A vehicle user preference prediction method based on historical table data

Abstract

This invention discloses a vehicle user preference prediction method based on historical tabular data, primarily addressing the difficulty in obtaining and quantifying subjective vehicle user preferences in resource allocation and service provision within SAGIVNs. The method includes: 1) constructing a vehicle user preference prediction architecture based on DDTs; 2) acquiring a basic publicly available dataset and performing preliminary preprocessing; 3) generating multi-dimensional vehicle user preference data, including latency, stability, and cost, based on the preprocessed dataset; 4) constructing a MO-TabPFN multi-objective regression prediction model comprising an input layer, a preprocessing layer, a model layer, and an output layer, constructing independent regression branches for the three types of preference data and integrating them; and 5) completing the prediction and evaluation process. This invention effectively improves the accuracy of multi-objective preference prediction and adaptability to small-sample scenarios, providing technical support for personalized resource allocation and service provision optimization in integrated air-space-ground vehicle networks.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and further relates to a method for predicting vehicle user preferences, specifically a method for predicting vehicle user preferences based on historical tabular data. It can be used in a vehicle-to-everything (V2X) environment to achieve quantitative representation and accurate prediction of subjective preferences for vehicle users in various computing tasks. Background Technology

[0002] In recent years, in-vehicle services have been undergoing an upgrade from "generalization" to "differentiation." Vehicle users, due to differences in driving habits, interests, and other personal characteristics, exhibit significant differences in preferences for computationally offloaded services such as route planning, content recommendation, and emergency response. Latency sensitivity, service stability, and cost acceptance constitute the three major preference dimensions affecting service experience. Currently, the resource allocation and service supply of Space-Air-Ground Integrated Vehicular Networks (SAGIVNs) still face a supply-demand mismatch dilemma: on the one hand, the subjective preferences of vehicle users are difficult to quantify, lacking effective data representation and mapping methods; on the other hand, traditional service models rely heavily on objective resource requirements (such as bandwidth and computing power) while neglecting the impact of subjective preferences on service satisfaction, resulting in insufficient personalization of resource allocation and difficulty in adapting to the dynamic needs of complex traffic scenarios. However, the scarcity of user preference data further exacerbates this dilemma. Vehicle user service interaction data involves private information such as location, communication status, and consumption habits, and publicly available datasets are limited; simultaneously, preference features have implicit correlations with other features, making direct modeling through traditional statistical methods difficult. These issues make it difficult for SAGIVNs to provide differentiated on-demand services, which not only reduces the service experience for vehicle users but also wastes resources.

[0003] To address the challenges posed by the explosive growth of differentiated in-vehicle services across various environments and needs, SAGIVNs deeply integrates with Digital Twin (DT) technology to provide vehicle users with intelligent, deterministic, and customized differentiated services. In scenarios where computational tasks are offloaded, by constructing Differentiated Digital Twins (DTs) for vehicle users, real-time data synchronization between physical entities and virtual models can be achieved. This not only securely stores historical data such as the vehicle's communication environment, demand preferences, and service effectiveness, protecting user privacy, but also leverages server computing power to perform in-depth data processing, providing data support for preference quantification and prediction. Furthermore, tabular data, as the primary data carrier for vehicle status, network parameters, and service demands in in-vehicle scenarios, presents new challenges to traditional machine learning and deep learning models due to its high dimensionality, uneven sample distribution, and significant small-sample scenarios. Existing prediction models are mostly designed for single objectives, and their adaptability to multi-objective tabular data prediction scenarios is poor. Their prediction accuracy is insufficient in scenarios with small samples and imbalanced samples, which cannot meet the actual needs of accurate prediction of multi-dimensional preferences of vehicle users. This also provides a clear technical requirement for the development of dedicated multi-objective tabular data prediction models.

[0004] In the prior art, Nanjing University of Posts and Telecommunications proposed a heterogeneous network caching decision-making method based on user preference prediction in its patent application "A Heterogeneous Network Caching Decision-Making Method Based on User Preference Prediction" (Patent Application No.: CN202010551762.6, Publication No.: CN111860595A). This method involves scenarios where macro base stations, small base stations, and D2D communication coexist. It predicts user preferences through machine learning, constructs a system cost optimization model, and solves it using a greedy algorithm. While this can reduce the computational complexity and system cost of caching decisions to some extent, it still has several shortcomings in preference prediction: First, this method only predicts user preferences for file content based on historical request data, resulting in a broad preference definition and a lack of effective data representation and mapping methods. Second, this method relies on users' historical request records for preference prediction, failing to consider the scarcity of vehicle user preference data itself, making it difficult to support model training and analysis in data-scarce scenarios. Third, this method uses traditional statistical methods such as K-means clustering and Zipf distribution fitting for preference prediction, without considering the characteristics of the dataset to model the implicit correlation between preference features and other features.

[0005] Existing technologies for predicting vehicle-to-everything (V2X) service preferences and allocating resources still suffer from problems such as inaccurate preference quantification, weak adaptability to sparse data, and insufficient feature association modeling. These issues make it difficult to meet the needs of vehicle users for personalized and highly reliable service experiences in an integrated air-space-ground V2X environment. There is an urgent need for a vehicle user preference prediction method that can achieve accurate preference representation, sparse data enhancement, and multi-feature association modeling, so as to provide technical support for optimizing resource allocation and improving the accuracy of service supply. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing a vehicle user preference prediction method based on historical tabular data. This method aims to solve the problems of difficulty in quantifying and representing subjective preferences of vehicle users in SAGIVNs resource allocation and service provision, as well as the weak predictive ability and low accuracy caused by sparse preference data. The method uses existing public datasets to generate vehicle user preference data in three dimensions: latency, stability, and cost, as historical data. Then, it uses this historical data to achieve multi-objective vehicle user preference prediction. First, a vehicle user preference prediction architecture based on DDTs is constructed to provide a secure and private storage environment for all vehicle user data. Second, the basic public dataset is acquired and pre-processed. Then, multi-dimensional vehicle user preference data is generated based on the public dataset. Next, a multi-objective tabular prior-data fitted network (MO-TabPFN) multi-objective regression prediction model is constructed. Finally, vehicle user preference prediction and evaluation are achieved. This invention effectively overcomes the bottleneck of preference data acquisition through a full-process approach of "data generation - model building - prediction verification," improves the accuracy of multi-objective preference prediction and adaptability to small sample scenarios, and can provide technical support for personalized resource allocation and service supply optimization of integrated air-space-ground vehicle networks.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by the present invention includes the following steps:

[0008] (1) Construct a vehicle user preference prediction architecture based on differentiated digital twins (DDTs), including vehicle user entities, cloud servers, and differentiated vehicle user twins (DDTs of Vehicle Users, DVUs); wherein the vehicle user entities request computing task offloading services and simultaneously synchronize service data to DVUs; the cloud servers are used to provide a reliable storage and interaction environment for DVUs; DVUs store and manage the data generated by the vehicle user entities during the service process in tabular data format;

[0009] (2) Obtain the basic public dataset through the Kaggle open source data platform and perform preliminary preprocessing on it, including deleting the Vehicle_ID column and the Timestamp column from the basic public dataset, and keeping the data in other columns to obtain two preprocessed datasets, denoted as the first feature dataset A and the second feature dataset B.

