Vehicle-mounted terminal travel navigation recommendation method based on federal learning
Through the federated learning of the vehicle terminal travel navigation recommendation method, the problem of user privacy data leakage is solved, and personalized travel recommendations are achieved while protecting user privacy.
Patent Information
- Application Number
- CN202510403922.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-01
- Publication Date
- 2025-07-04
AI Technical Summary
The existing vehicle terminal travel navigation recommendation system needs to upload user personal information during training, resulting in privacy data leakage.
The vehicle terminal travel navigation recommendation method based on federated learning is adopted. The user behavior feature matrix is initialized through the server terminal and sent to the user terminal. The user terminal uses the double-layer convolution probability matrix decomposition model for training, analyzes the feature vector, and returns the weight to the server terminal for aggregation until the model converges. The user terminal uses the trained model to recommend personalized travel methods.
Ensure that user data is always stored locally, avoiding the upload of private data, realizing personalized travel recommendations while protecting user privacy.
Smart Images
Figure CN120256730A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of in-vehicle terminal travel navigation, and more particularly, to a method for recommending in-vehicle terminal travel navigation based on federated learning. Background Art
[0002] In recent years, recommendation systems have been widely used to create personalized prediction models to help individuals identify content of interest. Recommendation systems typically collect users' personal information and interaction records to centrally train recommendation models, and capture users' interest preferences to recommend preferences to users. Currently, recommendation systems are widely applied in other fields such as e-commerce and mainstream media recommendation services. In the recommendation of driving routes for users, users' personal information and preferences need to be uploaded to a central server for training, which largely discloses relevant privacy data such as users' personal information. Therefore, it is necessary to construct a recommendation method for driving routes without disclosing users' private data. Summary of the Invention
[0003] To solve the above problems, the object of the present invention is to provide a technology for recommending in-vehicle terminal travel navigation based on federated learning, aiming to solve the privacy problem of in-vehicle user data in travel route recommendation.
[0004] To achieve the above technical object, the present application provides a method for recommending in-vehicle terminal travel navigation based on federated learning, including the following steps:
[0005] Initialize the user behavior pattern feature matrix through the server terminal and send it to the user terminals participating in the training;
[0006] Through the user terminal, use the double-layer convolutional probability matrix factorization model to train the received user behavior pattern feature matrix, analyze the user behavior feature vector, and send it to the server terminal;
[0007] After the server terminal receives the model weights, through the server terminal, aggregate the weights of the driver behavior feature vectors, update the user behavior pattern feature matrix, and return the updated user behavior pattern feature matrix to the user terminal for continued training until the model converges;
[0008] Through the user terminal, use the trained model to predict the user's personalized travel mode, and based on the server side, according to the obtained starting point and destination of the user's travel, return the personalized travel mode to the user.
[0009] Preferably, in the process of initializing the user behavior pattern feature matrix, obtain the user behavior pattern feature matrix through the user's basic information and user behavior information.
[0010] Preferably, in the process of parsing the user behavior feature vector, based on the double-layer convolutional probability matrix factorization model, by decomposing the user behavior pattern feature matrix into two low-dimensional matrices, the latent feature vector of the user behavior pattern feature matrix is obtained;
[0011] According to the latent feature vector, use the convolutional layer for feature extraction, and perform convolution and pooling operations;
[0012] In the fully connected layer, use the ReLU activation function to extract the high-level features after convolution and pooling, and take the output of the fully connected layer as the factor of the probability matrix algorithm, and return the user behavior feature vector.
[0013] Preferably, in the process of obtaining the latent feature vector, use the GloVe word embedding model to extract features from the user behavior pattern feature matrix to obtain the latent feature vector, where the latent feature vector is expressed as:
[0014]
[0015] In the formula, represents the embedding vector of the i-th word, e i represents the user word embedding matrix of user i, and l is the document length.
[0016] Preferably, in the process of parsing the user behavior feature vector, the user behavior feature vector is parsed by the least squares method, where the parsing process is expressed as:
[0017]
[0018] where, y ui ∈{0,1} is an indicator variable, 1 + λy ui is the confidence weight, α is the trade-off parameter on the regularization term, and I is the identity matrix.
[0019] Preferably, in the process of decomposing the user behavior pattern feature matrix, by obtaining the decomposition loss of the user behavior pattern feature matrix, the user behavior pattern feature matrix is decomposed into two low-dimensional matrices, where the decomposition loss is expressed as:
[0020]
[0021] where, r i,j represents the historical score of user i for item j, represents the embedding vector of user i at the t-th communication round, represents the embedding vector of item j at the t-th communication round.
