectively to solve the problems and obtain a user interest representation S'
S′＝S+Dropout(FFN(LayerNorm(S)))
S3.6: to predict the next item, the present invention uses the latent factor model to calculate the user's preference score for item i:
Figure BDA0003763589100000077
wherein R is i Is the user's preference score for item i,
Figure BDA0003763589100000078
is the embedded vector for item i obtained in step S3.2 and S' is the user interest representation obtained in step 3.5.
S3.7: the candidate items are ranked based on the user preference scores for the items calculated at S3.6. And selecting the K items with the highest scores as a recommendation list of the model. And calculating the cross entropy loss according to the recommendation list and the actual item list in the training data. Converging all losses obtained by calculation in the support set, and obtaining the loss of the current model parameter in the support set:
Figure BDA0003763589100000081
where l is the loss function, where,
Figure BDA0003763589100000082
is the current local recommendation model, (x, y) isAnd supporting a training sample in a centralized way.
S3.8: calculate gradients based on the losses calculated at S3.7 and update the local model:
Figure BDA0003763589100000083
wherein, theta m The parameters before the update of the local recommendation model are recommended,
Figure BDA0003763589100000084
for the updated parameters of the local recommendation model,
Figure BDA0003763589100000085
alpha is the local model learning rate for the gradient of the loss calculated in step S3.7.
S4: and verifying the updated recommendation model on the query set and calculating model loss and gradient.
The training based on the query set comprises the following specific steps:
s4.1: updating the local model S3
Figure BDA0003763589100000086
And (3) obtaining the loss of the current model in the query set according to the same step of S3:
Figure BDA0003763589100000087
s4.2: gradient was calculated based on the losses calculated at S4.2:
Figure BDA0003763589100000088
s5: and the server collects and averages the gradients calculated by all the clients participating in training, updates the global model based on the average gradient and performs the next round of training.
The specific steps of the server for updating the global model are as follows:
s5.1: the server collects the gradients calculated in step S4.3 for all local models of the user clients participating in the training;
s5.2: the server maps the global model parameters
Figure BDA0003763589100000089
Is updated to
Figure BDA00037635891000000810
Figure BDA00037635891000000811
Wherein β is the learning rate of the global model.
S5.3: and carrying out the next round of training, resampling the client and updating the global model. And when the global model reaches convergence or the iteration turns reach a set threshold value, stopping training.
According to the recommendation model training method provided by the invention, a common initial point of a model is obtained based on a training mode of the MAML, different users can quickly approach the optimal point of the respective model by executing one-step or several-step gradient descent based on the common initial point by using a small number of samples based on respective local data, so that quick learning and quick adaptation are realized. The introduction of federal learning enables the whole training process not to involve uploading and aggregation of user local data, thereby protecting user privacy. Compared with the traditional federal recommendation system, the invention provides a public recommendation model for all users, provides an initial point which can rapidly approach to the respective optimal models for all users, so that the recommendation model finally obtained by each user can be adapted to the local data of the user to the maximum extent, and the personalized enhancement of the recommendation model is realized.
In the actual application stage, based on the common initial point described in the present invention, the user m randomly extracts a small number of samples based on the local data, and executes the gradient descent update model in one or more steps in the manner of step S3.8, so as to quickly obtain a recommendation model approaching the optimal point, thereby obtaining the personalized recommendation result.
In the embodiment of the invention, the effectiveness and feasibility of the federate recommendation system based on model independent meta learning are verified through experiments, and the performance of the system is verified through two experiments.
The method provided by the invention is compared with the recommended performance of the existing method to verify the effectiveness. As shown in table 1, experimental results on six public data sets in different fields show that the recommended performance of the method proposed by the present invention is consistently better than other baseline models. Particularly, compared with other methods based on the federal learning architecture, the method provided by the invention is obviously improved.
Table 1 is a comparison table of effects
Figure BDA0003763589100000091
Next, the impact of the two-part assembly encompassed by the present invention is verified. The method provided by the invention comprises a joint meta-learning framework (denoted by FML) and a recommendation model (denoted by SA) based on an attention-free mechanism. To verify the validity of each component, multiple experiments were performed on the Beauty and ML-1m data sets, analyzing the contribution of each component. For the whole framework, the invention considers the standard federal learning framework FedAvg. In view of the more extensive framework analysis studies to be carried out by the present invention, the present invention also contemplates a Meta-learning variant of FedAvg, namely FedAvg (Meta). Before testing, fedAvg (Meta) uses the support set of the testing client to update the model initialization received from the server by a one-step stochastic gradient descent, which reflects the essence of Meta-learning, "learning and trimming". Both FedAvg and FedAvg (Meta) used all data of the training client during the training process. For the recommendation model, the invention considers BPR as a comparison, which is a traditional recommendation model based on matrix decomposition. As shown in FIG. 3, the joint meta-learning framework and the self-attention model of the present invention each play an important role in the final recommendation performance.
Finally, it should be noted that: the described embodiments are only some embodiments of the present application and not all embodiments. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments in the present application without making any creative effort belong to the protection scope of the present application.