[0010] (3) Generate multi-dimensional vehicle user preference data based on the pre-processed dataset, including vehicle user latency preference, stability preference and cost preference; among them, based on dataset A, the classification result of the target column Processing_Time_Category is converted into a quantitative value to obtain the latency preference column Preference_time; with dataset A as the main dataset and dataset B as the auxiliary dataset, the target column packet error rate PER of dataset B is used as the stability preference index of vehicle users to obtain the stability preference column Preference_stability; based on the communication transmission rate column Bandwidth_Mbps of dataset A, feature mapping calculation is performed to obtain the cost preference column Preference_cost;

[0011] (4) Construct a MO-TabPFN multi-objective regression prediction model adapted to vehicle user preference prediction. The model includes four layers: input layer, preprocessing layer, model layer and output layer. The MO-TabPFN prediction algorithm is used to fit multi-objective tabular prior data. The TabPFNRegressor is used as the basic regressor to achieve multi-objective regression.

[0012] (5) Construct independent regression branches for the three types of preference data (latency, stability, and cost) obtained in step (3), and finally integrate them into a unified multi-objective regression model to realize vehicle user preference prediction and evaluation.

[0013] Compared with the prior art, the present invention has the following advantages:

[0014] First, in the scenario of offloading computing tasks in the integrated air-space-ground vehicle network, this invention decomposes the vague subjective preferences of vehicle users into three core dimensions that are computable and measurable: latency sensitivity, service stability, and cost acceptability. It establishes an effective data representation and mapping method from user behavior to preference values, thereby breaking through the bottleneck of the traditional broad and difficult-to-quantify preference definition, and providing a precise preference measurement basis for the differentiated service supply and refined resource allocation of the vehicle network.

[0015] Secondly, this invention utilizes publicly available relevant datasets and employs methods such as Spearman rank correlation coefficient mining of feature correlations and SHAP interpreter quantification of feature contribution to explicitly generate three major preference data: latency, stability, and cost. This constructs a structured preference dataset containing 18 feature columns and 10,000 samples, effectively improving the current situation of scarce vehicle user preference data and insufficient sample size. It provides reliable data support for model training and analysis and significantly enhances the usability of the model in small sample scenarios.

[0016] Third, this invention addresses the characteristics of high-dimensionality, small sample size, and uneven distribution of preference data by proposing the MO-TabPFN multi-objective regression algorithm. Through tabular data modeling guided by prior knowledge, it achieves high-precision multi-dimensional preference prediction in small sample scenarios, outperforming traditional machine learning models and general deep learning models. This not only significantly improves the accuracy of vehicle user preference prediction but also enhances the model's generalization ability and robustness in different vehicle networking business scenarios.

[0017] Fourth, because this invention adopts a DDTs (Differentiated Digital Twins) architecture to achieve distributed storage and controllable access to preference data, it can efficiently support the training and inference of preference prediction models while ensuring user data privacy, security and compliance. It can balance privacy and security with business value, and provide a safe and reliable technical guarantee for personalized services of the air-space-ground integrated vehicle network. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the overall implementation of the method of the present invention;

[0019] Figure 2 This is a diagram of the vehicle user preference prediction architecture based on DDTs in this invention;

[0020] Figure 3 A distribution map of vehicle user preference data generated using this invention;

[0021] Figure 4 This is a comparison chart of simulation results between the method of the present invention and existing methods. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0023] Example 1: Refer to Appendix Figure 1 The present invention proposes a method for predicting vehicle user preferences based on historical table data, which specifically includes the following steps:

[0024] Step S1. Construct a vehicle user preference prediction architecture based on Differentiated Digital Twins (DDTs), including vehicle user entities, cloud servers, and differentiated vehicle user twins (DVUs). Each vehicle user entity requests a computational task offloading service and simultaneously synchronizes service data to the DVUs. The cloud server provides a reliable storage and interaction environment for the DVUs. The DVUs store and manage the data generated by the vehicle user entities during the service process in tabular form. In this embodiment, the computational task offloading service includes route planning, content recommendation, traffic congestion prediction, and metaverse entertainment experience. The service data includes at least the communication environment, service requirements, and service effects. The DVUs are virtual mappings of vehicle user entities; each vehicle user entity has a unique digital twin mapping DVU, and all vehicle user entity DVUs are deployed on the cloud server.

[0025] Step S2. Obtain the basic public dataset through the Kaggle open-source data platform and perform preliminary preprocessing on it, including deleting the Vehicle_ID and Timestamp columns from the basic public dataset, and keeping the data in other columns to obtain two preprocessed datasets, denoted as the first feature dataset A and the second feature dataset B.

[0026] Step S3. Generate multi-dimensional vehicle user preference data based on the pre-processed dataset, including vehicle user latency preference, stability preference, and cost preference. Specifically, based on dataset A, the classification result of the target column Processing_Time_Category is converted into a quantitative value to obtain the latency preference column Preference_time. Using dataset A as the primary dataset and dataset B as a secondary dataset, the target column Packet Error Rate (PER) of dataset B is used as the vehicle user's stability preference indicator to obtain the stability preference column Preference_stability. A higher packet error rate indicates poorer link stability, and vehicle users will be more sensitive to stability preferences; that is, a higher packet error rate indicates a higher stability preference. Based on the communication transmission rate column Bandwidth_Mbps of dataset A, feature mapping calculation is performed to obtain the cost preference column Preference_cost. This cost preference is related to the requested bandwidth, and bandwidth is related to the communication transmission rate.

[0027] In this embodiment, the generation process of the aforementioned vehicle user latency preference includes the following steps:

[0028] (3a1) Temporarily encode the target column Processing_Time_Category of dataset A as 0, 1, 2. For each feature, calculate its Spearman correlation coefficient with the temporarily encoded target column, obtain the correlation analysis results between the feature and the target column, and record them.

