Automobile user behavior prediction method based on data model

Abstract

The invention relates to the field of intelligent automobile data processing, and discloses an automobile user behavior prediction method based on a data model, and the method comprises the following steps: 1, collecting user multi-modal data which comprises a user driving track, a purchase record, a vehicle-mounted video and a voice instruction; 2, carrying out standardization processing on the multi-modal data, and constructing a data matrix; 3, calculating a covariance matrix of the data matrix, and improving the stability of the covariance matrix through an optimization algorithm; step 4, eliminating noise features in the covariance matrix based on a random matrix theory, and extracting features with relatively high signal intensity; and 5, optimizing the cooperative relationship between the signal features by adopting a feature interaction modeling method. According to the method, the technical scheme of combining multi-modal data acquisition and standardized processing is adopted, and the fusion of high-dimensional and diversified data under the same analysis framework is realized by unifying the scale and distribution of different modal data.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent automobile data processing, and in particular to a method for predicting automobile user behavior based on a data model. Background Art

[0002] In the current smart car field, user behavior analysis and prediction has become an important direction to promote personalized services and optimize user experience. With the popularization of smart vehicle-mounted devices, vehicles can collect multimodal data, including but not limited to driving trajectories, user purchase records, vehicle-mounted videos, and voice interaction information. By mining and analyzing these multimodal data, user behavior habits can be accurately predicted, thereby providing support for intelligent vehicle decision-making. However, the high dimensionality and heterogeneity of multimodal data bring many challenges to analysis and modeling.

[0003] In the prior art, single-modal data is usually used to build user behavior prediction models, such as using only driving trajectory data for route planning, or using purchase record data to analyze user preferences. Although these methods can achieve certain results in specific scenarios, the prediction accuracy of the model is limited due to the failure to effectively combine the rich information of multimodal data. In addition, when processing high-dimensional multimodal data, the prior art is easily disturbed by inconsistent data scales and noise characteristics, which further reduces the stability and generalization ability of the model.

[0004] The existing technology has the following main problems: In the process of processing multimodal data, due to the significant differences in the scale and distribution of different modal data, directly using these data for modeling can easily lead to uneven weight distribution between features, thereby affecting model performance. In addition, when processing high-dimensional small sample data, the instability of the covariance matrix estimation will lead to noise amplification in the feature extraction stage, thereby reducing the reliability of the prediction model. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a method for predicting automobile user behavior based on a data model, which solves the problems of inaccurate feature extraction and insufficient user behavior prediction accuracy caused by inconsistent scales and unstable covariance matrix estimation of multimodal data in the prior art.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for predicting automobile user behavior based on a data model, comprising the following steps: Step 1: Collect multimodal data of users, including driving trajectories, purchase records, in-vehicle videos and voice commands; Step 2: standardize the multimodal data to construct a data matrix; Step 3: Calculate the covariance matrix of the data matrix and improve the stability of the covariance matrix through an optimization algorithm; Step 4: Based on the random matrix theory, the noise features in the covariance matrix are eliminated to extract the features with higher signal strength; Step 5: Use feature interaction modeling to optimize the collaborative relationship between signal features and adjust feature weights through the attention mechanism; Step 6: Input the optimized features into the user behavior prediction model to complete the prediction of user behavior; Step 7: When new data arrives, dynamically update the covariance matrix and user behavior prediction model, and adjust the prediction results in real time.

[0007] Preferably, the standardization of the multimodal data includes normalizing the data of different modalities according to the mean and standard deviation of their characteristic dimensions so that the data are comparable within the same dimension.

[0008] Preferably, the optimization of the covariance matrix is ​​performed by Ledoit-Wolf shrinkage estimation, which weights the linear combination of the covariance matrix and the identity matrix to improve the stability and estimation accuracy of the covariance matrix.

[0009] Preferably, the step of eliminating noise features by random matrix theory includes the following: S1. Analyze the eigenvalue spectrum distribution of the covariance matrix; S2. According to The theorem determines the distribution range of noise characteristic values; S3. Eliminate eigenvalues ​​and their corresponding eigenvectors that fall within the noise range.

[0010] Preferably, the interaction modeling of the signal features is accomplished by constructing a quadratic function, which is composed of an interaction relationship matrix of feature vectors, linear weights, and bias terms.

