CN116089715A - Sequence recommendation method based on personalized federal technology - Google Patents

Abstract

The invention discloses a sequence recommending method based on personalized federal technology, which comprises the following steps: s1, preprocessing self-interaction data and attribute values of interaction items on local equipment by each client; s2, the client builds a hash index on the local equipment through self-interaction data; s3, the client performs enhancement operation on the training data on the local equipment based on a Bayesian training strategy; s4, combining the hash index with the enhanced data, constructing a local sequence recommendation frame based on multiple tasks on the local equipment by the client, and training a local sequence model in cooperation with the central server until the local sequence model converges; s5, the client acquires parameters of the embedded network of the user, and combines the converged local sequence model output to acquire a preference prediction result at the next moment so as to finish recommendation. The invention can effectively model the dynamic preference of the user on the premise of privacy protection, and has the advantages of expandability, portability and privacy protection.

Description

A sequence recommendation method based on personalized federation technology

Technical Field

The present invention relates to the field of computer technology, and in particular to a sequence recommendation method based on personalized federation technology.

Background Art

The main task of the sequential recommendation system is to mine the user's behavior patterns through a sequence of behaviors with a time sequence relationship, so as to model the client's dynamic preferences and predict the client's item selection at the next moment. Since sequential recommendation is likely to use the collected user data for malicious transactions, many users are worried about their privacy being leaked and are unwilling to share data. This can easily lead to serious problems such as "data islands" and "recommendation barriers". Therefore, the sequential recommendation system based on privacy protection has received great attention and extensive research and discussion from the industry and academia.

At present, the existing privacy-preserving sequential recommendation methods mainly protect data by introducing cryptographic knowledge, such as homomorphic encryption algorithms and differential privacy technologies. In the latest technical research, some researchers use federated learning to protect the privacy of sequential recommendation systems. Although the federated architecture can keep user data local, such distributed data and architecture reduce the effectiveness of the sequential recommendation model. Assuming that clients with certain attributes that can be classified into one category are called a domain, the above problems are specifically reflected in: (1) Inter-domain model imbalance: a single global model cannot adapt to the sequence characteristics of all domains. This may be due to the different attributes of each domain, such as habit bias caused by geographical factors. (2) Intra-domain model imbalance: different clients have different amounts of data. Clients with less data cannot effectively cope with complex sequence models, and are prone to model drift in the aggregation operation of the central server. Therefore, it is urgent to develop a new privacy-preserving sequential recommendation method to solve these problems existing in the current sequential recommendation method.

Summary of the invention

The purpose of the present invention is to provide a sequence recommendation method based on personalized federation technology. The present invention can solve the problems of inter-domain model imbalance and intra-domain model imbalance caused by distributed data and clients, can effectively model the dynamic preferences of users under the premise of privacy protection, and has the advantages of scalability, portability and privacy protection.

The technical solution of the present invention is a sequence recommendation method based on personalized federation technology, comprising the following steps:

Step S1: Each client maintains the data it holds on the local device, and pre-processes its own interaction data and the attribute values of the interaction items to remove the intervening items and abnormal values in the data; at the same time, the data structures of all clients are aligned in the distributed framework;

Step S2: The client constructs a hash index on the local device through its own interactive data and forms a hash storage table; after the hash index is constructed, the client uploads it to the central server;

Step S3: The client performs an enhancement operation on the training data based on the Bayesian training strategy on the local device to obtain enhanced data for self-supervised learning of the local sequence model to enhance the representation ability of the local sequence model;

Step S4: Combining the hash index constructed in step S2 with the enhanced data obtained in step S3, the client constructs a multi-task-based local sequence recommendation framework on the local device (the local sequence recommendation framework can effectively integrate the general sequence encoder to form a local sequence model), and performs distributed training on the local sequence model in collaboration with the central server until the local sequence model converges;

First, the client initializes and pre-trains the local sequence model; second, the local sequence model is uploaded to the central server to perform personalized aggregation operations based on local sensitive hashing; then, the central server sends a specific aggregation model to the client, and after the client receives the aggregation model, it performs the next round of model training tasks until the local sequence model is sent and converged;

Step S5: The client obtains the parameters of the user embedding network, combines the output of the converged local sequence model, obtains the preference prediction result at the next moment, and completes the personalized recommendation for the client.