[0029] (3a2) Feature weight design: Calculate the absolute value of the correlation coefficient of all features based on the correlation analysis results, and normalize it according to its proportion of the absolute value of the correlation coefficient of all features to assign weights to each feature. ;

[0030] (3a3) If the feature is positively correlated with the target, the larger the feature value, the higher the time delay preference; if the feature is negatively correlated with the target, the smaller the feature value, the higher the time delay preference; the formula for calculating time delay preference is as follows:

[0031] ,

[0032] in, To scale the eigenvalues ​​to the 0-1 interval, The correlation direction is indicated by a value greater than 0, which indicates a positive correlation between the feature and the target column, and a value less than 0, which indicates a negative correlation between the feature and the target column.

[0033] (3a4) The obtained time delay preference The mapping is completed based on the original target column Processing_Time_Category classification, and a new target column is generated.

[0034] In this embodiment, the generation process of the aforementioned vehicle user stability preferences includes the following steps:

[0035] (3b1) Multiple regression prediction algorithms are used to predict the target column Preference_stability in dataset B. Based on the prediction results, the optimal algorithm is selected to perform SHAP analysis on the target column, screening out three core features that have the most significant impact on stability preference: Speed_kmh, Signal-to-Noise Ratio (SNR)_dB, and Channel_Interference_% (Channel Interference Percentage). Other non-critical features are removed. The preprocessed dataset B is obtained, containing only the above three core features and the target column Preference_stability. The regression prediction algorithms used in this step of this embodiment include at least linear regression, random forest, XGBoost, and LightGBM.

[0036] (3b2) By comparing the preprocessed dataset B, it is determined that dataset A is missing two features, SNR_dB and Channel_Interference_%, and the missing features are filled in according to the preset rules or fitting formulas. The preprocessing of dataset A is then completed. The specific implementation process includes the following:

[0037] (3b2.1) Based on the original signal strength column in dataset A The SNR_dB feature is calculated using the following formula and added to dataset A:

[0038] ,

[0039] Among them, noise intensity column It is inversely proportional to the communication transmission rate. It is obtained by normalizing the original communication transmission rate column Bandwidth_Mbps in dataset A to the interval [-110, -90] according to the inverse proportional relationship, and the unit is dBm.

[0040] (3b2.2) Based on the five features of signal-to-noise ratio, traffic density, signal strength, noise intensity and speed in dataset A, channel interference data is generated according to the rule that "signal-to-noise ratio and signal strength positively affect channel interference, while traffic density, noise intensity and speed negatively affect channel interference". The data is then normalized to the interval [0,100] to obtain the Channel_Interference_% feature column, and added to dataset A.

[0041] (3b3) Generate vehicle user stability preference data based on the preprocessed dataset A.

[0042] In this embodiment, the generation process of the aforementioned vehicle user cost preference includes the following steps:

[0043] (3c1) Calculate the bandwidth characteristic column Bandwidth_Hz according to Shannon's formula:

[0044] ,

[0045] in, For communication transmission rate;

[0046] (3c2) Based on the bandwidth feature column Bandwidth_Hz, the cost preference column Preference_cost is obtained through reverse normalization.

[0047] Step S4. Construct a MO-TabPFN multi-objective regression prediction model adapted to predict vehicle user preferences. This model consists of four layers: input layer, preprocessing layer, model layer, and output layer. The MO-TabPFN prediction algorithm is used to fit multi-objective tabular prior data, with TabPFNRegressor as the basic regressor to achieve multi-objective regression.

[0048] In this embodiment, the MO-TabPFN multi-objective regression prediction model has the following four layers in its network structure:

[0049] The input layer receives input data as a two-dimensional table of vehicle user historical preferences stored in DVUs, including multiple feature columns and three target columns, where the target columns represent vehicle users' preferences for latency, stability, and cost, respectively.

[0050] The preprocessing layer divides the input data into training and testing sets, and performs OneHot encoding on the classification features.

[0051] The model layer uses TabPFNRegressor as the basic regressor and encapsulates TabPFNRegressor to achieve multi-objective regression. In this embodiment, MultiOutputRegressor is preferably used to encapsulate TabPFNRegressor. An independent TabPFNRegressor regression branch is constructed for each of the three preference objectives of time delay, stability, and cost, and finally integrated into a unified multi-objective regression model.

[0052] The output layer inputs the preprocessed test set feature matrix into the optimal model and outputs the prediction results through the prediction method. The output results include the predicted values ​​of vehicle user latency preference, stability preference, and cost preference.

[0053] Step S5. Construct independent regression branches for the three types of preference data (latency, stability, and cost) obtained in Step S3, and finally integrate them into a unified multi-objective regression model to achieve vehicle user preference prediction and evaluation.

[0054] Example 2: The overall implementation steps of the vehicle user preference prediction method proposed in this example are the same as in Example 1. A specific preferred example is given below to further illustrate the overall implementation idea of ​​this invention:

[0055] Step 1) Construct a vehicle user preference prediction architecture based on DDTs, including vehicle user entities, cloud servers, and DVUs; the vehicle user entities are vehicle user entities in the physical world; the cloud servers have abundant storage, communication, and computing resources, mainly providing a reliable storage and interaction environment for DVUs; the DVUs are virtual mappings of vehicle user entities.

[0056] Step 2) Obtain and pre-process the basic public datasets. The research data comes from two dedicated datasets released by user Ziya07 on the Kaggle open-source data platform: 5G-VANET MEC Dataset and Predictive Packet Error Rate Dataset. The 5G-VANET MEC Dataset is designed for the convergence scenario of 5G-enabled vehicular ad hoc networks (5G-VANET) and mobile edge computing (MEC), and its core service is to meet the real-time vehicular data processing needs of ultra-low latency applications in smart cities. The Predictive Packet Error Rate Dataset is designed for the packet error rate (PER) prediction task in the VANET scenario, focusing on capturing the core parameters affecting real-time data transmission in VANET, and can support the development of machine learning models for communication reliability optimization and packet loss suppression in vehicular networks.