Claims (9)

1. A federal recommendation method based on model independent meta learning comprises the following steps:
1) The server selects a plurality of clients and sends the selected recommendation model to each client;
2) Each client divides local data used for training the recommendation model into a support set and a query set;
3) Each client side is trained and updates the received recommendation model by adopting a model independent meta learning method based on a local support set;
4) Each client verifies the updated recommendation model on the query set and calculates the gradient of the recommendation model;
5) The server collects and averages the gradients calculated by the clients, updates the recommendation model based on the average gradient and sends the updated recommendation model to the clients for the next round of training;
6) Repeating the steps 3) -5) until a condition is set, and obtaining a common recommendation model by each client;
7) Each client trains the public recommendation model by using local data to obtain an individual recommendation model of each client;
8) The client m inputs interactive data in a local target application scene into the personalized recommendation model corresponding to the client m to obtain a corresponding personalized recommendation result; and the client m is the client corresponding to the user m selected by the server.
2. The method according to claim 1, wherein in step 1), the server selects a set of user clients
Figure FDA0003763589090000011
A subset of
Figure FDA0003763589090000012
Model parameters of the recommended model are then compared
Figure FDA0003763589090000013
Sent to the subset
Figure FDA0003763589090000014
Each client in (1); wherein, the local model parameters of the client m
Figure FDA0003763589090000015
3. The method of claim 2, wherein the method for the client m to divide the local data into the support set and the query set is as follows:
21 To convert local interaction sequence data of user m
Figure FDA0003763589090000016
Divided into k interactive sessions, i.e.
Figure FDA0003763589090000017
Wherein for the k-th session
Figure FDA0003763589090000018
Figure FDA0003763589090000019
Representing user m at time step
Figure FDA00037635890900000110
The items that are interacted with,
Figure FDA00037635890900000111
on behalf of the sessionA length;
22 The k interactive sessions are segmented into two parts which are not intersected with each other to obtain a support set
Figure FDA00037635890900000112
And query set
Figure FDA00037635890900000113
Wherein (x) i ,y i ) Represents the ith sample, x, in the support set i Representing a sequence of items, y, of an interactive session a except the last item i Representing the last item in the interactive session a; (x' i ,y′ i ) Query the ith sample in set, x' i Sequence of items, y ', representing the last item except in one interactive session b' i Representing the last item in the interactive session b,
Figure FDA00037635890900000114
to a support set
Figure FDA00037635890900000115
The total number of samples in (a),
Figure FDA00037635890900000116
as a set of queries
Figure FDA00037635890900000117
Total number of samples in (1).
4. The method of claim 3, wherein the method of training and updating the received recommendation model using a model independent meta learning method is:
31 The recommendation model comprises an embedding layer and an attention layer, wherein the attention layer comprises an attention mechanism and two layers of feedforward neural networks; the model parameters of the recommendation model include a parameter θ of the embedding layer e And the parameter theta of the attention layer a (ii) a Wherein, theta e ＝{M I ,P}，θ a ＝{W Q ,W K ,W V ,W (1) ,W (2) ,b (1) ,b (2) }；M I Embedding a matrix for an item, P is a learnable position matrix, W Q 、W K 、W V Weight matrices, W, corresponding to the query, key, value in the self-attention mechanism (1) As a weight matrix of a first layer feedforward neural network, b (1) Is a bias vector of a first layer feedforward neural network, W (2) As a weight matrix of a second layer feedforward neural network, b (2) A bias vector for a second layer feedforward neural network;
32 For supporting sets
Figure FDA0003763589090000021