The prediction results only exist in the client device. The central server does not access any training data source or recommendation results, and there is no communication between clients.

In the aforementioned sequence recommendation method based on personalized federation technology, the overall direction of the preprocessing in step S1 is to organize and record the client's access to all items and represent them through a vector matrix;

Preprocessing specifically includes cleaning up outliers and missing values in interactive data (to ensure data availability), archiving project attributes of projects that have been accessed by the client (such as project type), and archiving user attributes (such as region, age, occupation, hobbies, etc.); and aligning the data structures of project attributes and user attributes on all clients, representing the aligned project attributes and user attributes through vector matrices.

In the aforementioned sequence recommendation method based on personalized federation technology, step 2 includes the following sub-steps:

Sub-step S2.1, the client converts the historical interaction data into a binary feature vector, downloads the relevant hash data from the central server, and constructs a set of hash function clusters on the local device according to the data downloaded from the central server;

Sub-step S2.2, combining the binary feature vector with the hash function cluster to generate a hash index specific to the client and uploading it to a central server;

Sub-step S2.3: The central server receives the hash index of each client to construct a hash storage table.

In the aforementioned sequence recommendation method based on personalized federation technology, step 3 includes the following sub-steps:

Sub-step S3.1, constructing a subsequence set to generate training data, each training data includes a positive sample and a negative sample sequence pair, and the training data is used for Bayesian model optimization;

Sub-step S3.2, constructing enhanced positive samples based on positive samples in the training data;

Sub-step S3.3, constructing enhanced negative samples based on negative samples in the training data;

Sub-step S3.4: Enhanced positive samples and enhanced negative samples are paired to form enhanced data for self-supervised learning of the Bayesian model.

In the aforementioned sequence recommendation method based on personalized federated technology, in the sub-step S3.2, enhanced positive samples are constructed according to the correlation between items and the length of the positive sample sequence, and the degree of correlation between items is determined according to the rule of calculating the area of a triangle by its sides.

In the aforementioned sequence recommendation method based on personalized federated technology, in the sub-step S3.3, enhanced negative samples are constructed according to the length of the negative sample sequence.

In the aforementioned sequence recommendation method based on personalized federation technology, step 4 includes the following sub-steps:

Sub-step S4.1, constructing a sequence recommendation framework based on multi-task learning, the sequence recommendation framework is scalable and portable, the sequence recommendation framework is composed of a user attribute embedding network, a local contrast learning mechanism, an item embedding network and a universal sequence encoder, and after the universal sequence encoder is selected as needed, a local sequence model is formed;

Sub-step S4.2, initially updating the local sequence model by receiving initialization parameters from the central server, locally training the local sequence model using the training data and the enhanced data, and uploading the trained local sequence model to the central server to wait for the personalized aggregation model transmission from the central server;

Sub-step S4.3, the central server receives local sequence models from all clients, and obtains similar users of each client by querying the global hash storage table, and then performs personalized aggregation on all local sequence models according to the query results, so that a specific client corresponds to a specific aggregation model, and sends the aggregation model to the corresponding client;

Sub-step S4.4: The client receives the aggregate model and continues to update the aggregate model with the training data and enhanced data until the local sequence model converges.

In the aforementioned sequence recommendation method based on personalized federation technology, step 5 includes the following sub-steps:

Sub-step S5.1, extracting the parameters of the user attribute embedding network, using them as the feature representation of the user attribute, and performing an inner product operation on the parameters and the output vector of the local sequence model;

Sub-step S5.2: using the inner product result as a recommendation prediction result to predict the user's preference at the next moment; the prediction result is only retained in the client's local device and is not shared with the central server in order to protect user privacy;

Sub-step S5.3: The recommendation system maintains the global hash table online in real time; however, when a client exits the federated training framework, the recommendation system will no longer retain any model sending records related to the client.

