Federated Cross-view E-commerce Recommendation Method Based on Feature Rescaling

Abstract

The present invention proposes a federated cross-view e-commerce recommendation method based on feature re-adjustment, aiming to improve the privacy protection capability and multi-view data fusion effect of the recommendation system. This method collects data sets of multiple views on the user client and extracts feature representations of users and items through deep neural networks. On this basis, a feature re-adjustment mechanism is used to dynamically adjust the weights of feature channels, enhance the representation capability of important features, and suppress the interference of noise and redundant features. This method ensures that the user's original data remains locally, reduces the risk of data leakage, and realizes user privacy protection. At the same time, by optimizing the loss function and posterior probability calculation, the system can generate a personalized recommendation list. The global model parameter aggregation mechanism in federated learning improves the accuracy of recommendations and the system response speed, and is adapted to the personalized needs of large-scale users.

Description

Technical Field

[0001] The present invention relates to the technical field of e-commerce recommendation methods, and in particular to a federated cross-view e-commerce recommendation method based on feature re-adjustment. Background Art

[0002] Driven by the digital age, the booming development of e-commerce and online service platforms has made recommendation systems increasingly important in daily life. Recommendation systems aim to provide users with personalized recommendations of goods, services or content by analyzing their behavioral data and preferences. With the continuous expansion of the user scale and the surge in data volume, traditional recommendation systems face major challenges in dealing with data privacy and security issues.

[0003] 1. Limitations of Traditional Recommendation Systems

[0004] Traditional recommendation systems mainly rely on centralized data processing models, usually including collaborative filtering, content-based recommendation, and hybrid recommendation technologies. These methods collect historical user behavior data (such as clicks, purchases, and ratings) and analyze them on centralized servers to generate recommendation results.

[0005] However, the shortcomings of centralized recommendation systems are becoming increasingly apparent, mainly in the following aspects:

[0006] Data privacy issues: When users share their personal data, they face the risk of privacy leakage. In traditional recommendation systems, user behavior data must be uploaded to a central server for processing, which exposes users' personal information to potential data security risks. For example, data may be hacked, leading to the leakage of user sensitive information. As users' awareness of privacy protection increases, many users are less willing to share their personal data, which directly affects the effectiveness of the recommendation system.

[0007] Data silos: User data is often isolated in different platforms or systems, forming "data silos". For example, users' browsing and purchasing records on e-commerce platforms cannot be integrated with their behavior data on social media, which limits the performance of the recommendation system. The lack of a comprehensive view of user behavior makes it impossible for the system to fully understand the user's true preferences, reducing the accuracy of recommendations.

[0008] Computing resources and efficiency issues: As the user scale and data volume increase, the demand for computing resources in the recommendation system continues to rise. Centralized processing requires powerful server support. As the data scale grows, the computing and storage pressure on the server also increases. This has, to a certain extent, prolonged the system's response time and affected the user experience.

[0009] 2. Introduction of Federated Learning

[0010] In order to address the privacy issues and data silos in traditional recommendation systems, federated learning has emerged as an emerging distributed machine learning technology. Federated learning fundamentally solves the data privacy problem by training models locally on user devices without sharing the user's original data. Under this framework, the user's personal data is always retained locally, and the model parameters are aggregated on the central server to form a global model. The advantages of this method are mainly reflected in:

[0011] Privacy protection: Federated learning significantly reduces the risk of data leakage. The user's original data never leaves their device, and only the trained model parameters are shared. In this way, users can maintain data privacy while still enjoying personalized recommendation services.

[0012] Data integration and utilization: Federated learning can effectively integrate data from different platforms and devices. Since user behavior data is trained locally on each client, the system can make full use of each user's behavior data to form a more comprehensive user preference model, thereby improving the accuracy of recommendations.

[0013] Resource optimization and efficiency improvement: The distributed computing method of federated learning significantly reduces the pressure on the central server. Each client is responsible for training the local model, and the central server only needs to aggregate and update the model parameters. This method not only saves computing resources, but also improves the system's response speed and scalability.