X in the ith sample i Firstly, converting the sequence into a sequence with a fixed length, then converting each item in the sequence into a one-hot coded vector and embedding the vector and the item into a matrix M I Multiply to obtain x i The corresponding input is embedded into the matrix I; combining the learnable position matrix P with the input embedding matrix I to obtain the output E = I + P of the embedding layer;
33 Input the matrix E into the attention mechanism of the attention layer to obtain an interest representation S of the user m;
34 Input an interest representation S into a two-layer feedforward neural network connected in series and take Relu as an activation function to obtain an output FFN (S) = ReLU (SW) (1) +b (1) )W (2) +b (2) ；
35 Processing the FFN (S) obtained in the step 34) by a layer normalization unit, a residual error connection unit and a dropout unit in sequence to obtain a user interest representation S';
36 Predicting preference score of user m for item i according to the obtained S
Figure FDA0003763589090000022
Wherein the content of the first and second substances,
Figure FDA0003763589090000023
is the corresponding inlay of item i in the sequenceInputting a vector;
37 Rank the items based on the preference scores of user m for the items; selecting K items with the highest preference scores to obtain a recommended item list; calculating the cross entropy loss according to the recommended item list and the actual item list in the training data
Figure FDA0003763589090000024
38 Based on cross entropy loss
Figure FDA0003763589090000025
Calculating gradient and updating local recommendation model to obtain updated recommendation model
Figure FDA0003763589090000026
5. The method of claim 4, wherein the cross-entropy penalty is
Figure FDA0003763589090000027
Where, l is the loss function,
Figure FDA0003763589090000028
is the current local recommendation model of the client m, and (x, y) is an example of a training sample in the support set.
6. The method according to claim 4, wherein in step 4), the gradient method for calculating the recommendation model is:
41 Client m inputs training samples in the query set into the updated recommendation model
Figure FDA0003763589090000029
And calculating to obtain cross entropy loss
Figure FDA0003763589090000031
42 Based on cross entropy loss
Figure FDA0003763589090000032
Calculating gradients
Figure FDA0003763589090000033
7. A federal recommendation system based on model independent meta-learning is characterized by comprising a plurality of clients and a server; wherein the content of the first and second substances,
the client is used for dividing local data of the training recommendation model into a support set and a query set, and then training and updating the received recommendation model by adopting a model independent meta learning method based on the local support set; verifying the updated recommendation model on a query set and calculating the gradient of the recommendation model;
the server is used for sending the selected recommendation model to each client; collecting and averaging the gradients calculated by the clients, updating the recommendation model based on the average gradient, and sending the updated recommendation model to the clients for the next round of training; stopping training when a set condition is reached, and obtaining a common recommendation model by each client;
the client trains the public recommendation model by using local data to obtain an individualized recommendation model corresponding to the client; and inputting the interactive data in the local target application scene into the corresponding personalized recommendation model to obtain a personalized recommendation result.
8. A server, comprising a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program comprising instructions for carrying out the steps of the method of any one of claims 1 to 6.
9. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the steps of the method according to any one of claims 1 to 6.
CN202210879263.9A 2022-07-25 2022-07-25 Federal recommendation method and system based on model independent meta learning Pending CN115409203A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210879263.9A CN115409203A (en) 2022-07-25 2022-07-25 Federal recommendation method and system based on model independent meta learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210879263.9A CN115409203A (en) 2022-07-25 2022-07-25 Federal recommendation method and system based on model independent meta learning