[0057] Step 3) Generate multi-dimensional vehicle user preference data based on the publicly available dataset, including vehicle users' preferences for latency, stability, and cost. The latency preference column in the table data is named "Preference_time," referring to the vehicle user's subjective preference for latency during service requests, with a value ranging from 0.01 to 1. A latency preference closer to 0.01 indicates less concern about latency during service access, while a preference closer to 1 indicates greater sensitivity to latency. The stability preference column in the table data is named "Preference_stability," referring to the vehicle user's subjective preference for stability during service requests, with a value ranging from 0.01 to 1. A stability preference closer to 0.01 indicates less concern about stability during service access, while a latency preference closer to 1 indicates greater sensitivity to stability. The cost preference column in the table data is named "Preference_cost," referring to the vehicle user's subjective preference for cost during service requests, with a value ranging from 0.01 to 1. The closer the cost preference is to 0.01, the less concerned vehicle users are about the cost of enjoying the service; the closer the latency preference is to 1, the more sensitive vehicle users are to the cost of enjoying the service.

[0058] Step 4) Construct the MO-TabPFN multi-objective regression prediction model. Based on the characteristics of the generated vehicle user preference dataset, such as high-dimensional features, small sample size, and unbalanced distribution, a prediction algorithm of multi-objective table prior data fitting network (MO-TabPFN) is proposed. In this step, the MO-TabPFN algorithm is designed with a four-layer structure: input layer, preprocessing layer, model layer, and output layer.

[0059] Step 5) Implement vehicle user preference prediction and evaluation. To verify the prediction performance of the proposed MO-TabPFN on the vehicle user preference dataset, it is compared with traditional machine learning algorithms such as LR, RF, XGBoost, and MO-LightGBM, as well as deep learning machine learning algorithms such as TabTransformer, FT-Transformer, and MLP-Mixer-Tab. RMSE, MAE and The metrics provide a comprehensive evaluation of the model's performance.

[0060] Example 3: The overall implementation steps of the vehicle user preference prediction method proposed in this example are the same as in Example 1 or 2, and will now be referred to... Figure 1-3 The parameter settings are given, and the implementation process of the method of the present invention is described in further detail:

[0061] Step 1. Construct a vehicle user preference prediction architecture based on DDTs.

[0062] The predictive architecture constructed in this invention includes vehicle user entities, cloud servers, and DVUs;

[0063] The vehicle user entity refers to a vehicle user entity in the physical world. The vehicle user entity can request various computing task offloading services, such as route planning, content recommendation, traffic congestion prediction, metaverse entertainment experience, etc. During the service request process, the vehicle user entity can synchronize data related to the entire service, such as communication environment, service requirements, and service effects, to DVUs, providing data support for improving the service experience of the vehicle user entity in the future.

[0064] The cloud server is a powerful data center with abundant storage, communication, and computing resources, primarily providing a reliable storage and interactive environment for DVUs;

[0065] The DVUs are virtual mappings of vehicle user entities. Each vehicle user entity has one and only one DVU, and all DVUs of vehicle user entities are deployed on cloud servers. DVUs map, store, and manage the service data of the corresponding vehicle user entity and other data generated by the vehicle user entity, protecting the privacy and security of the vehicle user entity data. DVUs use tabular data to store and manage the data generated by the vehicle user entity during the service process. In addition, relying on the powerful computing resources of the cloud server, DVUs can perform in-depth data processing, analyze historical data, and predict future data to achieve intelligent decision-making across time and regions.

[0066] Step 2. Obtain and pre-process the basic public dataset.

[0067] In this invention, the research data comes from two dedicated datasets published by user Ziya07 on the Kaggle open-source data platform: 5G-VANET MEC Dataset and Predictive Packet Error Rate Dataset.

[0068] (2.1) Dataset 1

[0069] The 5G-VANET MEC Dataset is designed for 5G-VANET and MEC converged scenarios, primarily serving the real-time vehicle data processing needs of ultra-low latency applications in smart cities. From a data structure and functional perspective, this dataset is specifically designed for machine learning tasks, supporting research directions such as priority data processing, resource allocation optimization, and Quality of Service (QoS) enhancement in 5G-VANET scenarios. The dataset simulates actual service characteristics in 5G-VANET and MEC converged scenarios by simulating vehicle mobility, network conditions, and edge computing parameters. The dataset contains 10,000 data entries, with "Processing_Time_Category" as the target column and the others as feature columns, as detailed in Table 1.

[0070] Table 1. Detailed description of dataset 1

[0071]

[0072] (2.2) Dataset 2

[0073] The Predictive Packet Error Rate Dataset is designed for PER prediction tasks in VANET scenarios. It focuses on capturing the core parameters affecting real-time data transmission in VANET and can support the development of machine learning models for communication reliability optimization and packet loss suppression in vehicular networks. In terms of data content and feature dimensions, the dataset covers the key factors affecting PER in VANET communication scenarios. The dataset contains 10,000 data points, where "PER" is the target column and the others are feature columns, as detailed in Table 2.

[0074] Table 2. Detailed description of Dataset 2

[0075]

[0076] This invention studies the impact of environmental and task characteristics on vehicle user preferences. Therefore, in subsequent studies, we will delete the "Vehicle_ID" and "Timestamp" columns and retain the other columns in the two datasets mentioned above to generate vehicle user preferences.

[0077] Step 3. Generate multi-dimensional vehicle user preference data based on the pre-processed dataset, including vehicle users' preferences for latency, stability, and cost;

[0078] (3.1) Vehicle User Latency Preference Generation Scheme

[0079] In this invention, dataset 1 is used to generate vehicle users' latency preferences. The target column of this dataset is "Processing_Time_Category". We convert the classification results of the target column into specific latency preference values, thus mapping vehicle users' latency during the service experience from qualitative to quantitative to represent their latency preferences. The specific generation process is as follows:

[0080] (3.1.1) Objective: Generate a latency preference column “Preference_time” in the range of 0.01-1 based on dataset 1;

[0081] (3.1.2) Rule: For the classification of the target column "Processing_Time_Category" in the dataset, the corresponding latency preference value is specified as shown in Table 3; the higher the processing latency classification, the longer the processing time, and the less vehicle users care about latency, so the smaller the latency preference; conversely, the lower the processing latency classification, the shorter the processing time, and the more vehicle users are sensitive to latency, so the larger the latency preference.

[0082] Table 3. Correspondence between “Processing_Time_Category” category and “Preference_time” range

[0083]

[0084] (3.1.3) Specific steps:

[0085] Step 1: Correlation analysis. Temporarily encode the target column "Processing_Time_Category" as 0, 1, and 2. For each feature, calculate the Spearman correlation coefficient with the temporarily encoded target column to obtain the correlation analysis between the feature and the target column, and record the correlation strength.