[0014] 3. Multi-view recommendation system and feature re-adjustment

[0015] Although federated learning has significant advantages in protecting user privacy and processing large-scale data, it still faces certain challenges when dealing with multi-view data. Multi-view recommendation systems can better capture user preferences by integrating information from different data sources (such as user browsing history, purchase history, and social relationships). However, when processing these multi-view data, the performance of the recommendation system is often affected due to the heterogeneity and high dimensionality of the features.

[0016] Feature heterogeneity: In multi-view data, data from different sources may have different feature spaces, and the representation and processing methods of features vary. How to effectively fuse this information is an important challenge.

[0017] High-dimensional feature problem: Multi-view data often contains a large number of high-dimensional features. How to extract effective features and suppress redundant features is the key to improving recommendation performance.

[0018] To solve these problems, the feature rescaling mechanism is introduced into the federated learning framework. By dynamically adjusting the weights of feature channels, feature rescaling can enhance the model's focus on important features while suppressing the influence of noise and irrelevant features. This mechanism not only improves the feature fusion capability of multi-view data, but also improves the robustness of the model.

[0019] 4. Deficiencies of existing technologies

[0020] At present, although federated learning and multi-view recommendation technologies have solved the privacy and data integration problems in traditional recommendation systems to a certain extent, they have not completely overcome the challenges of feature processing and fusion. Existing technologies often show low accuracy and stability when facing noisy data, high-dimensional features and data heterogeneity. Therefore, it is particularly important to develop a federated cross-view recommendation method based on feature rescaling to improve the feature extraction efficiency and noise resistance of multi-view recommendation systems.

[0021] To this end, we design a federated cross-view e-commerce recommendation method based on feature rescaling to solve the above problems. Summary of the invention

[0022] The purpose of the present invention is to provide a federated cross-view e-commerce recommendation method based on feature re-adjustment to address the deficiencies of the prior art and to solve the problems raised in the background technology.

[0023] To achieve the above object, the present invention provides the following technical solution: a federated cross-view e-commerce recommendation method based on feature re-adjustment, comprising the following steps: on each user client, collecting a data set Dn = {(i, U1), (i, U2), ..., (i, U n )};

[0024] Where i is the feature vector of the target item, U i is the user's view feature vector, n is the number of views;

[0025] Each client extracts features from local view data and generates feature representations of items and users through a neural network model: item feature representation y I , used to represent the multidimensional feature vector of goods or services; user feature representation Used to represent the user's behavior characteristics in different views.

[0026] Using the feature re-adjustment mechanism, the feature channel weights are dynamically adjusted. Each feature channel x c It is compressed and stimulated to improve the model's ability to identify important features. The compression process is achieved through the following formula:

[0027] z c=GlobalPooling(x c )

[0028] Among them, z c is the global compression result of the feature channel, x c is the input feature channel, GlobalPooling is the global pooling operation;

[0029] Dynamically adjust the weight s of each feature channel through the excitation operation c , the formula is:

[0030] s=σ(W2·ReLU(W1·z))

[0031] Among them, W1 and W2 are the weight matrices of the fully connected layer, σ is the Sigmoid activation function, and ReLU is the rectified linear unit activation function.

[0032] As a preferred technical solution of the present invention, the system calculates the loss function through the following formula for joint optimization of multi-view data:

[0033]

[0034] Among them, y ij is the interaction label between user i and item j, P(x j |x U ) is the recommendation probability of item j corresponding to user i, N and M are the number of users and items respectively.

[0035] As a preferred technical solution of the present invention, the system calculates the posterior probability P by the following formula: ji To generate a list of recommendations:

[0036]

[0037] Among them, r(y I ,y U ) is the item vector y I and user vector y U is the similarity score between them, and m is the total number of candidate items.

[0038] As a preferred technical solution of the present invention, the user feature vector is calculated by the following formula: With the item feature vector y I The similarity between:

[0039]

[0040] Among them, y I is the feature vector of the item, For user view U iThe cosine similarity is used to evaluate the similarity between two vectors.

[0041] As a preferred technical solution of the present invention, the global model in federated learning performs parameter aggregation through the following weighted average formula:

[0042]

[0043] Among them, W t+1 is the update weight of the global model in the t+1th round, is the local model weight of client i in round t, n i is the number of samples of client i, and n is the total number of samples of all clients.