Compared with the prior art, the beneficial effects of the present invention are as follows: the personalized federation of the present invention is embodied in the process of collaborative training of sequence encoders by users and central servers. The central server generates a specific model based on the data distribution characteristics, model training conditions, and environmental location of the client, and can effectively integrate the commonly used sequence encoders into the federated distributed framework without obtaining user privacy data. Therefore, the federated sequence recommendation system based on the present invention has scalability, portability, and privacy protection.

Specifically, the present invention introduces local sensitive hashing to design a personalized federated aggregation strategy. This strategy can alleviate the problem that a single global model cannot adapt to the sequence characteristics of all domains. In addition, a data enhancement method based on Bayesian training is designed to improve the contrastive learning strategy, thereby enhancing the representation ability of the local sequence encoder. Furthermore, clients with a small amount of training data can also effectively cope with complex sequence models, and in a distributed scenario, can effectively participate in the model training process of other clients. Therefore, the introduction of personalized and representation-enhanced federated sequence technology can not only adapt to commonly used encoders, so that the general sequence model can effectively model the user's dynamic preferences under the premise of privacy protection, but also when predicting the user's item preferences at the next moment, it does not consume a lot of computing resources, thereby improving the recommendation performance.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flowchart of an implementation of a sequence recommendation method embodiment based on personalized federation technology provided by the present invention;

FIG2 is a schematic diagram of the overall framework of the local multi-task sequence model and its distributed training;

FIG3 is a diagram showing an example of a data enhancement method involved in a sequence recommendation method based on personalized federated technology provided by the present invention.

DETAILED DESCRIPTION

The present invention is further described below in conjunction with the accompanying drawings and embodiments, but they are not intended to limit the present invention.

Embodiment: In the present invention, the data preprocessing operation performed by the client can be selected according to the requirements of a specific scenario. For example, in the POI recommendation scenario, the distance attribute of the POI can be preprocessed. The overall framework of the personalized federation technology is shown in FIG2. The framework diagram shows the details of all the methods involved and the order of precedence.

In this embodiment, a sequence recommendation method based on personalized federation technology includes steps S1-S5 of FIG1 :

Step S1: Each client maintains the data it holds on the local device and pre-processes its own interaction data and attribute values of the interaction items;

Preprocessing includes cleaning up outliers and missing values in the interaction data, archiving project attributes of projects that the client has accessed, and archiving the user attributes of the client itself; and aligning the data structures of project attributes and user attributes on all clients, and representing the aligned project attributes and user attributes through vector matrices.

Step S2: The client converts the historical interaction data into a binary feature vector, and builds a set of hash function clusters according to the assistance instructions of the central server; combines the binary feature vector and the hash function family to generate a specific hash index, and uploads it to the central server to build a global hash storage table. The implementation of the entire step S2 specifically includes the following sub-steps S2.1-S2.3:

S2.1: Assume that there are n projects. For a single user u in federated learning, all its interaction data can be represented as an n-dimensional vector _Ru = ( _v1 , _v2 , …, _vi , …, _vn ); _vi represents the interaction of user u; if _vi = 0, it means that user u has not visited project i; otherwise, _vi = 1 means that user u has visited project i.

S2.2: The client obtains a random vector Q = (q ₁ ,q ₂ ,… _qi ,…,q _n ), where _qi ∈[-1,1] is a random number, and the hash function is defined as follows:

表示向量之间的点乘操作。Among them, the symbol

Represents the dot product operation between vectors.

Based on formula (1), the hash index is constructed as: Dex _u = [G ₁ , G ₂ ,.., _Gi ,…, G _K ], where each _Gi is composed of {g ₁ (R _u ), g ₂ (R _u ),…, g _r (R _u )}, where _Gi represents the hash value generated by a hash bucket, and K such hash values constitute the hash index Dex _u .

S2.3: Upload Dex _u to the central server and make it part of the global hash storage table.

Step S3: The client generates enhanced positive and negative sample pairs by training positive and negative sample pairs. The training data is used for Bayesian model optimization, and the enhanced data is used for self-supervised learning of the Bayesian model. In the data enhancement process, the correlation between items and the sequence length of the training samples are fully considered. An example of data enhancement is shown in Figure 3, where the training sample length l is set to 5. The implementation of the entire step S3 specifically includes the following sub-steps S3.1-S3.4:

S3.1: To achieve BPR optimization, the client constructs a subsequence set to generate training samples; for example, when l=5, the subsequences [ _v1 , _v2 , _v3 , _v4 , _v5 ] and [ _v1 , _v2 , _v3 , _v4 , _v6 ] are regarded as a pair of positive and negative sequences, i.e., training samples, where _v5 and _v6 are regarded as positive labels and negative labels, respectively.