[0086] Step 2: Feature weight design. Based on the correlation coefficients of all features, calculate their absolute values, and then normalize them according to the proportion of each feature's absolute correlation coefficient to the total absolute correlation coefficients of all features, assigning weights to each feature. ;

[0087] Step 3: Calculate the time delay preference. If the feature is positively correlated with the target, the larger the feature value, the higher the time delay preference; if the feature is negatively correlated with the target, the smaller the feature value, the higher the time delay preference. The formula for calculating the time delay preference is as follows:

[0088]

[0089] in, To scale the eigenvalues ​​to the 0-1 interval, The correlation direction is indicated by a value greater than 0, which indicates a positive correlation between the feature and the target column, and a value less than 0, which indicates a negative correlation between the feature and the target column.

[0090] Step 4: Result mapping, mapping the obtained latency preferences to... Based on the original target column "Processing_Time_Category" classification, it is mapped to the corresponding interval according to the above rules, and the result is as follows: Figure 3 As shown in (a), the distribution of the time delay preference values ​​conforms to the distribution of the target column categories in the original dataset. The generated time delay preference column named "Preference_time" replaces the original "Processing_Time_Category" as the new target column.

[0091] (3.2) Vehicle User Stability Preference Generation Scheme

[0092] This invention uses dataset 1 and dataset 2 to generate vehicle users' stability preferences. Dataset 1 is the primary dataset, and dataset 2 is used to simulate the stability preferences in dataset 1. The target column of dataset 2 is "PER," representing the packet error rate. The packet error rate is closely related to link stability; a high packet error rate indicates poor stability, while a low packet error rate indicates good stability. Link stability affects the service experience of vehicle users. When link stability is poor, vehicle users are more sensitive to stability preferences; that is, a higher packet error rate corresponds to a higher stability preference. Conversely, when link stability is good, vehicle users are less concerned about stability; that is, a lower packet error rate corresponds to a lower stability preference. Therefore, in this paper, the target column "PER" is considered as the vehicle user's stability preference, and in subsequent descriptions, "PER" will be replaced with "Preference_stability." The specific generation process of the stability preference column in dataset 1 is as follows:

[0093] (3.2.1) Objective: Based on the analysis of the target column of dataset 2, generate the stability preference column "Preference_stability" in dataset 1, with a range of 0.01-1;

[0094] (3.2.2) Specific steps:

[0095] Step 1: Preprocessing of Dataset 2. First, the four algorithms mentioned in Chapter 2—linear regression, random forest, XGBoost, and LightGBM—are used to predict the target column "Preference_stability" in Dataset 2. The prediction results are shown in Table 4. The table shows that the LightGBM algorithm... From the perspectives of evaluation metrics such as MAE and RMSE, this algorithm is the optimal one. Then, the best prediction model among the four algorithms, LightGBM, is selected to perform SHAP analysis on the target column. The results show that the columns "Speed_kmh", "SNR_dB", and "Channel_Interference_%" have a significant impact on the target column, while the impact of other feature columns is almost negligible. Finally, the important influencing factors affecting the stability preference of vehicle users during service requests are obtained, and other feature columns are removed to obtain the preprocessed dataset 2, which includes the feature columns "Speed_kmh", "SNR_dB", "Channel_Interference_%" and the target column "Preference_stability".

[0096] Table 4. Prediction results of different algorithms for the target column "Preference_stability"

[0097]

[0098] Step 2: Preprocessing of Dataset 1. Compared to the preprocessed Dataset 2, Dataset 1 lacks the feature columns "SNR_dB" and "Channel_Interference_%". Therefore, Dataset 1 will be preprocessed to supplement the two missing features according to certain rules or formulas, providing data support for generating the stability preference of Dataset 1 using Dataset 2 later. First, the feature column "SNR_dB" will be added to Dataset 1. This column is the decibel representation of the signal-to-noise ratio, and its calculation formula is as follows:

[0099]

[0100] in, To ensure the signal strength corresponds to the original feature column "Signal_Strength_dBm" in dataset 1, For noise intensity, according to Shannon's formula, its value is inversely proportional to the communication transmission rate. Therefore, it is obtained by reverse normalizing the original feature column "Bandwidth_Mbps" in Dataset 1 to the interval [-110dBm, -90dBm]. Then, a feature column "Channel_Interference_%" is added to Dataset 1. This column represents the channel interference ratio. Combining the content of Dataset 1, the factors affecting channel interference in this paper are listed as signal-to-noise ratio, traffic density, signal strength, noise intensity, and speed. Among them, signal-to-noise ratio and signal strength have a positive impact on channel interference, while traffic density, noise intensity, and speed have a negative impact on channel interference ratio. Based on the data of these 5 columns in Dataset 1, channel interference is generated according to the above influence rules and normalized to the interval [0, 100] to obtain the channel interference ratio data in Dataset 1.

[0101] Step 3: Generating stability preferences from dataset 1. Through the data preprocessing in the previous two steps, some features in dataset 1 are aligned with all features in dataset 2. First, this paper uses the KNN algorithm to train dataset 2, and obtains the optimal value by adjusting the nearest neighbor number. The algorithm achieves the highest model accuracy when the nearest neighbor number is 7. Then, the model with the optimal nearest neighbor number is used to predict the stability preferences in dataset 1, obtaining the "Preference_stability" target column in dataset 1. The distribution of stability preferences is as follows: Figure 3 As shown in (b);

[0102] (3.3) Vehicle User Cost Preference Generation Scheme

[0103] This invention uses Dataset 1 to generate vehicle users' cost preferences. In this dataset, vehicle users' cost preferences are related to requested bandwidth, and bandwidth is related to communication transmission rate. Therefore, the cost preference column "Preference_cost" of Dataset 1 is obtained based on the feature column "Bandwidth_Mbps" in Dataset 1. The specific generation process is as follows:

[0104] (3.3.1) Objective: Generate a cost preference column “Preference_cost” in the range of 0.01-1 based on dataset 1;

[0105] (3.3.2) Rule: The cost preference of vehicle users is negatively correlated with bandwidth. The more bandwidth a vehicle user requests while enjoying the service, the less concerned the vehicle user is about cost and the more willing they are to sacrifice some cost to obtain higher quality service. In this scenario, the cost preference of vehicle users is lower. Conversely, the less bandwidth a vehicle user requests while enjoying the service, the more sensitive the vehicle user is to cost and the more willing they are to sacrifice other indicators to obtain less expenditure. In this scenario, the cost preference of vehicle users is higher.