[0044] As a preferred technical solution of the present invention, the system updates the gradients of the project sub-model and the user sub-model through the following formula:

[0045]

[0046] Among them, W I and are the weights of the project and user sub-models, respectively. is the loss function.

[0047] As a preferred technical solution of the present invention, the item vector and the user vector are re-weighted by the following formula:

[0048]

[0049] in, is the weighted feature channel, s c is the dynamically adjusted feature channel weight.

[0050] As a preferred technical solution of the present invention, the system calculates the matching score between user features and project features by the following formula:

[0051]

[0052] Among them, P(x j |x U ) is the probability that item j is recommended to user U, r(y I ,y U ) is item y I and user y U The similarity score of .

[0053] As a preferred technical solution of the present invention, the system extracts the feature representation of users and items through a deep neural network model f(x). The specific formula is:

[0054] y = f(x) = ReLU(W x + b)

[0055] Among them, x is the input feature vector, W is the weight matrix, b is the bias term, and ReLU is the activation function.

[0056] As a preferred technical solution of the present invention, the client generates a personalized recommendation list for the user through the following formula, and the recommendation probability is calculated based on the posterior probability:

[0057]

[0058] Among them, P(x j |x U ) represents the probability that item j is recommended to user U, and the recommendation ranking is based on the similarity score r(y I ,y U ).

[0059] Compared with the prior art, the present invention has the following beneficial effects: Privacy protection and data integration: This method effectively avoids the leakage of user original data through the distributed training framework of federated learning, ensuring the protection of data privacy. At the same time, user behavior data of different clients are fully utilized, solving the data island problem in the traditional centralized recommendation system and making the recommendation results more accurate.

[0060] Efficient fusion of multi-view data: The feature re-adjustment mechanism is used to dynamically adjust the weights of multi-view features, and the representation ability of important features is enhanced by suppressing irrelevant and redundant features. This mechanism not only improves the robustness of the model, but also significantly improves the processing efficiency of the multi-view recommendation system for high-dimensional data.

[0061] Improved noise resistance and scalability: When faced with noisy data, the feature re-adjustment mechanism can effectively suppress the negative impact of noise features on the model, thereby improving recommendation accuracy and stability. At the same time, this method uses federated learning to reduce the computational burden of the central server, enhance the scalability and response speed of the system, and is suitable for personalized recommendations for large-scale users. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is the basic structure diagram of the federated multi-view framework;

[0063] Figure 2 Schematic diagram of multi-view model structure;

[0064] Figure 3 Schematic diagram of feature rescaling of the compressed excitation module;

[0065] Figure 4 Schematic diagram of the deep semantic matching model. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] The following is combined with Figure 1-Figure 4 With reference to the accompanying drawings and a number of embodiments, specific embodiments of the present invention are described in detail.

[0068] Example 1: Implementation of a federated cross-view recommendation system based on feature rescaling

[0069] This embodiment describes a specific implementation of a federated cross-view recommendation system based on feature re-adjustment. The system consists of multiple user clients and a central server.

[0070] Each user client collects its local user behavior data and project data to form a multi-view dataset Dn = {(i,U1),(i,U2),...,(i,U n )}, where i is the project feature, U i View data for users.

[0071] Data processing: Each client stores user behavior data and project feature data locally. These data include browsing history, purchase history, click behavior, etc.

[0072] The system extracts features from these views and generates feature vectors for each user. and generate feature vector y for the item I .

[0073] Feature rescaling: During the feature extraction process, the system uses a feature rescaling mechanism to enhance the representation capability of important features.

[0074] The specific process is:

[0075] For each feature channel x c To perform global compression:

[0076] z c =GlobalPooling(x c )

[0077] The feature weights are adjusted through the squeeze-excitation network (SENet), and the formula is as follows:

[0078] s=σ(W2·ReLU(W1·z))

[0079] Feature channel x c Adjusted to where s c is the weight coefficient.

[0080] Federated learning training: Each client trains the model locally and optimizes the loss function The formula is:

[0081]

[0082] Among them, y ij is the interaction label between user i and item j, P(x j |x U ) is the recommendation probability.