[0011] Preferably, the user behavior prediction model adopts a Transformer framework, including an input layer, a self-attention layer, a feedforward neural network layer and an output layer, wherein the self-attention layer is used to capture the correlation between signal features.

[0012] Preferably, the dynamic update includes the following contents: Step 1: Standardize the new data; Step 2: Calculate the updated covariance matrix and perform a weighted combination with the historical covariance matrix according to the time decay factor; Step 3: Adjust the parameters of the prediction model based on the updated covariance matrix.

[0013] Preferably, the extraction of the signal features includes vectorized representation of audio and video data, which are integrated with structured data to form a unified data matrix.

[0014] Preferably, the optimization goal of the prediction model is to adjust the model parameters by minimizing the classification error, and the optimization objective function includes the logarithmic error between the user behavior label and the predicted value.

[0015] The present invention provides a method for predicting automobile user behavior based on a data model. It has the following beneficial effects: 1. The present invention adopts a technical solution combining multimodal data acquisition with standardized processing, and realizes the integration of high-dimensional and diverse data under the same analysis framework by unifying the scale and distribution of different modal data. Compared with the existing solution of modeling only for single modal data, the present invention solves the problem of model performance degradation caused by inconsistent scales between multimodal data, and significantly improves the accuracy and consistency of data analysis.

[0016] 2. The present invention introduces the Ledoit-Wolf shrinkage estimation to optimize the covariance matrix, and improves the stability of the covariance matrix by using the weighted combination of the unit matrix. Unlike the traditional covariance matrix calculation method, which is prone to unstable eigenvalue distribution in small sample high-dimensional scenarios, the present invention effectively solves the shortcomings of the covariance matrix being prone to overfitting and large estimation deviation in high-dimensional data, and ensures the accuracy of subsequent feature extraction.

[0017] 3. The present invention analyzes the distribution range of the covariance matrix eigenvalues ​​through random matrix theory, and eliminates the noise eigenvalues ​​and their corresponding eigenvectors, significantly improving the purity of the signal features. Compared with the prior art that directly uses unoptimized features and is easily interfered by noise, the solution of the present invention makes the extracted features more accurate and provides high-quality input for subsequent user behavior modeling.

[0018] 4. The present invention designs a dynamic update mechanism that combines time decay factor and incremental learning, which can update the covariance matrix and prediction model parameters in real time. Compared with the solution in the prior art that requires retraining the model to adapt to new data, the present invention effectively solves the shortcomings of the model in terms of real-time performance and computational efficiency, and significantly improves the adaptability and efficiency of user behavior prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the method flow of the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the specification of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] Please see attached Figure 1 The embodiment of the present invention provides a method for predicting automobile user behavior based on a data model, comprising the following steps: Step 1: Collect multimodal user data, including user driving trajectories, purchase records, in-vehicle videos, and voice commands; Step 2: Standardize the multimodal data and construct a data matrix; Step 3: Calculate the covariance matrix of the data matrix and improve the stability of the covariance matrix through optimization algorithm; Step 4: Based on the random matrix theory, the noise features in the covariance matrix are eliminated to extract the features with higher signal strength; Step 5: Use feature interaction modeling to optimize the collaborative relationship between signal features and adjust feature weights through the attention mechanism; Step 6: Input the optimized features into the user behavior prediction model to complete the prediction of user behavior; Step 7: When new data arrives, dynamically update the covariance matrix and user behavior prediction model, and adjust the prediction results in real time.

[0022] The standardization of multimodal data includes normalizing the data of different modalities according to the mean and standard deviation of their characteristic dimensions to make the data comparable within the same dimension.

[0023] The optimization of the covariance matrix is ​​done through the Ledoit-Wolf shrinkage estimation, which weights the linear combination of the covariance matrix and the identity matrix to improve the stability and estimation accuracy of the covariance matrix.

[0024] The steps of random matrix theory to remove noise features include the following: S1. Analyze the eigenvalue spectrum distribution of the covariance matrix; S2. According to The theorem determines the distribution range of noise characteristic values; S3. Eliminate the eigenvalues ​​and their corresponding eigenvectors that fall within the noise range.