替换v_i，生成增强正样本

其中，v_i≠v_lp，

指代v_i的相关项(关联项)；v_i和

之间的关联程度由这两个兴趣点间的地理属性决定；具体地，本发明充分考虑了用户正样本S_p中v_i的上一个访问序列v_i-1和下一个访问序列v_i+1，以此计算

与v_i的关联程度；首先，如公式(2)所示，基于v_i选择一定地理范围内的POI集合。S3.2: Enhanced positive samples: Take the positive sample S _p = [v ₁ ,v ₂ ,…, _vi ,…,v _lp ] as an example, where v _lp represents the positive label item; randomly select _vi ∈ S _p and use the related item

Replace _vi to generate enhanced positive samples

Among them, _vi ≠v _lp ,

Refers to the related items (associated items) of _vi ; _vi and

The degree of association between them is determined by the geographical attributes between the two points of interest; specifically, the present invention fully considers the previous access sequence _vi-1 and the next access sequence vi ₊₁ of _vi in the user positive sample S _p , and calculates

The degree of association with _vi ; first, as shown in formula (2), a set of POIs within a certain geographical range is selected based on _vi .

则表示与v_i相距在α范围内的所有POI组成的集合。Among them, α is a constant, indicating a threshold of geographical distance; dis() indicates the size of geographical distance;

It represents the set of all POIs within the range of α from _vi .

的相关度：Next, based on the idea of calculating the area of a triangle by its side length, we calculate v _i and

Relevance:

a＝dis(vi ₊₁ ,vi _-1 ) (3)

b ₁ =dis(v _i ,v _i-1 ), c ₁ =dis(v _i ,v _i+1 ) (4)

与v_i-1间的距离，c₂表示

与v_i+1间的距离，p₁和p₂表示三段距离的平均长度，

可看作v_i和

的相关度。Where dis() represents the distance, which can be the cosine distance or the actual physical space distance; a represents the distance between _vi+1 and _vi-1 , _b1 represents the distance between _vi and _vi-1 , _c1 represents the distance between _vi and _vi+1 , _b2 represents

The distance between v _i-1 and c ₂ represents

The distance between and vi ₊₁ , _p1 and _p2 represent the average length of the three distances,

can be regarded as _vi and

's relevance.

In addition, in the operation of enhancing the positive sample, the number of replacement items generated is determined by the sequence length l, and the specific calculation is shown in formula (8).

S3.3: Enhanced negative samples: Take the negative sample _Sn = [ _v1 , _v2 , ..., _v1 , ..., _v1n ] as an example, where _v1n represents a negative label item; select the replacement item _v1 (≠ _v1n ), randomly select the interest point _v′i that the user has not interacted with as the replacement value, and generate the sequence _Sn = [ _v1 , _v2 , ..., _v′i , ..., _v1n ]. The number of replacements follows the calculation rule of formula (9):

Among them, fre _pos represents the number of replacements in the enhanced positive samples, and fre _neg represents the number of replacements in the enhanced negative samples.

S3.4: Enhanced positive samples and enhanced negative samples are paired to form enhanced data for self-supervised learning of the Bayesian model.

Step S4: The client builds a sequence recommendation framework based on multi-task learning. The sequence recommendation framework consists of a user attribute embedding network, a local contrast learning mechanism, an item embedding network, and a universal sequence encoder. After the sequence encoder is selected on demand, a local sequence model is formed; then, the central server combines the global hash table and the client to complete the personalized training of the local sequence model. The implementation of the entire step S4 specifically includes the following sub-steps S4.1-S4.4:

S4.1: The client u represents its own attributes as a one-dimensional vector U _u through a neural embedding network. The client constructs a local multi-task learning model in combination with the required sequence encoder. The multi-task includes recommendation tasks and contrastive learning tasks. For the recommendation task, given the user u, the sequence encoder f(·), the interest point embedding V, the user's sequence Seq ^u,t at timestamp t, the time embedding T, and the context feature I, the hidden layer h _t of the sequence encoder can be obtained. h _t can be expressed by the following formula (10):

h _t =f(V,T,I,Seq ^u,t ; θ _s ) (10)

Among them, _θs represents the parameter set of the encoder.