[0106] (3.3.3) Specific steps:

[0107] Step 1: Calculate the bandwidth. According to Shannon's formula, the formula for calculating the bandwidth feature column "Bandwidth_Hz" is as follows:

[0108] ;

[0109] in, The communication transmission rate corresponds to the original feature column "Bandwidth_Mbps" in dataset 1. The signal-to-noise ratio in decibels corresponds to the feature column "SNR_dB" obtained during the generation of stability preferences in dataset 1;

[0110] Step 2: Calculate the cost preference. The cost preference is obtained by inversely normalizing the value of the feature column "Bandwidth_Hz" generated in Step 1. Its preference distribution is as follows: Figure 3 As shown in (c).

[0111] Step 4. Construct the MO-TabPFN multi-objective regression prediction model. Based on the characteristics of the generated vehicle user preference dataset, such as high-dimensional features, small sample size, and unbalanced distribution, a prediction algorithm of multi-objective table prior data fitting network (MO-TabPFN) is proposed.

[0112] The MO-TabPFN algorithm, based on the TabPFN algorithm, accurately and efficiently predicts vehicle user preferences in small sample scenarios. The network structure of the MO-TabPFN prediction algorithm includes four layers:

[0113] (1) Input layer: The input data is a two-dimensional table of vehicle user historical preferences stored in DVU, which includes 10,000 data entries with 18 feature columns and 3 target columns. The target columns are vehicle users’ preferences for latency, stability and cost, respectively.

[0114] (2) Preprocessing layer: The training set and the test set are divided into two sets in an 8:2 ratio, and the features of the classification are OneHot encoded;

[0115] (3) Model layer: TabPFNRegressor is used as the basic regressor, which is optimized for tabular data regression tasks and has the characteristics of small sample size and high efficiency. At the same time, MultiOutputRegressor is used to encapsulate TabPFNRegressor to realize multi-objective regression. An independent TabPFNRegressor regression branch is built for each of the three preference objectives of time delay, stability and cost. Finally, they are integrated into a unified multi-objective regression model to solve the problem that a single model cannot output multi-dimensional continuous values ​​at the same time.

[0116] (4) Output layer: Input the preprocessed test set feature matrix into the optimal model, and output the prediction results through the prediction method. The output dimension is 2000×3, where each row corresponds to 1 test sample, and the three columns correspond to the predicted values ​​of vehicle user latency preference, stability preference and cost preference, respectively.

[0117] Step 5. Implement vehicle user preference prediction and evaluation. To verify the prediction performance of the proposed MO-TabPFN on the vehicle user preference dataset, it is compared with traditional machine learning algorithms such as LR, RF, XGBoost, and MO-LightGBM, as well as deep learning machine learning algorithms such as TabTransformer, FT-Transformer, and MLP-Mixer-Tab. RMSE, MAE and The metrics provide a comprehensive evaluation of the model's performance.

[0118] The technical effects of the present invention will be further explained below with reference to simulation experiments.

[0119] 1. Simulation conditions

[0120] This invention utilizes the AutoDL platform to train and infer a vehicle user preference prediction model, and its performance has been validated through experiments. This platform provides flexible and configurable computing resources and a standardized development environment, compensating for insufficient local hardware computing power, resolving inconsistencies in environment configuration, and improving computational efficiency during the experimental process.

[0121] 2. Simulation Content

[0122] In the above scenarios, the simulation results are shown in Table 5 and Figure 4 As shown, where:

[0123] Table 5 shows the evaluation index data of the algorithm of the present invention and other algorithms;

[0124] Figure 4 (a) shows the performance of the algorithm of this invention and other algorithms in predicting the "Preference_time" column at different levels;

[0125] Figure 4 (b) shows the performance of the algorithm of this invention and other algorithms in predicting the "Preference_stability" column at different levels;

[0126] Figure 4 (c) in the figure represents the performance of the algorithm of this invention and other algorithms in predicting the "Preference_cost" column at different levels;

[0127] The algorithm described in this invention is the MO-TabPFN algorithm;

[0128] The other algorithms mentioned are traditional machine learning algorithms such as LR, RF, XGBoost, and MO-LightGBM, as well as deep learning machine algorithms such as TabTransformer, FT-Transformer, and MLP-Mixer-Tab.

[0129] 3. Simulation Result Analysis

[0130] Table 5 Evaluation metrics for predictions by different algorithms

[0131]

[0132] Table 5 shows the prediction results of different algorithms. From the evaluation metrics, the traditional LR method performs the worst. The scores were all below 0.9 across the three preference dimensions of latency, stability, and cost. The LR algorithm has the lowest RMSE and MAE compared to other algorithms, indicating that the LR method is difficult to adapt to the nonlinear characteristics of preference prediction. Ensemble tree models such as RF, XGBoost, and MO-LightGBM achieve performance improvements compared to LR. The performance improved to over 0.95 because ensemble tree models capture non-linear features through the divide-and-conquer strategy of decision trees, while gradient boosting algorithms like XGBoost and LightGBM further reduce the risk of overfitting, resulting in superior performance across various metrics. TabTransformer, FT-Transformer, and other tabular deep models leverage feature encoding and attention mechanisms to further uncover long-range feature dependencies, while MLP-Mixer-Tab further enhances feature interactions through hybrid networks. These improvements demonstrate the superior performance of deep learning algorithms. All improved to above 0.98, and the error continued to decrease compared to traditional methods; while MO-TabPFN, as a dedicated multi-objective tabular model, performed best in the prediction results. The latency, stability, and cost scores were 0.9963, 0.9985, and 0.9999, respectively. The RMSE and MAE values ​​were 0.9986, and several RMSE and MAE indicators dropped to below 0.01, indicating that the algorithm has strong adaptability in multi-dimensional preference prediction scenarios compared with other comparison algorithms.

[0133] Figure 4 (a) shows the performance of different algorithms in predicting the "Preference_time" column at different levels. The performance of each method at different levels... The results show that high-level, small-sample scenarios are the key factor determining the performance of each model. Traditional machine learning models have poor adaptability to this prediction task, with the LR method only achieving positive results at the medium level. Both low and high levels are negative; the RF method only achieves a small positive value at low levels. The results were negative at both medium and high levels; while XGBoost showed good fitting performance at low and medium levels, it still failed to achieve effective prediction in high-level small-sample scenarios; although MO-LightGBM achieved effective prediction in high-level small-sample scenarios, its... With a score of only 0.293371, the predictive performance is limited. However, the three types of deep table models outperform traditional machine learning models, with TabTransformer being the best performing model among them, especially in low-level... The value is 0.944993, indicating that high prediction levels can be achieved even at medium and high levels. Furthermore, the MO-TabPFN method proposed in this paper achieves the highest prediction value across all levels. The values ​​were 0.967065 for low, 0.968546 for medium, and 0.805408 for high levels. It not only achieved a prediction accuracy close to 0.97 at low and medium levels, but also achieved a performance breakthrough in high-level scenarios with small sample sizes. The MO-TabPFN outperforms other comparative methods, and its prediction performance is relatively balanced across different levels. Compared to traditional machine learning models and deep tabular models, MO-TabPFN demonstrates stronger feature extraction capabilities for tabular data and better generalization ability in imbalanced scenarios, validating its effectiveness in the "Preference_time" level prediction task and proving to be a superior solution for predicting imbalanced tabular data.