[0025] The interaction modeling of signal features is completed by constructing a quadratic function, which consists of the interaction relationship matrix of the feature vector, linear weights and bias terms.

[0026] The attention mechanism adopts a multi-head attention mechanism to sort the importance of signal features by calculating the correlation weights between signal feature vectors.

[0027] The user behavior prediction model adopts the Transformer framework, which includes an input layer, a self-attention layer, a feedforward neural network layer, and an output layer. The self-attention layer is used to capture the correlation between signal features.

[0028] Dynamic updates include the following: Step 1: Standardize the new data; Step 2: Calculate the updated covariance matrix and perform a weighted combination with the historical covariance matrix according to the time decay factor; Step 3: Adjust the parameters of the prediction model based on the updated covariance matrix.

[0029] The extraction of signal features includes the vector representation of audio and video data, which is integrated with structured data to form a unified data matrix.

[0030] The optimization goal of the prediction model is to adjust the model parameters by minimizing the classification error. The optimization objective function includes the logarithmic error between the user behavior label and the predicted value.

[0031] In the technical solution of the present invention, the collection of multimodal data is the starting point of the entire user behavior prediction process, and its quality directly affects the subsequent feature extraction, modeling and prediction accuracy. In general, multimodal data includes but is not limited to information such as the user's driving trajectory, purchase records, in-vehicle video, and voice commands. These data come from a variety of hardware devices and platforms and have different modalities and data characteristics. In practical applications, the spatiotemporal correlation of different modal data also needs to be comprehensively considered to ensure data consistency and integrity.

[0032] As an option, these multimodal data can be collected from multiple data sources in real time and integrated into a unified storage structure. Specifically, the present invention provides high-quality original input for subsequent data processing by designing a reasonable data collection mechanism.

[0033] In one possible implementation, the driving trajectory data can be obtained through the vehicle's global positioning system (GPS) module. Specifically, the GPS module can record the vehicle's timestamp, longitude and latitude coordinates, and driving speed. These data are usually stored in the form of a time series, and each row in the data matrix represents the trajectory characteristics at a certain point in time. In order to ensure the accuracy and integrity of the trajectory data, in general, the missing points or duplicate points that may appear during the collection process are preprocessed, such as completing the missing points through an interpolation algorithm and removing duplicate points through a deduplication algorithm.

[0034] As an improvement, the collection of trajectory data can also be combined with the data of the inertial measurement unit (IMU). For example, the vehicle's accelerometer and gyroscope can be used to supplement the driving trajectory information to form a more accurate trajectory data matrix.

[0035] The formula is as follows: in: For the Trajectory characteristics at each time point; and Respectively The latitude and longitude of a point in time; For the vehicle The speed at a point in time; is the acceleration of the vehicle; is the yaw angle of the vehicle.

[0036] Specifically, user purchase record data can be obtained by connecting to the database of the car service platform. This type of data includes historical order information, transaction amount, transaction time, and product category generated by users on the platform. These data have structured characteristics and can be directly organized into a matrix form.

[0037] In some embodiments, the purchase record also includes the negotiation record between the user and the dealer during the car purchase process. These data are converted into numerical features through natural language processing (NLP) methods, such as keyword extraction, sentiment analysis scores, etc.

[0038] In order to unify the structured representation of these purchase records, they can be organized into the following data matrix: in: For the Characteristics of the purchase record; is the purchase amount; is the numerical code for the commodity category; It is the transaction timestamp; The language feature score extracted from the transaction record.

[0039] As an option, the collection of in-vehicle video data can be achieved through a camera installed in the car. Generally, the video data is divided into a frame sequence, and each frame image is pre-processed to extract features, such as object detection, driving scene analysis, etc. Further, the video frame features are converted into a fixed-length vector representation through dimensionality reduction technology to form a video feature matrix.

[0040] Voice data can be collected by the vehicle voice recognition system, such as voice commands when users interact with the vehicle. The collected voice data is first framed and Fourier transformed to extract spectral features, and then a convolutional neural network (CNN) is applied to generate high-dimensional voice features.