[0022] Preferably, in the process of predicting the user's personalized travel mode, the probability matrix factorization algorithm is used to perform an inner product on the user feature vector and the user behavior feature matrix, score the user behavior, and obtain the user's personalized travel mode.
[0023] Preferably, in the process of obtaining the user's personalized travel mode, the user's personalized travel mode includes risk avoidance, the shortest journey, or the least number of traffic lights.
[0024] Preferably, in the process of aggregating the weights of the driver behavior feature vectors, the weights of the driver behavior feature vectors are aggregated by obtaining the loss value of the client.
[0025] The present invention discloses the following technical effects:
[0026] The present invention trains the data in multiple rounds to ensure that the original data always remains on the user client and is not uploaded to the central server, fundamentally solving the problem of user privacy leakage. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.
[0028] Figure 1 is a schematic flow chart of the method described in the present invention;
[0029] Figure 2 is a schematic diagram of the model of the vehicle-mounted terminal described in the present invention training the data of the client. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some, rather than all, of the embodiments of the present application. Usually, the components of the embodiments of the present application described and illustrated in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the present application that is required to be protected, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present application.
[0031] As Figure 1-2As shown in the figure, due to the progress of Internet technology and machine learning technology, machine learning can be combined with Internet technology to provide personalized route recommendations for users. Existing recommendation systems need to transmit data from some devices that interact with users through sensors and store a large amount of private user information, which greatly violates the original intention of protecting user privacy. To overcome the problem of the central server collecting user personal privacy and protect user privacy, the present invention proposes a vehicle-mounted terminal travel navigation recommendation technology based on federated learning. Federated learning mainly involves the central server sending model information to the participating clients for training. The clients receive the model and use their own data for training, and then upload the trained local models to the central server. The central server updates the models using an aggregation algorithm until the models converge. Due to the characteristics of federated learning, the participating users no longer transmit data to the central server, but only need to transmit parameters back to the central server.
[0032] The vehicle-mounted terminal travel navigation recommendation technology based on federated learning provided by the present invention specifically includes the following content:
[0033] In the federated learning recommendation system, it is necessary to coordinate the central server and several participating clients to form the entire federated learning training environment. The central server needs to prepare the neural network training model for the initial distribution, and the participating clients need to prepare the corresponding data sets.
[0034] The vehicle-mounted terminal travel navigation recommendation method based on federated learning includes the following steps:
[0035] Step 1: The central server initializes the neural network model and the user behavior feature matrix; the central server selects the user clients participating in local training, and sends the model parameters to the participating clients. The participating clients receive the relevant parameter configurations and use them as the model parameters for federated learning.
[0036] Step 2: The participating clients receive the parameter configurations transmitted from the central server and use the local data sets for training. The local data sets include the user's personal information data set U u and the driver behavior data set V i .
[0037] The above data set, the user personal data set U u includes the user's gender, age, and occupation; the driving behavior data V i includes the historical travel routes, driving preferences, and environmental information. For example, the driving preference data set includes information such as "shortest distance", "fewest red lights", "whether on the highway", and "shortest time" selected by the driver during driving, and other information includes "driving during a specific time period", "weather conditions", "road conditions", "traffic information", etc. The interaction information between the user and the driver behavior data is retained locally on the client.
[0038] Step 3: The client uses the double-layer convolutional probability matrix factorization model to extract text information; the client extracts features of the user's personal information to establish a user feature model, and at the same time extracts the user driving habit features from the driver behavior data to establish a driver driving habit feature model, calculates the weights of the user driving habit features through the model, and uses the probability matrix factorization algorithm to score the results, and scores the items by taking the inner product of the user feature vector and the driver behavior feature vector:
[0039]
[0040] where U u represents the feature vector of user u, and V i represents the feature vector of driver behavior i.
[0041] Step 4: The client uses local data to calculate U based on the least squares method, updates the analytical solution of the user feature vector based on the least squares method, and updates the driver behavior feature vector matrix according to the method.
[0042] Step 5: The client uses the model for training, calculates the weights of the driver behavior mode feature vectors through the user feature matrix vectors, and uploads them to the central server.
[0043] Step 6: The central server receives the weights of the driver behavior mode feature vectors from the client, aggregates the weights, updates the user behavior mode feature matrix, and returns the results to the client.