Publications (1)

Publication Number Publication Date
CN115409203A true CN115409203A (en) 2022-11-29

Family

ID=84157549

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210879263.9A Pending CN115409203A (en) 2022-07-25 2022-07-25 Federal recommendation method and system based on model independent meta learning

Country Status (1)

Country Link
CN (1) CN115409203A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115952838A (en) * 2023-02-03 2023-04-11 黑盒科技(广州)有限公司 Recommendation system generation method and system based on adaptive learning
CN116226540A (en) * 2023-05-09 2023-06-06 浙江大学 End-to-end federation personalized recommendation method and system based on user interest domain
CN117494191A (en) * 2023-10-17 2024-02-02 南昌大学 Point-of-interest micro-service system and method for information physical security

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115952838A (en) * 2023-02-03 2023-04-11 黑盒科技(广州)有限公司 Recommendation system generation method and system based on adaptive learning
CN115952838B (en) * 2023-02-03 2024-01-30 黑盒科技(广州)有限公司 Self-adaptive learning recommendation system-based generation method and system
CN116226540A (en) * 2023-05-09 2023-06-06 浙江大学 End-to-end federation personalized recommendation method and system based on user interest domain
CN116226540B (en) * 2023-05-09 2023-09-26 浙江大学 End-to-end federation personalized recommendation method and system based on user interest domain
CN117494191A (en) * 2023-10-17 2024-02-02 南昌大学 Point-of-interest micro-service system and method for information physical security

Similar Documents

Publication Publication Date Title
Geva et al. Transformer feed-forward layers are key-value memories
CN111931062B (en) Training method and related device of information recommendation model
CN110457589B (en) Vehicle recommendation method, device, equipment and storage medium
US20210326674A1 (en) Content recommendation method and apparatus, device, and storage medium
Zarzour et al. A new collaborative filtering recommendation algorithm based on dimensionality reduction and clustering techniques
CN115409203A (en) Federal recommendation method and system based on model independent meta learning
CN110266745B (en) Information flow recommendation method, device, equipment and storage medium based on deep network
CN111489095B (en) Risk user management method, apparatus, computer device and storage medium
US20230162098A1 (en) Schema-Guided Response Generation
Okon et al. An improved online book recommender system using collaborative filtering algorithm
WO2021135701A1 (en) Information recommendation method and apparatus, electronic device, and storage medium
Wang et al. Attention-based CNN for personalized course recommendations for MOOC learners
CN109313540A (en) The two stages training of spoken dialogue system
Ding et al. Product color emotional design considering color layout
CN114510646A (en) Neural network collaborative filtering recommendation method based on federal learning
CN110502701B (en) Friend recommendation method, system and storage medium introducing attention mechanism
CN113641835A (en) Multimedia resource recommendation method and device, electronic equipment and medium
CN114357201B (en) Audio-visual recommendation method and system based on information perception
CN115409204A (en) Federal recommendation method based on fast Fourier transform and learnable filter
CN113643046B (en) Co-emotion strategy recommendation method, device, equipment and medium suitable for virtual reality
Shi et al. Cross-domain variational autoencoder for recommender systems
CN111949894B (en) Collaborative filtering personalized recommendation method based on multi-space interaction
CN116484105B (en) Service processing method, device, computer equipment, storage medium and program product
CN117035059A (en) Efficient privacy protection recommendation system and method for communication
WO2023087933A1 (en) Content recommendation method and apparatus, device, storage medium, and program product

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