[0041] In one possible implementation, video and voice features are integrated into a mixed modality feature matrix, as follows: in: For the Mixed modal characteristics at each time point; is the feature vector of the video frame; is the spectral feature vector of the speech signal.

[0042] In the design of the present invention, different modal data need to be integrated into a unified data storage format. Specifically, the multimodal data is aligned by timestamp to form a global feature matrix. The aligned data matrix can be expressed as follows: in: is the integrated data matrix; is the length of the time series; is the total number of features, including all features of trajectory, purchase history, video, and voice.

[0043] In a possible implementation, a distributed storage system may be used to store the integrated data to facilitate subsequent large-scale processing.

[0044] In the technical solution of the present invention, standardization is a key link for multimodal data to enter the analysis stage. Its purpose is to make the data comparable within a unified dimension by eliminating the scale differences of different modal data. In general, multimodal data comes from a wide range of sources, including but not limited to user driving trajectories, purchase records, in-vehicle videos and voice data. The characteristics and numerical ranges of these data are different. Direct calculation may cause the weights of some modal features to be too large or too small, thereby affecting the calculation and modeling accuracy of the subsequent covariance matrix.

[0045] As an option, all modal data can be normalized by normalization to map them to the same numerical range. Specifically, normalization usually combines the mean and standard deviation of each modal data to ensure that the eigenvalues ​​are on the same distribution scale.

[0046] In this embodiment, the multimodal data collected in step 1 is organized into a unified data matrix ,in .in, represents the number of samples, Represents the total dimension of the feature. Each row represents a sample, and each column corresponds to a feature. In general, each element of the data matrix For the original collected data, there may be large differences in the numerical ranges of different modal features. For example, the latitude and longitude coordinates in the driving trajectory may range from [0,180], while the vehicle speed may range from [0,200]. This difference may cause specific modal data to dominate in subsequent analysis, interfering with the accuracy of the covariance matrix.

[0047] In order to eliminate the above problems, the present invention performs standardization processing on each column of the data matrix, and the specific formula is: in: After standardization Row, No. The eigenvalues ​​of the columns; is the original eigenvalue before standardization; For the Mean of column features; For the The standard deviation of the column feature.

[0048] In general, the mean and standard deviation It is calculated over the entire range of sample data to ensure that the standardized features of all samples are consistent.

[0049] In some embodiments, for discrete features (such as commodity category codes in purchase records), standardization may not be applicable. As an option, one-hot encoding can be used to convert discrete features into multi-dimensional sparse vector form. For example, The encoding result can be expressed as: Each dimension indicates whether a product category is related to the current sample.

[0050] Specifically, the standardized processing of unstructured data such as speech and video requires the combination of specific modal features. For the spectral characteristics of speech signals.

[0051] For video data, it can be standardized by normalizing pixel values ​​and reducing the dimension after feature extraction. The processing is done according to the following formula: The pixel value range [0,255] is compressed to [0,1].

[0052] In one possible implementation, the normalization of multimodal data follows the following steps: first, all numerical features are normalized according to the mean and standard deviation; second, discrete features are expanded using one-hot encoding; finally, unstructured modal features (such as video and speech) are normalized within a specific modality range and integrated into the data matrix middle.

[0053] After the above processing is completed, all modal data are mapped to the same numerical range, and finally a standardized data matrix is ​​formed. , which is used for the subsequent calculation of the covariance matrix.

[0054] As an improvement, the present invention can also remove noise from the standardized data matrix. For example, for abnormal points that may exist in the trajectory data (such as sudden changes in speed or yaw angle), a threshold filtering method can be used to correct them. For example, if the acceleration of the vehicle exceeds a reasonable range (such as 10m / s 2 ), it is replaced by the mean acceleration of adjacent time points: In the technical solution of the present invention, the calculation and optimization of the covariance matrix is ​​an indispensable key step in the user behavior prediction process. It contains the feature relationship information of multimodal data, while the covariance matrix can fully reflect the linear correlation between features. As a commonly used statistical tool, the calculation of the covariance matrix can reveal the interaction between features. However, in high-dimensional small samples (i.e. ), directly using the traditional covariance matrix calculation method may lead to the problem of unstable estimation. To solve this problem, the present invention adopts an optimization algorithm to correct the covariance matrix, thereby improving its robustness and accuracy.