如下公式(11)所示：According to the user embedding representation U _u , evaluate the preference of user u for item j at timestamp t

As shown in the following formula (11):

如下公式(12)：Next, we apply pairwise Bayesian personalized ranking (BPR) to learn the parameters θ _r of the sequence encoder and embedding network, and the loss function of the recommendation task

The following formula (12):

Here, σ(·) represents the sigmoid() function, and (j, k) is a pair of positive and negative labels in the training subsequence.

S4.2: For contrastive learning tasks, the enhanced sample representation formula (13) and formula (14) are:

被看作一对正样本对，其余的2N条

被视为它的负例；序列对

被编码后的特征为

其对应的负例

被编码后的特征为h′_i。采用多分类交叉熵损失函数(NCE)学习来对比任务，如下公式(15)所示：Among them, Seq _aug-p refers to the enhanced positive samples, and Seq _aug-n refers to the enhanced negative samples;

is considered as a pair of positive samples, and the remaining 2N

is considered as its negative example; the sequence pair

The encoded features are

Its corresponding negative example

The encoded feature is h′ _i . The multi-classification cross entropy loss function (NCE) is used to learn the comparison task, as shown in the following formula (15):

是序列对

的编码特征，

是序列对

的编码特征；符合sim()表示求相似度；

表示对比任务的损失函数。Among them, τ is the temperature coefficient, and an optimal constant value is obtained after parameter adjustment experiments;

is a sequence pair

The encoding features of

is a sequence pair

The encoding features of ; sim() means to find the similarity;

Represents the loss function of the comparison task.

Furthermore, the recommendation task and the contrastive learning task are combined to obtain the final multi-task sequence model. The optimization loss of the multi-task sequence model is expressed as follows (16):

是推荐任务的损失函数，

是对比任务的损失函数，

是最终的多任务损失函数，客户端按上述操作进行本地模型训练。Among them, λ is a constant coefficient used to control the proportion of contrast tasks.

is the loss function for the recommendation task,

is the loss function of the comparison task,

is the final multi-task loss function. The client performs local model training according to the above operations.

其定义为

接着，中心服务器对u生成特定的个性化模型，表示如下公式(17)：S4.3: Client u uploads the local model Θ _u = {W _u } to the central server, where W _u represents the parameters of Θ _u . The central server queries the global hash storage table to find other users in the same hash bucket as u.

It is defined as

Next, the central server generates a specific personalized model for u, expressed as the following formula (17):

指为用户u生成的个性化本地模型。Where Avg(·) refers to the averaging operation, α is a constant used to control the influence of similar and dissimilar users on the local model of user u, _Wu refers to the local model of similar users of user u, _Wz refers to the local model of dissimilar users of user u, ln refers to the number of users similar to user u,

refers to the personalized local model generated for user u.

发送给u，u对

进行新一轮训练并重复上述过程，直到模型收敛。S4.4: The central server will specify

Send to u, u

A new round of training is performed and the above process is repeated until the model converges.

Step S5: Obtain the preference prediction for the next moment based on U _u and the sequence model output h _t .

The implementation of the entire step S5 specifically includes sub-steps S5.1-S5.3:

S5.1: Perform inner product operation on the trained U _u and h _t .

S5.2: Predict the user preference at the next moment based on the inner product results. All recommendation results are only retained on the client's local device and are not shared with the central server.

S5.3: The recommendation system maintains a global hash storage table online in real time. When a client exits training, the system will no longer retain any data records about the client.

The above are only preferred embodiments of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the concept of the present invention should also be regarded as the protection scope of the present invention.