[0134] Figure 4 (b) shows the performance of different algorithms in predicting the "Preference_stability" column under different levels and different numbers of test samples in the same batch. Among traditional machine learning models, the LR method is the worst performing method, achieving a positive result only at the medium level. Neither low nor high levels can fit the characteristic patterns of stability preferences; RF achieves effective predictions at low, medium, and high levels, with each level... All scores remained above 0.69, demonstrating basic tabular data prediction capabilities; XGBoost outperformed LR and RF at all levels. All achieved a score of 0.86, reaching 0.913118 in high-level, large-sample scenarios, demonstrating the fitting advantage of ensemble tree models; MO-LightGBM is the best among traditional machine learning models, performing well in low, medium, and high-level scenarios. The prediction performance of "Preference_stability" was 0.947528, 0.931806, and 0.969795 respectively, reaching or even surpassing the performance of the three types of deep table models. This is because the prediction scenario for "Preference_stability" is simpler than that for "Preference_time," with fewer influencing factors on the target variable. The feature extraction and fitting capabilities of the MO-LightGBM algorithm can capture the data patterns of stability preferences, thus achieving similar performance levels even without the complex feature modeling mechanisms of deep models. Deep table models have designed feature extraction mechanisms for tabular data, and their prediction results have achieved high accuracy. Among them, TabTransformer performs particularly well in the mid-level. The value is 0.951660, which is the best result for this level among the comparative methods of deep table models. Furthermore, the MO-TabPFN proposed in this paper achieved the highest value across all levels. The prediction accuracy values ​​for low, medium, and high levels are 0.980122, 0.970709, and 0.988990, respectively. It not only achieves a prediction accuracy close to 0.99 in high-level scenarios with large samples, surpassing all traditional machine learning models and deep table models, but also achieves performance breakthroughs in low-level scenarios with small samples, demonstrating strong generalization ability for small samples and relatively balanced prediction performance across all levels. Even in prediction tasks with simpler scenarios and fewer influencing factors, such as "Preference_stability," MO-TabPFN outperforms traditional models and deep table models, validating its superior ability in modeling tabular data features. It can also better adapt to tabular data prediction tasks with different complexities and sample size distributions.

[0135] Figure 4 (c) shows the performance of different algorithms in predicting the "Preference_cost" column under different levels and different numbers of test samples in the same batch. The prediction results of the four traditional machine learning models show significant differences, with LR being the worst-performing method in this category (low level). As low as -11.134556, linear modeling has limitations in learning effective feature patterns, resulting in poor prediction performance; RF achieves effective prediction across all levels, but at lower levels... The R² value is 0.546112, reaching 0.927446 for medium-level and 0.911616 for high-level. It demonstrates good predictive ability in large-sample scenarios, but its prediction performance is relatively average for low-level scenarios with small sample sizes. XGBoost achieves a significant performance improvement over LR and RF, with an R² value of 0.818851 for low-level and close to 0.99 for medium- and high-level. MO-LightGBM is the best performing algorithm in this category, achieving high prediction accuracy across all levels. Further improvements were made, with the prediction accuracy for medium and high-level data exceeding 0.99, and low-level data also performing well in scenarios with extremely small samples. Its predictive performance has reached or even surpassed that of deep table models. This is because the prediction scenario for "Preference_cost" is simpler than that for "Preference_time" and "Preference_stability," with the fewest influencing factors on the target variable and more structured feature relationships. Traditional models can fully capture the data features in this simple scenario, thus achieving predictive performance similar to deep table models. All three types of deep table models showed good performance in prediction, with performance varying with the number of samples: TabTransformer showed good fitting results for large samples, but low-level data... With a score of only 0.651653, it performs poorly in small sample scenarios; the FT-Transformer is of a low grade. With a score of only 0.275316, it lacks adaptability for prediction in small sample scenarios. MLP-Mixer-Tab is the best among the three types of deep table models, achieving excellent prediction results at all levels, and its performance is basically on par with the traditional optimized model MO-LightGBM. The MO-TabPFN method presented in this paper achieves high scores at all levels. The R² values ​​are 0.997097, 0.999990, and 0.987393 for low, medium, and high levels, respectively. It achieves the highest prediction accuracy at low and medium levels. Although its performance is slightly weaker than the MO-LightGBM model and the MLP-Mixer-Tab method at high levels, the difference in R² is within 0.01, and it is close to a performance level of 1.0, showing relatively balanced prediction performance across levels compared to other methods. The slight weakness of the MO-TabPFN method at high levels is due to the fact that prediction scenarios like "Preference_cost" are simple, have few influencing factors, and a large sample size. Other models can achieve good prediction results at high levels, but the MO-TabPFN method's strong inductive bias weakens in higher-level scenarios with larger samples, and the influence of noise is amplified, resulting in a slightly lower R². Overall, compared to traditional models and deep table models, MO-TabPFN demonstrates strong prediction and generalization capabilities across different sample size distribution scenarios.

[0136] The above simulation analysis proves the correctness and effectiveness of the method proposed in this invention.

[0137] The parts of this invention not described in detail are common knowledge to those skilled in the art. The above description is only a preferred embodiment of the present invention and is not intended to limit the invention. Obviously, those skilled in the art, after understanding the content and principles of the present invention, may make various modifications and changes in form and detail without departing from the principles and structure of the present invention. However, these modifications and changes based on the concept of the present invention are still within the protection scope of the claims of the present invention.