[0055] Specifically, the present invention designs a covariance matrix optimization method based on Ledoit-Wolf shrinkage estimation, which significantly improves the stability of the covariance matrix by combining the linear weights of the unit matrix, laying a solid foundation for subsequent feature extraction and modeling.

[0056] In this embodiment, the covariance matrix is ​​calculated based on the standardized data matrix obtained in step 2. Each element of the covariance matrix represents the linear correlation between features, which is defined as follows: in: represents the covariance matrix; is the standardized data matrix; represents the number of samples; Indicates the number of features.

[0057] In the above formula, Representation Matrix The transpose operation of the matrix Each element of Defined as: in and Respectively represent samples No. and feature values.

[0058] In general, the covariance matrix is a symmetric matrix whose diagonal elements Represents the variance of a single feature, non-diagonal elements Represents the covariance between two features.

[0059] In actual scenarios, especially when That is, when the number of samples is less than the number of features, directly calculating the covariance matrix may lead to the following problems: the distribution of the eigenvalues ​​of the covariance matrix may be unstable, some eigenvalues ​​approach zero or even become negative, and cannot accurately reflect the relationship between the features; the condition number of the covariance matrix may be too large, resulting in numerical instability in subsequent calculations.

[0060] As an option, the stability of the covariance matrix can be improved by introducing linear weighting of the identity matrix.

[0061] To solve the above problem, the present invention uses the Ledoit-Wolf shrinkage estimation method to optimize the covariance matrix. The optimized covariance matrix is ​​defined as follows: in: represents the optimized covariance matrix; represents the original covariance matrix; is the identity matrix; Indicates the strength of contraction.

[0062] In general, The value of is determined by the number of samples and the number of features, and its calculation formula is: in: Var represents the variance of the diagonal elements of the covariance matrix; Represents the squares of the diagonal elements.

[0063] Specifically, the shrinkage strength The introduction of can reduce the random fluctuation effect between features, thereby improving the numerical stability of the covariance matrix.

[0064] In some embodiments, the optimization of the covariance matrix can also be combined with the following improved methods: Eigenvalue truncation: For the eigenvalue spectrum of the covariance matrix, a lower limit value can be set to eliminate eigenvalues ​​below this value to avoid the influence of too small eigenvalues ​​on the matrix inverse operation; Principal component analysis (PCA): By extracting the main eigenvalues ​​of the covariance matrix and reducing the dimensionality of the original data matrix, the computational complexity of the covariance matrix can be further reduced.

[0065] In one possible implementation, the PCA method can be used to reduce the dimension of the data matrix, and then the covariance matrix after the dimension reduction is Perform shrinkage optimization.

[0066] After the above optimization, the covariance matrix It has the following characteristics: the diagonal elements still represent the optimized feature variance; the off-diagonal elements represent the optimized feature covariance, which reduces noise interference; the eigenvalue distribution of the covariance matrix is ​​smoother, providing a reliable basis for subsequent noise removal and modeling.

[0067] Specifically, the present invention adopts The theorem analyzes the eigenvalue distribution of the covariance matrix, and by identifying the range of noise eigenvalues, removes the corresponding eigenvectors, thereby retaining the part with significant signal characteristics. The optimized feature matrix provides reliable support for subsequent modeling and prediction.

[0068] In this embodiment, random matrix theory is used to analyze the eigenvalue spectrum distribution of the covariance matrix. In general, the eigenvalue distribution of the covariance matrix can be expressed as Theorem description. According to this theorem, in the sample size and feature dimensions Both tend to infinity and the ratio When constant, the eigenvalue distribution of the covariance matrix has the following range: in: Represents the ratio of sample size to feature dimension; and are the upper and lower bounds of the eigenvalues, respectively.

[0069] Specifically, when the eigenvalues ​​of the covariance matrix are located at When the eigenvalues ​​are within the range, they are generally considered to be caused by noise; while the eigenvalues ​​outside the range are considered to have significant signal strength.

[0070] In order to eliminate noise features, the following steps are designed in this embodiment: First, the eigenvalues ​​and eigenvectors of the covariance matrix are obtained by performing eigenvalue decomposition on the covariance matrix. It can be broken down into: in: is the eigenvector matrix, whose column vector represents the eigenvector of the covariance matrix; is a diagonal matrix of eigenvalues, where represents the eigenvalues ​​of the covariance matrix.