[0083] Global model aggregation: The client aggregates the weights W of the local model i Upload to the central server, central service

[0084] The weighted average is calculated using the following formula:

[0085]

[0086] Among them, n i is the number of samples of client i, and n is the total number of samples of all clients.

[0087] Personalized recommendation: After the global model is updated, the client generates a personalized recommendation list based on the global model and calculates the similarity between users and items. And sort them, and finally calculate the recommendation probability P(x j |x U ).

[0088] Example 2: Application of multi-view e-commerce recommendation scenario

[0089] This example demonstrates the application scenario of this method in multi-view e-commerce recommendation. Different user behaviors (such as browsing, clicking, and purchasing) are regarded as multiple views, and the system extracts features from each view and makes personalized recommendations based on these features.

[0090] Multi-view data processing: for different user behavior views V1, V2, ..., V n , such as browsing behavior V1 and purchasing behavior V2, the system extracts the user's features in each view And generate feature vector y for the item I .

[0091] Feature fusion and adjustment: The features in each view are dynamically weighted through the feature re-adjustment mechanism.

[0092] Use global pooling to compress each feature channel, and then adjust the weight by the following formula:

[0093] s=σ(W2·ReLU(W1·z))

[0094] Finally, the weighted feature vector is generated

[0095] Recommendation generation: Based on the feature-weighted user view and item view, the system calculates the similarity:

[0096]

[0097] And generate the recommendation list by the following formula:

[0098]

[0099] Example 3: Noise immunity based on feature rescaling

[0100] In practical applications, user data often carries certain noise. This embodiment demonstrates how the feature readjustment mechanism can effectively improve the system's ability to resist noise.

[0101] Noise data introduction: A certain proportion of noise features is added to the system's training data. For example, users' accidental clicks or random browsing behaviors in certain operations will be processed as noise data.

[0102] Feature rescaling and noise suppression: The feature rescaling mechanism dynamically adjusts the feature channel weights through a squeeze-excitation network, and the formula is:

[0103] For example, users' accidental clicks or random browsing behaviors in certain operations will be treated as noise data.

[0104] Feature rescaling and noise suppression: The feature rescaling mechanism dynamically adjusts the feature channel weights through a squeeze-excitation network, and the formula is

[0105] s=σ(W2·ReLU(W1·z))

[0106] The weights of important features c is enhanced, the noise feature is suppressed, and the final output feature vector Reduces the impact of noisy data.

[0107] Surprise test results: In the data sets with different levels of noise added, the system shows high robustness in indicators such as recommendation accuracy and recall rate, proving that feature re-adjustment can effectively suppress the impact of noise data on recommendation effects.

[0108] Example 4: Application in cross-platform e-commerce recommendation scenario

[0109] This example demonstrates the application of this method in cross-platform e-commerce. The behavior data of users on multiple e-commerce platforms are often independent. This method breaks the "data island" through federated learning and realizes cross-platform personalized recommendations.

[0110] Data fusion and processing: On multiple e-commerce platforms, users’ browsing, purchasing, and search behavior data are viewed as different views. The data of each platform is processed locally through the federated learning framework, and after feature extraction, their respective feature vectors are generated.

[0111] Cross-platform model aggregation: After each platform independently trains the local model, it uploads the encrypted model parameters W i To the central server, the central server uses the weighted average formula

[0112]

[0113] Aggregate model parameters to generate a global model.

[0114] Cross-platform recommendations: Through the global model, the platform can generate personalized recommendation lists for users, breaking the data isolation between platforms and enabling users to obtain accurate recommendations across platforms.

[0115] Example 5: Accurate recommendations for specific product categories

[0116] This embodiment demonstrates the application in a recommendation scenario specific to a product category, especially for a scenario where product features vary greatly, such as the difference in recommendations for electronic products and clothing products.

[0117] Product feature processing: For different categories of products (such as electronic products, clothing, etc.), the system adjusts the weights of product features differently. For example, the features of electronic products may include brand, specifications, etc., while the features of clothing include size, color, etc. Through the feature re-adjustment mechanism, the system can dynamically adjust the weights of product category features.