[0044] Step 7: The client repeats Steps 2 to 5 until the model converges.
[0045] Step 8: According to the destination provided by the user, obtain multiple route maps, check whether there are temporary events in the route maps, and obtain the scores of different route maps. According to the trained route scores, recommend the optimal route (the route with the highest route score) to the user.
[0046] Collect user basic information and driver behavior preference data. The central server initializes the neural network model and the driver behavior data feature matrix and distributes them. Collecting user basic information and driver behavior preference data includes the travel vehicle route, user personal information, environmental information, and travel choices. The user personal information mainly refers to information such as user age, gender, occupation, vehicle category, and vehicle type. The environmental information mainly refers to information such as travel time period, weather conditions, and road conditions. The travel choices mainly refer to information such as "shortest distance", "fewest red lights", "whether on the highway", and "shortest time". The travel time period is divided into "weekend" and "weekday". Furthermore, the central server initializes the double-layer convolutional matrix decomposition model and the driver behavior data and sends them to the participating parties P1, P2, P3...P m ; The participating client receives the training model parameters and data as the model parameters for local model training of the client.
[0047] The client uses the double-layer convolutional probability matrix model to decompose the (user-driver behavior data) matrix into two low-dimensional matrices to obtain the latent feature vectors. The user client receives the information and performs E rounds of local training. During each round of local training, the double-layer convolutional probability matrix model is used for training, which includes a matrix decomposition module, a user comparison module, and a driving behavior comparison module to calculate the loss.
[0048] Assume that for the t-th round of training, user client i receives the driver behavior matrix and the global variable w sent by the central server and uses the model for training. Since the network structures of user information and driver behavior data are similar, in the double-layer convolutional probability matrix model, the GloVe word embedding model is used to extract the features of user information and driver behavior information to obtain the latent feature vectors;
[0049]
[0050] Among them, the vector represents the embedding vector of the i-th word, and e i represents the user word embedding matrix of user i, and l is the document length.
[0051] The client inputs the obtained feature vectors into the model for training. The convolutional layer extracts the features of the document, performs convolutional and pooling operations to form new feature vectors. The ReLU activation function is used in the fully connected layer to extract the high-level features after convolution and pooling. Finally, the output of the fully connected layer is used as the factor of the probability matrix algorithm to return the feature vectors of the user and the driver behavior.
[0052] U i =cnn n (W1,X i )
[0053] Vj = cnn n (W2, Y j )
[0054] U i and V j are the user and driver behavior feature vectors respectively. W1 and W2 represent all weight and bias variables. X i and Y j are the original input documents of user i and driver behavior j.
[0055] Finally, the inner product of the user feature vector and the driver behavior feature matrix is calculated using the probabilistic matrix factorization algorithm, representing the scoring of the corresponding driver behavior for the user, i.e.:
[0056]
[0057] where, U u represents the feature vector of user u, and V i represents the feature vector of driver behavior i.
[0058] After each training, the client uses local data to update the analytical solution of the user feature vector based on the least squares method, i.e.:
[0059]
[0060] where, y ui ∈ {0, 1} is an indicator variable, 1 + λy ui is the confidence weight, α is the trade-off parameter on the regularization term, and I is the identity matrix; the analytical solution of the driver behavior feature vector is updated in the same way.
[0061] Assume that the driving behavior sequence of user i is V i = {V i,1 , V i,2 ....V i,n}, then according to the global feature embedding vector V t , the can be obtained as the driving behavior embedding vector of user i at the t-th round. For the embedding vector of user i, is dotted with to obtain the predicted score. The predicted score is compared with the true score to obtain the required matrix factorization loss:
[0062]
[0063] where, r i，j represents the historical score of user i for item j, represents the embedding vector of user i at the t-th communication round, Denote the embedding vector of item j at the t-th communication round.
[0064] The client returns the weights of the updated driver behavior feature vector to the central server. The central server uses an aggregation method to aggregate the feature vector weights returned by the clients, updates the driver behavior feature matrix, and sends the new parameters back to the clients.
[0065] For each client, the local updated weight is:
[0066]
[0067] w n (t + 1)=w n (t)-ηg n (t)
[0068] where L n is the loss value of the loss client, g n denotes the gradient value of client n, w n (t) represents the model parameter of client n at the t-th round, η represents the learning rate, w n (t + 1) represents the model parameter of client n at the (t + 1)-th round. The server aggregates this, and the global update is defined as:
[0069]
[0070] where, d n denotes the number of samples of client n, d represents the total number of samples of all clients, N represents the client number, and n represents the client number. The global loss function is defined as:
[0071]
[0072] where, Pk represents the number of samples of client k, P represents the total number of samples of all clients, L(θ t-1 ) represents the global loss function, denotes the gradient of client k, K represents the client number, and k represents the client number.