[0071] Subsequently, according to Theorem computes the lower and upper bounds of eigenvalues Specifically, for the sample size and the number of features , the calculation formulas of the upper and lower bounds are as follows: In the calculation After that, remove The eigenvalues ​​and their corresponding eigenvectors in the range. The optimized covariance matrix can be constructed by the following formula: in: Indicates eigenvalues ​​beyond the noise range; is the corresponding feature vector.

[0072] In some embodiments, the noise removal process can also be combined with the calculation of feature contribution rate to ensure that the features retained after noise removal are comprehensive enough. For example, the feature contribution rate can be defined as the ratio of a certain feature value to the total feature value: When the cumulative contribution rate When a certain threshold (such as 95%) is exceeded, subsequent features are stopped from being eliminated. This improved method can ensure the retention of signal features while further optimizing the effect of noise elimination.

[0073] In a possible implementation, we can also introduce a regularization parameter Adjust to suit different data distributions. For example, when the data is noisy, you can choose to reduce The threshold is set to remove more noise features.

[0074] After the above noise removal process, the dimension of the covariance matrix will be significantly reduced, and only the part with higher signal strength will be retained. At this time, the signal feature matrix can be expressed by the following formula: in: Representation and signal characteristic value The corresponding eigenvector; is the number of signal features.

[0075] In the technical solution of the present invention, the signal feature matrix extracted through the above steps is It mainly includes the part with significant signal features in the covariance matrix. However, there may be complex correlations and interactions between these signal features. Modeling based on a single feature alone may ignore the synergy between features and reduce prediction accuracy. In general, in order to make more comprehensive use of signal features, the present invention designs a feature interaction modeling method to optimize the collaborative relationship between signal features.

[0076] Specifically, the present invention describes the interaction between signal features by constructing a quadratic function, and further optimizes the weight of the importance of signal features by combining a multi-head attention mechanism. This method can accurately capture the relationship between features while highlighting features that are highly relevant to the prediction task, thereby improving the modeling effect.

[0077] In this embodiment, the signal feature matrix Each column of Represents the extracted signal feature vectors, signal feature matrices The dimension is ,in: represents the number of samples; Represents the number of signal features.

[0078] To model the interaction between signal features, this embodiment constructs the following quadratic function: in: is the interaction relationship matrix, which indicates the relationship strength between different signal features; is a linear weight vector, representing the direct contribution of a single feature to the target; is the bias term.

[0079] In general, the interaction matrix is a symmetric matrix, each element of which Representation characteristics and By minimizing the error function, we can optimize 、 and value so that it accurately reflects the actual relationship between features.

[0080] In this embodiment, a multi-head attention mechanism is further used to adjust the weights of signal features. Generally, the multi-head attention mechanism can capture the importance of features in a global scope, thereby giving more reasonable weights to the features. Specifically, the calculation formula of the multi-head attention mechanism is as follows: in: is the query matrix, which represents the correlation between features; Key Matrix; is the value matrix; is the feature dimension.

[0081] As an improvement, this embodiment also performs weighted averaging on the output of the attention mechanism to further highlight the features related to target prediction. The weighted feature matrix can be expressed as: in is the attention weight vector, each element of which is calculated by the softmax function.

[0082] In some embodiments, the quadratic function of feature interaction modeling can be expanded to a form with a regularization term to improve the generalization ability of the model. For example, Regularization term, so that the optimization objective becomes: in: Represents the true value of the predicted target; The interaction matrix The Frobenius norm of ; is the regularization coefficient, which is used to control the strength of regularization.

[0083] In addition, in the weight optimization process, the characteristics of different modal features can be combined to For example, for video modality features, the importance of features at different moments can be highlighted by introducing a time series attention mechanism.