[0118] Recommendation list generation: The system generates feature vector y based on the characteristics of different product categories I , and through the user feature vector y of a specific category U Calculate the similarity, the formula is:

[0119]

[0120] Finally, personalized recommendations are generated based on similarity calculation.

[0121] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in the field. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A federated cross-view e-commerce recommendation method based on feature rescaling, characterized in that: The following steps are included: On each user client, a data set Dn = {(i, U1), (i, U2), ..., (i, U n )}; Where i is the feature vector of the target item, U i is the user's view feature vector, n is the number of views; Each client extracts features from local view data and generates feature representations of items and users through a neural network model: item feature representation y I , used to represent the multidimensional feature vector of goods or services; household Feature Representation Used to represent the user's behavior characteristics in different views. Using the feature re-adjustment mechanism, the feature channel weights are dynamically adjusted. Each feature channel x c It is compressed and stimulated to improve the model's ability to identify important features. The compression process is achieved through the following formula: z c =GlobalPooling(x c ) Among them, z c is the global compression result of the feature channel, x c is the input feature channel, GlobalPooling is the global pooling operation; Dynamically adjust the weight s of each feature channel through the excitation operation c , the formula is: s=σ(W2·ReLU(W1·z)) Among them, W1 and W2 are the weight matrices of the fully connected layer, σ is the Sigmoid activation function, and ReLU is the rectified linear unit activation function.

2. The method for federated cross-view e-commerce recommendation based on feature re-adjustment according to claim 1, characterized in that: The system calculates the loss function using the following formula for joint optimization of multi-view data: Among them, y ij is the interaction label between user i and item j, P(x j |x U ) is the recommendation probability of item j corresponding to user i, N and M are the number of users and items respectively.

3. The method for federated cross-view e-commerce recommendation based on feature re-adjustment according to claim 1, characterized in that: The system calculates the posterior probability P by the following formula ji To generate a list of recommendations: Among them, r(y I ,y U ) is the item vector y I and user vector y U is the similarity score between them, and m is the total number of candidate items.

4. The method for federated cross-view e-commerce recommendation based on feature rescaling according to claim 1, characterized in that: The user feature vector is calculated by the following formula With the item feature vector y I The similarity between: Among them, y I is the feature vector of the item, For user view U i The cosine similarity is used to evaluate the similarity between two vectors.

5. The method for federated cross-view e-commerce recommendation based on feature re-adjustment according to claim 1, characterized in that: The global model in federated learning performs parameter aggregation through the following weighted average formula: Among them, W t+1 is the update weight of the global model in the t+1th round, is the local model weight of client i in round t, n i is the number of samples of client i, and n is the total number of samples of all clients.

6. The method for federated cross-view e-commerce recommendation based on feature rescaling according to claim 1, characterized in that: The system updates the gradients of the project sub-model and the user sub-model using the following formula: Among them, W I and are the weights of the project and user sub-models, respectively. is the loss function.

7. The method for federated cross-view e-commerce recommendation based on feature rescaling according to claim 1, characterized in that: The item vector and user vector are reweighted by the following formula: in, is the weighted feature channel, s c is the dynamically adjusted feature channel weight.

8. The method for federated cross-view e-commerce recommendation based on feature rescaling according to claim 1, characterized in that: The system calculates the matching score between user features and project features using the following formula: Among them, P(x j |x U ) is the probability that item j is recommended to user U, r(y I ,y U ) is item y I and user y U The similarity score of .

9. The method for federated cross-view e-commerce recommendation based on feature rescaling according to claim 1, characterized in that: The system extracts the feature representation of users and items through the deep neural network model f(x). The specific formula is: y = f(x) = ReLU(W x + b) Among them, x is the input feature vector, W is the weight matrix, b is the bias term, and ReLU is the activation function.

10. The method for federated cross-view e-commerce recommendation based on feature re-adjustment according to claim 1, characterized in that: The client generates a personalized recommendation list for the user using the following formula, and the recommendation probability is calculated based on the posterior probability: Among them, P(x j |x U ) represents the probability that item j is recommended to user U, and the recommendation ranking is based on the similarity score r(y I ,y U ).

Images