[0073] The user terminal uses the trained model to predict personalized travel modes. The central server returns personalized travel modes to the user according to the starting point and destination of the user's travel, such as routes with {risk avoidance, shortest distance, or fewest traffic lights}.
[0074] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing device produce means for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or means for implementing the functions specified in one block or multiple blocks.
[0075] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality of" means two or more unless otherwise specifically defined.
[0076] Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.
Claims
1. A vehicle terminal travel navigation recommendation method based on federated learning, characterized in that, It includes the following steps: Initialize the user behavior pattern feature matrix through the server terminal and send it to the user terminals participating in the training; Through the user terminal, use the double-layer convolutional probability matrix factorization model to train the received user behavior pattern feature matrix, analyze the user behavior feature vector, and send it to the server terminal; After the server terminal receives the model weights, through the server terminal, aggregate the weights of the driver behavior feature vector, update the user behavior pattern feature matrix, and return the updated user behavior pattern feature matrix to the user terminal for continuous training until the model converges; Through the user terminal, use the trained model to predict the user's personalized travel mode, and based on the server side, according to the obtained starting point and destination of the user's travel, return the personalized travel mode to the user.
2. The method for in-vehicle terminal travel navigation recommendation based on federated learning according to claim 1, wherein: In the process of initializing the user behavior pattern feature matrix, obtain the user behavior pattern feature matrix through the user's basic information and user behavior information.
3. The method for in-vehicle terminal travel navigation recommendation based on federated learning according to claim 2, wherein: In the process of analyzing the user behavior feature vector, based on the double-layer convolutional probability matrix factorization model, by decomposing the user behavior pattern feature matrix into two low-dimensional matrices, obtain the latent feature vector of the user behavior pattern feature matrix; According to the latent feature vector, use the convolutional layer to extract features and perform convolution and pooling operations; In the fully connected layer, use the ReLU activation function to extract the high-level features after convolution and pooling, and use the output of the fully connected layer as the factor of the probability matrix algorithm to return the user behavior feature vector.
4. The method for in-vehicle terminal travel navigation recommendation based on federated learning according to claim 3, wherein: In the process of obtaining the latent feature vector, use the GloVe word embedding model to extract features from the user behavior pattern feature matrix to obtain the latent feature vector, where the latent feature vector is expressed as: In the formula, represents the embedding vector of the i-th word, and e i represents the user word embedding matrix of user i, and l is the document length.
5. The method for in-vehicle terminal travel navigation recommendation based on federated learning according to claim 4, wherein: In the process of analyzing the user behavior feature vector, analyze the user behavior feature vector through the least squares method, where the analysis process is expressed as: where, y ui ∈ {0, 1} is an indicator variable, 1 + λy ui is the confidence weight, α is the trade-off parameter on the regularization term, and I is the identity matrix.
6. The method for in-vehicle terminal travel navigation recommendation based on federated learning according to claim 5, wherein: In the process of decomposing the user behavior pattern feature matrix, by obtaining the decomposition loss of the user behavior pattern feature matrix, decompose the user behavior pattern feature matrix into two low-dimensional matrices, where the decomposition loss is expressed as: Among them, r i,j represents the historical rating of item j by user i, represents the embedding vector of user i at the t-th communication round, represents the embedding vector of item j at the t-th communication round.
7. The method for in-vehicle terminal travel navigation recommendation based on federated learning according to claim 6, wherein: In the process of predicting the user's personalized travel mode, the probability matrix factorization algorithm is used to perform the inner product of the user feature vector and the user behavior feature matrix, score the user behavior, and obtain the user's personalized travel mode.
8. The method for in-vehicle terminal travel navigation recommendation based on federated learning according to claim 7, wherein: In the process of obtaining the user's personalized travel mode, the user's personalized travel mode includes risk avoidance, the shortest journey, or the fewest traffic lights.
9. The method for in-vehicle terminal travel navigation recommendation based on federated learning according to claim 8, wherein: In the process of aggregating the weights of the driver behavior feature vectors, the weights of the driver behavior feature vectors are aggregated by obtaining the loss value of the client.