[0084] After the above feature interaction modeling and attention weight optimization, an optimized signal feature matrix is ​​generated. This matrix can more accurately express the interaction relationship between features and highlight the features that are highly relevant to the prediction task. Specifically, the optimized feature matrix can be expressed as: In the technical solution of the present invention, the optimized signal feature matrix is ​​obtained through the above steps. , the matrix retains the part of the data with significant signal features, and further improves the expressiveness of the features through feature interaction modeling and attention weight optimization. On this basis, in order to achieve accurate prediction of user behavior, the present invention designs a behavior prediction model based on deep learning. In general, the optimized feature matrix It is used as the input of the model, and the data features are further extracted and modeled through a multi-layer structure to finally generate the prediction results of user behavior.

[0085] As an option, the present invention adopts the Transformer framework to build a prediction model, combining the global attention mechanism and the expressive power of the feedforward neural network to achieve accurate modeling and behavior prediction of high-dimensional multimodal data.

[0086] In this embodiment, the user behavior prediction model is composed of a multi-layer network structure, including an input layer, a self-attention layer, a feedforward neural network layer, and an output layer. The core of the model is to capture the high-order relationship between the optimization features through the self-attention mechanism, thereby improving the prediction accuracy.

[0087] Input layer design In general, the input layer is used to receive the optimized signal feature matrix The shape of the input matrix is ,in: is the sample size; is the number of signal features.

[0088] In the input layer, positional encoding is added to each feature to retain the contextual information of the feature. Positional encoding can be expressed as: in: Represents the position index of the feature; The dimension index representing the feature.

[0089] Input Matrix After adding the positional encoding, the final embedding matrix is ​​generated , as the input of the subsequent network layer.

[0090] Generally, in order to enhance the model's ability to capture different feature interaction patterns, the self-attention mechanism is extended to a multi-head attention mechanism. Its output can be expressed as: in: Attention Indicates The calculation results of the attention heads; is the number of attention heads; is the linear transformation weight matrix.

[0091] Based on the output of the self-attention layer, the feedforward neural network layer further extracts nonlinear features. Specifically, the feedforward neural network layer contains two fully connected layers, which are expressed as follows: in: is the output of the previous layer; is the weight matrix; is the bias vector; ReLU is the activation function, which represents the linear rectifier unit.

[0092] Generally, the output layer is responsible for generating the prediction results of user behavior. In this embodiment, the output layer uses the Sigmoid activation function to process the output of the last layer of the network to generate the probability prediction value of user behavior. The specific formula is as follows: in: Indicates The predicted value of samples; Represents the linear transformation output of the last layer of the model.

[0093] In some embodiments, the training objectives of the model can be adjusted based on specific task requirements. For example: For binary classification tasks, the cross entropy loss function can be used as the optimization objective: in is the true label, is the model's predicted value.

[0094] For multi-classification tasks, it can be extended to the Softmax loss function.

[0095] In addition, in order to improve the model's adaptability to unbalanced data, a weighted loss function can be introduced to give higher weights to minority category samples.

[0096] In the technical solution of the present invention, in order to adapt to the changing trend of user behavior and ensure the real-time and accuracy of the prediction model, it is necessary to dynamically update the covariance matrix and the user behavior prediction model. In general, the arrival of new data may change the user's behavior pattern, and the original model may not be able to fully reflect these changes. To this end, the present invention designs a dynamic update mechanism, which gradually updates the covariance matrix in combination with the time decay factor, and adjusts the parameters of the prediction model at the same time, so as to achieve rapid adaptation to new data.

[0097] Specifically, the present invention further ensures the accuracy and robustness of the user behavior prediction model by updating the covariance matrix online and optimizing signal features.

[0098] In this embodiment, when new data arrives, the new data is standardized and weightedly combined with the covariance matrix of the historical data. In general, the updated covariance matrix is ​​defined as follows: in: is the time decay factor, which controls the weight of new and old data; is the historical covariance matrix; Represents the covariance matrix of the new data; is the number of samples of new data.

[0099] In general, the time decay factor The value of can be adjusted according to the specific application scenario. For example, in a scenario with high real-time requirements, a smaller , to increase the weight of new data.

[0100] As an option, you can Dynamic adjustment is performed to automatically optimize the weight distribution according to the difference between new data and historical data. Specifically, the dynamic adjustment formula can be expressed as: in Represents the Frobenius norm of the matrix, which is used to measure the difference between matrices.