Claims (8)

1. A sequence recommendation method based on personalized federation technology, characterized in that it includes the following steps: Step S1: Each client maintains the data it holds on the local device and pre-processes its own interaction data and attribute values of the interaction items; Step S2: The client constructs a hash index on the local device through its own interaction data and forms a hash storage table; Step S3: The client performs an enhancement operation on the training data based on the Bayesian training strategy on the local device to obtain enhanced data; Step S4: Combining the hash index constructed in step S2 with the enhanced data obtained in step S3, the client constructs a multi-task-based local sequence recommendation framework on the local device, and performs distributed training on the local sequence model in collaboration with the central server until the local sequence model converges; Step S5: The client obtains the parameters of the user embedding network, combines the output of the converged local sequence model, obtains the preference prediction result at the next moment, and completes the personalized recommendation for the client. 2. According to claim 1, a sequence recommendation method based on personalized federation technology is characterized in that: the preprocessing in step S1 includes cleaning up abnormal values and missing values of the interactive data, archiving project attributes of the projects visited by the client, and archiving the user attributes of the client; and performing data structure alignment operations on the project attributes and user attributes on all clients, and representing the aligned project attributes and user attributes through vector matrices. 3. According to the sequence recommendation method based on personalized federation technology in claim 1, it is characterized in that: step 2 comprises the following sub-steps: Sub-step S2.1, the client converts the historical interaction data into a binary feature vector, downloads the relevant hash data from the central server, and constructs a set of hash function clusters on the local device according to the data downloaded from the central server; Sub-step S2.2, combining the binary feature vector with the hash function cluster to generate a hash index specific to the client and uploading it to a central server; Sub-step S2.3: The central server receives the hash index of each client to construct a hash storage table. 4. According to the sequence recommendation method based on personalized federation technology as claimed in claim 1, it is characterized in that: step 3 comprises the following sub-steps: Sub-step S3.1, constructing a subsequence set to generate training data, each training data includes a positive sample and a negative sample sequence pair, and the training data is used for Bayesian model optimization; Sub-step S3.2, constructing enhanced positive samples based on positive samples in the training data; Sub-step S3.3, constructing enhanced negative samples based on negative samples in the training data; Sub-step S3.4: Enhanced positive samples and enhanced negative samples are paired to form enhanced data for self-supervised learning of the Bayesian model. 5. According to claim 4, a sequence recommendation method based on personalized federation technology is characterized in that: in the sub-step S3.2, enhanced positive samples are constructed according to the correlation between items and the length of the positive sample sequence, and the degree of correlation between items is determined according to the rule of calculating the area of a triangle by its sides. 6. A sequence recommendation method based on personalized federation technology according to claim 4, characterized in that: in the sub-step S3.3, enhanced negative samples are constructed according to the length of the negative sample sequence. 7. The sequence recommendation method based on personalized federation technology according to claim 1, characterized in that: step 4 comprises the following sub-steps: Sub-step S4.1, constructing a sequence recommendation framework based on multi-task learning, wherein the sequence recommendation framework is composed of a user attribute embedding network, a local contrastive learning mechanism, an item embedding network and a universal sequence encoder, and after the universal sequence encoder is selected as needed, a local sequence model is formed; Sub-step S4.2, initially updating the local sequence model by receiving initialization parameters from the central server, locally training the local sequence model using the training data and the enhanced data, and uploading the trained local sequence model to the central server; Sub-step S4.3, the central server receives local sequence models from all clients, and obtains similar users of each client by querying the global hash storage table, and then performs personalized aggregation on all local sequence models according to the query results, so that a specific client corresponds to a specific aggregation model, and sends the aggregation model to the corresponding client; Sub-step S4.4: The client receives the aggregate model and continues to update the aggregate model with the training data and enhanced data until the local sequence model converges. 8. The sequence recommendation method based on personalized federation technology according to claim 1, characterized in that: step 5 comprises the following sub-steps: Sub-step S5.1, extracting the parameters of the user attribute embedding network, using them as the feature representation of the user attribute, and performing an inner product operation on the parameters and the output vector of the local sequence model; Sub-step S5.2: using the inner product result as a recommendation prediction result to predict the user's preference at the next moment; Sub-step S5.3: The recommendation system maintains the global hash table online in real time.

Images