Claims

1. A method for predicting vehicle user preferences based on historical tabular data, characterized in that, Includes the following steps: (1) Construct a vehicle user preference prediction architecture based on differentiated digital twins (DDTs), including vehicle user entities, cloud servers, and differentiated vehicle user twins (DVUs). Among them, the vehicle user entity requests the computing task to be unloaded, and at the same time the service data is synchronized to DVUs; Cloud servers provide a reliable storage and interaction environment for DVUs; DVUs store and manage data generated by vehicle user entities during the service process in tabular data format. (2) Obtain the basic public dataset through the Kaggle open source data platform and perform preliminary preprocessing on it, including deleting the Vehicle_ID column and the Timestamp column from the basic public dataset, and keeping the data in other columns to obtain two preprocessed datasets, denoted as the first feature dataset A and the second feature dataset B. (3) Generate multi-dimensional vehicle user preference data based on the pre-processed dataset, including vehicle user latency preference, stability preference and cost preference; Based on dataset A, the classification results of the target column Processing_Time_Category are converted into quantitative values ​​to obtain the latency preference column Preference_time. Using dataset A as the primary dataset and dataset B as the secondary dataset, the target column of dataset B, packet error rate (PER), is used as the stability preference index for vehicle users, resulting in the stability preference column Preference_stability. Based on the communication transmission rate column Bandwidth_Mbps of dataset A, feature mapping calculation is performed to obtain the cost preference column Preference_cost. (4) Construct a MO-TabPFN multi-objective regression prediction model adapted to vehicle user preference prediction. The model includes four layers: input layer, preprocessing layer, model layer and output layer. The MO-TabPFN prediction algorithm is used to fit multi-objective tabular prior data. The TabPFNRegressor is used as the basic regressor to achieve multi-objective regression. (5) Construct independent regression branches for the three types of preference data (latency, stability, and cost) obtained in step (3), and finally integrate them into a unified multi-objective regression model to realize vehicle user preference prediction and evaluation.

2. The method according to claim 1, characterized in that: The computational task offloading service in step (1) includes path planning, content recommendation, traffic congestion prediction and metaverse entertainment experience; the service data includes at least the communication environment, service requirements and service effects; the DVUs are virtual mappings of vehicle user entities, each vehicle user entity has a unique digital twin mapping DVU, and all vehicle user entities' DVUs are deployed on cloud servers.

3. The method according to claim 1, characterized in that: In step (3), the target column PER of dataset B represents the packet error rate. The higher the packet error rate, the worse the link stability. Vehicle users will be more sensitive to stability preferences, that is, the higher the packet error rate, the higher the stability preference of vehicle users. The cost preference of vehicle users is related to the requested bandwidth, and the bandwidth is related to the communication transmission rate.

4. The method according to claim 1, characterized in that: The generation process of the vehicle user latency preference mentioned in step (3) includes the following steps: (3a1) Temporarily encode the target column Processing_Time_Category of dataset A as 0, 1, 2. For each feature, calculate its Spearman correlation coefficient with the temporarily encoded target column, obtain the correlation analysis results between the feature and the target column, and record them. (3a2) Feature weight design: Calculate the absolute value of the correlation coefficient of all features based on the correlation analysis results, and normalize it according to its proportion of the absolute value of the correlation coefficient of all features to assign weights to each feature. ; (3a3) If the feature is positively correlated with the target, the larger the feature value, the higher the time delay preference; if the feature is negatively correlated with the target, the smaller the feature value, the higher the time delay preference; the formula for calculating time delay preference is as follows: ， in, To scale the eigenvalues ​​to the 0-1 interval, The correlation direction is indicated by a value greater than 0, which indicates a positive correlation between the feature and the target column, and a value less than 0, which indicates a negative correlation between the feature and the target column. (3a4) The obtained time delay preference The mapping is completed based on the original target column Processing_Time_Category classification, and a new target column is generated.

5. The method according to claim 1, characterized in that: The generation process of the vehicle user stability preference mentioned in step (3) includes the following steps: (3b1) Multiple regression prediction algorithms are used to predict the target column Preference_stability in dataset B. Based on the prediction results, the optimal algorithm is selected to perform SHAP analysis on the target column. The three core features that have the most significant impact on stability preference are Speed_kmh, SNR_dB, and Channel_Interference_%. Other non-critical features are removed. The preprocessed dataset B is obtained, which contains only the above three core features and the target column Preference_stability. (3b2) Compare the preprocessed dataset B to determine that dataset A is missing two features, SNR_dB and Channel_Interference_%, and perform completion processing on dataset A based on preset rules or fitting formulas to fill in the two missing features and complete the preprocessing of dataset A. (3b3) Generate vehicle user stability preference data based on the preprocessed dataset A.

6. The method according to claim 5, characterized in that: The regression prediction algorithms described in step (3b1) include at least linear regression, random forest, XGBoost, and LightGBM.

7. The method according to claim 5, characterized in that: The preprocessing of dataset A described in step (3b2) is implemented as follows: (3b2.1) Based on the original signal strength column in dataset A The SNR_dB feature is calculated using the following formula and added to dataset A: ， Among them, noise intensity column It is inversely proportional to the communication transmission rate. It is obtained by normalizing the original communication transmission rate column Bandwidth_Mbps in dataset A to the interval [-110, -90] according to the inverse proportional relationship, and the unit is dBm. (3b2.2) Based on the five features of signal-to-noise ratio, traffic density, signal strength, noise intensity and speed in dataset A, channel interference data is generated according to the rule that "signal-to-noise ratio and signal strength positively affect channel interference, while traffic density, noise intensity and speed negatively affect channel interference". The data is then normalized to the interval [0,100] to obtain the Channel_Interference_% feature column, and added to dataset A.

8. The method according to claim 7, characterized in that: The generation process of the vehicle user cost preference mentioned in step (3) includes the following steps: (3c1) Calculate the bandwidth characteristic column Bandwidth_Hz according to Shannon's formula: ， in, For communication transmission rate; (3c2) Based on the bandwidth feature column Bandwidth_Hz, the cost preference column Preference_cost is obtained through reverse normalization.

9. The method according to claim 1, characterized in that: The MO-TabPFN multi-objective regression prediction model described in step (4) has the following four layers in its network structure: The input layer receives input data as a two-dimensional table of vehicle user historical preferences stored in DVUs, including multiple feature columns and three target columns, where the target columns represent vehicle users' preferences for latency, stability, and cost, respectively. The preprocessing layer divides the input data into training and testing sets, and performs OneHot encoding on the classification features. The model layer uses TabPFNRegressor as the basic regressor and encapsulates TabPFNRegressor with MultiOutputRegressor to achieve multi-objective regression. An independent TabPFNRegressor regression branch is constructed for each of the three preference objectives of time delay, stability, and cost, and finally integrated into a unified multi-objective regression model. The output layer inputs the preprocessed test set feature matrix into the optimal model and outputs the prediction results through the prediction method. The output results include the predicted values ​​of vehicle user latency preference, stability preference, and cost preference.

Images