[0101] After the covariance matrix is ​​updated, the signal feature matrix The extraction method also needs to be adjusted accordingly. In general, based on the new covariance matrix Recalculate the eigenvalue distribution and remove the noise features. The specific steps include: First, the updated eigenvalue upper and lower bounds are calculated according to the Marčenko-Pastur theorem in , is the number of new samples, is the feature dimension.

[0102] Then, remove the noise range Only the signal features are retained.

[0103] Finally, reconstruct the signal feature matrix , as the dynamically updated feature input.

[0104] After the covariance matrix and signal features are updated, the parameters of the user behavior prediction model also need to be adjusted accordingly. The present invention adopts an incremental learning method to optimize the existing model using new data. In general, the core of incremental learning is to only update the model parameters related to the new data, thereby reducing training time.

[0105] As a possible implementation method, the model parameters can be fine-tuned through the gradient update method. The specific formula is as follows: in: and Respectively represent the parameters before and after the model update; is the learning rate; is the gradient of the loss function on the new data.

[0106] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting automobile user behavior based on a data model, characterized in that: The following steps are involved: Step 1: Collect multimodal data of users, including driving trajectories, purchase records, in-vehicle videos and voice commands; Step 2: standardize the multimodal data to construct a data matrix; Step 3: Calculate the covariance matrix of the data matrix and improve the stability of the covariance matrix through an optimization algorithm; Step 4: Based on the random matrix theory, the noise features in the covariance matrix are eliminated to extract the features with higher signal strength; Step 5: Use feature interaction modeling to optimize the collaborative relationship between signal features and adjust feature weights through the attention mechanism; Step 6: Input the optimized features into the user behavior prediction model to complete the prediction of user behavior; Step 7: When new data arrives, dynamically update the covariance matrix and user behavior prediction model, and adjust the prediction results in real time.

2. The method for predicting automobile user behavior based on a data model according to claim 1, characterized in that: The standardization of the multimodal data includes normalizing the data of different modalities according to the mean and standard deviation of their characteristic dimensions so as to make the data comparable within the same dimension.

3. The method for predicting automobile user behavior based on a data model according to claim 1, characterized in that: The optimization of the covariance matrix is ​​accomplished by Ledoit-Wolf shrinkage estimation, which weights the linear combination of the covariance matrix and the identity matrix to improve the stability and estimation accuracy of the covariance matrix.

4. The method for predicting automobile user behavior based on a data model according to claim 1, characterized in that: The steps of removing noise features by random matrix theory include the following: S1. Analyze the eigenvalue spectrum distribution of the covariance matrix; S2. The theorem determines the distribution range of noise characteristic values; S3. Eliminate eigenvalues ​​and their corresponding eigenvectors that fall within the noise range.

5. The method for predicting automobile user behavior based on a data model according to claim 1, characterized in that: The interaction modeling of the signal features is completed by constructing a quadratic function, which is composed of an interaction relationship matrix of feature vectors, linear weights and bias terms.

6. The method for predicting automobile user behavior based on a data model according to claim 1, characterized in that: The attention mechanism adopts a multi-head attention mechanism and completes the importance ranking of signal features by calculating the correlation weights between signal feature vectors.

7. The method for predicting automobile user behavior based on a data model according to claim 1, characterized in that: The user behavior prediction model adopts a Transformer framework, which includes an input layer, a self-attention layer, a feedforward neural network layer and an output layer, wherein the self-attention layer is used to capture the correlation between signal features.

8. The method for predicting automobile user behavior based on a data model according to claim 1, characterized in that: The dynamic update includes the following: Step 1: Standardize the new data; Step 2: Calculate the updated covariance matrix and perform a weighted combination with the historical covariance matrix according to the time decay factor; Step 3: Adjust the parameters of the prediction model based on the updated covariance matrix.

9. The method for predicting automobile user behavior based on a data model according to claim 1, characterized in that: The extraction of the signal features includes vector representation of audio and video data, which is integrated with structured data to form a unified data matrix.

10. The method for predicting automobile user behavior based on a data model according to claim 1, characterized in that: The optimization goal of the prediction model is to adjust the model parameters by minimizing the classification error, and the optimization objective function includes the logarithmic error between the user behavior label and the predicted value.

Images